1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
185 struct io_mapped_ubuf {
188 struct bio_vec *bvec;
189 unsigned int nr_bvecs;
192 struct fixed_file_table {
196 struct fixed_file_ref_node {
197 struct percpu_ref refs;
198 struct list_head node;
199 struct list_head file_list;
200 struct fixed_file_data *file_data;
201 struct llist_node llist;
204 struct fixed_file_data {
205 struct fixed_file_table *table;
206 struct io_ring_ctx *ctx;
208 struct percpu_ref *cur_refs;
209 struct percpu_ref refs;
210 struct completion done;
211 struct list_head ref_list;
216 struct list_head list;
224 struct percpu_ref refs;
225 } ____cacheline_aligned_in_smp;
229 unsigned int compat: 1;
230 unsigned int limit_mem: 1;
231 unsigned int cq_overflow_flushed: 1;
232 unsigned int drain_next: 1;
233 unsigned int eventfd_async: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head;
250 unsigned sq_thread_idle;
251 unsigned cached_sq_dropped;
252 atomic_t cached_cq_overflow;
253 unsigned long sq_check_overflow;
255 struct list_head defer_list;
256 struct list_head timeout_list;
257 struct list_head cq_overflow_list;
259 wait_queue_head_t inflight_wait;
260 struct io_uring_sqe *sq_sqes;
261 } ____cacheline_aligned_in_smp;
263 struct io_rings *rings;
267 struct task_struct *sqo_thread; /* if using sq thread polling */
268 struct mm_struct *sqo_mm;
269 wait_queue_head_t sqo_wait;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data *file_data;
277 unsigned nr_user_files;
279 struct file *ring_file;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs;
283 struct io_mapped_ubuf *user_bufs;
285 struct user_struct *user;
287 const struct cred *creds;
289 struct completion ref_comp;
290 struct completion sq_thread_comp;
292 /* if all else fails... */
293 struct io_kiocb *fallback_req;
295 #if defined(CONFIG_UNIX)
296 struct socket *ring_sock;
299 struct idr io_buffer_idr;
301 struct idr personality_idr;
304 unsigned cached_cq_tail;
307 atomic_t cq_timeouts;
308 unsigned long cq_check_overflow;
309 struct wait_queue_head cq_wait;
310 struct fasync_struct *cq_fasync;
311 struct eventfd_ctx *cq_ev_fd;
312 } ____cacheline_aligned_in_smp;
315 struct mutex uring_lock;
316 wait_queue_head_t wait;
317 } ____cacheline_aligned_in_smp;
320 spinlock_t completion_lock;
323 * ->iopoll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head iopoll_list;
329 struct hlist_head *cancel_hash;
330 unsigned cancel_hash_bits;
331 bool poll_multi_file;
333 spinlock_t inflight_lock;
334 struct list_head inflight_list;
335 } ____cacheline_aligned_in_smp;
337 struct delayed_work file_put_work;
338 struct llist_head file_put_llist;
340 struct work_struct exit_work;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb {
350 struct wait_queue_head *head;
356 struct wait_queue_entry wait;
361 struct file *put_file;
365 struct io_timeout_data {
366 struct io_kiocb *req;
367 struct hrtimer timer;
368 struct timespec64 ts;
369 enum hrtimer_mode mode;
374 struct sockaddr __user *addr;
375 int __user *addr_len;
377 unsigned long nofile;
399 struct list_head list;
403 /* NOTE: kiocb has the file as the first member, so don't do it here */
411 struct sockaddr __user *addr;
418 struct user_msghdr __user *umsg;
424 struct io_buffer *kbuf;
430 struct filename *filename;
432 unsigned long nofile;
435 struct io_files_update {
461 struct epoll_event event;
465 struct file *file_out;
466 struct file *file_in;
473 struct io_provide_buf {
487 const char __user *filename;
488 struct statx __user *buffer;
491 struct io_completion {
493 struct list_head list;
497 struct io_async_connect {
498 struct sockaddr_storage address;
501 struct io_async_msghdr {
502 struct iovec fast_iov[UIO_FASTIOV];
504 struct sockaddr __user *uaddr;
506 struct sockaddr_storage addr;
510 struct iovec fast_iov[UIO_FASTIOV];
511 const struct iovec *free_iovec;
512 struct iov_iter iter;
514 struct wait_page_queue wpq;
517 struct io_async_ctx {
519 struct io_async_rw rw;
520 struct io_async_msghdr msg;
521 struct io_async_connect connect;
522 struct io_timeout_data timeout;
527 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
528 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
529 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
530 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
531 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
532 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
539 REQ_F_LINK_TIMEOUT_BIT,
541 REQ_F_COMP_LOCKED_BIT,
542 REQ_F_NEED_CLEANUP_BIT,
544 REQ_F_BUFFER_SELECTED_BIT,
545 REQ_F_NO_FILE_TABLE_BIT,
546 REQ_F_WORK_INITIALIZED_BIT,
547 REQ_F_TASK_PINNED_BIT,
549 /* not a real bit, just to check we're not overflowing the space */
555 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
556 /* drain existing IO first */
557 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
559 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
560 /* doesn't sever on completion < 0 */
561 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
563 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
564 /* IOSQE_BUFFER_SELECT */
565 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
568 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
569 /* fail rest of links */
570 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
571 /* on inflight list */
572 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
573 /* read/write uses file position */
574 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
575 /* must not punt to workers */
576 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
577 /* has linked timeout */
578 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
580 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
581 /* completion under lock */
582 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
584 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
585 /* already went through poll handler */
586 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
587 /* buffer already selected */
588 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
589 /* doesn't need file table for this request */
590 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
591 /* io_wq_work is initialized */
592 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
593 /* req->task is refcounted */
594 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
598 struct io_poll_iocb poll;
599 struct io_poll_iocb *double_poll;
603 * NOTE! Each of the iocb union members has the file pointer
604 * as the first entry in their struct definition. So you can
605 * access the file pointer through any of the sub-structs,
606 * or directly as just 'ki_filp' in this struct.
612 struct io_poll_iocb poll;
613 struct io_accept accept;
615 struct io_cancel cancel;
616 struct io_timeout timeout;
617 struct io_connect connect;
618 struct io_sr_msg sr_msg;
620 struct io_close close;
621 struct io_files_update files_update;
622 struct io_fadvise fadvise;
623 struct io_madvise madvise;
624 struct io_epoll epoll;
625 struct io_splice splice;
626 struct io_provide_buf pbuf;
627 struct io_statx statx;
628 /* use only after cleaning per-op data, see io_clean_op() */
629 struct io_completion compl;
632 struct io_async_ctx *io;
634 /* polled IO has completed */
640 struct io_ring_ctx *ctx;
643 struct task_struct *task;
646 struct list_head link_list;
649 * 1. used with ctx->iopoll_list with reads/writes
650 * 2. to track reqs with ->files (see io_op_def::file_table)
652 struct list_head inflight_entry;
654 struct percpu_ref *fixed_file_refs;
655 struct callback_head task_work;
656 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
657 struct hlist_node hash_node;
658 struct async_poll *apoll;
659 struct io_wq_work work;
662 struct io_defer_entry {
663 struct list_head list;
664 struct io_kiocb *req;
668 #define IO_IOPOLL_BATCH 8
670 struct io_comp_state {
672 struct list_head list;
673 struct io_ring_ctx *ctx;
676 struct io_submit_state {
677 struct blk_plug plug;
680 * io_kiocb alloc cache
682 void *reqs[IO_IOPOLL_BATCH];
683 unsigned int free_reqs;
686 * Batch completion logic
688 struct io_comp_state comp;
691 * File reference cache
695 unsigned int has_refs;
696 unsigned int ios_left;
700 /* needs req->io allocated for deferral/async */
701 unsigned async_ctx : 1;
702 /* needs current->mm setup, does mm access */
703 unsigned needs_mm : 1;
704 /* needs req->file assigned */
705 unsigned needs_file : 1;
706 /* don't fail if file grab fails */
707 unsigned needs_file_no_error : 1;
708 /* hash wq insertion if file is a regular file */
709 unsigned hash_reg_file : 1;
710 /* unbound wq insertion if file is a non-regular file */
711 unsigned unbound_nonreg_file : 1;
712 /* opcode is not supported by this kernel */
713 unsigned not_supported : 1;
714 /* needs file table */
715 unsigned file_table : 1;
717 unsigned needs_fs : 1;
718 /* set if opcode supports polled "wait" */
720 unsigned pollout : 1;
721 /* op supports buffer selection */
722 unsigned buffer_select : 1;
723 unsigned needs_fsize : 1;
726 static const struct io_op_def io_op_defs[] = {
727 [IORING_OP_NOP] = {},
728 [IORING_OP_READV] = {
732 .unbound_nonreg_file = 1,
736 [IORING_OP_WRITEV] = {
741 .unbound_nonreg_file = 1,
745 [IORING_OP_FSYNC] = {
748 [IORING_OP_READ_FIXED] = {
750 .unbound_nonreg_file = 1,
753 [IORING_OP_WRITE_FIXED] = {
756 .unbound_nonreg_file = 1,
760 [IORING_OP_POLL_ADD] = {
762 .unbound_nonreg_file = 1,
764 [IORING_OP_POLL_REMOVE] = {},
765 [IORING_OP_SYNC_FILE_RANGE] = {
768 [IORING_OP_SENDMSG] = {
772 .unbound_nonreg_file = 1,
776 [IORING_OP_RECVMSG] = {
780 .unbound_nonreg_file = 1,
785 [IORING_OP_TIMEOUT] = {
789 [IORING_OP_TIMEOUT_REMOVE] = {},
790 [IORING_OP_ACCEPT] = {
793 .unbound_nonreg_file = 1,
797 [IORING_OP_ASYNC_CANCEL] = {},
798 [IORING_OP_LINK_TIMEOUT] = {
802 [IORING_OP_CONNECT] = {
806 .unbound_nonreg_file = 1,
809 [IORING_OP_FALLOCATE] = {
813 [IORING_OP_OPENAT] = {
817 [IORING_OP_CLOSE] = {
819 .needs_file_no_error = 1,
822 [IORING_OP_FILES_UPDATE] = {
826 [IORING_OP_STATX] = {
834 .unbound_nonreg_file = 1,
838 [IORING_OP_WRITE] = {
841 .unbound_nonreg_file = 1,
845 [IORING_OP_FADVISE] = {
848 [IORING_OP_MADVISE] = {
854 .unbound_nonreg_file = 1,
860 .unbound_nonreg_file = 1,
864 [IORING_OP_OPENAT2] = {
868 [IORING_OP_EPOLL_CTL] = {
869 .unbound_nonreg_file = 1,
872 [IORING_OP_SPLICE] = {
875 .unbound_nonreg_file = 1,
877 [IORING_OP_PROVIDE_BUFFERS] = {},
878 [IORING_OP_REMOVE_BUFFERS] = {},
882 .unbound_nonreg_file = 1,
886 enum io_mem_account {
891 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
892 struct io_comp_state *cs);
893 static void io_cqring_fill_event(struct io_kiocb *req, long res);
894 static void io_put_req(struct io_kiocb *req);
895 static void io_double_put_req(struct io_kiocb *req);
896 static void __io_double_put_req(struct io_kiocb *req);
897 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
898 static void __io_queue_linked_timeout(struct io_kiocb *req);
899 static void io_queue_linked_timeout(struct io_kiocb *req);
900 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
901 struct io_uring_files_update *ip,
903 static int io_prep_work_files(struct io_kiocb *req);
904 static void __io_clean_op(struct io_kiocb *req);
905 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
906 int fd, struct file **out_file, bool fixed);
907 static void __io_queue_sqe(struct io_kiocb *req,
908 const struct io_uring_sqe *sqe,
909 struct io_comp_state *cs);
910 static void io_file_put_work(struct work_struct *work);
912 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
913 struct iovec **iovec, struct iov_iter *iter,
915 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
916 const struct iovec *fast_iov,
917 struct iov_iter *iter, bool force);
919 static struct kmem_cache *req_cachep;
921 static const struct file_operations io_uring_fops;
923 struct sock *io_uring_get_socket(struct file *file)
925 #if defined(CONFIG_UNIX)
926 if (file->f_op == &io_uring_fops) {
927 struct io_ring_ctx *ctx = file->private_data;
929 return ctx->ring_sock->sk;
934 EXPORT_SYMBOL(io_uring_get_socket);
936 static void io_get_req_task(struct io_kiocb *req)
938 if (req->flags & REQ_F_TASK_PINNED)
940 get_task_struct(req->task);
941 req->flags |= REQ_F_TASK_PINNED;
944 static inline void io_clean_op(struct io_kiocb *req)
946 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
951 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
952 static void __io_put_req_task(struct io_kiocb *req)
954 if (req->flags & REQ_F_TASK_PINNED)
955 put_task_struct(req->task);
958 static void io_sq_thread_drop_mm(void)
960 struct mm_struct *mm = current->mm;
963 kthread_unuse_mm(mm);
968 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
971 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
972 !mmget_not_zero(ctx->sqo_mm)))
974 kthread_use_mm(ctx->sqo_mm);
980 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
981 struct io_kiocb *req)
983 if (!io_op_defs[req->opcode].needs_mm)
985 return __io_sq_thread_acquire_mm(ctx);
988 static inline void req_set_fail_links(struct io_kiocb *req)
990 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
991 req->flags |= REQ_F_FAIL_LINK;
995 * Note: must call io_req_init_async() for the first time you
996 * touch any members of io_wq_work.
998 static inline void io_req_init_async(struct io_kiocb *req)
1000 if (req->flags & REQ_F_WORK_INITIALIZED)
1003 memset(&req->work, 0, sizeof(req->work));
1004 req->flags |= REQ_F_WORK_INITIALIZED;
1007 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1009 return ctx->flags & IORING_SETUP_SQPOLL;
1012 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1014 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1016 complete(&ctx->ref_comp);
1019 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1021 return !req->timeout.off;
1024 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1026 struct io_ring_ctx *ctx;
1029 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1033 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1034 if (!ctx->fallback_req)
1038 * Use 5 bits less than the max cq entries, that should give us around
1039 * 32 entries per hash list if totally full and uniformly spread.
1041 hash_bits = ilog2(p->cq_entries);
1045 ctx->cancel_hash_bits = hash_bits;
1046 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1048 if (!ctx->cancel_hash)
1050 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1052 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1053 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1056 ctx->flags = p->flags;
1057 init_waitqueue_head(&ctx->sqo_wait);
1058 init_waitqueue_head(&ctx->cq_wait);
1059 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1060 init_completion(&ctx->ref_comp);
1061 init_completion(&ctx->sq_thread_comp);
1062 idr_init(&ctx->io_buffer_idr);
1063 idr_init(&ctx->personality_idr);
1064 mutex_init(&ctx->uring_lock);
1065 init_waitqueue_head(&ctx->wait);
1066 spin_lock_init(&ctx->completion_lock);
1067 INIT_LIST_HEAD(&ctx->iopoll_list);
1068 INIT_LIST_HEAD(&ctx->defer_list);
1069 INIT_LIST_HEAD(&ctx->timeout_list);
1070 init_waitqueue_head(&ctx->inflight_wait);
1071 spin_lock_init(&ctx->inflight_lock);
1072 INIT_LIST_HEAD(&ctx->inflight_list);
1073 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1074 init_llist_head(&ctx->file_put_llist);
1077 if (ctx->fallback_req)
1078 kmem_cache_free(req_cachep, ctx->fallback_req);
1079 kfree(ctx->cancel_hash);
1084 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1086 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1087 struct io_ring_ctx *ctx = req->ctx;
1089 return seq != ctx->cached_cq_tail
1090 + atomic_read(&ctx->cached_cq_overflow);
1096 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1098 struct io_rings *rings = ctx->rings;
1100 /* order cqe stores with ring update */
1101 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1103 if (wq_has_sleeper(&ctx->cq_wait)) {
1104 wake_up_interruptible(&ctx->cq_wait);
1105 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1110 * Returns true if we need to defer file table putting. This can only happen
1111 * from the error path with REQ_F_COMP_LOCKED set.
1113 static bool io_req_clean_work(struct io_kiocb *req)
1115 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1118 req->flags &= ~REQ_F_WORK_INITIALIZED;
1121 mmdrop(req->work.mm);
1122 req->work.mm = NULL;
1124 if (req->work.creds) {
1125 put_cred(req->work.creds);
1126 req->work.creds = NULL;
1129 struct fs_struct *fs = req->work.fs;
1131 if (req->flags & REQ_F_COMP_LOCKED)
1134 spin_lock(&req->work.fs->lock);
1137 spin_unlock(&req->work.fs->lock);
1140 req->work.fs = NULL;
1146 static void io_prep_async_work(struct io_kiocb *req)
1148 const struct io_op_def *def = &io_op_defs[req->opcode];
1150 io_req_init_async(req);
1152 if (req->flags & REQ_F_ISREG) {
1153 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1154 io_wq_hash_work(&req->work, file_inode(req->file));
1156 if (def->unbound_nonreg_file)
1157 req->work.flags |= IO_WQ_WORK_UNBOUND;
1159 if (!req->work.mm && def->needs_mm) {
1160 mmgrab(current->mm);
1161 req->work.mm = current->mm;
1163 if (!req->work.creds)
1164 req->work.creds = get_current_cred();
1165 if (!req->work.fs && def->needs_fs) {
1166 spin_lock(¤t->fs->lock);
1167 if (!current->fs->in_exec) {
1168 req->work.fs = current->fs;
1169 req->work.fs->users++;
1171 req->work.flags |= IO_WQ_WORK_CANCEL;
1173 spin_unlock(¤t->fs->lock);
1175 if (def->needs_fsize)
1176 req->work.fsize = rlimit(RLIMIT_FSIZE);
1178 req->work.fsize = RLIM_INFINITY;
1181 static void io_prep_async_link(struct io_kiocb *req)
1183 struct io_kiocb *cur;
1185 io_prep_async_work(req);
1186 if (req->flags & REQ_F_LINK_HEAD)
1187 list_for_each_entry(cur, &req->link_list, link_list)
1188 io_prep_async_work(cur);
1191 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1194 struct io_kiocb *link = io_prep_linked_timeout(req);
1196 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1197 &req->work, req->flags);
1198 io_wq_enqueue(ctx->io_wq, &req->work);
1202 static void io_queue_async_work(struct io_kiocb *req)
1204 struct io_kiocb *link;
1206 /* init ->work of the whole link before punting */
1207 io_prep_async_link(req);
1208 link = __io_queue_async_work(req);
1211 io_queue_linked_timeout(link);
1214 static void io_kill_timeout(struct io_kiocb *req)
1218 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1220 atomic_set(&req->ctx->cq_timeouts,
1221 atomic_read(&req->ctx->cq_timeouts) + 1);
1222 list_del_init(&req->timeout.list);
1223 req->flags |= REQ_F_COMP_LOCKED;
1224 io_cqring_fill_event(req, 0);
1229 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1231 struct io_kiocb *req, *tmp;
1233 spin_lock_irq(&ctx->completion_lock);
1234 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1235 io_kill_timeout(req);
1236 spin_unlock_irq(&ctx->completion_lock);
1239 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1242 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1243 struct io_defer_entry, list);
1244 struct io_kiocb *link;
1246 if (req_need_defer(de->req, de->seq))
1248 list_del_init(&de->list);
1249 /* punt-init is done before queueing for defer */
1250 link = __io_queue_async_work(de->req);
1252 __io_queue_linked_timeout(link);
1253 /* drop submission reference */
1254 link->flags |= REQ_F_COMP_LOCKED;
1258 } while (!list_empty(&ctx->defer_list));
1261 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1263 while (!list_empty(&ctx->timeout_list)) {
1264 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1265 struct io_kiocb, timeout.list);
1267 if (io_is_timeout_noseq(req))
1269 if (req->timeout.target_seq != ctx->cached_cq_tail
1270 - atomic_read(&ctx->cq_timeouts))
1273 list_del_init(&req->timeout.list);
1274 io_kill_timeout(req);
1278 static void io_commit_cqring(struct io_ring_ctx *ctx)
1280 io_flush_timeouts(ctx);
1281 __io_commit_cqring(ctx);
1283 if (unlikely(!list_empty(&ctx->defer_list)))
1284 __io_queue_deferred(ctx);
1287 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1289 struct io_rings *rings = ctx->rings;
1292 tail = ctx->cached_cq_tail;
1294 * writes to the cq entry need to come after reading head; the
1295 * control dependency is enough as we're using WRITE_ONCE to
1298 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1301 ctx->cached_cq_tail++;
1302 return &rings->cqes[tail & ctx->cq_mask];
1305 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1309 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1311 if (!ctx->eventfd_async)
1313 return io_wq_current_is_worker();
1316 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1318 if (waitqueue_active(&ctx->wait))
1319 wake_up(&ctx->wait);
1320 if (waitqueue_active(&ctx->sqo_wait))
1321 wake_up(&ctx->sqo_wait);
1322 if (io_should_trigger_evfd(ctx))
1323 eventfd_signal(ctx->cq_ev_fd, 1);
1326 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1328 if (list_empty(&ctx->cq_overflow_list)) {
1329 clear_bit(0, &ctx->sq_check_overflow);
1330 clear_bit(0, &ctx->cq_check_overflow);
1331 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1335 /* Returns true if there are no backlogged entries after the flush */
1336 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1338 struct io_rings *rings = ctx->rings;
1339 struct io_uring_cqe *cqe;
1340 struct io_kiocb *req;
1341 unsigned long flags;
1345 if (list_empty_careful(&ctx->cq_overflow_list))
1347 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1348 rings->cq_ring_entries))
1352 spin_lock_irqsave(&ctx->completion_lock, flags);
1354 /* if force is set, the ring is going away. always drop after that */
1356 ctx->cq_overflow_flushed = 1;
1359 while (!list_empty(&ctx->cq_overflow_list)) {
1360 cqe = io_get_cqring(ctx);
1364 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1366 list_move(&req->compl.list, &list);
1368 WRITE_ONCE(cqe->user_data, req->user_data);
1369 WRITE_ONCE(cqe->res, req->result);
1370 WRITE_ONCE(cqe->flags, req->compl.cflags);
1372 WRITE_ONCE(ctx->rings->cq_overflow,
1373 atomic_inc_return(&ctx->cached_cq_overflow));
1377 io_commit_cqring(ctx);
1378 io_cqring_mark_overflow(ctx);
1380 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1381 io_cqring_ev_posted(ctx);
1383 while (!list_empty(&list)) {
1384 req = list_first_entry(&list, struct io_kiocb, compl.list);
1385 list_del(&req->compl.list);
1392 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1394 struct io_ring_ctx *ctx = req->ctx;
1395 struct io_uring_cqe *cqe;
1397 trace_io_uring_complete(ctx, req->user_data, res);
1400 * If we can't get a cq entry, userspace overflowed the
1401 * submission (by quite a lot). Increment the overflow count in
1404 cqe = io_get_cqring(ctx);
1406 WRITE_ONCE(cqe->user_data, req->user_data);
1407 WRITE_ONCE(cqe->res, res);
1408 WRITE_ONCE(cqe->flags, cflags);
1409 } else if (ctx->cq_overflow_flushed) {
1410 WRITE_ONCE(ctx->rings->cq_overflow,
1411 atomic_inc_return(&ctx->cached_cq_overflow));
1413 if (list_empty(&ctx->cq_overflow_list)) {
1414 set_bit(0, &ctx->sq_check_overflow);
1415 set_bit(0, &ctx->cq_check_overflow);
1416 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1420 req->compl.cflags = cflags;
1421 refcount_inc(&req->refs);
1422 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1426 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1428 __io_cqring_fill_event(req, res, 0);
1431 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1433 struct io_ring_ctx *ctx = req->ctx;
1434 unsigned long flags;
1436 spin_lock_irqsave(&ctx->completion_lock, flags);
1437 __io_cqring_fill_event(req, res, cflags);
1438 io_commit_cqring(ctx);
1439 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1441 io_cqring_ev_posted(ctx);
1444 static void io_submit_flush_completions(struct io_comp_state *cs)
1446 struct io_ring_ctx *ctx = cs->ctx;
1448 spin_lock_irq(&ctx->completion_lock);
1449 while (!list_empty(&cs->list)) {
1450 struct io_kiocb *req;
1452 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1453 list_del(&req->compl.list);
1454 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1455 if (!(req->flags & REQ_F_LINK_HEAD)) {
1456 req->flags |= REQ_F_COMP_LOCKED;
1459 spin_unlock_irq(&ctx->completion_lock);
1461 spin_lock_irq(&ctx->completion_lock);
1464 io_commit_cqring(ctx);
1465 spin_unlock_irq(&ctx->completion_lock);
1467 io_cqring_ev_posted(ctx);
1471 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1472 struct io_comp_state *cs)
1475 io_cqring_add_event(req, res, cflags);
1480 req->compl.cflags = cflags;
1481 list_add_tail(&req->compl.list, &cs->list);
1483 io_submit_flush_completions(cs);
1487 static void io_req_complete(struct io_kiocb *req, long res)
1489 __io_req_complete(req, res, 0, NULL);
1492 static inline bool io_is_fallback_req(struct io_kiocb *req)
1494 return req == (struct io_kiocb *)
1495 ((unsigned long) req->ctx->fallback_req & ~1UL);
1498 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1500 struct io_kiocb *req;
1502 req = ctx->fallback_req;
1503 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1509 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1510 struct io_submit_state *state)
1512 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1513 struct io_kiocb *req;
1515 if (!state->free_reqs) {
1519 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1520 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1523 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1524 * retry single alloc to be on the safe side.
1526 if (unlikely(ret <= 0)) {
1527 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1528 if (!state->reqs[0])
1532 state->free_reqs = ret - 1;
1533 req = state->reqs[ret - 1];
1536 req = state->reqs[state->free_reqs];
1541 return io_get_fallback_req(ctx);
1544 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1548 percpu_ref_put(req->fixed_file_refs);
1553 static bool io_dismantle_req(struct io_kiocb *req)
1560 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1562 return io_req_clean_work(req);
1565 static void __io_free_req_finish(struct io_kiocb *req)
1567 struct io_ring_ctx *ctx = req->ctx;
1569 __io_put_req_task(req);
1570 if (likely(!io_is_fallback_req(req)))
1571 kmem_cache_free(req_cachep, req);
1573 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1574 percpu_ref_put(&ctx->refs);
1577 static void io_req_task_file_table_put(struct callback_head *cb)
1579 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1580 struct fs_struct *fs = req->work.fs;
1582 spin_lock(&req->work.fs->lock);
1585 spin_unlock(&req->work.fs->lock);
1588 req->work.fs = NULL;
1589 __io_free_req_finish(req);
1592 static void __io_free_req(struct io_kiocb *req)
1594 if (!io_dismantle_req(req)) {
1595 __io_free_req_finish(req);
1599 init_task_work(&req->task_work, io_req_task_file_table_put);
1600 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1601 if (unlikely(ret)) {
1602 struct task_struct *tsk;
1604 tsk = io_wq_get_task(req->ctx->io_wq);
1605 task_work_add(tsk, &req->task_work, 0);
1610 static bool io_link_cancel_timeout(struct io_kiocb *req)
1612 struct io_ring_ctx *ctx = req->ctx;
1615 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1617 io_cqring_fill_event(req, -ECANCELED);
1618 io_commit_cqring(ctx);
1619 req->flags &= ~REQ_F_LINK_HEAD;
1627 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1629 struct io_kiocb *link;
1632 if (list_empty(&req->link_list))
1634 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1635 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1638 list_del_init(&link->link_list);
1639 link->flags |= REQ_F_COMP_LOCKED;
1640 wake_ev = io_link_cancel_timeout(link);
1641 req->flags &= ~REQ_F_LINK_TIMEOUT;
1645 static void io_kill_linked_timeout(struct io_kiocb *req)
1647 struct io_ring_ctx *ctx = req->ctx;
1650 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1651 unsigned long flags;
1653 spin_lock_irqsave(&ctx->completion_lock, flags);
1654 wake_ev = __io_kill_linked_timeout(req);
1655 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1657 wake_ev = __io_kill_linked_timeout(req);
1661 io_cqring_ev_posted(ctx);
1664 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1666 struct io_kiocb *nxt;
1669 * The list should never be empty when we are called here. But could
1670 * potentially happen if the chain is messed up, check to be on the
1673 if (unlikely(list_empty(&req->link_list)))
1676 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1677 list_del_init(&req->link_list);
1678 if (!list_empty(&nxt->link_list))
1679 nxt->flags |= REQ_F_LINK_HEAD;
1684 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1686 static void __io_fail_links(struct io_kiocb *req)
1688 struct io_ring_ctx *ctx = req->ctx;
1690 while (!list_empty(&req->link_list)) {
1691 struct io_kiocb *link = list_first_entry(&req->link_list,
1692 struct io_kiocb, link_list);
1694 list_del_init(&link->link_list);
1695 trace_io_uring_fail_link(req, link);
1697 io_cqring_fill_event(link, -ECANCELED);
1698 link->flags |= REQ_F_COMP_LOCKED;
1699 __io_double_put_req(link);
1700 req->flags &= ~REQ_F_LINK_TIMEOUT;
1703 io_commit_cqring(ctx);
1704 io_cqring_ev_posted(ctx);
1707 static void io_fail_links(struct io_kiocb *req)
1709 struct io_ring_ctx *ctx = req->ctx;
1711 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1712 unsigned long flags;
1714 spin_lock_irqsave(&ctx->completion_lock, flags);
1715 __io_fail_links(req);
1716 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1718 __io_fail_links(req);
1721 io_cqring_ev_posted(ctx);
1724 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1726 req->flags &= ~REQ_F_LINK_HEAD;
1727 if (req->flags & REQ_F_LINK_TIMEOUT)
1728 io_kill_linked_timeout(req);
1731 * If LINK is set, we have dependent requests in this chain. If we
1732 * didn't fail this request, queue the first one up, moving any other
1733 * dependencies to the next request. In case of failure, fail the rest
1736 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1737 return io_req_link_next(req);
1742 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1744 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1746 return __io_req_find_next(req);
1749 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
1752 struct task_struct *tsk = req->task;
1753 struct io_ring_ctx *ctx = req->ctx;
1757 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1758 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1759 * processing task_work. There's no reliable way to tell if TWA_RESUME
1763 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1764 notify = TWA_SIGNAL;
1766 ret = task_work_add(tsk, cb, notify);
1768 wake_up_process(tsk);
1773 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1775 struct io_ring_ctx *ctx = req->ctx;
1777 spin_lock_irq(&ctx->completion_lock);
1778 io_cqring_fill_event(req, error);
1779 io_commit_cqring(ctx);
1780 spin_unlock_irq(&ctx->completion_lock);
1782 io_cqring_ev_posted(ctx);
1783 req_set_fail_links(req);
1784 io_double_put_req(req);
1787 static void io_req_task_cancel(struct callback_head *cb)
1789 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1791 __io_req_task_cancel(req, -ECANCELED);
1794 static void __io_req_task_submit(struct io_kiocb *req)
1796 struct io_ring_ctx *ctx = req->ctx;
1798 if (!__io_sq_thread_acquire_mm(ctx)) {
1799 mutex_lock(&ctx->uring_lock);
1800 __io_queue_sqe(req, NULL, NULL);
1801 mutex_unlock(&ctx->uring_lock);
1803 __io_req_task_cancel(req, -EFAULT);
1807 static void io_req_task_submit(struct callback_head *cb)
1809 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1810 struct io_ring_ctx *ctx = req->ctx;
1812 __io_req_task_submit(req);
1813 percpu_ref_put(&ctx->refs);
1816 static void io_req_task_queue(struct io_kiocb *req)
1820 init_task_work(&req->task_work, io_req_task_submit);
1821 percpu_ref_get(&req->ctx->refs);
1823 ret = io_req_task_work_add(req, &req->task_work, true);
1824 if (unlikely(ret)) {
1825 struct task_struct *tsk;
1827 init_task_work(&req->task_work, io_req_task_cancel);
1828 tsk = io_wq_get_task(req->ctx->io_wq);
1829 task_work_add(tsk, &req->task_work, 0);
1830 wake_up_process(tsk);
1834 static void io_queue_next(struct io_kiocb *req)
1836 struct io_kiocb *nxt = io_req_find_next(req);
1839 io_req_task_queue(nxt);
1842 static void io_free_req(struct io_kiocb *req)
1849 void *reqs[IO_IOPOLL_BATCH];
1852 struct task_struct *task;
1856 static inline void io_init_req_batch(struct req_batch *rb)
1863 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1864 struct req_batch *rb)
1866 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1867 percpu_ref_put_many(&ctx->refs, rb->to_free);
1871 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1872 struct req_batch *rb)
1875 __io_req_free_batch_flush(ctx, rb);
1877 put_task_struct_many(rb->task, rb->task_refs);
1882 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1884 if (unlikely(io_is_fallback_req(req))) {
1888 if (req->flags & REQ_F_LINK_HEAD)
1891 if (req->flags & REQ_F_TASK_PINNED) {
1892 if (req->task != rb->task) {
1894 put_task_struct_many(rb->task, rb->task_refs);
1895 rb->task = req->task;
1899 req->flags &= ~REQ_F_TASK_PINNED;
1902 WARN_ON_ONCE(io_dismantle_req(req));
1903 rb->reqs[rb->to_free++] = req;
1904 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1905 __io_req_free_batch_flush(req->ctx, rb);
1909 * Drop reference to request, return next in chain (if there is one) if this
1910 * was the last reference to this request.
1912 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1914 struct io_kiocb *nxt = NULL;
1916 if (refcount_dec_and_test(&req->refs)) {
1917 nxt = io_req_find_next(req);
1923 static void io_put_req(struct io_kiocb *req)
1925 if (refcount_dec_and_test(&req->refs))
1929 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1931 struct io_kiocb *nxt;
1934 * A ref is owned by io-wq in which context we're. So, if that's the
1935 * last one, it's safe to steal next work. False negatives are Ok,
1936 * it just will be re-punted async in io_put_work()
1938 if (refcount_read(&req->refs) != 1)
1941 nxt = io_req_find_next(req);
1942 return nxt ? &nxt->work : NULL;
1946 * Must only be used if we don't need to care about links, usually from
1947 * within the completion handling itself.
1949 static void __io_double_put_req(struct io_kiocb *req)
1951 /* drop both submit and complete references */
1952 if (refcount_sub_and_test(2, &req->refs))
1956 static void io_double_put_req(struct io_kiocb *req)
1958 /* drop both submit and complete references */
1959 if (refcount_sub_and_test(2, &req->refs))
1963 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1965 struct io_rings *rings = ctx->rings;
1967 if (test_bit(0, &ctx->cq_check_overflow)) {
1969 * noflush == true is from the waitqueue handler, just ensure
1970 * we wake up the task, and the next invocation will flush the
1971 * entries. We cannot safely to it from here.
1973 if (noflush && !list_empty(&ctx->cq_overflow_list))
1976 io_cqring_overflow_flush(ctx, false);
1979 /* See comment at the top of this file */
1981 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1984 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1986 struct io_rings *rings = ctx->rings;
1988 /* make sure SQ entry isn't read before tail */
1989 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1992 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1994 unsigned int cflags;
1996 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1997 cflags |= IORING_CQE_F_BUFFER;
1998 req->flags &= ~REQ_F_BUFFER_SELECTED;
2003 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2005 struct io_buffer *kbuf;
2007 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2008 return io_put_kbuf(req, kbuf);
2011 static inline bool io_run_task_work(void)
2013 if (current->task_works) {
2014 __set_current_state(TASK_RUNNING);
2022 static void io_iopoll_queue(struct list_head *again)
2024 struct io_kiocb *req;
2027 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2028 list_del(&req->inflight_entry);
2029 __io_complete_rw(req, -EAGAIN, 0, NULL);
2030 } while (!list_empty(again));
2034 * Find and free completed poll iocbs
2036 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2037 struct list_head *done)
2039 struct req_batch rb;
2040 struct io_kiocb *req;
2043 /* order with ->result store in io_complete_rw_iopoll() */
2046 io_init_req_batch(&rb);
2047 while (!list_empty(done)) {
2050 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2051 if (READ_ONCE(req->result) == -EAGAIN) {
2053 req->iopoll_completed = 0;
2054 list_move_tail(&req->inflight_entry, &again);
2057 list_del(&req->inflight_entry);
2059 if (req->flags & REQ_F_BUFFER_SELECTED)
2060 cflags = io_put_rw_kbuf(req);
2062 __io_cqring_fill_event(req, req->result, cflags);
2065 if (refcount_dec_and_test(&req->refs))
2066 io_req_free_batch(&rb, req);
2069 io_commit_cqring(ctx);
2070 if (ctx->flags & IORING_SETUP_SQPOLL)
2071 io_cqring_ev_posted(ctx);
2072 io_req_free_batch_finish(ctx, &rb);
2074 if (!list_empty(&again))
2075 io_iopoll_queue(&again);
2078 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2081 struct io_kiocb *req, *tmp;
2087 * Only spin for completions if we don't have multiple devices hanging
2088 * off our complete list, and we're under the requested amount.
2090 spin = !ctx->poll_multi_file && *nr_events < min;
2093 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2094 struct kiocb *kiocb = &req->rw.kiocb;
2097 * Move completed and retryable entries to our local lists.
2098 * If we find a request that requires polling, break out
2099 * and complete those lists first, if we have entries there.
2101 if (READ_ONCE(req->iopoll_completed)) {
2102 list_move_tail(&req->inflight_entry, &done);
2105 if (!list_empty(&done))
2108 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2112 /* iopoll may have completed current req */
2113 if (READ_ONCE(req->iopoll_completed))
2114 list_move_tail(&req->inflight_entry, &done);
2121 if (!list_empty(&done))
2122 io_iopoll_complete(ctx, nr_events, &done);
2128 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2129 * non-spinning poll check - we'll still enter the driver poll loop, but only
2130 * as a non-spinning completion check.
2132 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2135 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2138 ret = io_do_iopoll(ctx, nr_events, min);
2141 if (*nr_events >= min)
2149 * We can't just wait for polled events to come to us, we have to actively
2150 * find and complete them.
2152 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2154 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2157 mutex_lock(&ctx->uring_lock);
2158 while (!list_empty(&ctx->iopoll_list)) {
2159 unsigned int nr_events = 0;
2161 io_do_iopoll(ctx, &nr_events, 0);
2163 /* let it sleep and repeat later if can't complete a request */
2167 * Ensure we allow local-to-the-cpu processing to take place,
2168 * in this case we need to ensure that we reap all events.
2169 * Also let task_work, etc. to progress by releasing the mutex
2171 if (need_resched()) {
2172 mutex_unlock(&ctx->uring_lock);
2174 mutex_lock(&ctx->uring_lock);
2177 mutex_unlock(&ctx->uring_lock);
2180 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2182 unsigned int nr_events = 0;
2183 int iters = 0, ret = 0;
2186 * We disallow the app entering submit/complete with polling, but we
2187 * still need to lock the ring to prevent racing with polled issue
2188 * that got punted to a workqueue.
2190 mutex_lock(&ctx->uring_lock);
2193 * Don't enter poll loop if we already have events pending.
2194 * If we do, we can potentially be spinning for commands that
2195 * already triggered a CQE (eg in error).
2197 if (io_cqring_events(ctx, false))
2201 * If a submit got punted to a workqueue, we can have the
2202 * application entering polling for a command before it gets
2203 * issued. That app will hold the uring_lock for the duration
2204 * of the poll right here, so we need to take a breather every
2205 * now and then to ensure that the issue has a chance to add
2206 * the poll to the issued list. Otherwise we can spin here
2207 * forever, while the workqueue is stuck trying to acquire the
2210 if (!(++iters & 7)) {
2211 mutex_unlock(&ctx->uring_lock);
2213 mutex_lock(&ctx->uring_lock);
2216 ret = io_iopoll_getevents(ctx, &nr_events, min);
2220 } while (min && !nr_events && !need_resched());
2222 mutex_unlock(&ctx->uring_lock);
2226 static void kiocb_end_write(struct io_kiocb *req)
2229 * Tell lockdep we inherited freeze protection from submission
2232 if (req->flags & REQ_F_ISREG) {
2233 struct inode *inode = file_inode(req->file);
2235 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2237 file_end_write(req->file);
2240 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2241 struct io_comp_state *cs)
2243 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2246 if (kiocb->ki_flags & IOCB_WRITE)
2247 kiocb_end_write(req);
2249 if (res != req->result)
2250 req_set_fail_links(req);
2251 if (req->flags & REQ_F_BUFFER_SELECTED)
2252 cflags = io_put_rw_kbuf(req);
2253 __io_req_complete(req, res, cflags, cs);
2257 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2259 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2260 ssize_t ret = -ECANCELED;
2261 struct iov_iter iter;
2269 switch (req->opcode) {
2270 case IORING_OP_READV:
2271 case IORING_OP_READ_FIXED:
2272 case IORING_OP_READ:
2275 case IORING_OP_WRITEV:
2276 case IORING_OP_WRITE_FIXED:
2277 case IORING_OP_WRITE:
2281 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2286 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2289 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2294 req_set_fail_links(req);
2295 io_req_complete(req, ret);
2300 static bool io_rw_reissue(struct io_kiocb *req, long res)
2303 umode_t mode = file_inode(req->file)->i_mode;
2306 if (!S_ISBLK(mode) && !S_ISREG(mode))
2308 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2311 ret = io_sq_thread_acquire_mm(req->ctx, req);
2313 if (io_resubmit_prep(req, ret)) {
2314 refcount_inc(&req->refs);
2315 io_queue_async_work(req);
2323 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2324 struct io_comp_state *cs)
2326 if (!io_rw_reissue(req, res))
2327 io_complete_rw_common(&req->rw.kiocb, res, cs);
2330 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2332 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2334 __io_complete_rw(req, res, res2, NULL);
2337 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2339 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2341 if (kiocb->ki_flags & IOCB_WRITE)
2342 kiocb_end_write(req);
2344 if (res != -EAGAIN && res != req->result)
2345 req_set_fail_links(req);
2347 WRITE_ONCE(req->result, res);
2348 /* order with io_poll_complete() checking ->result */
2350 WRITE_ONCE(req->iopoll_completed, 1);
2354 * After the iocb has been issued, it's safe to be found on the poll list.
2355 * Adding the kiocb to the list AFTER submission ensures that we don't
2356 * find it from a io_iopoll_getevents() thread before the issuer is done
2357 * accessing the kiocb cookie.
2359 static void io_iopoll_req_issued(struct io_kiocb *req)
2361 struct io_ring_ctx *ctx = req->ctx;
2364 * Track whether we have multiple files in our lists. This will impact
2365 * how we do polling eventually, not spinning if we're on potentially
2366 * different devices.
2368 if (list_empty(&ctx->iopoll_list)) {
2369 ctx->poll_multi_file = false;
2370 } else if (!ctx->poll_multi_file) {
2371 struct io_kiocb *list_req;
2373 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2375 if (list_req->file != req->file)
2376 ctx->poll_multi_file = true;
2380 * For fast devices, IO may have already completed. If it has, add
2381 * it to the front so we find it first.
2383 if (READ_ONCE(req->iopoll_completed))
2384 list_add(&req->inflight_entry, &ctx->iopoll_list);
2386 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2388 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2389 wq_has_sleeper(&ctx->sqo_wait))
2390 wake_up(&ctx->sqo_wait);
2393 static void __io_state_file_put(struct io_submit_state *state)
2395 if (state->has_refs)
2396 fput_many(state->file, state->has_refs);
2400 static inline void io_state_file_put(struct io_submit_state *state)
2403 __io_state_file_put(state);
2407 * Get as many references to a file as we have IOs left in this submission,
2408 * assuming most submissions are for one file, or at least that each file
2409 * has more than one submission.
2411 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2417 if (state->fd == fd) {
2422 __io_state_file_put(state);
2424 state->file = fget_many(fd, state->ios_left);
2430 state->has_refs = state->ios_left;
2434 static bool io_bdev_nowait(struct block_device *bdev)
2437 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2444 * If we tracked the file through the SCM inflight mechanism, we could support
2445 * any file. For now, just ensure that anything potentially problematic is done
2448 static bool io_file_supports_async(struct file *file, int rw)
2450 umode_t mode = file_inode(file)->i_mode;
2452 if (S_ISBLK(mode)) {
2453 if (io_bdev_nowait(file->f_inode->i_bdev))
2457 if (S_ISCHR(mode) || S_ISSOCK(mode))
2459 if (S_ISREG(mode)) {
2460 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2461 file->f_op != &io_uring_fops)
2466 /* any ->read/write should understand O_NONBLOCK */
2467 if (file->f_flags & O_NONBLOCK)
2470 if (!(file->f_mode & FMODE_NOWAIT))
2474 return file->f_op->read_iter != NULL;
2476 return file->f_op->write_iter != NULL;
2479 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2480 bool force_nonblock)
2482 struct io_ring_ctx *ctx = req->ctx;
2483 struct kiocb *kiocb = &req->rw.kiocb;
2487 if (S_ISREG(file_inode(req->file)->i_mode))
2488 req->flags |= REQ_F_ISREG;
2490 kiocb->ki_pos = READ_ONCE(sqe->off);
2491 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2492 req->flags |= REQ_F_CUR_POS;
2493 kiocb->ki_pos = req->file->f_pos;
2495 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2496 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2497 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2501 ioprio = READ_ONCE(sqe->ioprio);
2503 ret = ioprio_check_cap(ioprio);
2507 kiocb->ki_ioprio = ioprio;
2509 kiocb->ki_ioprio = get_current_ioprio();
2511 /* don't allow async punt if RWF_NOWAIT was requested */
2512 if (kiocb->ki_flags & IOCB_NOWAIT)
2513 req->flags |= REQ_F_NOWAIT;
2515 if (kiocb->ki_flags & IOCB_DIRECT)
2516 io_get_req_task(req);
2519 kiocb->ki_flags |= IOCB_NOWAIT;
2521 if (ctx->flags & IORING_SETUP_IOPOLL) {
2522 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2523 !kiocb->ki_filp->f_op->iopoll)
2526 kiocb->ki_flags |= IOCB_HIPRI;
2527 kiocb->ki_complete = io_complete_rw_iopoll;
2528 req->iopoll_completed = 0;
2529 io_get_req_task(req);
2531 if (kiocb->ki_flags & IOCB_HIPRI)
2533 kiocb->ki_complete = io_complete_rw;
2536 req->rw.addr = READ_ONCE(sqe->addr);
2537 req->rw.len = READ_ONCE(sqe->len);
2538 req->buf_index = READ_ONCE(sqe->buf_index);
2542 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2548 case -ERESTARTNOINTR:
2549 case -ERESTARTNOHAND:
2550 case -ERESTART_RESTARTBLOCK:
2552 * We can't just restart the syscall, since previously
2553 * submitted sqes may already be in progress. Just fail this
2559 kiocb->ki_complete(kiocb, ret, 0);
2563 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2564 struct io_comp_state *cs)
2566 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2568 /* add previously done IO, if any */
2569 if (req->io && req->io->rw.bytes_done > 0) {
2571 ret = req->io->rw.bytes_done;
2573 ret += req->io->rw.bytes_done;
2576 if (req->flags & REQ_F_CUR_POS)
2577 req->file->f_pos = kiocb->ki_pos;
2578 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2579 __io_complete_rw(req, ret, 0, cs);
2581 io_rw_done(kiocb, ret);
2584 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2585 struct iov_iter *iter)
2587 struct io_ring_ctx *ctx = req->ctx;
2588 size_t len = req->rw.len;
2589 struct io_mapped_ubuf *imu;
2590 u16 index, buf_index;
2594 /* attempt to use fixed buffers without having provided iovecs */
2595 if (unlikely(!ctx->user_bufs))
2598 buf_index = req->buf_index;
2599 if (unlikely(buf_index >= ctx->nr_user_bufs))
2602 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2603 imu = &ctx->user_bufs[index];
2604 buf_addr = req->rw.addr;
2607 if (buf_addr + len < buf_addr)
2609 /* not inside the mapped region */
2610 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2614 * May not be a start of buffer, set size appropriately
2615 * and advance us to the beginning.
2617 offset = buf_addr - imu->ubuf;
2618 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2622 * Don't use iov_iter_advance() here, as it's really slow for
2623 * using the latter parts of a big fixed buffer - it iterates
2624 * over each segment manually. We can cheat a bit here, because
2627 * 1) it's a BVEC iter, we set it up
2628 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2629 * first and last bvec
2631 * So just find our index, and adjust the iterator afterwards.
2632 * If the offset is within the first bvec (or the whole first
2633 * bvec, just use iov_iter_advance(). This makes it easier
2634 * since we can just skip the first segment, which may not
2635 * be PAGE_SIZE aligned.
2637 const struct bio_vec *bvec = imu->bvec;
2639 if (offset <= bvec->bv_len) {
2640 iov_iter_advance(iter, offset);
2642 unsigned long seg_skip;
2644 /* skip first vec */
2645 offset -= bvec->bv_len;
2646 seg_skip = 1 + (offset >> PAGE_SHIFT);
2648 iter->bvec = bvec + seg_skip;
2649 iter->nr_segs -= seg_skip;
2650 iter->count -= bvec->bv_len + offset;
2651 iter->iov_offset = offset & ~PAGE_MASK;
2658 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2661 mutex_unlock(&ctx->uring_lock);
2664 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2667 * "Normal" inline submissions always hold the uring_lock, since we
2668 * grab it from the system call. Same is true for the SQPOLL offload.
2669 * The only exception is when we've detached the request and issue it
2670 * from an async worker thread, grab the lock for that case.
2673 mutex_lock(&ctx->uring_lock);
2676 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2677 int bgid, struct io_buffer *kbuf,
2680 struct io_buffer *head;
2682 if (req->flags & REQ_F_BUFFER_SELECTED)
2685 io_ring_submit_lock(req->ctx, needs_lock);
2687 lockdep_assert_held(&req->ctx->uring_lock);
2689 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2691 if (!list_empty(&head->list)) {
2692 kbuf = list_last_entry(&head->list, struct io_buffer,
2694 list_del(&kbuf->list);
2697 idr_remove(&req->ctx->io_buffer_idr, bgid);
2699 if (*len > kbuf->len)
2702 kbuf = ERR_PTR(-ENOBUFS);
2705 io_ring_submit_unlock(req->ctx, needs_lock);
2710 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2713 struct io_buffer *kbuf;
2716 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2717 bgid = req->buf_index;
2718 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2721 req->rw.addr = (u64) (unsigned long) kbuf;
2722 req->flags |= REQ_F_BUFFER_SELECTED;
2723 return u64_to_user_ptr(kbuf->addr);
2726 #ifdef CONFIG_COMPAT
2727 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2730 struct compat_iovec __user *uiov;
2731 compat_ssize_t clen;
2735 uiov = u64_to_user_ptr(req->rw.addr);
2736 if (!access_ok(uiov, sizeof(*uiov)))
2738 if (__get_user(clen, &uiov->iov_len))
2744 buf = io_rw_buffer_select(req, &len, needs_lock);
2746 return PTR_ERR(buf);
2747 iov[0].iov_base = buf;
2748 iov[0].iov_len = (compat_size_t) len;
2753 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2756 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2760 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2763 len = iov[0].iov_len;
2766 buf = io_rw_buffer_select(req, &len, needs_lock);
2768 return PTR_ERR(buf);
2769 iov[0].iov_base = buf;
2770 iov[0].iov_len = len;
2774 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2777 if (req->flags & REQ_F_BUFFER_SELECTED) {
2778 struct io_buffer *kbuf;
2780 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2781 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2782 iov[0].iov_len = kbuf->len;
2787 else if (req->rw.len > 1)
2790 #ifdef CONFIG_COMPAT
2791 if (req->ctx->compat)
2792 return io_compat_import(req, iov, needs_lock);
2795 return __io_iov_buffer_select(req, iov, needs_lock);
2798 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2799 struct iovec **iovec, struct iov_iter *iter,
2802 void __user *buf = u64_to_user_ptr(req->rw.addr);
2803 size_t sqe_len = req->rw.len;
2807 opcode = req->opcode;
2808 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2810 return io_import_fixed(req, rw, iter);
2813 /* buffer index only valid with fixed read/write, or buffer select */
2814 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2817 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2818 if (req->flags & REQ_F_BUFFER_SELECT) {
2819 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2821 return PTR_ERR(buf);
2822 req->rw.len = sqe_len;
2825 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2827 return ret < 0 ? ret : sqe_len;
2830 if (req->flags & REQ_F_BUFFER_SELECT) {
2831 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2833 ret = (*iovec)->iov_len;
2834 iov_iter_init(iter, rw, *iovec, 1, ret);
2840 #ifdef CONFIG_COMPAT
2841 if (req->ctx->compat)
2842 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2846 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2849 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2850 struct iovec **iovec, struct iov_iter *iter,
2854 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2856 return iov_iter_count(&req->io->rw.iter);
2859 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2861 return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos;
2865 * For files that don't have ->read_iter() and ->write_iter(), handle them
2866 * by looping over ->read() or ->write() manually.
2868 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2869 struct iov_iter *iter)
2874 * Don't support polled IO through this interface, and we can't
2875 * support non-blocking either. For the latter, this just causes
2876 * the kiocb to be handled from an async context.
2878 if (kiocb->ki_flags & IOCB_HIPRI)
2880 if (kiocb->ki_flags & IOCB_NOWAIT)
2883 while (iov_iter_count(iter)) {
2887 if (!iov_iter_is_bvec(iter)) {
2888 iovec = iov_iter_iovec(iter);
2890 /* fixed buffers import bvec */
2891 iovec.iov_base = kmap(iter->bvec->bv_page)
2893 iovec.iov_len = min(iter->count,
2894 iter->bvec->bv_len - iter->iov_offset);
2898 nr = file->f_op->read(file, iovec.iov_base,
2899 iovec.iov_len, io_kiocb_ppos(kiocb));
2901 nr = file->f_op->write(file, iovec.iov_base,
2902 iovec.iov_len, io_kiocb_ppos(kiocb));
2905 if (iov_iter_is_bvec(iter))
2906 kunmap(iter->bvec->bv_page);
2914 if (nr != iovec.iov_len)
2916 iov_iter_advance(iter, nr);
2922 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2923 const struct iovec *fast_iov, struct iov_iter *iter)
2925 struct io_async_rw *rw = &req->io->rw;
2927 memcpy(&rw->iter, iter, sizeof(*iter));
2928 rw->free_iovec = NULL;
2930 /* can only be fixed buffers, no need to do anything */
2931 if (iter->type == ITER_BVEC)
2934 unsigned iov_off = 0;
2936 rw->iter.iov = rw->fast_iov;
2937 if (iter->iov != fast_iov) {
2938 iov_off = iter->iov - fast_iov;
2939 rw->iter.iov += iov_off;
2941 if (rw->fast_iov != fast_iov)
2942 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
2943 sizeof(struct iovec) * iter->nr_segs);
2945 rw->free_iovec = iovec;
2946 req->flags |= REQ_F_NEED_CLEANUP;
2950 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2952 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2953 return req->io == NULL;
2956 static int io_alloc_async_ctx(struct io_kiocb *req)
2958 if (!io_op_defs[req->opcode].async_ctx)
2961 return __io_alloc_async_ctx(req);
2964 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
2965 const struct iovec *fast_iov,
2966 struct iov_iter *iter, bool force)
2968 if (!force && !io_op_defs[req->opcode].async_ctx)
2971 if (__io_alloc_async_ctx(req))
2974 io_req_map_rw(req, iovec, fast_iov, iter);
2979 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2980 bool force_nonblock)
2982 struct io_async_rw *iorw = &req->io->rw;
2985 iorw->iter.iov = iorw->fast_iov;
2986 ret = __io_import_iovec(rw, req, (struct iovec **) &iorw->iter.iov,
2987 &iorw->iter, !force_nonblock);
2988 if (unlikely(ret < 0))
2991 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
2995 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2996 bool force_nonblock)
3000 ret = io_prep_rw(req, sqe, force_nonblock);
3004 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3007 /* either don't need iovec imported or already have it */
3008 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3010 return io_rw_prep_async(req, READ, force_nonblock);
3014 * This is our waitqueue callback handler, registered through lock_page_async()
3015 * when we initially tried to do the IO with the iocb armed our waitqueue.
3016 * This gets called when the page is unlocked, and we generally expect that to
3017 * happen when the page IO is completed and the page is now uptodate. This will
3018 * queue a task_work based retry of the operation, attempting to copy the data
3019 * again. If the latter fails because the page was NOT uptodate, then we will
3020 * do a thread based blocking retry of the operation. That's the unexpected
3023 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3024 int sync, void *arg)
3026 struct wait_page_queue *wpq;
3027 struct io_kiocb *req = wait->private;
3028 struct wait_page_key *key = arg;
3031 wpq = container_of(wait, struct wait_page_queue, wait);
3033 if (!wake_page_match(wpq, key))
3036 list_del_init(&wait->entry);
3038 init_task_work(&req->task_work, io_req_task_submit);
3039 percpu_ref_get(&req->ctx->refs);
3041 /* submit ref gets dropped, acquire a new one */
3042 refcount_inc(&req->refs);
3043 ret = io_req_task_work_add(req, &req->task_work, true);
3044 if (unlikely(ret)) {
3045 struct task_struct *tsk;
3047 /* queue just for cancelation */
3048 init_task_work(&req->task_work, io_req_task_cancel);
3049 tsk = io_wq_get_task(req->ctx->io_wq);
3050 task_work_add(tsk, &req->task_work, 0);
3051 wake_up_process(tsk);
3057 * This controls whether a given IO request should be armed for async page
3058 * based retry. If we return false here, the request is handed to the async
3059 * worker threads for retry. If we're doing buffered reads on a regular file,
3060 * we prepare a private wait_page_queue entry and retry the operation. This
3061 * will either succeed because the page is now uptodate and unlocked, or it
3062 * will register a callback when the page is unlocked at IO completion. Through
3063 * that callback, io_uring uses task_work to setup a retry of the operation.
3064 * That retry will attempt the buffered read again. The retry will generally
3065 * succeed, or in rare cases where it fails, we then fall back to using the
3066 * async worker threads for a blocking retry.
3068 static bool io_rw_should_retry(struct io_kiocb *req)
3070 struct wait_page_queue *wait = &req->io->rw.wpq;
3071 struct kiocb *kiocb = &req->rw.kiocb;
3073 /* never retry for NOWAIT, we just complete with -EAGAIN */
3074 if (req->flags & REQ_F_NOWAIT)
3077 /* Only for buffered IO */
3078 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3082 * just use poll if we can, and don't attempt if the fs doesn't
3083 * support callback based unlocks
3085 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3088 wait->wait.func = io_async_buf_func;
3089 wait->wait.private = req;
3090 wait->wait.flags = 0;
3091 INIT_LIST_HEAD(&wait->wait.entry);
3092 kiocb->ki_flags |= IOCB_WAITQ;
3093 kiocb->ki_waitq = wait;
3095 io_get_req_task(req);
3099 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3101 if (req->file->f_op->read_iter)
3102 return call_read_iter(req->file, &req->rw.kiocb, iter);
3103 else if (req->file->f_op->read)
3104 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3109 static int io_read(struct io_kiocb *req, bool force_nonblock,
3110 struct io_comp_state *cs)
3112 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3113 struct kiocb *kiocb = &req->rw.kiocb;
3114 struct iov_iter __iter, *iter = &__iter;
3115 ssize_t io_size, ret, ret2;
3119 iter = &req->io->rw.iter;
3121 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3124 iov_count = iov_iter_count(iter);
3126 req->result = io_size;
3129 /* Ensure we clear previously set non-block flag */
3130 if (!force_nonblock)
3131 kiocb->ki_flags &= ~IOCB_NOWAIT;
3133 /* If the file doesn't support async, just async punt */
3134 if (force_nonblock && !io_file_supports_async(req->file, READ))
3137 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3141 ret = io_iter_do_read(req, iter);
3145 } else if (ret == -EIOCBQUEUED) {
3148 } else if (ret == -EAGAIN) {
3149 /* IOPOLL retry should happen for io-wq threads */
3150 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3152 /* no retry on NONBLOCK marked file */
3153 if (req->file->f_flags & O_NONBLOCK)
3155 /* some cases will consume bytes even on error returns */
3156 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3157 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3161 } else if (ret < 0) {
3162 /* make sure -ERESTARTSYS -> -EINTR is done */
3166 /* read it all, or we did blocking attempt. no retry. */
3167 if (!iov_iter_count(iter) || !force_nonblock ||
3168 (req->file->f_flags & O_NONBLOCK))
3173 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3178 /* it's copied and will be cleaned with ->io */
3180 /* now use our persistent iterator, if we aren't already */
3181 iter = &req->io->rw.iter;
3183 req->io->rw.bytes_done += ret;
3184 /* if we can retry, do so with the callbacks armed */
3185 if (!io_rw_should_retry(req)) {
3186 kiocb->ki_flags &= ~IOCB_WAITQ;
3191 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3192 * get -EIOCBQUEUED, then we'll get a notification when the desired
3193 * page gets unlocked. We can also get a partial read here, and if we
3194 * do, then just retry at the new offset.
3196 ret = io_iter_do_read(req, iter);
3197 if (ret == -EIOCBQUEUED) {
3200 } else if (ret > 0 && ret < io_size) {
3201 /* we got some bytes, but not all. retry. */
3205 kiocb_done(kiocb, ret, cs);
3208 /* it's reportedly faster than delegating the null check to kfree() */
3214 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3215 bool force_nonblock)
3219 ret = io_prep_rw(req, sqe, force_nonblock);
3223 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3226 /* either don't need iovec imported or already have it */
3227 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3229 return io_rw_prep_async(req, WRITE, force_nonblock);
3232 static int io_write(struct io_kiocb *req, bool force_nonblock,
3233 struct io_comp_state *cs)
3235 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3236 struct kiocb *kiocb = &req->rw.kiocb;
3237 struct iov_iter __iter, *iter = &__iter;
3239 ssize_t ret, ret2, io_size;
3242 iter = &req->io->rw.iter;
3244 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3247 iov_count = iov_iter_count(iter);
3249 req->result = io_size;
3251 /* Ensure we clear previously set non-block flag */
3252 if (!force_nonblock)
3253 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3255 /* If the file doesn't support async, just async punt */
3256 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3259 /* file path doesn't support NOWAIT for non-direct_IO */
3260 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3261 (req->flags & REQ_F_ISREG))
3264 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3269 * Open-code file_start_write here to grab freeze protection,
3270 * which will be released by another thread in
3271 * io_complete_rw(). Fool lockdep by telling it the lock got
3272 * released so that it doesn't complain about the held lock when
3273 * we return to userspace.
3275 if (req->flags & REQ_F_ISREG) {
3276 __sb_start_write(file_inode(req->file)->i_sb,
3277 SB_FREEZE_WRITE, true);
3278 __sb_writers_release(file_inode(req->file)->i_sb,
3281 kiocb->ki_flags |= IOCB_WRITE;
3283 if (req->file->f_op->write_iter)
3284 ret2 = call_write_iter(req->file, kiocb, iter);
3285 else if (req->file->f_op->write)
3286 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3291 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3292 * retry them without IOCB_NOWAIT.
3294 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3296 /* no retry on NONBLOCK marked file */
3297 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3299 if (!force_nonblock || ret2 != -EAGAIN) {
3300 /* IOPOLL retry should happen for io-wq threads */
3301 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3304 kiocb_done(kiocb, ret2, cs);
3307 /* some cases will consume bytes even on error returns */
3308 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3309 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3314 /* it's reportedly faster than delegating the null check to kfree() */
3320 static int __io_splice_prep(struct io_kiocb *req,
3321 const struct io_uring_sqe *sqe)
3323 struct io_splice* sp = &req->splice;
3324 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3327 if (req->flags & REQ_F_NEED_CLEANUP)
3329 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3333 sp->len = READ_ONCE(sqe->len);
3334 sp->flags = READ_ONCE(sqe->splice_flags);
3336 if (unlikely(sp->flags & ~valid_flags))
3339 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3340 (sp->flags & SPLICE_F_FD_IN_FIXED));
3343 req->flags |= REQ_F_NEED_CLEANUP;
3345 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3347 * Splice operation will be punted aync, and here need to
3348 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3350 io_req_init_async(req);
3351 req->work.flags |= IO_WQ_WORK_UNBOUND;
3357 static int io_tee_prep(struct io_kiocb *req,
3358 const struct io_uring_sqe *sqe)
3360 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3362 return __io_splice_prep(req, sqe);
3365 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3367 struct io_splice *sp = &req->splice;
3368 struct file *in = sp->file_in;
3369 struct file *out = sp->file_out;
3370 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3376 ret = do_tee(in, out, sp->len, flags);
3378 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3379 req->flags &= ~REQ_F_NEED_CLEANUP;
3382 req_set_fail_links(req);
3383 io_req_complete(req, ret);
3387 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3389 struct io_splice* sp = &req->splice;
3391 sp->off_in = READ_ONCE(sqe->splice_off_in);
3392 sp->off_out = READ_ONCE(sqe->off);
3393 return __io_splice_prep(req, sqe);
3396 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3398 struct io_splice *sp = &req->splice;
3399 struct file *in = sp->file_in;
3400 struct file *out = sp->file_out;
3401 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3402 loff_t *poff_in, *poff_out;
3408 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3409 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3412 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3414 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3415 req->flags &= ~REQ_F_NEED_CLEANUP;
3418 req_set_fail_links(req);
3419 io_req_complete(req, ret);
3424 * IORING_OP_NOP just posts a completion event, nothing else.
3426 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3428 struct io_ring_ctx *ctx = req->ctx;
3430 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3433 __io_req_complete(req, 0, 0, cs);
3437 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3439 struct io_ring_ctx *ctx = req->ctx;
3444 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3446 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3449 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3450 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3453 req->sync.off = READ_ONCE(sqe->off);
3454 req->sync.len = READ_ONCE(sqe->len);
3458 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3460 loff_t end = req->sync.off + req->sync.len;
3463 /* fsync always requires a blocking context */
3467 ret = vfs_fsync_range(req->file, req->sync.off,
3468 end > 0 ? end : LLONG_MAX,
3469 req->sync.flags & IORING_FSYNC_DATASYNC);
3471 req_set_fail_links(req);
3472 io_req_complete(req, ret);
3476 static int io_fallocate_prep(struct io_kiocb *req,
3477 const struct io_uring_sqe *sqe)
3479 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3481 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3484 req->sync.off = READ_ONCE(sqe->off);
3485 req->sync.len = READ_ONCE(sqe->addr);
3486 req->sync.mode = READ_ONCE(sqe->len);
3490 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3494 /* fallocate always requiring blocking context */
3497 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3500 req_set_fail_links(req);
3501 io_req_complete(req, ret);
3505 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3507 const char __user *fname;
3510 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3512 if (unlikely(sqe->ioprio || sqe->buf_index))
3514 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3517 /* open.how should be already initialised */
3518 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3519 req->open.how.flags |= O_LARGEFILE;
3521 req->open.dfd = READ_ONCE(sqe->fd);
3522 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3523 req->open.filename = getname(fname);
3524 if (IS_ERR(req->open.filename)) {
3525 ret = PTR_ERR(req->open.filename);
3526 req->open.filename = NULL;
3529 req->open.nofile = rlimit(RLIMIT_NOFILE);
3530 req->flags |= REQ_F_NEED_CLEANUP;
3534 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3538 if (req->flags & REQ_F_NEED_CLEANUP)
3540 mode = READ_ONCE(sqe->len);
3541 flags = READ_ONCE(sqe->open_flags);
3542 req->open.how = build_open_how(flags, mode);
3543 return __io_openat_prep(req, sqe);
3546 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3548 struct open_how __user *how;
3552 if (req->flags & REQ_F_NEED_CLEANUP)
3554 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3555 len = READ_ONCE(sqe->len);
3556 if (len < OPEN_HOW_SIZE_VER0)
3559 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3564 return __io_openat_prep(req, sqe);
3567 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3569 struct open_flags op;
3576 ret = build_open_flags(&req->open.how, &op);
3580 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3584 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3587 ret = PTR_ERR(file);
3589 fsnotify_open(file);
3590 fd_install(ret, file);
3593 putname(req->open.filename);
3594 req->flags &= ~REQ_F_NEED_CLEANUP;
3596 req_set_fail_links(req);
3597 io_req_complete(req, ret);
3601 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3603 return io_openat2(req, force_nonblock);
3606 static int io_remove_buffers_prep(struct io_kiocb *req,
3607 const struct io_uring_sqe *sqe)
3609 struct io_provide_buf *p = &req->pbuf;
3612 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3615 tmp = READ_ONCE(sqe->fd);
3616 if (!tmp || tmp > USHRT_MAX)
3619 memset(p, 0, sizeof(*p));
3621 p->bgid = READ_ONCE(sqe->buf_group);
3625 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3626 int bgid, unsigned nbufs)
3630 /* shouldn't happen */
3634 /* the head kbuf is the list itself */
3635 while (!list_empty(&buf->list)) {
3636 struct io_buffer *nxt;
3638 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3639 list_del(&nxt->list);
3646 idr_remove(&ctx->io_buffer_idr, bgid);
3651 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3652 struct io_comp_state *cs)
3654 struct io_provide_buf *p = &req->pbuf;
3655 struct io_ring_ctx *ctx = req->ctx;
3656 struct io_buffer *head;
3659 io_ring_submit_lock(ctx, !force_nonblock);
3661 lockdep_assert_held(&ctx->uring_lock);
3664 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3666 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3668 io_ring_submit_lock(ctx, !force_nonblock);
3670 req_set_fail_links(req);
3671 __io_req_complete(req, ret, 0, cs);
3675 static int io_provide_buffers_prep(struct io_kiocb *req,
3676 const struct io_uring_sqe *sqe)
3678 struct io_provide_buf *p = &req->pbuf;
3681 if (sqe->ioprio || sqe->rw_flags)
3684 tmp = READ_ONCE(sqe->fd);
3685 if (!tmp || tmp > USHRT_MAX)
3688 p->addr = READ_ONCE(sqe->addr);
3689 p->len = READ_ONCE(sqe->len);
3691 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3694 p->bgid = READ_ONCE(sqe->buf_group);
3695 tmp = READ_ONCE(sqe->off);
3696 if (tmp > USHRT_MAX)
3702 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3704 struct io_buffer *buf;
3705 u64 addr = pbuf->addr;
3706 int i, bid = pbuf->bid;
3708 for (i = 0; i < pbuf->nbufs; i++) {
3709 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3714 buf->len = pbuf->len;
3719 INIT_LIST_HEAD(&buf->list);
3722 list_add_tail(&buf->list, &(*head)->list);
3726 return i ? i : -ENOMEM;
3729 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3730 struct io_comp_state *cs)
3732 struct io_provide_buf *p = &req->pbuf;
3733 struct io_ring_ctx *ctx = req->ctx;
3734 struct io_buffer *head, *list;
3737 io_ring_submit_lock(ctx, !force_nonblock);
3739 lockdep_assert_held(&ctx->uring_lock);
3741 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3743 ret = io_add_buffers(p, &head);
3748 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3751 __io_remove_buffers(ctx, head, p->bgid, -1U);
3756 io_ring_submit_unlock(ctx, !force_nonblock);
3758 req_set_fail_links(req);
3759 __io_req_complete(req, ret, 0, cs);
3763 static int io_epoll_ctl_prep(struct io_kiocb *req,
3764 const struct io_uring_sqe *sqe)
3766 #if defined(CONFIG_EPOLL)
3767 if (sqe->ioprio || sqe->buf_index)
3769 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3772 req->epoll.epfd = READ_ONCE(sqe->fd);
3773 req->epoll.op = READ_ONCE(sqe->len);
3774 req->epoll.fd = READ_ONCE(sqe->off);
3776 if (ep_op_has_event(req->epoll.op)) {
3777 struct epoll_event __user *ev;
3779 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3780 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3790 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3791 struct io_comp_state *cs)
3793 #if defined(CONFIG_EPOLL)
3794 struct io_epoll *ie = &req->epoll;
3797 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3798 if (force_nonblock && ret == -EAGAIN)
3802 req_set_fail_links(req);
3803 __io_req_complete(req, ret, 0, cs);
3810 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3812 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3813 if (sqe->ioprio || sqe->buf_index || sqe->off)
3815 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3818 req->madvise.addr = READ_ONCE(sqe->addr);
3819 req->madvise.len = READ_ONCE(sqe->len);
3820 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3827 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3829 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3830 struct io_madvise *ma = &req->madvise;
3836 ret = do_madvise(ma->addr, ma->len, ma->advice);
3838 req_set_fail_links(req);
3839 io_req_complete(req, ret);
3846 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3848 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3850 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3853 req->fadvise.offset = READ_ONCE(sqe->off);
3854 req->fadvise.len = READ_ONCE(sqe->len);
3855 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3859 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3861 struct io_fadvise *fa = &req->fadvise;
3864 if (force_nonblock) {
3865 switch (fa->advice) {
3866 case POSIX_FADV_NORMAL:
3867 case POSIX_FADV_RANDOM:
3868 case POSIX_FADV_SEQUENTIAL:
3875 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3877 req_set_fail_links(req);
3878 io_req_complete(req, ret);
3882 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3884 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3886 if (sqe->ioprio || sqe->buf_index)
3888 if (req->flags & REQ_F_FIXED_FILE)
3891 req->statx.dfd = READ_ONCE(sqe->fd);
3892 req->statx.mask = READ_ONCE(sqe->len);
3893 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3894 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3895 req->statx.flags = READ_ONCE(sqe->statx_flags);
3900 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3902 struct io_statx *ctx = &req->statx;
3905 if (force_nonblock) {
3906 /* only need file table for an actual valid fd */
3907 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3908 req->flags |= REQ_F_NO_FILE_TABLE;
3912 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3916 req_set_fail_links(req);
3917 io_req_complete(req, ret);
3921 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3924 * If we queue this for async, it must not be cancellable. That would
3925 * leave the 'file' in an undeterminate state, and here need to modify
3926 * io_wq_work.flags, so initialize io_wq_work firstly.
3928 io_req_init_async(req);
3929 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3931 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3933 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3934 sqe->rw_flags || sqe->buf_index)
3936 if (req->flags & REQ_F_FIXED_FILE)
3939 req->close.fd = READ_ONCE(sqe->fd);
3940 if ((req->file && req->file->f_op == &io_uring_fops) ||
3941 req->close.fd == req->ctx->ring_fd)
3944 req->close.put_file = NULL;
3948 static int io_close(struct io_kiocb *req, bool force_nonblock,
3949 struct io_comp_state *cs)
3951 struct io_close *close = &req->close;
3954 /* might be already done during nonblock submission */
3955 if (!close->put_file) {
3956 ret = __close_fd_get_file(close->fd, &close->put_file);
3958 return (ret == -ENOENT) ? -EBADF : ret;
3961 /* if the file has a flush method, be safe and punt to async */
3962 if (close->put_file->f_op->flush && force_nonblock) {
3963 /* was never set, but play safe */
3964 req->flags &= ~REQ_F_NOWAIT;
3965 /* avoid grabbing files - we don't need the files */
3966 req->flags |= REQ_F_NO_FILE_TABLE;
3970 /* No ->flush() or already async, safely close from here */
3971 ret = filp_close(close->put_file, req->work.files);
3973 req_set_fail_links(req);
3974 fput(close->put_file);
3975 close->put_file = NULL;
3976 __io_req_complete(req, ret, 0, cs);
3980 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3982 struct io_ring_ctx *ctx = req->ctx;
3987 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3989 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3992 req->sync.off = READ_ONCE(sqe->off);
3993 req->sync.len = READ_ONCE(sqe->len);
3994 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3998 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4002 /* sync_file_range always requires a blocking context */
4006 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4009 req_set_fail_links(req);
4010 io_req_complete(req, ret);
4014 #if defined(CONFIG_NET)
4015 static int io_setup_async_msg(struct io_kiocb *req,
4016 struct io_async_msghdr *kmsg)
4020 if (io_alloc_async_ctx(req)) {
4021 if (kmsg->iov != kmsg->fast_iov)
4025 req->flags |= REQ_F_NEED_CLEANUP;
4026 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4030 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4031 struct io_async_msghdr *iomsg)
4033 iomsg->iov = iomsg->fast_iov;
4034 iomsg->msg.msg_name = &iomsg->addr;
4035 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4036 req->sr_msg.msg_flags, &iomsg->iov);
4039 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4041 struct io_sr_msg *sr = &req->sr_msg;
4042 struct io_async_ctx *io = req->io;
4045 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4048 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4049 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4050 sr->len = READ_ONCE(sqe->len);
4052 #ifdef CONFIG_COMPAT
4053 if (req->ctx->compat)
4054 sr->msg_flags |= MSG_CMSG_COMPAT;
4057 if (!io || req->opcode == IORING_OP_SEND)
4059 /* iovec is already imported */
4060 if (req->flags & REQ_F_NEED_CLEANUP)
4063 ret = io_sendmsg_copy_hdr(req, &io->msg);
4065 req->flags |= REQ_F_NEED_CLEANUP;
4069 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4070 struct io_comp_state *cs)
4072 struct io_async_msghdr iomsg, *kmsg;
4073 struct socket *sock;
4077 sock = sock_from_file(req->file, &ret);
4078 if (unlikely(!sock))
4082 kmsg = &req->io->msg;
4083 kmsg->msg.msg_name = &req->io->msg.addr;
4084 /* if iov is set, it's allocated already */
4086 kmsg->iov = kmsg->fast_iov;
4087 kmsg->msg.msg_iter.iov = kmsg->iov;
4089 ret = io_sendmsg_copy_hdr(req, &iomsg);
4095 flags = req->sr_msg.msg_flags;
4096 if (flags & MSG_DONTWAIT)
4097 req->flags |= REQ_F_NOWAIT;
4098 else if (force_nonblock)
4099 flags |= MSG_DONTWAIT;
4101 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4102 if (force_nonblock && ret == -EAGAIN)
4103 return io_setup_async_msg(req, kmsg);
4104 if (ret == -ERESTARTSYS)
4107 if (kmsg->iov != kmsg->fast_iov)
4109 req->flags &= ~REQ_F_NEED_CLEANUP;
4111 req_set_fail_links(req);
4112 __io_req_complete(req, ret, 0, cs);
4116 static int io_send(struct io_kiocb *req, bool force_nonblock,
4117 struct io_comp_state *cs)
4119 struct io_sr_msg *sr = &req->sr_msg;
4122 struct socket *sock;
4126 sock = sock_from_file(req->file, &ret);
4127 if (unlikely(!sock))
4130 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4134 msg.msg_name = NULL;
4135 msg.msg_control = NULL;
4136 msg.msg_controllen = 0;
4137 msg.msg_namelen = 0;
4139 flags = req->sr_msg.msg_flags;
4140 if (flags & MSG_DONTWAIT)
4141 req->flags |= REQ_F_NOWAIT;
4142 else if (force_nonblock)
4143 flags |= MSG_DONTWAIT;
4145 msg.msg_flags = flags;
4146 ret = sock_sendmsg(sock, &msg);
4147 if (force_nonblock && ret == -EAGAIN)
4149 if (ret == -ERESTARTSYS)
4153 req_set_fail_links(req);
4154 __io_req_complete(req, ret, 0, cs);
4158 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4159 struct io_async_msghdr *iomsg)
4161 struct io_sr_msg *sr = &req->sr_msg;
4162 struct iovec __user *uiov;
4166 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4167 &iomsg->uaddr, &uiov, &iov_len);
4171 if (req->flags & REQ_F_BUFFER_SELECT) {
4174 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4176 sr->len = iomsg->iov[0].iov_len;
4177 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4181 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4182 &iomsg->iov, &iomsg->msg.msg_iter);
4190 #ifdef CONFIG_COMPAT
4191 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4192 struct io_async_msghdr *iomsg)
4194 struct compat_msghdr __user *msg_compat;
4195 struct io_sr_msg *sr = &req->sr_msg;
4196 struct compat_iovec __user *uiov;
4201 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4202 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4207 uiov = compat_ptr(ptr);
4208 if (req->flags & REQ_F_BUFFER_SELECT) {
4209 compat_ssize_t clen;
4213 if (!access_ok(uiov, sizeof(*uiov)))
4215 if (__get_user(clen, &uiov->iov_len))
4219 sr->len = iomsg->iov[0].iov_len;
4222 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4224 &iomsg->msg.msg_iter);
4233 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4234 struct io_async_msghdr *iomsg)
4236 iomsg->msg.msg_name = &iomsg->addr;
4237 iomsg->iov = iomsg->fast_iov;
4239 #ifdef CONFIG_COMPAT
4240 if (req->ctx->compat)
4241 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4244 return __io_recvmsg_copy_hdr(req, iomsg);
4247 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4250 struct io_sr_msg *sr = &req->sr_msg;
4251 struct io_buffer *kbuf;
4253 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4258 req->flags |= REQ_F_BUFFER_SELECTED;
4262 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4264 return io_put_kbuf(req, req->sr_msg.kbuf);
4267 static int io_recvmsg_prep(struct io_kiocb *req,
4268 const struct io_uring_sqe *sqe)
4270 struct io_sr_msg *sr = &req->sr_msg;
4271 struct io_async_ctx *io = req->io;
4274 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4277 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4278 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4279 sr->len = READ_ONCE(sqe->len);
4280 sr->bgid = READ_ONCE(sqe->buf_group);
4282 #ifdef CONFIG_COMPAT
4283 if (req->ctx->compat)
4284 sr->msg_flags |= MSG_CMSG_COMPAT;
4287 if (!io || req->opcode == IORING_OP_RECV)
4289 /* iovec is already imported */
4290 if (req->flags & REQ_F_NEED_CLEANUP)
4293 ret = io_recvmsg_copy_hdr(req, &io->msg);
4295 req->flags |= REQ_F_NEED_CLEANUP;
4299 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4300 struct io_comp_state *cs)
4302 struct io_async_msghdr iomsg, *kmsg;
4303 struct socket *sock;
4304 struct io_buffer *kbuf;
4306 int ret, cflags = 0;
4308 sock = sock_from_file(req->file, &ret);
4309 if (unlikely(!sock))
4313 kmsg = &req->io->msg;
4314 kmsg->msg.msg_name = &req->io->msg.addr;
4315 /* if iov is set, it's allocated already */
4317 kmsg->iov = kmsg->fast_iov;
4318 kmsg->msg.msg_iter.iov = kmsg->iov;
4320 ret = io_recvmsg_copy_hdr(req, &iomsg);
4326 if (req->flags & REQ_F_BUFFER_SELECT) {
4327 kbuf = io_recv_buffer_select(req, !force_nonblock);
4329 return PTR_ERR(kbuf);
4330 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4331 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4332 1, req->sr_msg.len);
4335 flags = req->sr_msg.msg_flags;
4336 if (flags & MSG_DONTWAIT)
4337 req->flags |= REQ_F_NOWAIT;
4338 else if (force_nonblock)
4339 flags |= MSG_DONTWAIT;
4341 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4342 kmsg->uaddr, flags);
4343 if (force_nonblock && ret == -EAGAIN)
4344 return io_setup_async_msg(req, kmsg);
4345 if (ret == -ERESTARTSYS)
4348 if (req->flags & REQ_F_BUFFER_SELECTED)
4349 cflags = io_put_recv_kbuf(req);
4350 if (kmsg->iov != kmsg->fast_iov)
4352 req->flags &= ~REQ_F_NEED_CLEANUP;
4354 req_set_fail_links(req);
4355 __io_req_complete(req, ret, cflags, cs);
4359 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4360 struct io_comp_state *cs)
4362 struct io_buffer *kbuf;
4363 struct io_sr_msg *sr = &req->sr_msg;
4365 void __user *buf = sr->buf;
4366 struct socket *sock;
4369 int ret, cflags = 0;
4371 sock = sock_from_file(req->file, &ret);
4372 if (unlikely(!sock))
4375 if (req->flags & REQ_F_BUFFER_SELECT) {
4376 kbuf = io_recv_buffer_select(req, !force_nonblock);
4378 return PTR_ERR(kbuf);
4379 buf = u64_to_user_ptr(kbuf->addr);
4382 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4386 msg.msg_name = NULL;
4387 msg.msg_control = NULL;
4388 msg.msg_controllen = 0;
4389 msg.msg_namelen = 0;
4390 msg.msg_iocb = NULL;
4393 flags = req->sr_msg.msg_flags;
4394 if (flags & MSG_DONTWAIT)
4395 req->flags |= REQ_F_NOWAIT;
4396 else if (force_nonblock)
4397 flags |= MSG_DONTWAIT;
4399 ret = sock_recvmsg(sock, &msg, flags);
4400 if (force_nonblock && ret == -EAGAIN)
4402 if (ret == -ERESTARTSYS)
4405 if (req->flags & REQ_F_BUFFER_SELECTED)
4406 cflags = io_put_recv_kbuf(req);
4408 req_set_fail_links(req);
4409 __io_req_complete(req, ret, cflags, cs);
4413 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4415 struct io_accept *accept = &req->accept;
4417 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4419 if (sqe->ioprio || sqe->len || sqe->buf_index)
4422 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4423 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4424 accept->flags = READ_ONCE(sqe->accept_flags);
4425 accept->nofile = rlimit(RLIMIT_NOFILE);
4429 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4430 struct io_comp_state *cs)
4432 struct io_accept *accept = &req->accept;
4433 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4436 if (req->file->f_flags & O_NONBLOCK)
4437 req->flags |= REQ_F_NOWAIT;
4439 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4440 accept->addr_len, accept->flags,
4442 if (ret == -EAGAIN && force_nonblock)
4445 if (ret == -ERESTARTSYS)
4447 req_set_fail_links(req);
4449 __io_req_complete(req, ret, 0, cs);
4453 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4455 struct io_connect *conn = &req->connect;
4456 struct io_async_ctx *io = req->io;
4458 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4460 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4463 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4464 conn->addr_len = READ_ONCE(sqe->addr2);
4469 return move_addr_to_kernel(conn->addr, conn->addr_len,
4470 &io->connect.address);
4473 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4474 struct io_comp_state *cs)
4476 struct io_async_ctx __io, *io;
4477 unsigned file_flags;
4483 ret = move_addr_to_kernel(req->connect.addr,
4484 req->connect.addr_len,
4485 &__io.connect.address);
4491 file_flags = force_nonblock ? O_NONBLOCK : 0;
4493 ret = __sys_connect_file(req->file, &io->connect.address,
4494 req->connect.addr_len, file_flags);
4495 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4498 if (io_alloc_async_ctx(req)) {
4502 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4505 if (ret == -ERESTARTSYS)
4509 req_set_fail_links(req);
4510 __io_req_complete(req, ret, 0, cs);
4513 #else /* !CONFIG_NET */
4514 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4519 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4520 struct io_comp_state *cs)
4525 static int io_send(struct io_kiocb *req, bool force_nonblock,
4526 struct io_comp_state *cs)
4531 static int io_recvmsg_prep(struct io_kiocb *req,
4532 const struct io_uring_sqe *sqe)
4537 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4538 struct io_comp_state *cs)
4543 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4544 struct io_comp_state *cs)
4549 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4554 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4555 struct io_comp_state *cs)
4560 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4565 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4566 struct io_comp_state *cs)
4570 #endif /* CONFIG_NET */
4572 struct io_poll_table {
4573 struct poll_table_struct pt;
4574 struct io_kiocb *req;
4578 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4579 __poll_t mask, task_work_func_t func)
4584 /* for instances that support it check for an event match first: */
4585 if (mask && !(mask & poll->events))
4588 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4590 list_del_init(&poll->wait.entry);
4593 init_task_work(&req->task_work, func);
4594 percpu_ref_get(&req->ctx->refs);
4597 * If we using the signalfd wait_queue_head for this wakeup, then
4598 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4599 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4600 * either, as the normal wakeup will suffice.
4602 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4605 * If this fails, then the task is exiting. When a task exits, the
4606 * work gets canceled, so just cancel this request as well instead
4607 * of executing it. We can't safely execute it anyway, as we may not
4608 * have the needed state needed for it anyway.
4610 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4611 if (unlikely(ret)) {
4612 struct task_struct *tsk;
4614 WRITE_ONCE(poll->canceled, true);
4615 tsk = io_wq_get_task(req->ctx->io_wq);
4616 task_work_add(tsk, &req->task_work, 0);
4617 wake_up_process(tsk);
4622 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4623 __acquires(&req->ctx->completion_lock)
4625 struct io_ring_ctx *ctx = req->ctx;
4627 if (!req->result && !READ_ONCE(poll->canceled)) {
4628 struct poll_table_struct pt = { ._key = poll->events };
4630 req->result = vfs_poll(req->file, &pt) & poll->events;
4633 spin_lock_irq(&ctx->completion_lock);
4634 if (!req->result && !READ_ONCE(poll->canceled)) {
4635 add_wait_queue(poll->head, &poll->wait);
4642 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4644 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4645 if (req->opcode == IORING_OP_POLL_ADD)
4646 return (struct io_poll_iocb *) req->io;
4647 return req->apoll->double_poll;
4650 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4652 if (req->opcode == IORING_OP_POLL_ADD)
4654 return &req->apoll->poll;
4657 static void io_poll_remove_double(struct io_kiocb *req)
4659 struct io_poll_iocb *poll = io_poll_get_double(req);
4661 lockdep_assert_held(&req->ctx->completion_lock);
4663 if (poll && poll->head) {
4664 struct wait_queue_head *head = poll->head;
4666 spin_lock(&head->lock);
4667 list_del_init(&poll->wait.entry);
4668 if (poll->wait.private)
4669 refcount_dec(&req->refs);
4671 spin_unlock(&head->lock);
4675 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4677 struct io_ring_ctx *ctx = req->ctx;
4679 io_poll_remove_double(req);
4680 req->poll.done = true;
4681 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4682 io_commit_cqring(ctx);
4685 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4687 struct io_ring_ctx *ctx = req->ctx;
4689 if (io_poll_rewait(req, &req->poll)) {
4690 spin_unlock_irq(&ctx->completion_lock);
4694 hash_del(&req->hash_node);
4695 io_poll_complete(req, req->result, 0);
4696 req->flags |= REQ_F_COMP_LOCKED;
4697 *nxt = io_put_req_find_next(req);
4698 spin_unlock_irq(&ctx->completion_lock);
4700 io_cqring_ev_posted(ctx);
4703 static void io_poll_task_func(struct callback_head *cb)
4705 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4706 struct io_ring_ctx *ctx = req->ctx;
4707 struct io_kiocb *nxt = NULL;
4709 io_poll_task_handler(req, &nxt);
4711 __io_req_task_submit(nxt);
4712 percpu_ref_put(&ctx->refs);
4715 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4716 int sync, void *key)
4718 struct io_kiocb *req = wait->private;
4719 struct io_poll_iocb *poll = io_poll_get_single(req);
4720 __poll_t mask = key_to_poll(key);
4722 /* for instances that support it check for an event match first: */
4723 if (mask && !(mask & poll->events))
4726 if (poll && poll->head) {
4729 spin_lock(&poll->head->lock);
4730 done = list_empty(&poll->wait.entry);
4732 list_del_init(&poll->wait.entry);
4733 /* make sure double remove sees this as being gone */
4734 wait->private = NULL;
4735 spin_unlock(&poll->head->lock);
4737 __io_async_wake(req, poll, mask, io_poll_task_func);
4739 refcount_dec(&req->refs);
4743 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4744 wait_queue_func_t wake_func)
4748 poll->canceled = false;
4749 poll->events = events;
4750 INIT_LIST_HEAD(&poll->wait.entry);
4751 init_waitqueue_func_entry(&poll->wait, wake_func);
4754 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4755 struct wait_queue_head *head,
4756 struct io_poll_iocb **poll_ptr)
4758 struct io_kiocb *req = pt->req;
4761 * If poll->head is already set, it's because the file being polled
4762 * uses multiple waitqueues for poll handling (eg one for read, one
4763 * for write). Setup a separate io_poll_iocb if this happens.
4765 if (unlikely(poll->head)) {
4766 /* already have a 2nd entry, fail a third attempt */
4768 pt->error = -EINVAL;
4771 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4773 pt->error = -ENOMEM;
4776 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4777 refcount_inc(&req->refs);
4778 poll->wait.private = req;
4785 if (poll->events & EPOLLEXCLUSIVE)
4786 add_wait_queue_exclusive(head, &poll->wait);
4788 add_wait_queue(head, &poll->wait);
4791 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4792 struct poll_table_struct *p)
4794 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4795 struct async_poll *apoll = pt->req->apoll;
4797 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4800 static void io_async_task_func(struct callback_head *cb)
4802 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4803 struct async_poll *apoll = req->apoll;
4804 struct io_ring_ctx *ctx = req->ctx;
4806 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4808 if (io_poll_rewait(req, &apoll->poll)) {
4809 spin_unlock_irq(&ctx->completion_lock);
4810 percpu_ref_put(&ctx->refs);
4814 /* If req is still hashed, it cannot have been canceled. Don't check. */
4815 if (hash_hashed(&req->hash_node))
4816 hash_del(&req->hash_node);
4818 io_poll_remove_double(req);
4819 spin_unlock_irq(&ctx->completion_lock);
4821 if (!READ_ONCE(apoll->poll.canceled))
4822 __io_req_task_submit(req);
4824 __io_req_task_cancel(req, -ECANCELED);
4826 percpu_ref_put(&ctx->refs);
4827 kfree(apoll->double_poll);
4831 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4834 struct io_kiocb *req = wait->private;
4835 struct io_poll_iocb *poll = &req->apoll->poll;
4837 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4840 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4843 static void io_poll_req_insert(struct io_kiocb *req)
4845 struct io_ring_ctx *ctx = req->ctx;
4846 struct hlist_head *list;
4848 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4849 hlist_add_head(&req->hash_node, list);
4852 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4853 struct io_poll_iocb *poll,
4854 struct io_poll_table *ipt, __poll_t mask,
4855 wait_queue_func_t wake_func)
4856 __acquires(&ctx->completion_lock)
4858 struct io_ring_ctx *ctx = req->ctx;
4859 bool cancel = false;
4861 io_init_poll_iocb(poll, mask, wake_func);
4862 poll->file = req->file;
4863 poll->wait.private = req;
4865 ipt->pt._key = mask;
4867 ipt->error = -EINVAL;
4869 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4871 spin_lock_irq(&ctx->completion_lock);
4872 if (likely(poll->head)) {
4873 spin_lock(&poll->head->lock);
4874 if (unlikely(list_empty(&poll->wait.entry))) {
4880 if (mask || ipt->error)
4881 list_del_init(&poll->wait.entry);
4883 WRITE_ONCE(poll->canceled, true);
4884 else if (!poll->done) /* actually waiting for an event */
4885 io_poll_req_insert(req);
4886 spin_unlock(&poll->head->lock);
4892 static bool io_arm_poll_handler(struct io_kiocb *req)
4894 const struct io_op_def *def = &io_op_defs[req->opcode];
4895 struct io_ring_ctx *ctx = req->ctx;
4896 struct async_poll *apoll;
4897 struct io_poll_table ipt;
4901 if (!req->file || !file_can_poll(req->file))
4903 if (req->flags & REQ_F_POLLED)
4907 else if (def->pollout)
4911 /* if we can't nonblock try, then no point in arming a poll handler */
4912 if (!io_file_supports_async(req->file, rw))
4915 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4916 if (unlikely(!apoll))
4918 apoll->double_poll = NULL;
4920 req->flags |= REQ_F_POLLED;
4921 io_get_req_task(req);
4923 INIT_HLIST_NODE(&req->hash_node);
4927 mask |= POLLIN | POLLRDNORM;
4929 mask |= POLLOUT | POLLWRNORM;
4930 mask |= POLLERR | POLLPRI;
4932 ipt.pt._qproc = io_async_queue_proc;
4934 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4936 if (ret || ipt.error) {
4937 io_poll_remove_double(req);
4938 spin_unlock_irq(&ctx->completion_lock);
4939 kfree(apoll->double_poll);
4943 spin_unlock_irq(&ctx->completion_lock);
4944 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4945 apoll->poll.events);
4949 static bool __io_poll_remove_one(struct io_kiocb *req,
4950 struct io_poll_iocb *poll)
4952 bool do_complete = false;
4954 spin_lock(&poll->head->lock);
4955 WRITE_ONCE(poll->canceled, true);
4956 if (!list_empty(&poll->wait.entry)) {
4957 list_del_init(&poll->wait.entry);
4960 spin_unlock(&poll->head->lock);
4961 hash_del(&req->hash_node);
4965 static bool io_poll_remove_one(struct io_kiocb *req)
4969 io_poll_remove_double(req);
4971 if (req->opcode == IORING_OP_POLL_ADD) {
4972 do_complete = __io_poll_remove_one(req, &req->poll);
4974 struct async_poll *apoll = req->apoll;
4976 /* non-poll requests have submit ref still */
4977 do_complete = __io_poll_remove_one(req, &apoll->poll);
4980 kfree(apoll->double_poll);
4986 io_cqring_fill_event(req, -ECANCELED);
4987 io_commit_cqring(req->ctx);
4988 req->flags |= REQ_F_COMP_LOCKED;
4989 req_set_fail_links(req);
4996 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4998 struct hlist_node *tmp;
4999 struct io_kiocb *req;
5002 spin_lock_irq(&ctx->completion_lock);
5003 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5004 struct hlist_head *list;
5006 list = &ctx->cancel_hash[i];
5007 hlist_for_each_entry_safe(req, tmp, list, hash_node)
5008 posted += io_poll_remove_one(req);
5010 spin_unlock_irq(&ctx->completion_lock);
5013 io_cqring_ev_posted(ctx);
5016 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5018 struct hlist_head *list;
5019 struct io_kiocb *req;
5021 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5022 hlist_for_each_entry(req, list, hash_node) {
5023 if (sqe_addr != req->user_data)
5025 if (io_poll_remove_one(req))
5033 static int io_poll_remove_prep(struct io_kiocb *req,
5034 const struct io_uring_sqe *sqe)
5036 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5038 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5042 req->poll.addr = READ_ONCE(sqe->addr);
5047 * Find a running poll command that matches one specified in sqe->addr,
5048 * and remove it if found.
5050 static int io_poll_remove(struct io_kiocb *req)
5052 struct io_ring_ctx *ctx = req->ctx;
5056 addr = req->poll.addr;
5057 spin_lock_irq(&ctx->completion_lock);
5058 ret = io_poll_cancel(ctx, addr);
5059 spin_unlock_irq(&ctx->completion_lock);
5062 req_set_fail_links(req);
5063 io_req_complete(req, ret);
5067 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5070 struct io_kiocb *req = wait->private;
5071 struct io_poll_iocb *poll = &req->poll;
5073 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5076 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5077 struct poll_table_struct *p)
5079 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5081 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5084 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5086 struct io_poll_iocb *poll = &req->poll;
5089 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5091 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5096 events = READ_ONCE(sqe->poll32_events);
5098 events = swahw32(events);
5100 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5101 (events & EPOLLEXCLUSIVE);
5103 io_get_req_task(req);
5107 static int io_poll_add(struct io_kiocb *req)
5109 struct io_poll_iocb *poll = &req->poll;
5110 struct io_ring_ctx *ctx = req->ctx;
5111 struct io_poll_table ipt;
5114 INIT_HLIST_NODE(&req->hash_node);
5115 ipt.pt._qproc = io_poll_queue_proc;
5117 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5120 if (mask) { /* no async, we'd stolen it */
5122 io_poll_complete(req, mask, 0);
5124 spin_unlock_irq(&ctx->completion_lock);
5127 io_cqring_ev_posted(ctx);
5133 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5135 struct io_timeout_data *data = container_of(timer,
5136 struct io_timeout_data, timer);
5137 struct io_kiocb *req = data->req;
5138 struct io_ring_ctx *ctx = req->ctx;
5139 unsigned long flags;
5141 spin_lock_irqsave(&ctx->completion_lock, flags);
5142 atomic_set(&req->ctx->cq_timeouts,
5143 atomic_read(&req->ctx->cq_timeouts) + 1);
5146 * We could be racing with timeout deletion. If the list is empty,
5147 * then timeout lookup already found it and will be handling it.
5149 if (!list_empty(&req->timeout.list))
5150 list_del_init(&req->timeout.list);
5152 io_cqring_fill_event(req, -ETIME);
5153 io_commit_cqring(ctx);
5154 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5156 io_cqring_ev_posted(ctx);
5157 req_set_fail_links(req);
5159 return HRTIMER_NORESTART;
5162 static int __io_timeout_cancel(struct io_kiocb *req)
5166 list_del_init(&req->timeout.list);
5168 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5172 req_set_fail_links(req);
5173 req->flags |= REQ_F_COMP_LOCKED;
5174 io_cqring_fill_event(req, -ECANCELED);
5179 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5181 struct io_kiocb *req;
5184 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5185 if (user_data == req->user_data) {
5194 return __io_timeout_cancel(req);
5197 static int io_timeout_remove_prep(struct io_kiocb *req,
5198 const struct io_uring_sqe *sqe)
5200 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5202 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5204 if (sqe->ioprio || sqe->buf_index || sqe->len)
5207 req->timeout.addr = READ_ONCE(sqe->addr);
5208 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5209 if (req->timeout.flags)
5216 * Remove or update an existing timeout command
5218 static int io_timeout_remove(struct io_kiocb *req)
5220 struct io_ring_ctx *ctx = req->ctx;
5223 spin_lock_irq(&ctx->completion_lock);
5224 ret = io_timeout_cancel(ctx, req->timeout.addr);
5226 io_cqring_fill_event(req, ret);
5227 io_commit_cqring(ctx);
5228 spin_unlock_irq(&ctx->completion_lock);
5229 io_cqring_ev_posted(ctx);
5231 req_set_fail_links(req);
5236 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5237 bool is_timeout_link)
5239 struct io_timeout_data *data;
5241 u32 off = READ_ONCE(sqe->off);
5243 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5245 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5247 if (off && is_timeout_link)
5249 flags = READ_ONCE(sqe->timeout_flags);
5250 if (flags & ~IORING_TIMEOUT_ABS)
5253 req->timeout.off = off;
5255 if (!req->io && io_alloc_async_ctx(req))
5258 data = &req->io->timeout;
5261 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5264 if (flags & IORING_TIMEOUT_ABS)
5265 data->mode = HRTIMER_MODE_ABS;
5267 data->mode = HRTIMER_MODE_REL;
5269 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5273 static int io_timeout(struct io_kiocb *req)
5275 struct io_ring_ctx *ctx = req->ctx;
5276 struct io_timeout_data *data = &req->io->timeout;
5277 struct list_head *entry;
5278 u32 tail, off = req->timeout.off;
5280 spin_lock_irq(&ctx->completion_lock);
5283 * sqe->off holds how many events that need to occur for this
5284 * timeout event to be satisfied. If it isn't set, then this is
5285 * a pure timeout request, sequence isn't used.
5287 if (io_is_timeout_noseq(req)) {
5288 entry = ctx->timeout_list.prev;
5292 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5293 req->timeout.target_seq = tail + off;
5296 * Insertion sort, ensuring the first entry in the list is always
5297 * the one we need first.
5299 list_for_each_prev(entry, &ctx->timeout_list) {
5300 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5303 if (io_is_timeout_noseq(nxt))
5305 /* nxt.seq is behind @tail, otherwise would've been completed */
5306 if (off >= nxt->timeout.target_seq - tail)
5310 list_add(&req->timeout.list, entry);
5311 data->timer.function = io_timeout_fn;
5312 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5313 spin_unlock_irq(&ctx->completion_lock);
5317 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5319 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5321 return req->user_data == (unsigned long) data;
5324 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5326 enum io_wq_cancel cancel_ret;
5329 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5330 switch (cancel_ret) {
5331 case IO_WQ_CANCEL_OK:
5334 case IO_WQ_CANCEL_RUNNING:
5337 case IO_WQ_CANCEL_NOTFOUND:
5345 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5346 struct io_kiocb *req, __u64 sqe_addr,
5349 unsigned long flags;
5352 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5353 if (ret != -ENOENT) {
5354 spin_lock_irqsave(&ctx->completion_lock, flags);
5358 spin_lock_irqsave(&ctx->completion_lock, flags);
5359 ret = io_timeout_cancel(ctx, sqe_addr);
5362 ret = io_poll_cancel(ctx, sqe_addr);
5366 io_cqring_fill_event(req, ret);
5367 io_commit_cqring(ctx);
5368 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5369 io_cqring_ev_posted(ctx);
5372 req_set_fail_links(req);
5376 static int io_async_cancel_prep(struct io_kiocb *req,
5377 const struct io_uring_sqe *sqe)
5379 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5381 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5383 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5386 req->cancel.addr = READ_ONCE(sqe->addr);
5390 static int io_async_cancel(struct io_kiocb *req)
5392 struct io_ring_ctx *ctx = req->ctx;
5394 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5398 static int io_files_update_prep(struct io_kiocb *req,
5399 const struct io_uring_sqe *sqe)
5401 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5403 if (sqe->ioprio || sqe->rw_flags)
5406 req->files_update.offset = READ_ONCE(sqe->off);
5407 req->files_update.nr_args = READ_ONCE(sqe->len);
5408 if (!req->files_update.nr_args)
5410 req->files_update.arg = READ_ONCE(sqe->addr);
5414 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5415 struct io_comp_state *cs)
5417 struct io_ring_ctx *ctx = req->ctx;
5418 struct io_uring_files_update up;
5424 up.offset = req->files_update.offset;
5425 up.fds = req->files_update.arg;
5427 mutex_lock(&ctx->uring_lock);
5428 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5429 mutex_unlock(&ctx->uring_lock);
5432 req_set_fail_links(req);
5433 __io_req_complete(req, ret, 0, cs);
5437 static int io_req_defer_prep(struct io_kiocb *req,
5438 const struct io_uring_sqe *sqe)
5445 if (io_alloc_async_ctx(req))
5447 ret = io_prep_work_files(req);
5451 switch (req->opcode) {
5454 case IORING_OP_READV:
5455 case IORING_OP_READ_FIXED:
5456 case IORING_OP_READ:
5457 ret = io_read_prep(req, sqe, true);
5459 case IORING_OP_WRITEV:
5460 case IORING_OP_WRITE_FIXED:
5461 case IORING_OP_WRITE:
5462 ret = io_write_prep(req, sqe, true);
5464 case IORING_OP_POLL_ADD:
5465 ret = io_poll_add_prep(req, sqe);
5467 case IORING_OP_POLL_REMOVE:
5468 ret = io_poll_remove_prep(req, sqe);
5470 case IORING_OP_FSYNC:
5471 ret = io_prep_fsync(req, sqe);
5473 case IORING_OP_SYNC_FILE_RANGE:
5474 ret = io_prep_sfr(req, sqe);
5476 case IORING_OP_SENDMSG:
5477 case IORING_OP_SEND:
5478 ret = io_sendmsg_prep(req, sqe);
5480 case IORING_OP_RECVMSG:
5481 case IORING_OP_RECV:
5482 ret = io_recvmsg_prep(req, sqe);
5484 case IORING_OP_CONNECT:
5485 ret = io_connect_prep(req, sqe);
5487 case IORING_OP_TIMEOUT:
5488 ret = io_timeout_prep(req, sqe, false);
5490 case IORING_OP_TIMEOUT_REMOVE:
5491 ret = io_timeout_remove_prep(req, sqe);
5493 case IORING_OP_ASYNC_CANCEL:
5494 ret = io_async_cancel_prep(req, sqe);
5496 case IORING_OP_LINK_TIMEOUT:
5497 ret = io_timeout_prep(req, sqe, true);
5499 case IORING_OP_ACCEPT:
5500 ret = io_accept_prep(req, sqe);
5502 case IORING_OP_FALLOCATE:
5503 ret = io_fallocate_prep(req, sqe);
5505 case IORING_OP_OPENAT:
5506 ret = io_openat_prep(req, sqe);
5508 case IORING_OP_CLOSE:
5509 ret = io_close_prep(req, sqe);
5511 case IORING_OP_FILES_UPDATE:
5512 ret = io_files_update_prep(req, sqe);
5514 case IORING_OP_STATX:
5515 ret = io_statx_prep(req, sqe);
5517 case IORING_OP_FADVISE:
5518 ret = io_fadvise_prep(req, sqe);
5520 case IORING_OP_MADVISE:
5521 ret = io_madvise_prep(req, sqe);
5523 case IORING_OP_OPENAT2:
5524 ret = io_openat2_prep(req, sqe);
5526 case IORING_OP_EPOLL_CTL:
5527 ret = io_epoll_ctl_prep(req, sqe);
5529 case IORING_OP_SPLICE:
5530 ret = io_splice_prep(req, sqe);
5532 case IORING_OP_PROVIDE_BUFFERS:
5533 ret = io_provide_buffers_prep(req, sqe);
5535 case IORING_OP_REMOVE_BUFFERS:
5536 ret = io_remove_buffers_prep(req, sqe);
5539 ret = io_tee_prep(req, sqe);
5542 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5551 static u32 io_get_sequence(struct io_kiocb *req)
5553 struct io_kiocb *pos;
5554 struct io_ring_ctx *ctx = req->ctx;
5555 u32 total_submitted, nr_reqs = 1;
5557 if (req->flags & REQ_F_LINK_HEAD)
5558 list_for_each_entry(pos, &req->link_list, link_list)
5561 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5562 return total_submitted - nr_reqs;
5565 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5567 struct io_ring_ctx *ctx = req->ctx;
5568 struct io_defer_entry *de;
5572 /* Still need defer if there is pending req in defer list. */
5573 if (likely(list_empty_careful(&ctx->defer_list) &&
5574 !(req->flags & REQ_F_IO_DRAIN)))
5577 seq = io_get_sequence(req);
5578 /* Still a chance to pass the sequence check */
5579 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5583 ret = io_req_defer_prep(req, sqe);
5587 io_prep_async_link(req);
5588 de = kmalloc(sizeof(*de), GFP_KERNEL);
5592 spin_lock_irq(&ctx->completion_lock);
5593 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5594 spin_unlock_irq(&ctx->completion_lock);
5596 io_queue_async_work(req);
5597 return -EIOCBQUEUED;
5600 trace_io_uring_defer(ctx, req, req->user_data);
5603 list_add_tail(&de->list, &ctx->defer_list);
5604 spin_unlock_irq(&ctx->completion_lock);
5605 return -EIOCBQUEUED;
5608 static void __io_clean_op(struct io_kiocb *req)
5610 struct io_async_ctx *io = req->io;
5612 if (req->flags & REQ_F_BUFFER_SELECTED) {
5613 switch (req->opcode) {
5614 case IORING_OP_READV:
5615 case IORING_OP_READ_FIXED:
5616 case IORING_OP_READ:
5617 kfree((void *)(unsigned long)req->rw.addr);
5619 case IORING_OP_RECVMSG:
5620 case IORING_OP_RECV:
5621 kfree(req->sr_msg.kbuf);
5624 req->flags &= ~REQ_F_BUFFER_SELECTED;
5627 if (req->flags & REQ_F_NEED_CLEANUP) {
5628 switch (req->opcode) {
5629 case IORING_OP_READV:
5630 case IORING_OP_READ_FIXED:
5631 case IORING_OP_READ:
5632 case IORING_OP_WRITEV:
5633 case IORING_OP_WRITE_FIXED:
5634 case IORING_OP_WRITE:
5635 if (io->rw.free_iovec)
5636 kfree(io->rw.free_iovec);
5638 case IORING_OP_RECVMSG:
5639 case IORING_OP_SENDMSG:
5640 if (io->msg.iov != io->msg.fast_iov)
5643 case IORING_OP_SPLICE:
5645 io_put_file(req, req->splice.file_in,
5646 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5649 req->flags &= ~REQ_F_NEED_CLEANUP;
5652 if (req->flags & REQ_F_INFLIGHT) {
5653 struct io_ring_ctx *ctx = req->ctx;
5654 unsigned long flags;
5656 spin_lock_irqsave(&ctx->inflight_lock, flags);
5657 list_del(&req->inflight_entry);
5658 if (waitqueue_active(&ctx->inflight_wait))
5659 wake_up(&ctx->inflight_wait);
5660 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5661 req->flags &= ~REQ_F_INFLIGHT;
5665 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5666 bool force_nonblock, struct io_comp_state *cs)
5668 struct io_ring_ctx *ctx = req->ctx;
5671 switch (req->opcode) {
5673 ret = io_nop(req, cs);
5675 case IORING_OP_READV:
5676 case IORING_OP_READ_FIXED:
5677 case IORING_OP_READ:
5679 ret = io_read_prep(req, sqe, force_nonblock);
5683 ret = io_read(req, force_nonblock, cs);
5685 case IORING_OP_WRITEV:
5686 case IORING_OP_WRITE_FIXED:
5687 case IORING_OP_WRITE:
5689 ret = io_write_prep(req, sqe, force_nonblock);
5693 ret = io_write(req, force_nonblock, cs);
5695 case IORING_OP_FSYNC:
5697 ret = io_prep_fsync(req, sqe);
5701 ret = io_fsync(req, force_nonblock);
5703 case IORING_OP_POLL_ADD:
5705 ret = io_poll_add_prep(req, sqe);
5709 ret = io_poll_add(req);
5711 case IORING_OP_POLL_REMOVE:
5713 ret = io_poll_remove_prep(req, sqe);
5717 ret = io_poll_remove(req);
5719 case IORING_OP_SYNC_FILE_RANGE:
5721 ret = io_prep_sfr(req, sqe);
5725 ret = io_sync_file_range(req, force_nonblock);
5727 case IORING_OP_SENDMSG:
5728 case IORING_OP_SEND:
5730 ret = io_sendmsg_prep(req, sqe);
5734 if (req->opcode == IORING_OP_SENDMSG)
5735 ret = io_sendmsg(req, force_nonblock, cs);
5737 ret = io_send(req, force_nonblock, cs);
5739 case IORING_OP_RECVMSG:
5740 case IORING_OP_RECV:
5742 ret = io_recvmsg_prep(req, sqe);
5746 if (req->opcode == IORING_OP_RECVMSG)
5747 ret = io_recvmsg(req, force_nonblock, cs);
5749 ret = io_recv(req, force_nonblock, cs);
5751 case IORING_OP_TIMEOUT:
5753 ret = io_timeout_prep(req, sqe, false);
5757 ret = io_timeout(req);
5759 case IORING_OP_TIMEOUT_REMOVE:
5761 ret = io_timeout_remove_prep(req, sqe);
5765 ret = io_timeout_remove(req);
5767 case IORING_OP_ACCEPT:
5769 ret = io_accept_prep(req, sqe);
5773 ret = io_accept(req, force_nonblock, cs);
5775 case IORING_OP_CONNECT:
5777 ret = io_connect_prep(req, sqe);
5781 ret = io_connect(req, force_nonblock, cs);
5783 case IORING_OP_ASYNC_CANCEL:
5785 ret = io_async_cancel_prep(req, sqe);
5789 ret = io_async_cancel(req);
5791 case IORING_OP_FALLOCATE:
5793 ret = io_fallocate_prep(req, sqe);
5797 ret = io_fallocate(req, force_nonblock);
5799 case IORING_OP_OPENAT:
5801 ret = io_openat_prep(req, sqe);
5805 ret = io_openat(req, force_nonblock);
5807 case IORING_OP_CLOSE:
5809 ret = io_close_prep(req, sqe);
5813 ret = io_close(req, force_nonblock, cs);
5815 case IORING_OP_FILES_UPDATE:
5817 ret = io_files_update_prep(req, sqe);
5821 ret = io_files_update(req, force_nonblock, cs);
5823 case IORING_OP_STATX:
5825 ret = io_statx_prep(req, sqe);
5829 ret = io_statx(req, force_nonblock);
5831 case IORING_OP_FADVISE:
5833 ret = io_fadvise_prep(req, sqe);
5837 ret = io_fadvise(req, force_nonblock);
5839 case IORING_OP_MADVISE:
5841 ret = io_madvise_prep(req, sqe);
5845 ret = io_madvise(req, force_nonblock);
5847 case IORING_OP_OPENAT2:
5849 ret = io_openat2_prep(req, sqe);
5853 ret = io_openat2(req, force_nonblock);
5855 case IORING_OP_EPOLL_CTL:
5857 ret = io_epoll_ctl_prep(req, sqe);
5861 ret = io_epoll_ctl(req, force_nonblock, cs);
5863 case IORING_OP_SPLICE:
5865 ret = io_splice_prep(req, sqe);
5869 ret = io_splice(req, force_nonblock);
5871 case IORING_OP_PROVIDE_BUFFERS:
5873 ret = io_provide_buffers_prep(req, sqe);
5877 ret = io_provide_buffers(req, force_nonblock, cs);
5879 case IORING_OP_REMOVE_BUFFERS:
5881 ret = io_remove_buffers_prep(req, sqe);
5885 ret = io_remove_buffers(req, force_nonblock, cs);
5889 ret = io_tee_prep(req, sqe);
5893 ret = io_tee(req, force_nonblock);
5903 /* If the op doesn't have a file, we're not polling for it */
5904 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5905 const bool in_async = io_wq_current_is_worker();
5907 /* workqueue context doesn't hold uring_lock, grab it now */
5909 mutex_lock(&ctx->uring_lock);
5911 io_iopoll_req_issued(req);
5914 mutex_unlock(&ctx->uring_lock);
5920 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5922 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5923 struct io_kiocb *timeout;
5926 timeout = io_prep_linked_timeout(req);
5928 io_queue_linked_timeout(timeout);
5930 /* if NO_CANCEL is set, we must still run the work */
5931 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5932 IO_WQ_WORK_CANCEL) {
5938 ret = io_issue_sqe(req, NULL, false, NULL);
5940 * We can get EAGAIN for polled IO even though we're
5941 * forcing a sync submission from here, since we can't
5942 * wait for request slots on the block side.
5951 req_set_fail_links(req);
5952 io_req_complete(req, ret);
5955 return io_steal_work(req);
5958 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5961 struct fixed_file_table *table;
5963 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5964 return table->files[index & IORING_FILE_TABLE_MASK];
5967 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5968 int fd, struct file **out_file, bool fixed)
5970 struct io_ring_ctx *ctx = req->ctx;
5974 if (unlikely(!ctx->file_data ||
5975 (unsigned) fd >= ctx->nr_user_files))
5977 fd = array_index_nospec(fd, ctx->nr_user_files);
5978 file = io_file_from_index(ctx, fd);
5980 req->fixed_file_refs = ctx->file_data->cur_refs;
5981 percpu_ref_get(req->fixed_file_refs);
5984 trace_io_uring_file_get(ctx, fd);
5985 file = __io_file_get(state, fd);
5988 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5995 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6000 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6001 if (unlikely(!fixed && io_async_submit(req->ctx)))
6004 return io_file_get(state, req, fd, &req->file, fixed);
6007 static int io_grab_files(struct io_kiocb *req)
6010 struct io_ring_ctx *ctx = req->ctx;
6012 io_req_init_async(req);
6014 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6016 if (!ctx->ring_file)
6020 spin_lock_irq(&ctx->inflight_lock);
6022 * We use the f_ops->flush() handler to ensure that we can flush
6023 * out work accessing these files if the fd is closed. Check if
6024 * the fd has changed since we started down this path, and disallow
6025 * this operation if it has.
6027 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
6028 list_add(&req->inflight_entry, &ctx->inflight_list);
6029 req->flags |= REQ_F_INFLIGHT;
6030 req->work.files = current->files;
6033 spin_unlock_irq(&ctx->inflight_lock);
6039 static inline int io_prep_work_files(struct io_kiocb *req)
6041 if (!io_op_defs[req->opcode].file_table)
6043 return io_grab_files(req);
6046 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6048 struct io_timeout_data *data = container_of(timer,
6049 struct io_timeout_data, timer);
6050 struct io_kiocb *req = data->req;
6051 struct io_ring_ctx *ctx = req->ctx;
6052 struct io_kiocb *prev = NULL;
6053 unsigned long flags;
6055 spin_lock_irqsave(&ctx->completion_lock, flags);
6058 * We don't expect the list to be empty, that will only happen if we
6059 * race with the completion of the linked work.
6061 if (!list_empty(&req->link_list)) {
6062 prev = list_entry(req->link_list.prev, struct io_kiocb,
6064 if (refcount_inc_not_zero(&prev->refs)) {
6065 list_del_init(&req->link_list);
6066 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6071 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6074 req_set_fail_links(prev);
6075 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6078 io_req_complete(req, -ETIME);
6080 return HRTIMER_NORESTART;
6083 static void __io_queue_linked_timeout(struct io_kiocb *req)
6086 * If the list is now empty, then our linked request finished before
6087 * we got a chance to setup the timer
6089 if (!list_empty(&req->link_list)) {
6090 struct io_timeout_data *data = &req->io->timeout;
6092 data->timer.function = io_link_timeout_fn;
6093 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6098 static void io_queue_linked_timeout(struct io_kiocb *req)
6100 struct io_ring_ctx *ctx = req->ctx;
6102 spin_lock_irq(&ctx->completion_lock);
6103 __io_queue_linked_timeout(req);
6104 spin_unlock_irq(&ctx->completion_lock);
6106 /* drop submission reference */
6110 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6112 struct io_kiocb *nxt;
6114 if (!(req->flags & REQ_F_LINK_HEAD))
6116 if (req->flags & REQ_F_LINK_TIMEOUT)
6119 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6121 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6124 req->flags |= REQ_F_LINK_TIMEOUT;
6128 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6129 struct io_comp_state *cs)
6131 struct io_kiocb *linked_timeout;
6132 struct io_kiocb *nxt;
6133 const struct cred *old_creds = NULL;
6137 linked_timeout = io_prep_linked_timeout(req);
6139 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6140 req->work.creds != current_cred()) {
6142 revert_creds(old_creds);
6143 if (old_creds == req->work.creds)
6144 old_creds = NULL; /* restored original creds */
6146 old_creds = override_creds(req->work.creds);
6149 ret = io_issue_sqe(req, sqe, true, cs);
6152 * We async punt it if the file wasn't marked NOWAIT, or if the file
6153 * doesn't support non-blocking read/write attempts
6155 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6156 if (!io_arm_poll_handler(req)) {
6158 ret = io_prep_work_files(req);
6162 * Queued up for async execution, worker will release
6163 * submit reference when the iocb is actually submitted.
6165 io_queue_async_work(req);
6169 io_queue_linked_timeout(linked_timeout);
6173 if (unlikely(ret)) {
6175 /* un-prep timeout, so it'll be killed as any other linked */
6176 req->flags &= ~REQ_F_LINK_TIMEOUT;
6177 req_set_fail_links(req);
6179 io_req_complete(req, ret);
6183 /* drop submission reference */
6184 nxt = io_put_req_find_next(req);
6186 io_queue_linked_timeout(linked_timeout);
6191 if (req->flags & REQ_F_FORCE_ASYNC)
6197 revert_creds(old_creds);
6200 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6201 struct io_comp_state *cs)
6205 ret = io_req_defer(req, sqe);
6207 if (ret != -EIOCBQUEUED) {
6209 req_set_fail_links(req);
6211 io_req_complete(req, ret);
6213 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6215 ret = io_req_defer_prep(req, sqe);
6221 * Never try inline submit of IOSQE_ASYNC is set, go straight
6222 * to async execution.
6224 io_req_init_async(req);
6225 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6226 io_queue_async_work(req);
6228 __io_queue_sqe(req, sqe, cs);
6232 static inline void io_queue_link_head(struct io_kiocb *req,
6233 struct io_comp_state *cs)
6235 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6237 io_req_complete(req, -ECANCELED);
6239 io_queue_sqe(req, NULL, cs);
6242 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6243 struct io_kiocb **link, struct io_comp_state *cs)
6245 struct io_ring_ctx *ctx = req->ctx;
6249 * If we already have a head request, queue this one for async
6250 * submittal once the head completes. If we don't have a head but
6251 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6252 * submitted sync once the chain is complete. If none of those
6253 * conditions are true (normal request), then just queue it.
6256 struct io_kiocb *head = *link;
6259 * Taking sequential execution of a link, draining both sides
6260 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6261 * requests in the link. So, it drains the head and the
6262 * next after the link request. The last one is done via
6263 * drain_next flag to persist the effect across calls.
6265 if (req->flags & REQ_F_IO_DRAIN) {
6266 head->flags |= REQ_F_IO_DRAIN;
6267 ctx->drain_next = 1;
6269 ret = io_req_defer_prep(req, sqe);
6270 if (unlikely(ret)) {
6271 /* fail even hard links since we don't submit */
6272 head->flags |= REQ_F_FAIL_LINK;
6275 trace_io_uring_link(ctx, req, head);
6276 io_get_req_task(req);
6277 list_add_tail(&req->link_list, &head->link_list);
6279 /* last request of a link, enqueue the link */
6280 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6281 io_queue_link_head(head, cs);
6285 if (unlikely(ctx->drain_next)) {
6286 req->flags |= REQ_F_IO_DRAIN;
6287 ctx->drain_next = 0;
6289 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6290 req->flags |= REQ_F_LINK_HEAD;
6291 INIT_LIST_HEAD(&req->link_list);
6293 ret = io_req_defer_prep(req, sqe);
6295 req->flags |= REQ_F_FAIL_LINK;
6298 io_queue_sqe(req, sqe, cs);
6306 * Batched submission is done, ensure local IO is flushed out.
6308 static void io_submit_state_end(struct io_submit_state *state)
6310 if (!list_empty(&state->comp.list))
6311 io_submit_flush_completions(&state->comp);
6312 blk_finish_plug(&state->plug);
6313 io_state_file_put(state);
6314 if (state->free_reqs)
6315 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6319 * Start submission side cache.
6321 static void io_submit_state_start(struct io_submit_state *state,
6322 struct io_ring_ctx *ctx, unsigned int max_ios)
6324 blk_start_plug(&state->plug);
6326 state->plug.nowait = true;
6329 INIT_LIST_HEAD(&state->comp.list);
6330 state->comp.ctx = ctx;
6331 state->free_reqs = 0;
6333 state->ios_left = max_ios;
6336 static void io_commit_sqring(struct io_ring_ctx *ctx)
6338 struct io_rings *rings = ctx->rings;
6341 * Ensure any loads from the SQEs are done at this point,
6342 * since once we write the new head, the application could
6343 * write new data to them.
6345 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6349 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6350 * that is mapped by userspace. This means that care needs to be taken to
6351 * ensure that reads are stable, as we cannot rely on userspace always
6352 * being a good citizen. If members of the sqe are validated and then later
6353 * used, it's important that those reads are done through READ_ONCE() to
6354 * prevent a re-load down the line.
6356 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6358 u32 *sq_array = ctx->sq_array;
6362 * The cached sq head (or cq tail) serves two purposes:
6364 * 1) allows us to batch the cost of updating the user visible
6366 * 2) allows the kernel side to track the head on its own, even
6367 * though the application is the one updating it.
6369 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6370 if (likely(head < ctx->sq_entries))
6371 return &ctx->sq_sqes[head];
6373 /* drop invalid entries */
6374 ctx->cached_sq_dropped++;
6375 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6379 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6381 ctx->cached_sq_head++;
6384 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6385 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6386 IOSQE_BUFFER_SELECT)
6388 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6389 const struct io_uring_sqe *sqe,
6390 struct io_submit_state *state)
6392 unsigned int sqe_flags;
6395 req->opcode = READ_ONCE(sqe->opcode);
6396 req->user_data = READ_ONCE(sqe->user_data);
6401 /* one is dropped after submission, the other at completion */
6402 refcount_set(&req->refs, 2);
6403 req->task = current;
6406 if (unlikely(req->opcode >= IORING_OP_LAST))
6409 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6412 sqe_flags = READ_ONCE(sqe->flags);
6413 /* enforce forwards compatibility on users */
6414 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6417 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6418 !io_op_defs[req->opcode].buffer_select)
6421 id = READ_ONCE(sqe->personality);
6423 io_req_init_async(req);
6424 req->work.creds = idr_find(&ctx->personality_idr, id);
6425 if (unlikely(!req->work.creds))
6427 get_cred(req->work.creds);
6430 /* same numerical values with corresponding REQ_F_*, safe to copy */
6431 req->flags |= sqe_flags;
6433 if (!io_op_defs[req->opcode].needs_file)
6436 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6439 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6440 struct file *ring_file, int ring_fd)
6442 struct io_submit_state state;
6443 struct io_kiocb *link = NULL;
6444 int i, submitted = 0;
6446 /* if we have a backlog and couldn't flush it all, return BUSY */
6447 if (test_bit(0, &ctx->sq_check_overflow)) {
6448 if (!list_empty(&ctx->cq_overflow_list) &&
6449 !io_cqring_overflow_flush(ctx, false))
6453 /* make sure SQ entry isn't read before tail */
6454 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6456 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6459 io_submit_state_start(&state, ctx, nr);
6461 ctx->ring_fd = ring_fd;
6462 ctx->ring_file = ring_file;
6464 for (i = 0; i < nr; i++) {
6465 const struct io_uring_sqe *sqe;
6466 struct io_kiocb *req;
6469 sqe = io_get_sqe(ctx);
6470 if (unlikely(!sqe)) {
6471 io_consume_sqe(ctx);
6474 req = io_alloc_req(ctx, &state);
6475 if (unlikely(!req)) {
6477 submitted = -EAGAIN;
6481 err = io_init_req(ctx, req, sqe, &state);
6482 io_consume_sqe(ctx);
6483 /* will complete beyond this point, count as submitted */
6486 if (unlikely(err)) {
6489 io_req_complete(req, err);
6493 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6494 true, io_async_submit(ctx));
6495 err = io_submit_sqe(req, sqe, &link, &state.comp);
6500 if (unlikely(submitted != nr)) {
6501 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6503 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6506 io_queue_link_head(link, &state.comp);
6507 io_submit_state_end(&state);
6509 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6510 io_commit_sqring(ctx);
6515 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6517 /* Tell userspace we may need a wakeup call */
6518 spin_lock_irq(&ctx->completion_lock);
6519 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6520 spin_unlock_irq(&ctx->completion_lock);
6523 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6525 spin_lock_irq(&ctx->completion_lock);
6526 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6527 spin_unlock_irq(&ctx->completion_lock);
6530 static int io_sq_thread(void *data)
6532 struct io_ring_ctx *ctx = data;
6533 const struct cred *old_cred;
6535 unsigned long timeout;
6538 complete(&ctx->sq_thread_comp);
6540 old_cred = override_creds(ctx->creds);
6542 timeout = jiffies + ctx->sq_thread_idle;
6543 while (!kthread_should_park()) {
6544 unsigned int to_submit;
6546 if (!list_empty(&ctx->iopoll_list)) {
6547 unsigned nr_events = 0;
6549 mutex_lock(&ctx->uring_lock);
6550 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6551 io_do_iopoll(ctx, &nr_events, 0);
6553 timeout = jiffies + ctx->sq_thread_idle;
6554 mutex_unlock(&ctx->uring_lock);
6557 to_submit = io_sqring_entries(ctx);
6560 * If submit got -EBUSY, flag us as needing the application
6561 * to enter the kernel to reap and flush events.
6563 if (!to_submit || ret == -EBUSY || need_resched()) {
6565 * Drop cur_mm before scheduling, we can't hold it for
6566 * long periods (or over schedule()). Do this before
6567 * adding ourselves to the waitqueue, as the unuse/drop
6570 io_sq_thread_drop_mm();
6573 * We're polling. If we're within the defined idle
6574 * period, then let us spin without work before going
6575 * to sleep. The exception is if we got EBUSY doing
6576 * more IO, we should wait for the application to
6577 * reap events and wake us up.
6579 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6580 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6581 !percpu_ref_is_dying(&ctx->refs))) {
6587 prepare_to_wait(&ctx->sqo_wait, &wait,
6588 TASK_INTERRUPTIBLE);
6591 * While doing polled IO, before going to sleep, we need
6592 * to check if there are new reqs added to iopoll_list,
6593 * it is because reqs may have been punted to io worker
6594 * and will be added to iopoll_list later, hence check
6595 * the iopoll_list again.
6597 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6598 !list_empty_careful(&ctx->iopoll_list)) {
6599 finish_wait(&ctx->sqo_wait, &wait);
6603 io_ring_set_wakeup_flag(ctx);
6605 to_submit = io_sqring_entries(ctx);
6606 if (!to_submit || ret == -EBUSY) {
6607 if (kthread_should_park()) {
6608 finish_wait(&ctx->sqo_wait, &wait);
6611 if (io_run_task_work()) {
6612 finish_wait(&ctx->sqo_wait, &wait);
6613 io_ring_clear_wakeup_flag(ctx);
6616 if (signal_pending(current))
6617 flush_signals(current);
6619 finish_wait(&ctx->sqo_wait, &wait);
6621 io_ring_clear_wakeup_flag(ctx);
6625 finish_wait(&ctx->sqo_wait, &wait);
6627 io_ring_clear_wakeup_flag(ctx);
6630 mutex_lock(&ctx->uring_lock);
6631 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6632 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6633 mutex_unlock(&ctx->uring_lock);
6634 timeout = jiffies + ctx->sq_thread_idle;
6639 io_sq_thread_drop_mm();
6640 revert_creds(old_cred);
6647 struct io_wait_queue {
6648 struct wait_queue_entry wq;
6649 struct io_ring_ctx *ctx;
6651 unsigned nr_timeouts;
6654 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6656 struct io_ring_ctx *ctx = iowq->ctx;
6659 * Wake up if we have enough events, or if a timeout occurred since we
6660 * started waiting. For timeouts, we always want to return to userspace,
6661 * regardless of event count.
6663 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6664 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6667 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6668 int wake_flags, void *key)
6670 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6673 /* use noflush == true, as we can't safely rely on locking context */
6674 if (!io_should_wake(iowq, true))
6677 return autoremove_wake_function(curr, mode, wake_flags, key);
6681 * Wait until events become available, if we don't already have some. The
6682 * application must reap them itself, as they reside on the shared cq ring.
6684 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6685 const sigset_t __user *sig, size_t sigsz)
6687 struct io_wait_queue iowq = {
6690 .func = io_wake_function,
6691 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6694 .to_wait = min_events,
6696 struct io_rings *rings = ctx->rings;
6700 if (io_cqring_events(ctx, false) >= min_events)
6702 if (!io_run_task_work())
6707 #ifdef CONFIG_COMPAT
6708 if (in_compat_syscall())
6709 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6713 ret = set_user_sigmask(sig, sigsz);
6719 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6720 trace_io_uring_cqring_wait(ctx, min_events);
6722 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6723 TASK_INTERRUPTIBLE);
6724 /* make sure we run task_work before checking for signals */
6725 if (io_run_task_work())
6727 if (signal_pending(current)) {
6728 if (current->jobctl & JOBCTL_TASK_WORK) {
6729 spin_lock_irq(¤t->sighand->siglock);
6730 current->jobctl &= ~JOBCTL_TASK_WORK;
6731 recalc_sigpending();
6732 spin_unlock_irq(¤t->sighand->siglock);
6738 if (io_should_wake(&iowq, false))
6742 finish_wait(&ctx->wait, &iowq.wq);
6744 restore_saved_sigmask_unless(ret == -EINTR);
6746 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6749 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6751 #if defined(CONFIG_UNIX)
6752 if (ctx->ring_sock) {
6753 struct sock *sock = ctx->ring_sock->sk;
6754 struct sk_buff *skb;
6756 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6762 for (i = 0; i < ctx->nr_user_files; i++) {
6765 file = io_file_from_index(ctx, i);
6772 static void io_file_ref_kill(struct percpu_ref *ref)
6774 struct fixed_file_data *data;
6776 data = container_of(ref, struct fixed_file_data, refs);
6777 complete(&data->done);
6780 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6782 struct fixed_file_data *data = ctx->file_data;
6783 struct fixed_file_ref_node *ref_node = NULL;
6784 unsigned nr_tables, i;
6789 spin_lock(&data->lock);
6790 if (!list_empty(&data->ref_list))
6791 ref_node = list_first_entry(&data->ref_list,
6792 struct fixed_file_ref_node, node);
6793 spin_unlock(&data->lock);
6795 percpu_ref_kill(&ref_node->refs);
6797 percpu_ref_kill(&data->refs);
6799 /* wait for all refs nodes to complete */
6800 flush_delayed_work(&ctx->file_put_work);
6801 wait_for_completion(&data->done);
6803 __io_sqe_files_unregister(ctx);
6804 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6805 for (i = 0; i < nr_tables; i++)
6806 kfree(data->table[i].files);
6808 percpu_ref_exit(&data->refs);
6810 ctx->file_data = NULL;
6811 ctx->nr_user_files = 0;
6815 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6817 if (ctx->sqo_thread) {
6818 wait_for_completion(&ctx->sq_thread_comp);
6820 * The park is a bit of a work-around, without it we get
6821 * warning spews on shutdown with SQPOLL set and affinity
6822 * set to a single CPU.
6824 kthread_park(ctx->sqo_thread);
6825 kthread_stop(ctx->sqo_thread);
6826 ctx->sqo_thread = NULL;
6830 static void io_finish_async(struct io_ring_ctx *ctx)
6832 io_sq_thread_stop(ctx);
6835 io_wq_destroy(ctx->io_wq);
6840 #if defined(CONFIG_UNIX)
6842 * Ensure the UNIX gc is aware of our file set, so we are certain that
6843 * the io_uring can be safely unregistered on process exit, even if we have
6844 * loops in the file referencing.
6846 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6848 struct sock *sk = ctx->ring_sock->sk;
6849 struct scm_fp_list *fpl;
6850 struct sk_buff *skb;
6853 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6857 skb = alloc_skb(0, GFP_KERNEL);
6866 fpl->user = get_uid(ctx->user);
6867 for (i = 0; i < nr; i++) {
6868 struct file *file = io_file_from_index(ctx, i + offset);
6872 fpl->fp[nr_files] = get_file(file);
6873 unix_inflight(fpl->user, fpl->fp[nr_files]);
6878 fpl->max = SCM_MAX_FD;
6879 fpl->count = nr_files;
6880 UNIXCB(skb).fp = fpl;
6881 skb->destructor = unix_destruct_scm;
6882 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6883 skb_queue_head(&sk->sk_receive_queue, skb);
6885 for (i = 0; i < nr_files; i++)
6896 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6897 * causes regular reference counting to break down. We rely on the UNIX
6898 * garbage collection to take care of this problem for us.
6900 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6902 unsigned left, total;
6906 left = ctx->nr_user_files;
6908 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6910 ret = __io_sqe_files_scm(ctx, this_files, total);
6914 total += this_files;
6920 while (total < ctx->nr_user_files) {
6921 struct file *file = io_file_from_index(ctx, total);
6931 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6937 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6942 for (i = 0; i < nr_tables; i++) {
6943 struct fixed_file_table *table = &ctx->file_data->table[i];
6944 unsigned this_files;
6946 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6947 table->files = kcalloc(this_files, sizeof(struct file *),
6951 nr_files -= this_files;
6957 for (i = 0; i < nr_tables; i++) {
6958 struct fixed_file_table *table = &ctx->file_data->table[i];
6959 kfree(table->files);
6964 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6966 #if defined(CONFIG_UNIX)
6967 struct sock *sock = ctx->ring_sock->sk;
6968 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6969 struct sk_buff *skb;
6972 __skb_queue_head_init(&list);
6975 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6976 * remove this entry and rearrange the file array.
6978 skb = skb_dequeue(head);
6980 struct scm_fp_list *fp;
6982 fp = UNIXCB(skb).fp;
6983 for (i = 0; i < fp->count; i++) {
6986 if (fp->fp[i] != file)
6989 unix_notinflight(fp->user, fp->fp[i]);
6990 left = fp->count - 1 - i;
6992 memmove(&fp->fp[i], &fp->fp[i + 1],
6993 left * sizeof(struct file *));
7000 __skb_queue_tail(&list, skb);
7010 __skb_queue_tail(&list, skb);
7012 skb = skb_dequeue(head);
7015 if (skb_peek(&list)) {
7016 spin_lock_irq(&head->lock);
7017 while ((skb = __skb_dequeue(&list)) != NULL)
7018 __skb_queue_tail(head, skb);
7019 spin_unlock_irq(&head->lock);
7026 struct io_file_put {
7027 struct list_head list;
7031 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7033 struct fixed_file_data *file_data = ref_node->file_data;
7034 struct io_ring_ctx *ctx = file_data->ctx;
7035 struct io_file_put *pfile, *tmp;
7037 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7038 list_del(&pfile->list);
7039 io_ring_file_put(ctx, pfile->file);
7043 spin_lock(&file_data->lock);
7044 list_del(&ref_node->node);
7045 spin_unlock(&file_data->lock);
7047 percpu_ref_exit(&ref_node->refs);
7049 percpu_ref_put(&file_data->refs);
7052 static void io_file_put_work(struct work_struct *work)
7054 struct io_ring_ctx *ctx;
7055 struct llist_node *node;
7057 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7058 node = llist_del_all(&ctx->file_put_llist);
7061 struct fixed_file_ref_node *ref_node;
7062 struct llist_node *next = node->next;
7064 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7065 __io_file_put_work(ref_node);
7070 static void io_file_data_ref_zero(struct percpu_ref *ref)
7072 struct fixed_file_ref_node *ref_node;
7073 struct io_ring_ctx *ctx;
7077 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7078 ctx = ref_node->file_data->ctx;
7080 if (percpu_ref_is_dying(&ctx->file_data->refs))
7083 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7085 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7087 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7090 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7091 struct io_ring_ctx *ctx)
7093 struct fixed_file_ref_node *ref_node;
7095 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7097 return ERR_PTR(-ENOMEM);
7099 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7102 return ERR_PTR(-ENOMEM);
7104 INIT_LIST_HEAD(&ref_node->node);
7105 INIT_LIST_HEAD(&ref_node->file_list);
7106 ref_node->file_data = ctx->file_data;
7110 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7112 percpu_ref_exit(&ref_node->refs);
7116 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7119 __s32 __user *fds = (__s32 __user *) arg;
7124 struct fixed_file_ref_node *ref_node;
7130 if (nr_args > IORING_MAX_FIXED_FILES)
7133 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7134 if (!ctx->file_data)
7136 ctx->file_data->ctx = ctx;
7137 init_completion(&ctx->file_data->done);
7138 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7139 spin_lock_init(&ctx->file_data->lock);
7141 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7142 ctx->file_data->table = kcalloc(nr_tables,
7143 sizeof(struct fixed_file_table),
7145 if (!ctx->file_data->table) {
7146 kfree(ctx->file_data);
7147 ctx->file_data = NULL;
7151 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7152 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7153 kfree(ctx->file_data->table);
7154 kfree(ctx->file_data);
7155 ctx->file_data = NULL;
7159 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7160 percpu_ref_exit(&ctx->file_data->refs);
7161 kfree(ctx->file_data->table);
7162 kfree(ctx->file_data);
7163 ctx->file_data = NULL;
7167 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7168 struct fixed_file_table *table;
7172 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7174 /* allow sparse sets */
7180 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7181 index = i & IORING_FILE_TABLE_MASK;
7189 * Don't allow io_uring instances to be registered. If UNIX
7190 * isn't enabled, then this causes a reference cycle and this
7191 * instance can never get freed. If UNIX is enabled we'll
7192 * handle it just fine, but there's still no point in allowing
7193 * a ring fd as it doesn't support regular read/write anyway.
7195 if (file->f_op == &io_uring_fops) {
7200 table->files[index] = file;
7204 for (i = 0; i < ctx->nr_user_files; i++) {
7205 file = io_file_from_index(ctx, i);
7209 for (i = 0; i < nr_tables; i++)
7210 kfree(ctx->file_data->table[i].files);
7212 percpu_ref_exit(&ctx->file_data->refs);
7213 kfree(ctx->file_data->table);
7214 kfree(ctx->file_data);
7215 ctx->file_data = NULL;
7216 ctx->nr_user_files = 0;
7220 ret = io_sqe_files_scm(ctx);
7222 io_sqe_files_unregister(ctx);
7226 ref_node = alloc_fixed_file_ref_node(ctx);
7227 if (IS_ERR(ref_node)) {
7228 io_sqe_files_unregister(ctx);
7229 return PTR_ERR(ref_node);
7232 ctx->file_data->cur_refs = &ref_node->refs;
7233 spin_lock(&ctx->file_data->lock);
7234 list_add(&ref_node->node, &ctx->file_data->ref_list);
7235 spin_unlock(&ctx->file_data->lock);
7236 percpu_ref_get(&ctx->file_data->refs);
7240 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7243 #if defined(CONFIG_UNIX)
7244 struct sock *sock = ctx->ring_sock->sk;
7245 struct sk_buff_head *head = &sock->sk_receive_queue;
7246 struct sk_buff *skb;
7249 * See if we can merge this file into an existing skb SCM_RIGHTS
7250 * file set. If there's no room, fall back to allocating a new skb
7251 * and filling it in.
7253 spin_lock_irq(&head->lock);
7254 skb = skb_peek(head);
7256 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7258 if (fpl->count < SCM_MAX_FD) {
7259 __skb_unlink(skb, head);
7260 spin_unlock_irq(&head->lock);
7261 fpl->fp[fpl->count] = get_file(file);
7262 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7264 spin_lock_irq(&head->lock);
7265 __skb_queue_head(head, skb);
7270 spin_unlock_irq(&head->lock);
7277 return __io_sqe_files_scm(ctx, 1, index);
7283 static int io_queue_file_removal(struct fixed_file_data *data,
7286 struct io_file_put *pfile;
7287 struct percpu_ref *refs = data->cur_refs;
7288 struct fixed_file_ref_node *ref_node;
7290 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7294 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7296 list_add(&pfile->list, &ref_node->file_list);
7301 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7302 struct io_uring_files_update *up,
7305 struct fixed_file_data *data = ctx->file_data;
7306 struct fixed_file_ref_node *ref_node;
7311 bool needs_switch = false;
7313 if (check_add_overflow(up->offset, nr_args, &done))
7315 if (done > ctx->nr_user_files)
7318 ref_node = alloc_fixed_file_ref_node(ctx);
7319 if (IS_ERR(ref_node))
7320 return PTR_ERR(ref_node);
7323 fds = u64_to_user_ptr(up->fds);
7325 struct fixed_file_table *table;
7329 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7333 i = array_index_nospec(up->offset, ctx->nr_user_files);
7334 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7335 index = i & IORING_FILE_TABLE_MASK;
7336 if (table->files[index]) {
7337 file = table->files[index];
7338 err = io_queue_file_removal(data, file);
7341 table->files[index] = NULL;
7342 needs_switch = true;
7351 * Don't allow io_uring instances to be registered. If
7352 * UNIX isn't enabled, then this causes a reference
7353 * cycle and this instance can never get freed. If UNIX
7354 * is enabled we'll handle it just fine, but there's
7355 * still no point in allowing a ring fd as it doesn't
7356 * support regular read/write anyway.
7358 if (file->f_op == &io_uring_fops) {
7363 table->files[index] = file;
7364 err = io_sqe_file_register(ctx, file, i);
7366 table->files[index] = NULL;
7377 percpu_ref_kill(data->cur_refs);
7378 spin_lock(&data->lock);
7379 list_add(&ref_node->node, &data->ref_list);
7380 data->cur_refs = &ref_node->refs;
7381 spin_unlock(&data->lock);
7382 percpu_ref_get(&ctx->file_data->refs);
7384 destroy_fixed_file_ref_node(ref_node);
7386 return done ? done : err;
7389 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7392 struct io_uring_files_update up;
7394 if (!ctx->file_data)
7398 if (copy_from_user(&up, arg, sizeof(up)))
7403 return __io_sqe_files_update(ctx, &up, nr_args);
7406 static void io_free_work(struct io_wq_work *work)
7408 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7410 /* Consider that io_steal_work() relies on this ref */
7414 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7415 struct io_uring_params *p)
7417 struct io_wq_data data;
7419 struct io_ring_ctx *ctx_attach;
7420 unsigned int concurrency;
7423 data.user = ctx->user;
7424 data.free_work = io_free_work;
7425 data.do_work = io_wq_submit_work;
7427 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7428 /* Do QD, or 4 * CPUS, whatever is smallest */
7429 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7431 ctx->io_wq = io_wq_create(concurrency, &data);
7432 if (IS_ERR(ctx->io_wq)) {
7433 ret = PTR_ERR(ctx->io_wq);
7439 f = fdget(p->wq_fd);
7443 if (f.file->f_op != &io_uring_fops) {
7448 ctx_attach = f.file->private_data;
7449 /* @io_wq is protected by holding the fd */
7450 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7455 ctx->io_wq = ctx_attach->io_wq;
7461 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7462 struct io_uring_params *p)
7466 if (ctx->flags & IORING_SETUP_SQPOLL) {
7468 if (!capable(CAP_SYS_ADMIN))
7471 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7472 if (!ctx->sq_thread_idle)
7473 ctx->sq_thread_idle = HZ;
7475 if (p->flags & IORING_SETUP_SQ_AFF) {
7476 int cpu = p->sq_thread_cpu;
7479 if (cpu >= nr_cpu_ids)
7481 if (!cpu_online(cpu))
7484 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7488 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7491 if (IS_ERR(ctx->sqo_thread)) {
7492 ret = PTR_ERR(ctx->sqo_thread);
7493 ctx->sqo_thread = NULL;
7496 wake_up_process(ctx->sqo_thread);
7497 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7498 /* Can't have SQ_AFF without SQPOLL */
7503 ret = io_init_wq_offload(ctx, p);
7509 io_finish_async(ctx);
7513 static inline void __io_unaccount_mem(struct user_struct *user,
7514 unsigned long nr_pages)
7516 atomic_long_sub(nr_pages, &user->locked_vm);
7519 static inline int __io_account_mem(struct user_struct *user,
7520 unsigned long nr_pages)
7522 unsigned long page_limit, cur_pages, new_pages;
7524 /* Don't allow more pages than we can safely lock */
7525 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7528 cur_pages = atomic_long_read(&user->locked_vm);
7529 new_pages = cur_pages + nr_pages;
7530 if (new_pages > page_limit)
7532 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7533 new_pages) != cur_pages);
7538 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7539 enum io_mem_account acct)
7542 __io_unaccount_mem(ctx->user, nr_pages);
7545 if (acct == ACCT_LOCKED)
7546 ctx->sqo_mm->locked_vm -= nr_pages;
7547 else if (acct == ACCT_PINNED)
7548 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7552 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7553 enum io_mem_account acct)
7557 if (ctx->limit_mem) {
7558 ret = __io_account_mem(ctx->user, nr_pages);
7564 if (acct == ACCT_LOCKED)
7565 ctx->sqo_mm->locked_vm += nr_pages;
7566 else if (acct == ACCT_PINNED)
7567 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7573 static void io_mem_free(void *ptr)
7580 page = virt_to_head_page(ptr);
7581 if (put_page_testzero(page))
7582 free_compound_page(page);
7585 static void *io_mem_alloc(size_t size)
7587 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7590 return (void *) __get_free_pages(gfp_flags, get_order(size));
7593 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7596 struct io_rings *rings;
7597 size_t off, sq_array_size;
7599 off = struct_size(rings, cqes, cq_entries);
7600 if (off == SIZE_MAX)
7604 off = ALIGN(off, SMP_CACHE_BYTES);
7612 sq_array_size = array_size(sizeof(u32), sq_entries);
7613 if (sq_array_size == SIZE_MAX)
7616 if (check_add_overflow(off, sq_array_size, &off))
7622 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7626 pages = (size_t)1 << get_order(
7627 rings_size(sq_entries, cq_entries, NULL));
7628 pages += (size_t)1 << get_order(
7629 array_size(sizeof(struct io_uring_sqe), sq_entries));
7634 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7638 if (!ctx->user_bufs)
7641 for (i = 0; i < ctx->nr_user_bufs; i++) {
7642 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7644 for (j = 0; j < imu->nr_bvecs; j++)
7645 unpin_user_page(imu->bvec[j].bv_page);
7647 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7652 kfree(ctx->user_bufs);
7653 ctx->user_bufs = NULL;
7654 ctx->nr_user_bufs = 0;
7658 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7659 void __user *arg, unsigned index)
7661 struct iovec __user *src;
7663 #ifdef CONFIG_COMPAT
7665 struct compat_iovec __user *ciovs;
7666 struct compat_iovec ciov;
7668 ciovs = (struct compat_iovec __user *) arg;
7669 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7672 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7673 dst->iov_len = ciov.iov_len;
7677 src = (struct iovec __user *) arg;
7678 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7683 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7686 struct vm_area_struct **vmas = NULL;
7687 struct page **pages = NULL;
7688 int i, j, got_pages = 0;
7693 if (!nr_args || nr_args > UIO_MAXIOV)
7696 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7698 if (!ctx->user_bufs)
7701 for (i = 0; i < nr_args; i++) {
7702 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7703 unsigned long off, start, end, ubuf;
7708 ret = io_copy_iov(ctx, &iov, arg, i);
7713 * Don't impose further limits on the size and buffer
7714 * constraints here, we'll -EINVAL later when IO is
7715 * submitted if they are wrong.
7718 if (!iov.iov_base || !iov.iov_len)
7721 /* arbitrary limit, but we need something */
7722 if (iov.iov_len > SZ_1G)
7725 ubuf = (unsigned long) iov.iov_base;
7726 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7727 start = ubuf >> PAGE_SHIFT;
7728 nr_pages = end - start;
7730 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7735 if (!pages || nr_pages > got_pages) {
7738 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7740 vmas = kvmalloc_array(nr_pages,
7741 sizeof(struct vm_area_struct *),
7743 if (!pages || !vmas) {
7745 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7748 got_pages = nr_pages;
7751 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7755 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7760 mmap_read_lock(current->mm);
7761 pret = pin_user_pages(ubuf, nr_pages,
7762 FOLL_WRITE | FOLL_LONGTERM,
7764 if (pret == nr_pages) {
7765 /* don't support file backed memory */
7766 for (j = 0; j < nr_pages; j++) {
7767 struct vm_area_struct *vma = vmas[j];
7770 !is_file_hugepages(vma->vm_file)) {
7776 ret = pret < 0 ? pret : -EFAULT;
7778 mmap_read_unlock(current->mm);
7781 * if we did partial map, or found file backed vmas,
7782 * release any pages we did get
7785 unpin_user_pages(pages, pret);
7786 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7791 off = ubuf & ~PAGE_MASK;
7793 for (j = 0; j < nr_pages; j++) {
7796 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7797 imu->bvec[j].bv_page = pages[j];
7798 imu->bvec[j].bv_len = vec_len;
7799 imu->bvec[j].bv_offset = off;
7803 /* store original address for later verification */
7805 imu->len = iov.iov_len;
7806 imu->nr_bvecs = nr_pages;
7808 ctx->nr_user_bufs++;
7816 io_sqe_buffer_unregister(ctx);
7820 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7822 __s32 __user *fds = arg;
7828 if (copy_from_user(&fd, fds, sizeof(*fds)))
7831 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7832 if (IS_ERR(ctx->cq_ev_fd)) {
7833 int ret = PTR_ERR(ctx->cq_ev_fd);
7834 ctx->cq_ev_fd = NULL;
7841 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7843 if (ctx->cq_ev_fd) {
7844 eventfd_ctx_put(ctx->cq_ev_fd);
7845 ctx->cq_ev_fd = NULL;
7852 static int __io_destroy_buffers(int id, void *p, void *data)
7854 struct io_ring_ctx *ctx = data;
7855 struct io_buffer *buf = p;
7857 __io_remove_buffers(ctx, buf, id, -1U);
7861 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7863 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7864 idr_destroy(&ctx->io_buffer_idr);
7867 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7869 io_finish_async(ctx);
7870 io_sqe_buffer_unregister(ctx);
7872 mmdrop(ctx->sqo_mm);
7876 io_sqe_files_unregister(ctx);
7877 io_eventfd_unregister(ctx);
7878 io_destroy_buffers(ctx);
7879 idr_destroy(&ctx->personality_idr);
7881 #if defined(CONFIG_UNIX)
7882 if (ctx->ring_sock) {
7883 ctx->ring_sock->file = NULL; /* so that iput() is called */
7884 sock_release(ctx->ring_sock);
7888 io_mem_free(ctx->rings);
7889 io_mem_free(ctx->sq_sqes);
7891 percpu_ref_exit(&ctx->refs);
7892 free_uid(ctx->user);
7893 put_cred(ctx->creds);
7894 kfree(ctx->cancel_hash);
7895 kmem_cache_free(req_cachep, ctx->fallback_req);
7899 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7901 struct io_ring_ctx *ctx = file->private_data;
7904 poll_wait(file, &ctx->cq_wait, wait);
7906 * synchronizes with barrier from wq_has_sleeper call in
7910 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7911 ctx->rings->sq_ring_entries)
7912 mask |= EPOLLOUT | EPOLLWRNORM;
7913 if (io_cqring_events(ctx, false))
7914 mask |= EPOLLIN | EPOLLRDNORM;
7919 static int io_uring_fasync(int fd, struct file *file, int on)
7921 struct io_ring_ctx *ctx = file->private_data;
7923 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7926 static int io_remove_personalities(int id, void *p, void *data)
7928 struct io_ring_ctx *ctx = data;
7929 const struct cred *cred;
7931 cred = idr_remove(&ctx->personality_idr, id);
7937 static void io_ring_exit_work(struct work_struct *work)
7939 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7943 * If we're doing polled IO and end up having requests being
7944 * submitted async (out-of-line), then completions can come in while
7945 * we're waiting for refs to drop. We need to reap these manually,
7946 * as nobody else will be looking for them.
7950 io_cqring_overflow_flush(ctx, true);
7951 io_iopoll_try_reap_events(ctx);
7952 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7953 io_ring_ctx_free(ctx);
7956 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7958 mutex_lock(&ctx->uring_lock);
7959 percpu_ref_kill(&ctx->refs);
7960 mutex_unlock(&ctx->uring_lock);
7962 io_kill_timeouts(ctx);
7963 io_poll_remove_all(ctx);
7966 io_wq_cancel_all(ctx->io_wq);
7968 /* if we failed setting up the ctx, we might not have any rings */
7970 io_cqring_overflow_flush(ctx, true);
7971 io_iopoll_try_reap_events(ctx);
7972 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7975 * Do this upfront, so we won't have a grace period where the ring
7976 * is closed but resources aren't reaped yet. This can cause
7977 * spurious failure in setting up a new ring.
7979 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7982 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7984 * Use system_unbound_wq to avoid spawning tons of event kworkers
7985 * if we're exiting a ton of rings at the same time. It just adds
7986 * noise and overhead, there's no discernable change in runtime
7987 * over using system_wq.
7989 queue_work(system_unbound_wq, &ctx->exit_work);
7992 static int io_uring_release(struct inode *inode, struct file *file)
7994 struct io_ring_ctx *ctx = file->private_data;
7996 file->private_data = NULL;
7997 io_ring_ctx_wait_and_kill(ctx);
8001 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8003 struct files_struct *files = data;
8005 return work->files == files;
8009 * Returns true if 'preq' is the link parent of 'req'
8011 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8013 struct io_kiocb *link;
8015 if (!(preq->flags & REQ_F_LINK_HEAD))
8018 list_for_each_entry(link, &preq->link_list, link_list) {
8027 * We're looking to cancel 'req' because it's holding on to our files, but
8028 * 'req' could be a link to another request. See if it is, and cancel that
8029 * parent request if so.
8031 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8033 struct hlist_node *tmp;
8034 struct io_kiocb *preq;
8038 spin_lock_irq(&ctx->completion_lock);
8039 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8040 struct hlist_head *list;
8042 list = &ctx->cancel_hash[i];
8043 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8044 found = io_match_link(preq, req);
8046 io_poll_remove_one(preq);
8051 spin_unlock_irq(&ctx->completion_lock);
8055 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8056 struct io_kiocb *req)
8058 struct io_kiocb *preq;
8061 spin_lock_irq(&ctx->completion_lock);
8062 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8063 found = io_match_link(preq, req);
8065 __io_timeout_cancel(preq);
8069 spin_unlock_irq(&ctx->completion_lock);
8073 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8075 return io_match_link(container_of(work, struct io_kiocb, work), data);
8078 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8080 enum io_wq_cancel cret;
8082 /* cancel this particular work, if it's running */
8083 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8084 if (cret != IO_WQ_CANCEL_NOTFOUND)
8087 /* find links that hold this pending, cancel those */
8088 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8089 if (cret != IO_WQ_CANCEL_NOTFOUND)
8092 /* if we have a poll link holding this pending, cancel that */
8093 if (io_poll_remove_link(ctx, req))
8096 /* final option, timeout link is holding this req pending */
8097 io_timeout_remove_link(ctx, req);
8100 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8101 struct files_struct *files)
8103 if (list_empty_careful(&ctx->inflight_list))
8106 /* cancel all at once, should be faster than doing it one by one*/
8107 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8109 while (!list_empty_careful(&ctx->inflight_list)) {
8110 struct io_kiocb *cancel_req = NULL, *req;
8113 spin_lock_irq(&ctx->inflight_lock);
8114 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8115 if (req->work.files != files)
8117 /* req is being completed, ignore */
8118 if (!refcount_inc_not_zero(&req->refs))
8124 prepare_to_wait(&ctx->inflight_wait, &wait,
8125 TASK_UNINTERRUPTIBLE);
8126 spin_unlock_irq(&ctx->inflight_lock);
8128 /* We need to keep going until we don't find a matching req */
8131 /* cancel this request, or head link requests */
8132 io_attempt_cancel(ctx, cancel_req);
8133 io_put_req(cancel_req);
8135 finish_wait(&ctx->inflight_wait, &wait);
8139 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8141 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8142 struct task_struct *task = data;
8144 return req->task == task;
8147 static int io_uring_flush(struct file *file, void *data)
8149 struct io_ring_ctx *ctx = file->private_data;
8151 io_uring_cancel_files(ctx, data);
8154 * If the task is going away, cancel work it may have pending
8156 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8157 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
8162 static void *io_uring_validate_mmap_request(struct file *file,
8163 loff_t pgoff, size_t sz)
8165 struct io_ring_ctx *ctx = file->private_data;
8166 loff_t offset = pgoff << PAGE_SHIFT;
8171 case IORING_OFF_SQ_RING:
8172 case IORING_OFF_CQ_RING:
8175 case IORING_OFF_SQES:
8179 return ERR_PTR(-EINVAL);
8182 page = virt_to_head_page(ptr);
8183 if (sz > page_size(page))
8184 return ERR_PTR(-EINVAL);
8191 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8193 size_t sz = vma->vm_end - vma->vm_start;
8197 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8199 return PTR_ERR(ptr);
8201 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8202 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8205 #else /* !CONFIG_MMU */
8207 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8209 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8212 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8214 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8217 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8218 unsigned long addr, unsigned long len,
8219 unsigned long pgoff, unsigned long flags)
8223 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8225 return PTR_ERR(ptr);
8227 return (unsigned long) ptr;
8230 #endif /* !CONFIG_MMU */
8232 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8233 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8236 struct io_ring_ctx *ctx;
8243 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8251 if (f.file->f_op != &io_uring_fops)
8255 ctx = f.file->private_data;
8256 if (!percpu_ref_tryget(&ctx->refs))
8260 * For SQ polling, the thread will do all submissions and completions.
8261 * Just return the requested submit count, and wake the thread if
8265 if (ctx->flags & IORING_SETUP_SQPOLL) {
8266 if (!list_empty_careful(&ctx->cq_overflow_list))
8267 io_cqring_overflow_flush(ctx, false);
8268 if (flags & IORING_ENTER_SQ_WAKEUP)
8269 wake_up(&ctx->sqo_wait);
8270 submitted = to_submit;
8271 } else if (to_submit) {
8272 mutex_lock(&ctx->uring_lock);
8273 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8274 mutex_unlock(&ctx->uring_lock);
8276 if (submitted != to_submit)
8279 if (flags & IORING_ENTER_GETEVENTS) {
8280 min_complete = min(min_complete, ctx->cq_entries);
8283 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8284 * space applications don't need to do io completion events
8285 * polling again, they can rely on io_sq_thread to do polling
8286 * work, which can reduce cpu usage and uring_lock contention.
8288 if (ctx->flags & IORING_SETUP_IOPOLL &&
8289 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8290 ret = io_iopoll_check(ctx, min_complete);
8292 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8297 percpu_ref_put(&ctx->refs);
8300 return submitted ? submitted : ret;
8303 #ifdef CONFIG_PROC_FS
8304 static int io_uring_show_cred(int id, void *p, void *data)
8306 const struct cred *cred = p;
8307 struct seq_file *m = data;
8308 struct user_namespace *uns = seq_user_ns(m);
8309 struct group_info *gi;
8314 seq_printf(m, "%5d\n", id);
8315 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8316 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8317 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8318 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8319 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8320 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8321 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8322 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8323 seq_puts(m, "\n\tGroups:\t");
8324 gi = cred->group_info;
8325 for (g = 0; g < gi->ngroups; g++) {
8326 seq_put_decimal_ull(m, g ? " " : "",
8327 from_kgid_munged(uns, gi->gid[g]));
8329 seq_puts(m, "\n\tCapEff:\t");
8330 cap = cred->cap_effective;
8331 CAP_FOR_EACH_U32(__capi)
8332 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8337 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8341 mutex_lock(&ctx->uring_lock);
8342 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8343 for (i = 0; i < ctx->nr_user_files; i++) {
8344 struct fixed_file_table *table;
8347 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8348 f = table->files[i & IORING_FILE_TABLE_MASK];
8350 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8352 seq_printf(m, "%5u: <none>\n", i);
8354 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8355 for (i = 0; i < ctx->nr_user_bufs; i++) {
8356 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8358 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8359 (unsigned int) buf->len);
8361 if (!idr_is_empty(&ctx->personality_idr)) {
8362 seq_printf(m, "Personalities:\n");
8363 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8365 seq_printf(m, "PollList:\n");
8366 spin_lock_irq(&ctx->completion_lock);
8367 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8368 struct hlist_head *list = &ctx->cancel_hash[i];
8369 struct io_kiocb *req;
8371 hlist_for_each_entry(req, list, hash_node)
8372 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8373 req->task->task_works != NULL);
8375 spin_unlock_irq(&ctx->completion_lock);
8376 mutex_unlock(&ctx->uring_lock);
8379 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8381 struct io_ring_ctx *ctx = f->private_data;
8383 if (percpu_ref_tryget(&ctx->refs)) {
8384 __io_uring_show_fdinfo(ctx, m);
8385 percpu_ref_put(&ctx->refs);
8390 static const struct file_operations io_uring_fops = {
8391 .release = io_uring_release,
8392 .flush = io_uring_flush,
8393 .mmap = io_uring_mmap,
8395 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8396 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8398 .poll = io_uring_poll,
8399 .fasync = io_uring_fasync,
8400 #ifdef CONFIG_PROC_FS
8401 .show_fdinfo = io_uring_show_fdinfo,
8405 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8406 struct io_uring_params *p)
8408 struct io_rings *rings;
8409 size_t size, sq_array_offset;
8411 /* make sure these are sane, as we already accounted them */
8412 ctx->sq_entries = p->sq_entries;
8413 ctx->cq_entries = p->cq_entries;
8415 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8416 if (size == SIZE_MAX)
8419 rings = io_mem_alloc(size);
8424 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8425 rings->sq_ring_mask = p->sq_entries - 1;
8426 rings->cq_ring_mask = p->cq_entries - 1;
8427 rings->sq_ring_entries = p->sq_entries;
8428 rings->cq_ring_entries = p->cq_entries;
8429 ctx->sq_mask = rings->sq_ring_mask;
8430 ctx->cq_mask = rings->cq_ring_mask;
8432 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8433 if (size == SIZE_MAX) {
8434 io_mem_free(ctx->rings);
8439 ctx->sq_sqes = io_mem_alloc(size);
8440 if (!ctx->sq_sqes) {
8441 io_mem_free(ctx->rings);
8450 * Allocate an anonymous fd, this is what constitutes the application
8451 * visible backing of an io_uring instance. The application mmaps this
8452 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8453 * we have to tie this fd to a socket for file garbage collection purposes.
8455 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8460 #if defined(CONFIG_UNIX)
8461 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8467 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8471 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8472 O_RDWR | O_CLOEXEC);
8475 ret = PTR_ERR(file);
8479 #if defined(CONFIG_UNIX)
8480 ctx->ring_sock->file = file;
8482 fd_install(ret, file);
8485 #if defined(CONFIG_UNIX)
8486 sock_release(ctx->ring_sock);
8487 ctx->ring_sock = NULL;
8492 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8493 struct io_uring_params __user *params)
8495 struct user_struct *user = NULL;
8496 struct io_ring_ctx *ctx;
8502 if (entries > IORING_MAX_ENTRIES) {
8503 if (!(p->flags & IORING_SETUP_CLAMP))
8505 entries = IORING_MAX_ENTRIES;
8509 * Use twice as many entries for the CQ ring. It's possible for the
8510 * application to drive a higher depth than the size of the SQ ring,
8511 * since the sqes are only used at submission time. This allows for
8512 * some flexibility in overcommitting a bit. If the application has
8513 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8514 * of CQ ring entries manually.
8516 p->sq_entries = roundup_pow_of_two(entries);
8517 if (p->flags & IORING_SETUP_CQSIZE) {
8519 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8520 * to a power-of-two, if it isn't already. We do NOT impose
8521 * any cq vs sq ring sizing.
8523 if (p->cq_entries < p->sq_entries)
8525 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8526 if (!(p->flags & IORING_SETUP_CLAMP))
8528 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8530 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8532 p->cq_entries = 2 * p->sq_entries;
8535 user = get_uid(current_user());
8536 limit_mem = !capable(CAP_IPC_LOCK);
8539 ret = __io_account_mem(user,
8540 ring_pages(p->sq_entries, p->cq_entries));
8547 ctx = io_ring_ctx_alloc(p);
8550 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8555 ctx->compat = in_compat_syscall();
8557 ctx->creds = get_current_cred();
8559 mmgrab(current->mm);
8560 ctx->sqo_mm = current->mm;
8563 * Account memory _before_ installing the file descriptor. Once
8564 * the descriptor is installed, it can get closed at any time. Also
8565 * do this before hitting the general error path, as ring freeing
8566 * will un-account as well.
8568 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8570 ctx->limit_mem = limit_mem;
8572 ret = io_allocate_scq_urings(ctx, p);
8576 ret = io_sq_offload_start(ctx, p);
8580 memset(&p->sq_off, 0, sizeof(p->sq_off));
8581 p->sq_off.head = offsetof(struct io_rings, sq.head);
8582 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8583 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8584 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8585 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8586 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8587 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8589 memset(&p->cq_off, 0, sizeof(p->cq_off));
8590 p->cq_off.head = offsetof(struct io_rings, cq.head);
8591 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8592 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8593 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8594 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8595 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8596 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8598 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8599 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8600 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8601 IORING_FEAT_POLL_32BITS;
8603 if (copy_to_user(params, p, sizeof(*p))) {
8609 * Install ring fd as the very last thing, so we don't risk someone
8610 * having closed it before we finish setup
8612 ret = io_uring_get_fd(ctx);
8616 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8619 io_ring_ctx_wait_and_kill(ctx);
8624 * Sets up an aio uring context, and returns the fd. Applications asks for a
8625 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8626 * params structure passed in.
8628 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8630 struct io_uring_params p;
8633 if (copy_from_user(&p, params, sizeof(p)))
8635 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8640 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8641 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8642 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8645 return io_uring_create(entries, &p, params);
8648 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8649 struct io_uring_params __user *, params)
8651 return io_uring_setup(entries, params);
8654 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8656 struct io_uring_probe *p;
8660 size = struct_size(p, ops, nr_args);
8661 if (size == SIZE_MAX)
8663 p = kzalloc(size, GFP_KERNEL);
8668 if (copy_from_user(p, arg, size))
8671 if (memchr_inv(p, 0, size))
8674 p->last_op = IORING_OP_LAST - 1;
8675 if (nr_args > IORING_OP_LAST)
8676 nr_args = IORING_OP_LAST;
8678 for (i = 0; i < nr_args; i++) {
8680 if (!io_op_defs[i].not_supported)
8681 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8686 if (copy_to_user(arg, p, size))
8693 static int io_register_personality(struct io_ring_ctx *ctx)
8695 const struct cred *creds = get_current_cred();
8698 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8699 USHRT_MAX, GFP_KERNEL);
8705 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8707 const struct cred *old_creds;
8709 old_creds = idr_remove(&ctx->personality_idr, id);
8711 put_cred(old_creds);
8718 static bool io_register_op_must_quiesce(int op)
8721 case IORING_UNREGISTER_FILES:
8722 case IORING_REGISTER_FILES_UPDATE:
8723 case IORING_REGISTER_PROBE:
8724 case IORING_REGISTER_PERSONALITY:
8725 case IORING_UNREGISTER_PERSONALITY:
8732 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8733 void __user *arg, unsigned nr_args)
8734 __releases(ctx->uring_lock)
8735 __acquires(ctx->uring_lock)
8740 * We're inside the ring mutex, if the ref is already dying, then
8741 * someone else killed the ctx or is already going through
8742 * io_uring_register().
8744 if (percpu_ref_is_dying(&ctx->refs))
8747 if (io_register_op_must_quiesce(opcode)) {
8748 percpu_ref_kill(&ctx->refs);
8751 * Drop uring mutex before waiting for references to exit. If
8752 * another thread is currently inside io_uring_enter() it might
8753 * need to grab the uring_lock to make progress. If we hold it
8754 * here across the drain wait, then we can deadlock. It's safe
8755 * to drop the mutex here, since no new references will come in
8756 * after we've killed the percpu ref.
8758 mutex_unlock(&ctx->uring_lock);
8759 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8760 mutex_lock(&ctx->uring_lock);
8762 percpu_ref_resurrect(&ctx->refs);
8769 case IORING_REGISTER_BUFFERS:
8770 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8772 case IORING_UNREGISTER_BUFFERS:
8776 ret = io_sqe_buffer_unregister(ctx);
8778 case IORING_REGISTER_FILES:
8779 ret = io_sqe_files_register(ctx, arg, nr_args);
8781 case IORING_UNREGISTER_FILES:
8785 ret = io_sqe_files_unregister(ctx);
8787 case IORING_REGISTER_FILES_UPDATE:
8788 ret = io_sqe_files_update(ctx, arg, nr_args);
8790 case IORING_REGISTER_EVENTFD:
8791 case IORING_REGISTER_EVENTFD_ASYNC:
8795 ret = io_eventfd_register(ctx, arg);
8798 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8799 ctx->eventfd_async = 1;
8801 ctx->eventfd_async = 0;
8803 case IORING_UNREGISTER_EVENTFD:
8807 ret = io_eventfd_unregister(ctx);
8809 case IORING_REGISTER_PROBE:
8811 if (!arg || nr_args > 256)
8813 ret = io_probe(ctx, arg, nr_args);
8815 case IORING_REGISTER_PERSONALITY:
8819 ret = io_register_personality(ctx);
8821 case IORING_UNREGISTER_PERSONALITY:
8825 ret = io_unregister_personality(ctx, nr_args);
8832 if (io_register_op_must_quiesce(opcode)) {
8833 /* bring the ctx back to life */
8834 percpu_ref_reinit(&ctx->refs);
8836 reinit_completion(&ctx->ref_comp);
8841 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8842 void __user *, arg, unsigned int, nr_args)
8844 struct io_ring_ctx *ctx;
8853 if (f.file->f_op != &io_uring_fops)
8856 ctx = f.file->private_data;
8858 mutex_lock(&ctx->uring_lock);
8859 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8860 mutex_unlock(&ctx->uring_lock);
8861 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8862 ctx->cq_ev_fd != NULL, ret);
8868 static int __init io_uring_init(void)
8870 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8871 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8872 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8875 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8876 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8877 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8878 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8879 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8880 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8881 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8882 BUILD_BUG_SQE_ELEM(8, __u64, off);
8883 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8884 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8885 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8886 BUILD_BUG_SQE_ELEM(24, __u32, len);
8887 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8888 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8889 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8890 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8891 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8892 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8893 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8894 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8895 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8896 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8897 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8898 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8899 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8900 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8901 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8902 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8903 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8904 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8905 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8907 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8908 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8909 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8912 __initcall(io_uring_init);