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>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
107 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
112 u32 head ____cacheline_aligned_in_smp;
113 u32 tail ____cacheline_aligned_in_smp;
117 * This data is shared with the application through the mmap at offsets
118 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
120 * The offsets to the member fields are published through struct
121 * io_sqring_offsets when calling io_uring_setup.
125 * Head and tail offsets into the ring; the offsets need to be
126 * masked to get valid indices.
128 * The kernel controls head of the sq ring and the tail of the cq ring,
129 * and the application controls tail of the sq ring and the head of the
132 struct io_uring sq, cq;
134 * Bitmasks to apply to head and tail offsets (constant, equals
137 u32 sq_ring_mask, cq_ring_mask;
138 /* Ring sizes (constant, power of 2) */
139 u32 sq_ring_entries, cq_ring_entries;
141 * Number of invalid entries dropped by the kernel due to
142 * invalid index stored in array
144 * Written by the kernel, shouldn't be modified by the
145 * application (i.e. get number of "new events" by comparing to
148 * After a new SQ head value was read by the application this
149 * counter includes all submissions that were dropped reaching
150 * the new SQ head (and possibly more).
156 * Written by the kernel, shouldn't be modified by the
159 * The application needs a full memory barrier before checking
160 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
166 * Written by the application, shouldn't be modified by the
171 * Number of completion events lost because the queue was full;
172 * this should be avoided by the application by making sure
173 * there are not more requests pending than there is space in
174 * the completion queue.
176 * Written by the kernel, shouldn't be modified by the
177 * application (i.e. get number of "new events" by comparing to
180 * As completion events come in out of order this counter is not
181 * ordered with any other data.
185 * Ring buffer of completion events.
187 * The kernel writes completion events fresh every time they are
188 * produced, so the application is allowed to modify pending
191 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
194 enum io_uring_cmd_flags {
195 IO_URING_F_NONBLOCK = 1,
196 IO_URING_F_COMPLETE_DEFER = 2,
199 struct io_mapped_ubuf {
202 struct bio_vec *bvec;
203 unsigned int nr_bvecs;
204 unsigned long acct_pages;
210 struct list_head list;
217 struct fixed_rsrc_table {
221 struct fixed_rsrc_ref_node {
222 struct percpu_ref refs;
223 struct list_head node;
224 struct list_head rsrc_list;
225 struct fixed_rsrc_data *rsrc_data;
226 void (*rsrc_put)(struct io_ring_ctx *ctx,
227 struct io_rsrc_put *prsrc);
228 struct llist_node llist;
232 struct fixed_rsrc_data {
233 struct fixed_rsrc_table *table;
234 struct io_ring_ctx *ctx;
236 struct fixed_rsrc_ref_node *node;
237 struct percpu_ref refs;
238 struct completion done;
243 struct list_head list;
249 struct io_restriction {
250 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
251 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
252 u8 sqe_flags_allowed;
253 u8 sqe_flags_required;
261 /* ctx's that are using this sqd */
262 struct list_head ctx_list;
263 struct list_head ctx_new_list;
264 struct mutex ctx_lock;
266 struct task_struct *thread;
267 struct wait_queue_head wait;
269 unsigned sq_thread_idle;
272 #define IO_IOPOLL_BATCH 8
273 #define IO_COMPL_BATCH 32
274 #define IO_REQ_CACHE_SIZE 32
275 #define IO_REQ_ALLOC_BATCH 8
277 struct io_comp_state {
278 struct io_kiocb *reqs[IO_COMPL_BATCH];
280 unsigned int locked_free_nr;
281 /* inline/task_work completion list, under ->uring_lock */
282 struct list_head free_list;
283 /* IRQ completion list, under ->completion_lock */
284 struct list_head locked_free_list;
287 struct io_submit_link {
288 struct io_kiocb *head;
289 struct io_kiocb *last;
292 struct io_submit_state {
293 struct blk_plug plug;
294 struct io_submit_link link;
297 * io_kiocb alloc cache
299 void *reqs[IO_REQ_CACHE_SIZE];
300 unsigned int free_reqs;
305 * Batch completion logic
307 struct io_comp_state comp;
310 * File reference cache
314 unsigned int file_refs;
315 unsigned int ios_left;
320 struct percpu_ref refs;
321 } ____cacheline_aligned_in_smp;
325 unsigned int compat: 1;
326 unsigned int limit_mem: 1;
327 unsigned int cq_overflow_flushed: 1;
328 unsigned int drain_next: 1;
329 unsigned int eventfd_async: 1;
330 unsigned int restricted: 1;
331 unsigned int sqo_dead: 1;
334 * Ring buffer of indices into array of io_uring_sqe, which is
335 * mmapped by the application using the IORING_OFF_SQES offset.
337 * This indirection could e.g. be used to assign fixed
338 * io_uring_sqe entries to operations and only submit them to
339 * the queue when needed.
341 * The kernel modifies neither the indices array nor the entries
345 unsigned cached_sq_head;
348 unsigned sq_thread_idle;
349 unsigned cached_sq_dropped;
350 unsigned cached_cq_overflow;
351 unsigned long sq_check_overflow;
353 struct list_head defer_list;
354 struct list_head timeout_list;
355 struct list_head cq_overflow_list;
357 struct io_uring_sqe *sq_sqes;
358 } ____cacheline_aligned_in_smp;
361 struct mutex uring_lock;
362 wait_queue_head_t wait;
363 } ____cacheline_aligned_in_smp;
365 struct io_submit_state submit_state;
367 struct io_rings *rings;
373 * For SQPOLL usage - we hold a reference to the parent task, so we
374 * have access to the ->files
376 struct task_struct *sqo_task;
378 /* Only used for accounting purposes */
379 struct mm_struct *mm_account;
381 #ifdef CONFIG_BLK_CGROUP
382 struct cgroup_subsys_state *sqo_blkcg_css;
385 struct io_sq_data *sq_data; /* if using sq thread polling */
387 struct wait_queue_head sqo_sq_wait;
388 struct list_head sqd_list;
391 * If used, fixed file set. Writers must ensure that ->refs is dead,
392 * readers must ensure that ->refs is alive as long as the file* is
393 * used. Only updated through io_uring_register(2).
395 struct fixed_rsrc_data *file_data;
396 unsigned nr_user_files;
398 /* if used, fixed mapped user buffers */
399 unsigned nr_user_bufs;
400 struct io_mapped_ubuf *user_bufs;
402 struct user_struct *user;
404 const struct cred *creds;
408 unsigned int sessionid;
411 struct completion ref_comp;
412 struct completion sq_thread_comp;
414 #if defined(CONFIG_UNIX)
415 struct socket *ring_sock;
418 struct idr io_buffer_idr;
420 struct idr personality_idr;
423 unsigned cached_cq_tail;
426 atomic_t cq_timeouts;
427 unsigned cq_last_tm_flush;
428 unsigned long cq_check_overflow;
429 struct wait_queue_head cq_wait;
430 struct fasync_struct *cq_fasync;
431 struct eventfd_ctx *cq_ev_fd;
432 } ____cacheline_aligned_in_smp;
435 spinlock_t completion_lock;
438 * ->iopoll_list is protected by the ctx->uring_lock for
439 * io_uring instances that don't use IORING_SETUP_SQPOLL.
440 * For SQPOLL, only the single threaded io_sq_thread() will
441 * manipulate the list, hence no extra locking is needed there.
443 struct list_head iopoll_list;
444 struct hlist_head *cancel_hash;
445 unsigned cancel_hash_bits;
446 bool poll_multi_file;
448 spinlock_t inflight_lock;
449 struct list_head inflight_list;
450 } ____cacheline_aligned_in_smp;
452 struct delayed_work rsrc_put_work;
453 struct llist_head rsrc_put_llist;
454 struct list_head rsrc_ref_list;
455 spinlock_t rsrc_ref_lock;
457 struct io_restriction restrictions;
459 /* Keep this last, we don't need it for the fast path */
460 struct work_struct exit_work;
464 * First field must be the file pointer in all the
465 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
467 struct io_poll_iocb {
469 struct wait_queue_head *head;
473 struct wait_queue_entry wait;
476 struct io_poll_remove {
486 struct io_timeout_data {
487 struct io_kiocb *req;
488 struct hrtimer timer;
489 struct timespec64 ts;
490 enum hrtimer_mode mode;
495 struct sockaddr __user *addr;
496 int __user *addr_len;
498 unsigned long nofile;
518 struct list_head list;
519 /* head of the link, used by linked timeouts only */
520 struct io_kiocb *head;
523 struct io_timeout_rem {
528 struct timespec64 ts;
533 /* NOTE: kiocb has the file as the first member, so don't do it here */
541 struct sockaddr __user *addr;
548 struct user_msghdr __user *umsg;
554 struct io_buffer *kbuf;
560 struct filename *filename;
562 unsigned long nofile;
565 struct io_rsrc_update {
591 struct epoll_event event;
595 struct file *file_out;
596 struct file *file_in;
603 struct io_provide_buf {
617 const char __user *filename;
618 struct statx __user *buffer;
630 struct filename *oldpath;
631 struct filename *newpath;
639 struct filename *filename;
642 struct io_completion {
644 struct list_head list;
648 struct io_async_connect {
649 struct sockaddr_storage address;
652 struct io_async_msghdr {
653 struct iovec fast_iov[UIO_FASTIOV];
654 /* points to an allocated iov, if NULL we use fast_iov instead */
655 struct iovec *free_iov;
656 struct sockaddr __user *uaddr;
658 struct sockaddr_storage addr;
662 struct iovec fast_iov[UIO_FASTIOV];
663 const struct iovec *free_iovec;
664 struct iov_iter iter;
666 struct wait_page_queue wpq;
670 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
671 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
672 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
673 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
674 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
675 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
681 REQ_F_LINK_TIMEOUT_BIT,
683 REQ_F_NEED_CLEANUP_BIT,
685 REQ_F_BUFFER_SELECTED_BIT,
686 REQ_F_NO_FILE_TABLE_BIT,
687 REQ_F_WORK_INITIALIZED_BIT,
688 REQ_F_LTIMEOUT_ACTIVE_BIT,
689 REQ_F_COMPLETE_INLINE_BIT,
691 /* not a real bit, just to check we're not overflowing the space */
697 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
698 /* drain existing IO first */
699 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
701 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
702 /* doesn't sever on completion < 0 */
703 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
705 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
706 /* IOSQE_BUFFER_SELECT */
707 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
709 /* fail rest of links */
710 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
711 /* on inflight list */
712 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
713 /* read/write uses file position */
714 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
715 /* must not punt to workers */
716 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
717 /* has or had linked timeout */
718 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
720 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
722 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
723 /* already went through poll handler */
724 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
725 /* buffer already selected */
726 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
727 /* doesn't need file table for this request */
728 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
729 /* io_wq_work is initialized */
730 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
731 /* linked timeout is active, i.e. prepared by link's head */
732 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
733 /* completion is deferred through io_comp_state */
734 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
738 struct io_poll_iocb poll;
739 struct io_poll_iocb *double_poll;
742 struct io_task_work {
743 struct io_wq_work_node node;
744 task_work_func_t func;
748 * NOTE! Each of the iocb union members has the file pointer
749 * as the first entry in their struct definition. So you can
750 * access the file pointer through any of the sub-structs,
751 * or directly as just 'ki_filp' in this struct.
757 struct io_poll_iocb poll;
758 struct io_poll_remove poll_remove;
759 struct io_accept accept;
761 struct io_cancel cancel;
762 struct io_timeout timeout;
763 struct io_timeout_rem timeout_rem;
764 struct io_connect connect;
765 struct io_sr_msg sr_msg;
767 struct io_close close;
768 struct io_rsrc_update rsrc_update;
769 struct io_fadvise fadvise;
770 struct io_madvise madvise;
771 struct io_epoll epoll;
772 struct io_splice splice;
773 struct io_provide_buf pbuf;
774 struct io_statx statx;
775 struct io_shutdown shutdown;
776 struct io_rename rename;
777 struct io_unlink unlink;
778 /* use only after cleaning per-op data, see io_clean_op() */
779 struct io_completion compl;
782 /* opcode allocated if it needs to store data for async defer */
785 /* polled IO has completed */
791 struct io_ring_ctx *ctx;
794 struct task_struct *task;
797 struct io_kiocb *link;
798 struct percpu_ref *fixed_rsrc_refs;
801 * 1. used with ctx->iopoll_list with reads/writes
802 * 2. to track reqs with ->files (see io_op_def::file_table)
804 struct list_head inflight_entry;
806 struct io_task_work io_task_work;
807 struct callback_head task_work;
809 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
810 struct hlist_node hash_node;
811 struct async_poll *apoll;
812 struct io_wq_work work;
815 struct io_defer_entry {
816 struct list_head list;
817 struct io_kiocb *req;
822 /* needs req->file assigned */
823 unsigned needs_file : 1;
824 /* hash wq insertion if file is a regular file */
825 unsigned hash_reg_file : 1;
826 /* unbound wq insertion if file is a non-regular file */
827 unsigned unbound_nonreg_file : 1;
828 /* opcode is not supported by this kernel */
829 unsigned not_supported : 1;
830 /* set if opcode supports polled "wait" */
832 unsigned pollout : 1;
833 /* op supports buffer selection */
834 unsigned buffer_select : 1;
835 /* must always have async data allocated */
836 unsigned needs_async_data : 1;
837 /* should block plug */
839 /* size of async data needed, if any */
840 unsigned short async_size;
844 static const struct io_op_def io_op_defs[] = {
845 [IORING_OP_NOP] = {},
846 [IORING_OP_READV] = {
848 .unbound_nonreg_file = 1,
851 .needs_async_data = 1,
853 .async_size = sizeof(struct io_async_rw),
854 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
856 [IORING_OP_WRITEV] = {
859 .unbound_nonreg_file = 1,
861 .needs_async_data = 1,
863 .async_size = sizeof(struct io_async_rw),
864 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
867 [IORING_OP_FSYNC] = {
869 .work_flags = IO_WQ_WORK_BLKCG,
871 [IORING_OP_READ_FIXED] = {
873 .unbound_nonreg_file = 1,
876 .async_size = sizeof(struct io_async_rw),
877 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
879 [IORING_OP_WRITE_FIXED] = {
882 .unbound_nonreg_file = 1,
885 .async_size = sizeof(struct io_async_rw),
886 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
889 [IORING_OP_POLL_ADD] = {
891 .unbound_nonreg_file = 1,
893 [IORING_OP_POLL_REMOVE] = {},
894 [IORING_OP_SYNC_FILE_RANGE] = {
896 .work_flags = IO_WQ_WORK_BLKCG,
898 [IORING_OP_SENDMSG] = {
900 .unbound_nonreg_file = 1,
902 .needs_async_data = 1,
903 .async_size = sizeof(struct io_async_msghdr),
904 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
907 [IORING_OP_RECVMSG] = {
909 .unbound_nonreg_file = 1,
912 .needs_async_data = 1,
913 .async_size = sizeof(struct io_async_msghdr),
914 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
917 [IORING_OP_TIMEOUT] = {
918 .needs_async_data = 1,
919 .async_size = sizeof(struct io_timeout_data),
920 .work_flags = IO_WQ_WORK_MM,
922 [IORING_OP_TIMEOUT_REMOVE] = {
923 /* used by timeout updates' prep() */
924 .work_flags = IO_WQ_WORK_MM,
926 [IORING_OP_ACCEPT] = {
928 .unbound_nonreg_file = 1,
930 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
932 [IORING_OP_ASYNC_CANCEL] = {},
933 [IORING_OP_LINK_TIMEOUT] = {
934 .needs_async_data = 1,
935 .async_size = sizeof(struct io_timeout_data),
936 .work_flags = IO_WQ_WORK_MM,
938 [IORING_OP_CONNECT] = {
940 .unbound_nonreg_file = 1,
942 .needs_async_data = 1,
943 .async_size = sizeof(struct io_async_connect),
944 .work_flags = IO_WQ_WORK_MM,
946 [IORING_OP_FALLOCATE] = {
948 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
950 [IORING_OP_OPENAT] = {
951 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
952 IO_WQ_WORK_FS | IO_WQ_WORK_MM,
954 [IORING_OP_CLOSE] = {
955 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
957 [IORING_OP_FILES_UPDATE] = {
958 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
960 [IORING_OP_STATX] = {
961 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
962 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
966 .unbound_nonreg_file = 1,
970 .async_size = sizeof(struct io_async_rw),
971 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
973 [IORING_OP_WRITE] = {
975 .unbound_nonreg_file = 1,
978 .async_size = sizeof(struct io_async_rw),
979 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
982 [IORING_OP_FADVISE] = {
984 .work_flags = IO_WQ_WORK_BLKCG,
986 [IORING_OP_MADVISE] = {
987 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
991 .unbound_nonreg_file = 1,
993 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
997 .unbound_nonreg_file = 1,
1000 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
1002 [IORING_OP_OPENAT2] = {
1003 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
1004 IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
1006 [IORING_OP_EPOLL_CTL] = {
1007 .unbound_nonreg_file = 1,
1008 .work_flags = IO_WQ_WORK_FILES,
1010 [IORING_OP_SPLICE] = {
1013 .unbound_nonreg_file = 1,
1014 .work_flags = IO_WQ_WORK_BLKCG,
1016 [IORING_OP_PROVIDE_BUFFERS] = {},
1017 [IORING_OP_REMOVE_BUFFERS] = {},
1021 .unbound_nonreg_file = 1,
1023 [IORING_OP_SHUTDOWN] = {
1026 [IORING_OP_RENAMEAT] = {
1027 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
1028 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
1030 [IORING_OP_UNLINKAT] = {
1031 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
1032 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
1036 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1037 struct task_struct *task,
1038 struct files_struct *files);
1039 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1040 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1041 struct io_ring_ctx *ctx);
1042 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1044 static bool io_rw_reissue(struct io_kiocb *req);
1045 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1046 static void io_put_req(struct io_kiocb *req);
1047 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1048 static void io_double_put_req(struct io_kiocb *req);
1049 static void io_dismantle_req(struct io_kiocb *req);
1050 static void io_put_task(struct task_struct *task, int nr);
1051 static void io_queue_next(struct io_kiocb *req);
1052 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1053 static void __io_queue_linked_timeout(struct io_kiocb *req);
1054 static void io_queue_linked_timeout(struct io_kiocb *req);
1055 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1056 struct io_uring_rsrc_update *ip,
1058 static void __io_clean_op(struct io_kiocb *req);
1059 static struct file *io_file_get(struct io_submit_state *state,
1060 struct io_kiocb *req, int fd, bool fixed);
1061 static void __io_queue_sqe(struct io_kiocb *req);
1062 static void io_rsrc_put_work(struct work_struct *work);
1064 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1065 struct iov_iter *iter, bool needs_lock);
1066 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1067 const struct iovec *fast_iov,
1068 struct iov_iter *iter, bool force);
1069 static void io_req_task_queue(struct io_kiocb *req);
1070 static void io_submit_flush_completions(struct io_comp_state *cs,
1071 struct io_ring_ctx *ctx);
1073 static struct kmem_cache *req_cachep;
1075 static const struct file_operations io_uring_fops;
1077 struct sock *io_uring_get_socket(struct file *file)
1079 #if defined(CONFIG_UNIX)
1080 if (file->f_op == &io_uring_fops) {
1081 struct io_ring_ctx *ctx = file->private_data;
1083 return ctx->ring_sock->sk;
1088 EXPORT_SYMBOL(io_uring_get_socket);
1090 #define io_for_each_link(pos, head) \
1091 for (pos = (head); pos; pos = pos->link)
1093 static inline void io_clean_op(struct io_kiocb *req)
1095 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1099 static inline void io_set_resource_node(struct io_kiocb *req)
1101 struct io_ring_ctx *ctx = req->ctx;
1103 if (!req->fixed_rsrc_refs) {
1104 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1105 percpu_ref_get(req->fixed_rsrc_refs);
1109 static bool io_match_task(struct io_kiocb *head,
1110 struct task_struct *task,
1111 struct files_struct *files)
1113 struct io_kiocb *req;
1115 if (task && head->task != task) {
1116 /* in terms of cancelation, always match if req task is dead */
1117 if (head->task->flags & PF_EXITING)
1124 io_for_each_link(req, head) {
1125 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1127 if (req->file && req->file->f_op == &io_uring_fops)
1129 if ((req->work.flags & IO_WQ_WORK_FILES) &&
1130 req->work.identity->files == files)
1136 static void io_sq_thread_drop_mm_files(void)
1138 struct files_struct *files = current->files;
1139 struct mm_struct *mm = current->mm;
1142 kthread_unuse_mm(mm);
1147 struct nsproxy *nsproxy = current->nsproxy;
1150 current->files = NULL;
1151 current->nsproxy = NULL;
1152 task_unlock(current);
1153 put_files_struct(files);
1154 put_nsproxy(nsproxy);
1158 static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1160 if (!current->files) {
1161 struct files_struct *files;
1162 struct nsproxy *nsproxy;
1164 task_lock(ctx->sqo_task);
1165 files = ctx->sqo_task->files;
1167 task_unlock(ctx->sqo_task);
1170 atomic_inc(&files->count);
1171 get_nsproxy(ctx->sqo_task->nsproxy);
1172 nsproxy = ctx->sqo_task->nsproxy;
1173 task_unlock(ctx->sqo_task);
1176 current->files = files;
1177 current->nsproxy = nsproxy;
1178 task_unlock(current);
1183 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1185 struct mm_struct *mm;
1190 task_lock(ctx->sqo_task);
1191 mm = ctx->sqo_task->mm;
1192 if (unlikely(!mm || !mmget_not_zero(mm)))
1194 task_unlock(ctx->sqo_task);
1204 static int __io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1205 struct io_kiocb *req)
1207 const struct io_op_def *def = &io_op_defs[req->opcode];
1210 if (def->work_flags & IO_WQ_WORK_MM) {
1211 ret = __io_sq_thread_acquire_mm(ctx);
1216 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) {
1217 ret = __io_sq_thread_acquire_files(ctx);
1225 static inline int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1226 struct io_kiocb *req)
1228 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1230 return __io_sq_thread_acquire_mm_files(ctx, req);
1233 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1234 struct cgroup_subsys_state **cur_css)
1237 #ifdef CONFIG_BLK_CGROUP
1238 /* puts the old one when swapping */
1239 if (*cur_css != ctx->sqo_blkcg_css) {
1240 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1241 *cur_css = ctx->sqo_blkcg_css;
1246 static void io_sq_thread_unassociate_blkcg(void)
1248 #ifdef CONFIG_BLK_CGROUP
1249 kthread_associate_blkcg(NULL);
1253 static inline void req_set_fail_links(struct io_kiocb *req)
1255 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1256 req->flags |= REQ_F_FAIL_LINK;
1260 * None of these are dereferenced, they are simply used to check if any of
1261 * them have changed. If we're under current and check they are still the
1262 * same, we're fine to grab references to them for actual out-of-line use.
1264 static void io_init_identity(struct io_identity *id)
1266 id->files = current->files;
1267 id->mm = current->mm;
1268 #ifdef CONFIG_BLK_CGROUP
1270 id->blkcg_css = blkcg_css();
1273 id->creds = current_cred();
1274 id->nsproxy = current->nsproxy;
1275 id->fs = current->fs;
1276 id->fsize = rlimit(RLIMIT_FSIZE);
1278 id->loginuid = current->loginuid;
1279 id->sessionid = current->sessionid;
1281 refcount_set(&id->count, 1);
1284 static inline void __io_req_init_async(struct io_kiocb *req)
1286 memset(&req->work, 0, sizeof(req->work));
1287 req->flags |= REQ_F_WORK_INITIALIZED;
1291 * Note: must call io_req_init_async() for the first time you
1292 * touch any members of io_wq_work.
1294 static inline void io_req_init_async(struct io_kiocb *req)
1296 struct io_uring_task *tctx = current->io_uring;
1298 if (req->flags & REQ_F_WORK_INITIALIZED)
1301 __io_req_init_async(req);
1303 /* Grab a ref if this isn't our static identity */
1304 req->work.identity = tctx->identity;
1305 if (tctx->identity != &tctx->__identity)
1306 refcount_inc(&req->work.identity->count);
1309 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1311 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1313 complete(&ctx->ref_comp);
1316 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1318 return !req->timeout.off;
1321 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1323 struct io_ring_ctx *ctx;
1326 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1331 * Use 5 bits less than the max cq entries, that should give us around
1332 * 32 entries per hash list if totally full and uniformly spread.
1334 hash_bits = ilog2(p->cq_entries);
1338 ctx->cancel_hash_bits = hash_bits;
1339 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1341 if (!ctx->cancel_hash)
1343 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1345 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1346 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1349 ctx->flags = p->flags;
1350 init_waitqueue_head(&ctx->sqo_sq_wait);
1351 INIT_LIST_HEAD(&ctx->sqd_list);
1352 init_waitqueue_head(&ctx->cq_wait);
1353 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1354 init_completion(&ctx->ref_comp);
1355 init_completion(&ctx->sq_thread_comp);
1356 idr_init(&ctx->io_buffer_idr);
1357 idr_init(&ctx->personality_idr);
1358 mutex_init(&ctx->uring_lock);
1359 init_waitqueue_head(&ctx->wait);
1360 spin_lock_init(&ctx->completion_lock);
1361 INIT_LIST_HEAD(&ctx->iopoll_list);
1362 INIT_LIST_HEAD(&ctx->defer_list);
1363 INIT_LIST_HEAD(&ctx->timeout_list);
1364 spin_lock_init(&ctx->inflight_lock);
1365 INIT_LIST_HEAD(&ctx->inflight_list);
1366 spin_lock_init(&ctx->rsrc_ref_lock);
1367 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1368 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1369 init_llist_head(&ctx->rsrc_put_llist);
1370 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1371 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1374 kfree(ctx->cancel_hash);
1379 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1381 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1382 struct io_ring_ctx *ctx = req->ctx;
1384 return seq != ctx->cached_cq_tail
1385 + READ_ONCE(ctx->cached_cq_overflow);
1391 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1393 if (req->work.identity == &tctx->__identity)
1395 if (refcount_dec_and_test(&req->work.identity->count))
1396 kfree(req->work.identity);
1399 static void io_req_clean_work(struct io_kiocb *req)
1401 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1404 if (req->work.flags & IO_WQ_WORK_MM)
1405 mmdrop(req->work.identity->mm);
1406 #ifdef CONFIG_BLK_CGROUP
1407 if (req->work.flags & IO_WQ_WORK_BLKCG)
1408 css_put(req->work.identity->blkcg_css);
1410 if (req->work.flags & IO_WQ_WORK_CREDS)
1411 put_cred(req->work.identity->creds);
1412 if (req->work.flags & IO_WQ_WORK_FS) {
1413 struct fs_struct *fs = req->work.identity->fs;
1415 spin_lock(&req->work.identity->fs->lock);
1418 spin_unlock(&req->work.identity->fs->lock);
1422 if (req->work.flags & IO_WQ_WORK_FILES) {
1423 put_files_struct(req->work.identity->files);
1424 put_nsproxy(req->work.identity->nsproxy);
1426 if (req->flags & REQ_F_INFLIGHT) {
1427 struct io_ring_ctx *ctx = req->ctx;
1428 struct io_uring_task *tctx = req->task->io_uring;
1429 unsigned long flags;
1431 spin_lock_irqsave(&ctx->inflight_lock, flags);
1432 list_del(&req->inflight_entry);
1433 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1434 req->flags &= ~REQ_F_INFLIGHT;
1435 if (atomic_read(&tctx->in_idle))
1436 wake_up(&tctx->wait);
1439 req->flags &= ~REQ_F_WORK_INITIALIZED;
1440 req->work.flags &= ~(IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG | IO_WQ_WORK_FS |
1441 IO_WQ_WORK_CREDS | IO_WQ_WORK_FILES);
1442 io_put_identity(req->task->io_uring, req);
1446 * Create a private copy of io_identity, since some fields don't match
1447 * the current context.
1449 static bool io_identity_cow(struct io_kiocb *req)
1451 struct io_uring_task *tctx = current->io_uring;
1452 const struct cred *creds = NULL;
1453 struct io_identity *id;
1455 if (req->work.flags & IO_WQ_WORK_CREDS)
1456 creds = req->work.identity->creds;
1458 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1459 if (unlikely(!id)) {
1460 req->work.flags |= IO_WQ_WORK_CANCEL;
1465 * We can safely just re-init the creds we copied Either the field
1466 * matches the current one, or we haven't grabbed it yet. The only
1467 * exception is ->creds, through registered personalities, so handle
1468 * that one separately.
1470 io_init_identity(id);
1474 /* add one for this request */
1475 refcount_inc(&id->count);
1477 /* drop tctx and req identity references, if needed */
1478 if (tctx->identity != &tctx->__identity &&
1479 refcount_dec_and_test(&tctx->identity->count))
1480 kfree(tctx->identity);
1481 if (req->work.identity != &tctx->__identity &&
1482 refcount_dec_and_test(&req->work.identity->count))
1483 kfree(req->work.identity);
1485 req->work.identity = id;
1486 tctx->identity = id;
1490 static void io_req_track_inflight(struct io_kiocb *req)
1492 struct io_ring_ctx *ctx = req->ctx;
1494 if (!(req->flags & REQ_F_INFLIGHT)) {
1495 io_req_init_async(req);
1496 req->flags |= REQ_F_INFLIGHT;
1498 spin_lock_irq(&ctx->inflight_lock);
1499 list_add(&req->inflight_entry, &ctx->inflight_list);
1500 spin_unlock_irq(&ctx->inflight_lock);
1504 static bool io_grab_identity(struct io_kiocb *req)
1506 const struct io_op_def *def = &io_op_defs[req->opcode];
1507 struct io_identity *id = req->work.identity;
1509 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1510 if (id->fsize != rlimit(RLIMIT_FSIZE))
1512 req->work.flags |= IO_WQ_WORK_FSIZE;
1514 #ifdef CONFIG_BLK_CGROUP
1515 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1516 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1518 if (id->blkcg_css != blkcg_css()) {
1523 * This should be rare, either the cgroup is dying or the task
1524 * is moving cgroups. Just punt to root for the handful of ios.
1526 if (css_tryget_online(id->blkcg_css))
1527 req->work.flags |= IO_WQ_WORK_BLKCG;
1531 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1532 if (id->creds != current_cred())
1534 get_cred(id->creds);
1535 req->work.flags |= IO_WQ_WORK_CREDS;
1538 if (!uid_eq(current->loginuid, id->loginuid) ||
1539 current->sessionid != id->sessionid)
1542 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1543 (def->work_flags & IO_WQ_WORK_FS)) {
1544 if (current->fs != id->fs)
1546 spin_lock(&id->fs->lock);
1547 if (!id->fs->in_exec) {
1549 req->work.flags |= IO_WQ_WORK_FS;
1551 req->work.flags |= IO_WQ_WORK_CANCEL;
1553 spin_unlock(¤t->fs->lock);
1555 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1556 (def->work_flags & IO_WQ_WORK_FILES) &&
1557 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1558 if (id->files != current->files ||
1559 id->nsproxy != current->nsproxy)
1561 atomic_inc(&id->files->count);
1562 get_nsproxy(id->nsproxy);
1563 req->work.flags |= IO_WQ_WORK_FILES;
1564 io_req_track_inflight(req);
1566 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1567 (def->work_flags & IO_WQ_WORK_MM)) {
1568 if (id->mm != current->mm)
1571 req->work.flags |= IO_WQ_WORK_MM;
1577 static void io_prep_async_work(struct io_kiocb *req)
1579 const struct io_op_def *def = &io_op_defs[req->opcode];
1580 struct io_ring_ctx *ctx = req->ctx;
1582 io_req_init_async(req);
1584 if (req->flags & REQ_F_FORCE_ASYNC)
1585 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1587 if (req->flags & REQ_F_ISREG) {
1588 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1589 io_wq_hash_work(&req->work, file_inode(req->file));
1591 if (def->unbound_nonreg_file)
1592 req->work.flags |= IO_WQ_WORK_UNBOUND;
1595 /* if we fail grabbing identity, we must COW, regrab, and retry */
1596 if (io_grab_identity(req))
1599 if (!io_identity_cow(req))
1602 /* can't fail at this point */
1603 if (!io_grab_identity(req))
1607 static void io_prep_async_link(struct io_kiocb *req)
1609 struct io_kiocb *cur;
1611 io_for_each_link(cur, req)
1612 io_prep_async_work(cur);
1615 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1617 struct io_ring_ctx *ctx = req->ctx;
1618 struct io_kiocb *link = io_prep_linked_timeout(req);
1620 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1621 &req->work, req->flags);
1622 io_wq_enqueue(ctx->io_wq, &req->work);
1626 static void io_queue_async_work(struct io_kiocb *req)
1628 struct io_kiocb *link;
1630 /* init ->work of the whole link before punting */
1631 io_prep_async_link(req);
1632 link = __io_queue_async_work(req);
1635 io_queue_linked_timeout(link);
1638 static void io_kill_timeout(struct io_kiocb *req)
1640 struct io_timeout_data *io = req->async_data;
1643 ret = hrtimer_try_to_cancel(&io->timer);
1645 atomic_set(&req->ctx->cq_timeouts,
1646 atomic_read(&req->ctx->cq_timeouts) + 1);
1647 list_del_init(&req->timeout.list);
1648 io_cqring_fill_event(req, 0);
1649 io_put_req_deferred(req, 1);
1654 * Returns true if we found and killed one or more timeouts
1656 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1657 struct files_struct *files)
1659 struct io_kiocb *req, *tmp;
1662 spin_lock_irq(&ctx->completion_lock);
1663 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1664 if (io_match_task(req, tsk, files)) {
1665 io_kill_timeout(req);
1669 spin_unlock_irq(&ctx->completion_lock);
1670 return canceled != 0;
1673 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1676 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1677 struct io_defer_entry, list);
1679 if (req_need_defer(de->req, de->seq))
1681 list_del_init(&de->list);
1682 io_req_task_queue(de->req);
1684 } while (!list_empty(&ctx->defer_list));
1687 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1691 if (list_empty(&ctx->timeout_list))
1694 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1697 u32 events_needed, events_got;
1698 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1699 struct io_kiocb, timeout.list);
1701 if (io_is_timeout_noseq(req))
1705 * Since seq can easily wrap around over time, subtract
1706 * the last seq at which timeouts were flushed before comparing.
1707 * Assuming not more than 2^31-1 events have happened since,
1708 * these subtractions won't have wrapped, so we can check if
1709 * target is in [last_seq, current_seq] by comparing the two.
1711 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1712 events_got = seq - ctx->cq_last_tm_flush;
1713 if (events_got < events_needed)
1716 list_del_init(&req->timeout.list);
1717 io_kill_timeout(req);
1718 } while (!list_empty(&ctx->timeout_list));
1720 ctx->cq_last_tm_flush = seq;
1723 static void io_commit_cqring(struct io_ring_ctx *ctx)
1725 io_flush_timeouts(ctx);
1727 /* order cqe stores with ring update */
1728 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1730 if (unlikely(!list_empty(&ctx->defer_list)))
1731 __io_queue_deferred(ctx);
1734 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1736 struct io_rings *r = ctx->rings;
1738 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1741 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1743 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1746 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1748 struct io_rings *rings = ctx->rings;
1752 * writes to the cq entry need to come after reading head; the
1753 * control dependency is enough as we're using WRITE_ONCE to
1756 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1759 tail = ctx->cached_cq_tail++;
1760 return &rings->cqes[tail & ctx->cq_mask];
1763 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1767 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1769 if (!ctx->eventfd_async)
1771 return io_wq_current_is_worker();
1774 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1776 /* see waitqueue_active() comment */
1779 if (waitqueue_active(&ctx->wait))
1780 wake_up(&ctx->wait);
1781 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1782 wake_up(&ctx->sq_data->wait);
1783 if (io_should_trigger_evfd(ctx))
1784 eventfd_signal(ctx->cq_ev_fd, 1);
1785 if (waitqueue_active(&ctx->cq_wait)) {
1786 wake_up_interruptible(&ctx->cq_wait);
1787 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1791 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1793 /* see waitqueue_active() comment */
1796 if (ctx->flags & IORING_SETUP_SQPOLL) {
1797 if (waitqueue_active(&ctx->wait))
1798 wake_up(&ctx->wait);
1800 if (io_should_trigger_evfd(ctx))
1801 eventfd_signal(ctx->cq_ev_fd, 1);
1802 if (waitqueue_active(&ctx->cq_wait)) {
1803 wake_up_interruptible(&ctx->cq_wait);
1804 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1808 /* Returns true if there are no backlogged entries after the flush */
1809 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1810 struct task_struct *tsk,
1811 struct files_struct *files)
1813 struct io_rings *rings = ctx->rings;
1814 struct io_kiocb *req, *tmp;
1815 struct io_uring_cqe *cqe;
1816 unsigned long flags;
1817 bool all_flushed, posted;
1820 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1824 spin_lock_irqsave(&ctx->completion_lock, flags);
1825 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1826 if (!io_match_task(req, tsk, files))
1829 cqe = io_get_cqring(ctx);
1833 list_move(&req->compl.list, &list);
1835 WRITE_ONCE(cqe->user_data, req->user_data);
1836 WRITE_ONCE(cqe->res, req->result);
1837 WRITE_ONCE(cqe->flags, req->compl.cflags);
1839 ctx->cached_cq_overflow++;
1840 WRITE_ONCE(ctx->rings->cq_overflow,
1841 ctx->cached_cq_overflow);
1846 all_flushed = list_empty(&ctx->cq_overflow_list);
1848 clear_bit(0, &ctx->sq_check_overflow);
1849 clear_bit(0, &ctx->cq_check_overflow);
1850 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1854 io_commit_cqring(ctx);
1855 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1857 io_cqring_ev_posted(ctx);
1859 while (!list_empty(&list)) {
1860 req = list_first_entry(&list, struct io_kiocb, compl.list);
1861 list_del(&req->compl.list);
1868 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1869 struct task_struct *tsk,
1870 struct files_struct *files)
1872 if (test_bit(0, &ctx->cq_check_overflow)) {
1873 /* iopoll syncs against uring_lock, not completion_lock */
1874 if (ctx->flags & IORING_SETUP_IOPOLL)
1875 mutex_lock(&ctx->uring_lock);
1876 __io_cqring_overflow_flush(ctx, force, tsk, files);
1877 if (ctx->flags & IORING_SETUP_IOPOLL)
1878 mutex_unlock(&ctx->uring_lock);
1882 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1884 struct io_ring_ctx *ctx = req->ctx;
1885 struct io_uring_cqe *cqe;
1887 trace_io_uring_complete(ctx, req->user_data, res);
1890 * If we can't get a cq entry, userspace overflowed the
1891 * submission (by quite a lot). Increment the overflow count in
1894 cqe = io_get_cqring(ctx);
1896 WRITE_ONCE(cqe->user_data, req->user_data);
1897 WRITE_ONCE(cqe->res, res);
1898 WRITE_ONCE(cqe->flags, cflags);
1899 } else if (ctx->cq_overflow_flushed ||
1900 atomic_read(&req->task->io_uring->in_idle)) {
1902 * If we're in ring overflow flush mode, or in task cancel mode,
1903 * then we cannot store the request for later flushing, we need
1904 * to drop it on the floor.
1906 ctx->cached_cq_overflow++;
1907 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1909 if (list_empty(&ctx->cq_overflow_list)) {
1910 set_bit(0, &ctx->sq_check_overflow);
1911 set_bit(0, &ctx->cq_check_overflow);
1912 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1916 req->compl.cflags = cflags;
1917 refcount_inc(&req->refs);
1918 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1922 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1924 __io_cqring_fill_event(req, res, 0);
1927 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1928 unsigned int cflags)
1930 struct io_ring_ctx *ctx = req->ctx;
1931 unsigned long flags;
1933 spin_lock_irqsave(&ctx->completion_lock, flags);
1934 __io_cqring_fill_event(req, res, cflags);
1935 io_commit_cqring(ctx);
1937 * If we're the last reference to this request, add to our locked
1940 if (refcount_dec_and_test(&req->refs)) {
1941 struct io_comp_state *cs = &ctx->submit_state.comp;
1943 io_dismantle_req(req);
1944 io_put_task(req->task, 1);
1945 list_add(&req->compl.list, &cs->locked_free_list);
1946 cs->locked_free_nr++;
1949 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1951 io_cqring_ev_posted(ctx);
1954 percpu_ref_put(&ctx->refs);
1958 static void io_req_complete_state(struct io_kiocb *req, long res,
1959 unsigned int cflags)
1963 req->compl.cflags = cflags;
1964 req->flags |= REQ_F_COMPLETE_INLINE;
1967 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1968 long res, unsigned cflags)
1970 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1971 io_req_complete_state(req, res, cflags);
1973 io_req_complete_post(req, res, cflags);
1976 static inline void io_req_complete(struct io_kiocb *req, long res)
1978 __io_req_complete(req, 0, res, 0);
1981 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1983 struct io_submit_state *state = &ctx->submit_state;
1984 struct io_comp_state *cs = &state->comp;
1985 struct io_kiocb *req = NULL;
1988 * If we have more than a batch's worth of requests in our IRQ side
1989 * locked cache, grab the lock and move them over to our submission
1992 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1993 spin_lock_irq(&ctx->completion_lock);
1994 list_splice_init(&cs->locked_free_list, &cs->free_list);
1995 cs->locked_free_nr = 0;
1996 spin_unlock_irq(&ctx->completion_lock);
1999 while (!list_empty(&cs->free_list)) {
2000 req = list_first_entry(&cs->free_list, struct io_kiocb,
2002 list_del(&req->compl.list);
2003 state->reqs[state->free_reqs++] = req;
2004 if (state->free_reqs == ARRAY_SIZE(state->reqs))
2011 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
2013 struct io_submit_state *state = &ctx->submit_state;
2015 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
2017 if (!state->free_reqs) {
2018 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
2021 if (io_flush_cached_reqs(ctx))
2024 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
2028 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2029 * retry single alloc to be on the safe side.
2031 if (unlikely(ret <= 0)) {
2032 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
2033 if (!state->reqs[0])
2037 state->free_reqs = ret;
2041 return state->reqs[state->free_reqs];
2044 static inline void io_put_file(struct io_kiocb *req, struct file *file,
2051 static void io_dismantle_req(struct io_kiocb *req)
2055 if (req->async_data)
2056 kfree(req->async_data);
2058 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
2059 if (req->fixed_rsrc_refs)
2060 percpu_ref_put(req->fixed_rsrc_refs);
2061 io_req_clean_work(req);
2064 static inline void io_put_task(struct task_struct *task, int nr)
2066 struct io_uring_task *tctx = task->io_uring;
2068 percpu_counter_sub(&tctx->inflight, nr);
2069 if (unlikely(atomic_read(&tctx->in_idle)))
2070 wake_up(&tctx->wait);
2071 put_task_struct_many(task, nr);
2074 static void __io_free_req(struct io_kiocb *req)
2076 struct io_ring_ctx *ctx = req->ctx;
2078 io_dismantle_req(req);
2079 io_put_task(req->task, 1);
2081 kmem_cache_free(req_cachep, req);
2082 percpu_ref_put(&ctx->refs);
2085 static inline void io_remove_next_linked(struct io_kiocb *req)
2087 struct io_kiocb *nxt = req->link;
2089 req->link = nxt->link;
2093 static void io_kill_linked_timeout(struct io_kiocb *req)
2095 struct io_ring_ctx *ctx = req->ctx;
2096 struct io_kiocb *link;
2097 bool cancelled = false;
2098 unsigned long flags;
2100 spin_lock_irqsave(&ctx->completion_lock, flags);
2104 * Can happen if a linked timeout fired and link had been like
2105 * req -> link t-out -> link t-out [-> ...]
2107 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
2108 struct io_timeout_data *io = link->async_data;
2111 io_remove_next_linked(req);
2112 link->timeout.head = NULL;
2113 ret = hrtimer_try_to_cancel(&io->timer);
2115 io_cqring_fill_event(link, -ECANCELED);
2116 io_commit_cqring(ctx);
2120 req->flags &= ~REQ_F_LINK_TIMEOUT;
2121 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2124 io_cqring_ev_posted(ctx);
2130 static void io_fail_links(struct io_kiocb *req)
2132 struct io_kiocb *link, *nxt;
2133 struct io_ring_ctx *ctx = req->ctx;
2134 unsigned long flags;
2136 spin_lock_irqsave(&ctx->completion_lock, flags);
2144 trace_io_uring_fail_link(req, link);
2145 io_cqring_fill_event(link, -ECANCELED);
2148 * It's ok to free under spinlock as they're not linked anymore,
2149 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2152 if (link->flags & REQ_F_WORK_INITIALIZED)
2153 io_put_req_deferred(link, 2);
2155 io_double_put_req(link);
2158 io_commit_cqring(ctx);
2159 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2161 io_cqring_ev_posted(ctx);
2164 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2166 if (req->flags & REQ_F_LINK_TIMEOUT)
2167 io_kill_linked_timeout(req);
2170 * If LINK is set, we have dependent requests in this chain. If we
2171 * didn't fail this request, queue the first one up, moving any other
2172 * dependencies to the next request. In case of failure, fail the rest
2175 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2176 struct io_kiocb *nxt = req->link;
2185 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2187 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2189 return __io_req_find_next(req);
2192 static bool __tctx_task_work(struct io_uring_task *tctx)
2194 struct io_ring_ctx *ctx = NULL;
2195 struct io_wq_work_list list;
2196 struct io_wq_work_node *node;
2198 if (wq_list_empty(&tctx->task_list))
2201 spin_lock_irq(&tctx->task_lock);
2202 list = tctx->task_list;
2203 INIT_WQ_LIST(&tctx->task_list);
2204 spin_unlock_irq(&tctx->task_lock);
2208 struct io_wq_work_node *next = node->next;
2209 struct io_ring_ctx *this_ctx;
2210 struct io_kiocb *req;
2212 req = container_of(node, struct io_kiocb, io_task_work.node);
2213 this_ctx = req->ctx;
2214 req->task_work.func(&req->task_work);
2219 } else if (ctx != this_ctx) {
2220 mutex_lock(&ctx->uring_lock);
2221 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
2222 mutex_unlock(&ctx->uring_lock);
2227 if (ctx && ctx->submit_state.comp.nr) {
2228 mutex_lock(&ctx->uring_lock);
2229 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
2230 mutex_unlock(&ctx->uring_lock);
2233 return list.first != NULL;
2236 static void tctx_task_work(struct callback_head *cb)
2238 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
2240 while (__tctx_task_work(tctx))
2243 clear_bit(0, &tctx->task_state);
2246 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
2247 enum task_work_notify_mode notify)
2249 struct io_uring_task *tctx = tsk->io_uring;
2250 struct io_wq_work_node *node, *prev;
2251 unsigned long flags;
2254 WARN_ON_ONCE(!tctx);
2256 spin_lock_irqsave(&tctx->task_lock, flags);
2257 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2258 spin_unlock_irqrestore(&tctx->task_lock, flags);
2260 /* task_work already pending, we're done */
2261 if (test_bit(0, &tctx->task_state) ||
2262 test_and_set_bit(0, &tctx->task_state))
2265 if (!task_work_add(tsk, &tctx->task_work, notify))
2269 * Slow path - we failed, find and delete work. if the work is not
2270 * in the list, it got run and we're fine.
2273 spin_lock_irqsave(&tctx->task_lock, flags);
2274 wq_list_for_each(node, prev, &tctx->task_list) {
2275 if (&req->io_task_work.node == node) {
2276 wq_list_del(&tctx->task_list, node, prev);
2281 spin_unlock_irqrestore(&tctx->task_lock, flags);
2282 clear_bit(0, &tctx->task_state);
2286 static int io_req_task_work_add(struct io_kiocb *req)
2288 struct task_struct *tsk = req->task;
2289 struct io_ring_ctx *ctx = req->ctx;
2290 enum task_work_notify_mode notify;
2293 if (tsk->flags & PF_EXITING)
2297 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2298 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2299 * processing task_work. There's no reliable way to tell if TWA_RESUME
2303 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2304 notify = TWA_SIGNAL;
2306 ret = io_task_work_add(tsk, req, notify);
2308 wake_up_process(tsk);
2313 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2314 task_work_func_t cb)
2316 struct task_struct *tsk = io_wq_get_task(req->ctx->io_wq);
2318 init_task_work(&req->task_work, cb);
2319 task_work_add(tsk, &req->task_work, TWA_NONE);
2320 wake_up_process(tsk);
2323 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2325 struct io_ring_ctx *ctx = req->ctx;
2327 spin_lock_irq(&ctx->completion_lock);
2328 io_cqring_fill_event(req, error);
2329 io_commit_cqring(ctx);
2330 spin_unlock_irq(&ctx->completion_lock);
2332 io_cqring_ev_posted(ctx);
2333 req_set_fail_links(req);
2334 io_double_put_req(req);
2337 static void io_req_task_cancel(struct callback_head *cb)
2339 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2340 struct io_ring_ctx *ctx = req->ctx;
2342 mutex_lock(&ctx->uring_lock);
2343 __io_req_task_cancel(req, req->result);
2344 mutex_unlock(&ctx->uring_lock);
2345 percpu_ref_put(&ctx->refs);
2348 static void __io_req_task_submit(struct io_kiocb *req)
2350 struct io_ring_ctx *ctx = req->ctx;
2352 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2353 mutex_lock(&ctx->uring_lock);
2354 if (!ctx->sqo_dead && !(current->flags & PF_EXITING) &&
2355 !io_sq_thread_acquire_mm_files(ctx, req))
2356 __io_queue_sqe(req);
2358 __io_req_task_cancel(req, -EFAULT);
2359 mutex_unlock(&ctx->uring_lock);
2361 if (ctx->flags & IORING_SETUP_SQPOLL)
2362 io_sq_thread_drop_mm_files();
2365 static void io_req_task_submit(struct callback_head *cb)
2367 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2369 __io_req_task_submit(req);
2372 static void io_req_task_queue(struct io_kiocb *req)
2376 req->task_work.func = io_req_task_submit;
2377 ret = io_req_task_work_add(req);
2378 if (unlikely(ret)) {
2379 req->result = -ECANCELED;
2380 percpu_ref_get(&req->ctx->refs);
2381 io_req_task_work_add_fallback(req, io_req_task_cancel);
2385 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2387 percpu_ref_get(&req->ctx->refs);
2389 req->task_work.func = io_req_task_cancel;
2391 if (unlikely(io_req_task_work_add(req)))
2392 io_req_task_work_add_fallback(req, io_req_task_cancel);
2395 static inline void io_queue_next(struct io_kiocb *req)
2397 struct io_kiocb *nxt = io_req_find_next(req);
2400 io_req_task_queue(nxt);
2403 static void io_free_req(struct io_kiocb *req)
2410 struct task_struct *task;
2415 static inline void io_init_req_batch(struct req_batch *rb)
2422 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2423 struct req_batch *rb)
2426 io_put_task(rb->task, rb->task_refs);
2428 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2431 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2432 struct io_submit_state *state)
2436 if (req->task != rb->task) {
2438 io_put_task(rb->task, rb->task_refs);
2439 rb->task = req->task;
2445 io_dismantle_req(req);
2446 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2447 state->reqs[state->free_reqs++] = req;
2449 list_add(&req->compl.list, &state->comp.free_list);
2452 static void io_submit_flush_completions(struct io_comp_state *cs,
2453 struct io_ring_ctx *ctx)
2456 struct io_kiocb *req;
2457 struct req_batch rb;
2459 io_init_req_batch(&rb);
2460 spin_lock_irq(&ctx->completion_lock);
2461 for (i = 0; i < nr; i++) {
2463 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2465 io_commit_cqring(ctx);
2466 spin_unlock_irq(&ctx->completion_lock);
2468 io_cqring_ev_posted(ctx);
2469 for (i = 0; i < nr; i++) {
2472 /* submission and completion refs */
2473 if (refcount_sub_and_test(2, &req->refs))
2474 io_req_free_batch(&rb, req, &ctx->submit_state);
2477 io_req_free_batch_finish(ctx, &rb);
2482 * Drop reference to request, return next in chain (if there is one) if this
2483 * was the last reference to this request.
2485 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2487 struct io_kiocb *nxt = NULL;
2489 if (refcount_dec_and_test(&req->refs)) {
2490 nxt = io_req_find_next(req);
2496 static void io_put_req(struct io_kiocb *req)
2498 if (refcount_dec_and_test(&req->refs))
2502 static void io_put_req_deferred_cb(struct callback_head *cb)
2504 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2509 static void io_free_req_deferred(struct io_kiocb *req)
2513 req->task_work.func = io_put_req_deferred_cb;
2514 ret = io_req_task_work_add(req);
2516 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2519 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2521 if (refcount_sub_and_test(refs, &req->refs))
2522 io_free_req_deferred(req);
2525 static void io_double_put_req(struct io_kiocb *req)
2527 /* drop both submit and complete references */
2528 if (refcount_sub_and_test(2, &req->refs))
2532 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2534 /* See comment at the top of this file */
2536 return __io_cqring_events(ctx);
2539 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2541 struct io_rings *rings = ctx->rings;
2543 /* make sure SQ entry isn't read before tail */
2544 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2547 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2549 unsigned int cflags;
2551 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2552 cflags |= IORING_CQE_F_BUFFER;
2553 req->flags &= ~REQ_F_BUFFER_SELECTED;
2558 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2560 struct io_buffer *kbuf;
2562 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2563 return io_put_kbuf(req, kbuf);
2566 static inline bool io_run_task_work(void)
2569 * Not safe to run on exiting task, and the task_work handling will
2570 * not add work to such a task.
2572 if (unlikely(current->flags & PF_EXITING))
2574 if (current->task_works) {
2575 __set_current_state(TASK_RUNNING);
2584 * Find and free completed poll iocbs
2586 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2587 struct list_head *done)
2589 struct req_batch rb;
2590 struct io_kiocb *req;
2592 /* order with ->result store in io_complete_rw_iopoll() */
2595 io_init_req_batch(&rb);
2596 while (!list_empty(done)) {
2599 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2600 list_del(&req->inflight_entry);
2602 if (READ_ONCE(req->result) == -EAGAIN) {
2603 req->iopoll_completed = 0;
2604 if (io_rw_reissue(req))
2608 if (req->flags & REQ_F_BUFFER_SELECTED)
2609 cflags = io_put_rw_kbuf(req);
2611 __io_cqring_fill_event(req, req->result, cflags);
2614 if (refcount_dec_and_test(&req->refs))
2615 io_req_free_batch(&rb, req, &ctx->submit_state);
2618 io_commit_cqring(ctx);
2619 io_cqring_ev_posted_iopoll(ctx);
2620 io_req_free_batch_finish(ctx, &rb);
2623 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2626 struct io_kiocb *req, *tmp;
2632 * Only spin for completions if we don't have multiple devices hanging
2633 * off our complete list, and we're under the requested amount.
2635 spin = !ctx->poll_multi_file && *nr_events < min;
2638 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2639 struct kiocb *kiocb = &req->rw.kiocb;
2642 * Move completed and retryable entries to our local lists.
2643 * If we find a request that requires polling, break out
2644 * and complete those lists first, if we have entries there.
2646 if (READ_ONCE(req->iopoll_completed)) {
2647 list_move_tail(&req->inflight_entry, &done);
2650 if (!list_empty(&done))
2653 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2657 /* iopoll may have completed current req */
2658 if (READ_ONCE(req->iopoll_completed))
2659 list_move_tail(&req->inflight_entry, &done);
2666 if (!list_empty(&done))
2667 io_iopoll_complete(ctx, nr_events, &done);
2673 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2674 * non-spinning poll check - we'll still enter the driver poll loop, but only
2675 * as a non-spinning completion check.
2677 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2680 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2683 ret = io_do_iopoll(ctx, nr_events, min);
2686 if (*nr_events >= min)
2694 * We can't just wait for polled events to come to us, we have to actively
2695 * find and complete them.
2697 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2699 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2702 mutex_lock(&ctx->uring_lock);
2703 while (!list_empty(&ctx->iopoll_list)) {
2704 unsigned int nr_events = 0;
2706 io_do_iopoll(ctx, &nr_events, 0);
2708 /* let it sleep and repeat later if can't complete a request */
2712 * Ensure we allow local-to-the-cpu processing to take place,
2713 * in this case we need to ensure that we reap all events.
2714 * Also let task_work, etc. to progress by releasing the mutex
2716 if (need_resched()) {
2717 mutex_unlock(&ctx->uring_lock);
2719 mutex_lock(&ctx->uring_lock);
2722 mutex_unlock(&ctx->uring_lock);
2725 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2727 unsigned int nr_events = 0;
2728 int iters = 0, ret = 0;
2731 * We disallow the app entering submit/complete with polling, but we
2732 * still need to lock the ring to prevent racing with polled issue
2733 * that got punted to a workqueue.
2735 mutex_lock(&ctx->uring_lock);
2738 * Don't enter poll loop if we already have events pending.
2739 * If we do, we can potentially be spinning for commands that
2740 * already triggered a CQE (eg in error).
2742 if (test_bit(0, &ctx->cq_check_overflow))
2743 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2744 if (io_cqring_events(ctx))
2748 * If a submit got punted to a workqueue, we can have the
2749 * application entering polling for a command before it gets
2750 * issued. That app will hold the uring_lock for the duration
2751 * of the poll right here, so we need to take a breather every
2752 * now and then to ensure that the issue has a chance to add
2753 * the poll to the issued list. Otherwise we can spin here
2754 * forever, while the workqueue is stuck trying to acquire the
2757 if (!(++iters & 7)) {
2758 mutex_unlock(&ctx->uring_lock);
2760 mutex_lock(&ctx->uring_lock);
2763 ret = io_iopoll_getevents(ctx, &nr_events, min);
2767 } while (min && !nr_events && !need_resched());
2769 mutex_unlock(&ctx->uring_lock);
2773 static void kiocb_end_write(struct io_kiocb *req)
2776 * Tell lockdep we inherited freeze protection from submission
2779 if (req->flags & REQ_F_ISREG) {
2780 struct inode *inode = file_inode(req->file);
2782 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2784 file_end_write(req->file);
2788 static bool io_resubmit_prep(struct io_kiocb *req)
2790 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2792 struct iov_iter iter;
2794 /* already prepared */
2795 if (req->async_data)
2798 switch (req->opcode) {
2799 case IORING_OP_READV:
2800 case IORING_OP_READ_FIXED:
2801 case IORING_OP_READ:
2804 case IORING_OP_WRITEV:
2805 case IORING_OP_WRITE_FIXED:
2806 case IORING_OP_WRITE:
2810 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2815 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2818 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2822 static bool io_rw_reissue(struct io_kiocb *req)
2825 umode_t mode = file_inode(req->file)->i_mode;
2828 if (!S_ISBLK(mode) && !S_ISREG(mode))
2830 if ((req->flags & REQ_F_NOWAIT) || io_wq_current_is_worker())
2833 * If ref is dying, we might be running poll reap from the exit work.
2834 * Don't attempt to reissue from that path, just let it fail with
2837 if (percpu_ref_is_dying(&req->ctx->refs))
2840 lockdep_assert_held(&req->ctx->uring_lock);
2842 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2844 if (!ret && io_resubmit_prep(req)) {
2845 refcount_inc(&req->refs);
2846 io_queue_async_work(req);
2849 req_set_fail_links(req);
2854 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2855 unsigned int issue_flags)
2859 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2861 if (res != req->result)
2862 req_set_fail_links(req);
2864 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2865 kiocb_end_write(req);
2866 if (req->flags & REQ_F_BUFFER_SELECTED)
2867 cflags = io_put_rw_kbuf(req);
2868 __io_req_complete(req, issue_flags, res, cflags);
2871 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2873 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2875 __io_complete_rw(req, res, res2, 0);
2878 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2880 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2882 if (kiocb->ki_flags & IOCB_WRITE)
2883 kiocb_end_write(req);
2885 if (res != -EAGAIN && res != req->result)
2886 req_set_fail_links(req);
2888 WRITE_ONCE(req->result, res);
2889 /* order with io_poll_complete() checking ->result */
2891 WRITE_ONCE(req->iopoll_completed, 1);
2895 * After the iocb has been issued, it's safe to be found on the poll list.
2896 * Adding the kiocb to the list AFTER submission ensures that we don't
2897 * find it from a io_iopoll_getevents() thread before the issuer is done
2898 * accessing the kiocb cookie.
2900 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2902 struct io_ring_ctx *ctx = req->ctx;
2905 * Track whether we have multiple files in our lists. This will impact
2906 * how we do polling eventually, not spinning if we're on potentially
2907 * different devices.
2909 if (list_empty(&ctx->iopoll_list)) {
2910 ctx->poll_multi_file = false;
2911 } else if (!ctx->poll_multi_file) {
2912 struct io_kiocb *list_req;
2914 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2916 if (list_req->file != req->file)
2917 ctx->poll_multi_file = true;
2921 * For fast devices, IO may have already completed. If it has, add
2922 * it to the front so we find it first.
2924 if (READ_ONCE(req->iopoll_completed))
2925 list_add(&req->inflight_entry, &ctx->iopoll_list);
2927 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2930 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2931 * task context or in io worker task context. If current task context is
2932 * sq thread, we don't need to check whether should wake up sq thread.
2934 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2935 wq_has_sleeper(&ctx->sq_data->wait))
2936 wake_up(&ctx->sq_data->wait);
2939 static inline void io_state_file_put(struct io_submit_state *state)
2941 if (state->file_refs) {
2942 fput_many(state->file, state->file_refs);
2943 state->file_refs = 0;
2948 * Get as many references to a file as we have IOs left in this submission,
2949 * assuming most submissions are for one file, or at least that each file
2950 * has more than one submission.
2952 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2957 if (state->file_refs) {
2958 if (state->fd == fd) {
2962 io_state_file_put(state);
2964 state->file = fget_many(fd, state->ios_left);
2965 if (unlikely(!state->file))
2969 state->file_refs = state->ios_left - 1;
2973 static bool io_bdev_nowait(struct block_device *bdev)
2975 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2979 * If we tracked the file through the SCM inflight mechanism, we could support
2980 * any file. For now, just ensure that anything potentially problematic is done
2983 static bool io_file_supports_async(struct file *file, int rw)
2985 umode_t mode = file_inode(file)->i_mode;
2987 if (S_ISBLK(mode)) {
2988 if (IS_ENABLED(CONFIG_BLOCK) &&
2989 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2993 if (S_ISCHR(mode) || S_ISSOCK(mode))
2995 if (S_ISREG(mode)) {
2996 if (IS_ENABLED(CONFIG_BLOCK) &&
2997 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2998 file->f_op != &io_uring_fops)
3003 /* any ->read/write should understand O_NONBLOCK */
3004 if (file->f_flags & O_NONBLOCK)
3007 if (!(file->f_mode & FMODE_NOWAIT))
3011 return file->f_op->read_iter != NULL;
3013 return file->f_op->write_iter != NULL;
3016 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3018 struct io_ring_ctx *ctx = req->ctx;
3019 struct kiocb *kiocb = &req->rw.kiocb;
3020 struct file *file = req->file;
3024 if (S_ISREG(file_inode(file)->i_mode))
3025 req->flags |= REQ_F_ISREG;
3027 kiocb->ki_pos = READ_ONCE(sqe->off);
3028 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
3029 req->flags |= REQ_F_CUR_POS;
3030 kiocb->ki_pos = file->f_pos;
3032 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
3033 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
3034 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
3038 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
3039 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
3040 req->flags |= REQ_F_NOWAIT;
3042 ioprio = READ_ONCE(sqe->ioprio);
3044 ret = ioprio_check_cap(ioprio);
3048 kiocb->ki_ioprio = ioprio;
3050 kiocb->ki_ioprio = get_current_ioprio();
3052 if (ctx->flags & IORING_SETUP_IOPOLL) {
3053 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
3054 !kiocb->ki_filp->f_op->iopoll)
3057 kiocb->ki_flags |= IOCB_HIPRI;
3058 kiocb->ki_complete = io_complete_rw_iopoll;
3059 req->iopoll_completed = 0;
3061 if (kiocb->ki_flags & IOCB_HIPRI)
3063 kiocb->ki_complete = io_complete_rw;
3066 req->rw.addr = READ_ONCE(sqe->addr);
3067 req->rw.len = READ_ONCE(sqe->len);
3068 req->buf_index = READ_ONCE(sqe->buf_index);
3072 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
3078 case -ERESTARTNOINTR:
3079 case -ERESTARTNOHAND:
3080 case -ERESTART_RESTARTBLOCK:
3082 * We can't just restart the syscall, since previously
3083 * submitted sqes may already be in progress. Just fail this
3089 kiocb->ki_complete(kiocb, ret, 0);
3093 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
3094 unsigned int issue_flags)
3096 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
3097 struct io_async_rw *io = req->async_data;
3099 /* add previously done IO, if any */
3100 if (io && io->bytes_done > 0) {
3102 ret = io->bytes_done;
3104 ret += io->bytes_done;
3107 if (req->flags & REQ_F_CUR_POS)
3108 req->file->f_pos = kiocb->ki_pos;
3109 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
3110 __io_complete_rw(req, ret, 0, issue_flags);
3112 io_rw_done(kiocb, ret);
3115 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3117 struct io_ring_ctx *ctx = req->ctx;
3118 size_t len = req->rw.len;
3119 struct io_mapped_ubuf *imu;
3120 u16 index, buf_index = req->buf_index;
3124 if (unlikely(buf_index >= ctx->nr_user_bufs))
3126 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3127 imu = &ctx->user_bufs[index];
3128 buf_addr = req->rw.addr;
3131 if (buf_addr + len < buf_addr)
3133 /* not inside the mapped region */
3134 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
3138 * May not be a start of buffer, set size appropriately
3139 * and advance us to the beginning.
3141 offset = buf_addr - imu->ubuf;
3142 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3146 * Don't use iov_iter_advance() here, as it's really slow for
3147 * using the latter parts of a big fixed buffer - it iterates
3148 * over each segment manually. We can cheat a bit here, because
3151 * 1) it's a BVEC iter, we set it up
3152 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3153 * first and last bvec
3155 * So just find our index, and adjust the iterator afterwards.
3156 * If the offset is within the first bvec (or the whole first
3157 * bvec, just use iov_iter_advance(). This makes it easier
3158 * since we can just skip the first segment, which may not
3159 * be PAGE_SIZE aligned.
3161 const struct bio_vec *bvec = imu->bvec;
3163 if (offset <= bvec->bv_len) {
3164 iov_iter_advance(iter, offset);
3166 unsigned long seg_skip;
3168 /* skip first vec */
3169 offset -= bvec->bv_len;
3170 seg_skip = 1 + (offset >> PAGE_SHIFT);
3172 iter->bvec = bvec + seg_skip;
3173 iter->nr_segs -= seg_skip;
3174 iter->count -= bvec->bv_len + offset;
3175 iter->iov_offset = offset & ~PAGE_MASK;
3182 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3185 mutex_unlock(&ctx->uring_lock);
3188 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3191 * "Normal" inline submissions always hold the uring_lock, since we
3192 * grab it from the system call. Same is true for the SQPOLL offload.
3193 * The only exception is when we've detached the request and issue it
3194 * from an async worker thread, grab the lock for that case.
3197 mutex_lock(&ctx->uring_lock);
3200 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3201 int bgid, struct io_buffer *kbuf,
3204 struct io_buffer *head;
3206 if (req->flags & REQ_F_BUFFER_SELECTED)
3209 io_ring_submit_lock(req->ctx, needs_lock);
3211 lockdep_assert_held(&req->ctx->uring_lock);
3213 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3215 if (!list_empty(&head->list)) {
3216 kbuf = list_last_entry(&head->list, struct io_buffer,
3218 list_del(&kbuf->list);
3221 idr_remove(&req->ctx->io_buffer_idr, bgid);
3223 if (*len > kbuf->len)
3226 kbuf = ERR_PTR(-ENOBUFS);
3229 io_ring_submit_unlock(req->ctx, needs_lock);
3234 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3237 struct io_buffer *kbuf;
3240 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3241 bgid = req->buf_index;
3242 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3245 req->rw.addr = (u64) (unsigned long) kbuf;
3246 req->flags |= REQ_F_BUFFER_SELECTED;
3247 return u64_to_user_ptr(kbuf->addr);
3250 #ifdef CONFIG_COMPAT
3251 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3254 struct compat_iovec __user *uiov;
3255 compat_ssize_t clen;
3259 uiov = u64_to_user_ptr(req->rw.addr);
3260 if (!access_ok(uiov, sizeof(*uiov)))
3262 if (__get_user(clen, &uiov->iov_len))
3268 buf = io_rw_buffer_select(req, &len, needs_lock);
3270 return PTR_ERR(buf);
3271 iov[0].iov_base = buf;
3272 iov[0].iov_len = (compat_size_t) len;
3277 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3280 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3284 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3287 len = iov[0].iov_len;
3290 buf = io_rw_buffer_select(req, &len, needs_lock);
3292 return PTR_ERR(buf);
3293 iov[0].iov_base = buf;
3294 iov[0].iov_len = len;
3298 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3301 if (req->flags & REQ_F_BUFFER_SELECTED) {
3302 struct io_buffer *kbuf;
3304 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3305 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3306 iov[0].iov_len = kbuf->len;
3309 if (req->rw.len != 1)
3312 #ifdef CONFIG_COMPAT
3313 if (req->ctx->compat)
3314 return io_compat_import(req, iov, needs_lock);
3317 return __io_iov_buffer_select(req, iov, needs_lock);
3320 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3321 struct iov_iter *iter, bool needs_lock)
3323 void __user *buf = u64_to_user_ptr(req->rw.addr);
3324 size_t sqe_len = req->rw.len;
3325 u8 opcode = req->opcode;
3328 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3330 return io_import_fixed(req, rw, iter);
3333 /* buffer index only valid with fixed read/write, or buffer select */
3334 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3337 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3338 if (req->flags & REQ_F_BUFFER_SELECT) {
3339 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3341 return PTR_ERR(buf);
3342 req->rw.len = sqe_len;
3345 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3350 if (req->flags & REQ_F_BUFFER_SELECT) {
3351 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3353 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3358 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3362 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3364 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3368 * For files that don't have ->read_iter() and ->write_iter(), handle them
3369 * by looping over ->read() or ->write() manually.
3371 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3373 struct kiocb *kiocb = &req->rw.kiocb;
3374 struct file *file = req->file;
3378 * Don't support polled IO through this interface, and we can't
3379 * support non-blocking either. For the latter, this just causes
3380 * the kiocb to be handled from an async context.
3382 if (kiocb->ki_flags & IOCB_HIPRI)
3384 if (kiocb->ki_flags & IOCB_NOWAIT)
3387 while (iov_iter_count(iter)) {
3391 if (!iov_iter_is_bvec(iter)) {
3392 iovec = iov_iter_iovec(iter);
3394 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3395 iovec.iov_len = req->rw.len;
3399 nr = file->f_op->read(file, iovec.iov_base,
3400 iovec.iov_len, io_kiocb_ppos(kiocb));
3402 nr = file->f_op->write(file, iovec.iov_base,
3403 iovec.iov_len, io_kiocb_ppos(kiocb));
3412 if (nr != iovec.iov_len)
3416 iov_iter_advance(iter, nr);
3422 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3423 const struct iovec *fast_iov, struct iov_iter *iter)
3425 struct io_async_rw *rw = req->async_data;
3427 memcpy(&rw->iter, iter, sizeof(*iter));
3428 rw->free_iovec = iovec;
3430 /* can only be fixed buffers, no need to do anything */
3431 if (iov_iter_is_bvec(iter))
3434 unsigned iov_off = 0;
3436 rw->iter.iov = rw->fast_iov;
3437 if (iter->iov != fast_iov) {
3438 iov_off = iter->iov - fast_iov;
3439 rw->iter.iov += iov_off;
3441 if (rw->fast_iov != fast_iov)
3442 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3443 sizeof(struct iovec) * iter->nr_segs);
3445 req->flags |= REQ_F_NEED_CLEANUP;
3449 static inline int __io_alloc_async_data(struct io_kiocb *req)
3451 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3452 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3453 return req->async_data == NULL;
3456 static int io_alloc_async_data(struct io_kiocb *req)
3458 if (!io_op_defs[req->opcode].needs_async_data)
3461 return __io_alloc_async_data(req);
3464 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3465 const struct iovec *fast_iov,
3466 struct iov_iter *iter, bool force)
3468 if (!force && !io_op_defs[req->opcode].needs_async_data)
3470 if (!req->async_data) {
3471 if (__io_alloc_async_data(req)) {
3476 io_req_map_rw(req, iovec, fast_iov, iter);
3481 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3483 struct io_async_rw *iorw = req->async_data;
3484 struct iovec *iov = iorw->fast_iov;
3487 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3488 if (unlikely(ret < 0))
3491 iorw->bytes_done = 0;
3492 iorw->free_iovec = iov;
3494 req->flags |= REQ_F_NEED_CLEANUP;
3498 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3500 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3502 return io_prep_rw(req, sqe);
3506 * This is our waitqueue callback handler, registered through lock_page_async()
3507 * when we initially tried to do the IO with the iocb armed our waitqueue.
3508 * This gets called when the page is unlocked, and we generally expect that to
3509 * happen when the page IO is completed and the page is now uptodate. This will
3510 * queue a task_work based retry of the operation, attempting to copy the data
3511 * again. If the latter fails because the page was NOT uptodate, then we will
3512 * do a thread based blocking retry of the operation. That's the unexpected
3515 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3516 int sync, void *arg)
3518 struct wait_page_queue *wpq;
3519 struct io_kiocb *req = wait->private;
3520 struct wait_page_key *key = arg;
3522 wpq = container_of(wait, struct wait_page_queue, wait);
3524 if (!wake_page_match(wpq, key))
3527 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3528 list_del_init(&wait->entry);
3530 /* submit ref gets dropped, acquire a new one */
3531 refcount_inc(&req->refs);
3532 io_req_task_queue(req);
3537 * This controls whether a given IO request should be armed for async page
3538 * based retry. If we return false here, the request is handed to the async
3539 * worker threads for retry. If we're doing buffered reads on a regular file,
3540 * we prepare a private wait_page_queue entry and retry the operation. This
3541 * will either succeed because the page is now uptodate and unlocked, or it
3542 * will register a callback when the page is unlocked at IO completion. Through
3543 * that callback, io_uring uses task_work to setup a retry of the operation.
3544 * That retry will attempt the buffered read again. The retry will generally
3545 * succeed, or in rare cases where it fails, we then fall back to using the
3546 * async worker threads for a blocking retry.
3548 static bool io_rw_should_retry(struct io_kiocb *req)
3550 struct io_async_rw *rw = req->async_data;
3551 struct wait_page_queue *wait = &rw->wpq;
3552 struct kiocb *kiocb = &req->rw.kiocb;
3554 /* never retry for NOWAIT, we just complete with -EAGAIN */
3555 if (req->flags & REQ_F_NOWAIT)
3558 /* Only for buffered IO */
3559 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3563 * just use poll if we can, and don't attempt if the fs doesn't
3564 * support callback based unlocks
3566 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3569 wait->wait.func = io_async_buf_func;
3570 wait->wait.private = req;
3571 wait->wait.flags = 0;
3572 INIT_LIST_HEAD(&wait->wait.entry);
3573 kiocb->ki_flags |= IOCB_WAITQ;
3574 kiocb->ki_flags &= ~IOCB_NOWAIT;
3575 kiocb->ki_waitq = wait;
3579 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3581 if (req->file->f_op->read_iter)
3582 return call_read_iter(req->file, &req->rw.kiocb, iter);
3583 else if (req->file->f_op->read)
3584 return loop_rw_iter(READ, req, iter);
3589 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3591 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3592 struct kiocb *kiocb = &req->rw.kiocb;
3593 struct iov_iter __iter, *iter = &__iter;
3594 struct io_async_rw *rw = req->async_data;
3595 ssize_t io_size, ret, ret2;
3596 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3602 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3606 io_size = iov_iter_count(iter);
3607 req->result = io_size;
3609 /* Ensure we clear previously set non-block flag */
3610 if (!force_nonblock)
3611 kiocb->ki_flags &= ~IOCB_NOWAIT;
3613 kiocb->ki_flags |= IOCB_NOWAIT;
3615 /* If the file doesn't support async, just async punt */
3616 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3617 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3618 return ret ?: -EAGAIN;
3621 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3622 if (unlikely(ret)) {
3627 ret = io_iter_do_read(req, iter);
3629 if (ret == -EIOCBQUEUED) {
3631 } else if (ret == -EAGAIN) {
3632 /* IOPOLL retry should happen for io-wq threads */
3633 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3635 /* no retry on NONBLOCK nor RWF_NOWAIT */
3636 if (req->flags & REQ_F_NOWAIT)
3638 /* some cases will consume bytes even on error returns */
3639 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3641 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3642 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3643 /* read all, failed, already did sync or don't want to retry */
3647 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3652 rw = req->async_data;
3653 /* now use our persistent iterator, if we aren't already */
3658 rw->bytes_done += ret;
3659 /* if we can retry, do so with the callbacks armed */
3660 if (!io_rw_should_retry(req)) {
3661 kiocb->ki_flags &= ~IOCB_WAITQ;
3666 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3667 * we get -EIOCBQUEUED, then we'll get a notification when the
3668 * desired page gets unlocked. We can also get a partial read
3669 * here, and if we do, then just retry at the new offset.
3671 ret = io_iter_do_read(req, iter);
3672 if (ret == -EIOCBQUEUED)
3674 /* we got some bytes, but not all. retry. */
3675 } while (ret > 0 && ret < io_size);
3677 kiocb_done(kiocb, ret, issue_flags);
3679 /* it's faster to check here then delegate to kfree */
3685 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3687 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3689 return io_prep_rw(req, sqe);
3692 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3694 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3695 struct kiocb *kiocb = &req->rw.kiocb;
3696 struct iov_iter __iter, *iter = &__iter;
3697 struct io_async_rw *rw = req->async_data;
3698 ssize_t ret, ret2, io_size;
3699 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3705 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3709 io_size = iov_iter_count(iter);
3710 req->result = io_size;
3712 /* Ensure we clear previously set non-block flag */
3713 if (!force_nonblock)
3714 kiocb->ki_flags &= ~IOCB_NOWAIT;
3716 kiocb->ki_flags |= IOCB_NOWAIT;
3718 /* If the file doesn't support async, just async punt */
3719 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3722 /* file path doesn't support NOWAIT for non-direct_IO */
3723 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3724 (req->flags & REQ_F_ISREG))
3727 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3732 * Open-code file_start_write here to grab freeze protection,
3733 * which will be released by another thread in
3734 * io_complete_rw(). Fool lockdep by telling it the lock got
3735 * released so that it doesn't complain about the held lock when
3736 * we return to userspace.
3738 if (req->flags & REQ_F_ISREG) {
3739 sb_start_write(file_inode(req->file)->i_sb);
3740 __sb_writers_release(file_inode(req->file)->i_sb,
3743 kiocb->ki_flags |= IOCB_WRITE;
3745 if (req->file->f_op->write_iter)
3746 ret2 = call_write_iter(req->file, kiocb, iter);
3747 else if (req->file->f_op->write)
3748 ret2 = loop_rw_iter(WRITE, req, iter);
3753 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3754 * retry them without IOCB_NOWAIT.
3756 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3758 /* no retry on NONBLOCK nor RWF_NOWAIT */
3759 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3761 if (!force_nonblock || ret2 != -EAGAIN) {
3762 /* IOPOLL retry should happen for io-wq threads */
3763 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3766 kiocb_done(kiocb, ret2, issue_flags);
3769 /* some cases will consume bytes even on error returns */
3770 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3771 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3772 return ret ?: -EAGAIN;
3775 /* it's reportedly faster than delegating the null check to kfree() */
3781 static int io_renameat_prep(struct io_kiocb *req,
3782 const struct io_uring_sqe *sqe)
3784 struct io_rename *ren = &req->rename;
3785 const char __user *oldf, *newf;
3787 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3790 ren->old_dfd = READ_ONCE(sqe->fd);
3791 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3792 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3793 ren->new_dfd = READ_ONCE(sqe->len);
3794 ren->flags = READ_ONCE(sqe->rename_flags);
3796 ren->oldpath = getname(oldf);
3797 if (IS_ERR(ren->oldpath))
3798 return PTR_ERR(ren->oldpath);
3800 ren->newpath = getname(newf);
3801 if (IS_ERR(ren->newpath)) {
3802 putname(ren->oldpath);
3803 return PTR_ERR(ren->newpath);
3806 req->flags |= REQ_F_NEED_CLEANUP;
3810 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3812 struct io_rename *ren = &req->rename;
3815 if (issue_flags & IO_URING_F_NONBLOCK)
3818 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3819 ren->newpath, ren->flags);
3821 req->flags &= ~REQ_F_NEED_CLEANUP;
3823 req_set_fail_links(req);
3824 io_req_complete(req, ret);
3828 static int io_unlinkat_prep(struct io_kiocb *req,
3829 const struct io_uring_sqe *sqe)
3831 struct io_unlink *un = &req->unlink;
3832 const char __user *fname;
3834 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3837 un->dfd = READ_ONCE(sqe->fd);
3839 un->flags = READ_ONCE(sqe->unlink_flags);
3840 if (un->flags & ~AT_REMOVEDIR)
3843 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3844 un->filename = getname(fname);
3845 if (IS_ERR(un->filename))
3846 return PTR_ERR(un->filename);
3848 req->flags |= REQ_F_NEED_CLEANUP;
3852 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3854 struct io_unlink *un = &req->unlink;
3857 if (issue_flags & IO_URING_F_NONBLOCK)
3860 if (un->flags & AT_REMOVEDIR)
3861 ret = do_rmdir(un->dfd, un->filename);
3863 ret = do_unlinkat(un->dfd, un->filename);
3865 req->flags &= ~REQ_F_NEED_CLEANUP;
3867 req_set_fail_links(req);
3868 io_req_complete(req, ret);
3872 static int io_shutdown_prep(struct io_kiocb *req,
3873 const struct io_uring_sqe *sqe)
3875 #if defined(CONFIG_NET)
3876 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3878 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3882 req->shutdown.how = READ_ONCE(sqe->len);
3889 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3891 #if defined(CONFIG_NET)
3892 struct socket *sock;
3895 if (issue_flags & IO_URING_F_NONBLOCK)
3898 sock = sock_from_file(req->file);
3899 if (unlikely(!sock))
3902 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3904 req_set_fail_links(req);
3905 io_req_complete(req, ret);
3912 static int __io_splice_prep(struct io_kiocb *req,
3913 const struct io_uring_sqe *sqe)
3915 struct io_splice* sp = &req->splice;
3916 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3918 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3922 sp->len = READ_ONCE(sqe->len);
3923 sp->flags = READ_ONCE(sqe->splice_flags);
3925 if (unlikely(sp->flags & ~valid_flags))
3928 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3929 (sp->flags & SPLICE_F_FD_IN_FIXED));
3932 req->flags |= REQ_F_NEED_CLEANUP;
3934 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3936 * Splice operation will be punted aync, and here need to
3937 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3939 io_req_init_async(req);
3940 req->work.flags |= IO_WQ_WORK_UNBOUND;
3946 static int io_tee_prep(struct io_kiocb *req,
3947 const struct io_uring_sqe *sqe)
3949 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3951 return __io_splice_prep(req, sqe);
3954 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3956 struct io_splice *sp = &req->splice;
3957 struct file *in = sp->file_in;
3958 struct file *out = sp->file_out;
3959 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3962 if (issue_flags & IO_URING_F_NONBLOCK)
3965 ret = do_tee(in, out, sp->len, flags);
3967 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3968 req->flags &= ~REQ_F_NEED_CLEANUP;
3971 req_set_fail_links(req);
3972 io_req_complete(req, ret);
3976 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3978 struct io_splice* sp = &req->splice;
3980 sp->off_in = READ_ONCE(sqe->splice_off_in);
3981 sp->off_out = READ_ONCE(sqe->off);
3982 return __io_splice_prep(req, sqe);
3985 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3987 struct io_splice *sp = &req->splice;
3988 struct file *in = sp->file_in;
3989 struct file *out = sp->file_out;
3990 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3991 loff_t *poff_in, *poff_out;
3994 if (issue_flags & IO_URING_F_NONBLOCK)
3997 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3998 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4001 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4003 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
4004 req->flags &= ~REQ_F_NEED_CLEANUP;
4007 req_set_fail_links(req);
4008 io_req_complete(req, ret);
4013 * IORING_OP_NOP just posts a completion event, nothing else.
4015 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4017 struct io_ring_ctx *ctx = req->ctx;
4019 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4022 __io_req_complete(req, issue_flags, 0, 0);
4026 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4028 struct io_ring_ctx *ctx = req->ctx;
4033 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4035 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4038 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4039 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4042 req->sync.off = READ_ONCE(sqe->off);
4043 req->sync.len = READ_ONCE(sqe->len);
4047 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4049 loff_t end = req->sync.off + req->sync.len;
4052 /* fsync always requires a blocking context */
4053 if (issue_flags & IO_URING_F_NONBLOCK)
4056 ret = vfs_fsync_range(req->file, req->sync.off,
4057 end > 0 ? end : LLONG_MAX,
4058 req->sync.flags & IORING_FSYNC_DATASYNC);
4060 req_set_fail_links(req);
4061 io_req_complete(req, ret);
4065 static int io_fallocate_prep(struct io_kiocb *req,
4066 const struct io_uring_sqe *sqe)
4068 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
4070 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4073 req->sync.off = READ_ONCE(sqe->off);
4074 req->sync.len = READ_ONCE(sqe->addr);
4075 req->sync.mode = READ_ONCE(sqe->len);
4079 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4083 /* fallocate always requiring blocking context */
4084 if (issue_flags & IO_URING_F_NONBLOCK)
4086 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4089 req_set_fail_links(req);
4090 io_req_complete(req, ret);
4094 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4096 const char __user *fname;
4099 if (unlikely(sqe->ioprio || sqe->buf_index))
4101 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4104 /* open.how should be already initialised */
4105 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4106 req->open.how.flags |= O_LARGEFILE;
4108 req->open.dfd = READ_ONCE(sqe->fd);
4109 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4110 req->open.filename = getname(fname);
4111 if (IS_ERR(req->open.filename)) {
4112 ret = PTR_ERR(req->open.filename);
4113 req->open.filename = NULL;
4116 req->open.nofile = rlimit(RLIMIT_NOFILE);
4117 req->flags |= REQ_F_NEED_CLEANUP;
4121 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4125 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4127 mode = READ_ONCE(sqe->len);
4128 flags = READ_ONCE(sqe->open_flags);
4129 req->open.how = build_open_how(flags, mode);
4130 return __io_openat_prep(req, sqe);
4133 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4135 struct open_how __user *how;
4139 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4141 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4142 len = READ_ONCE(sqe->len);
4143 if (len < OPEN_HOW_SIZE_VER0)
4146 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4151 return __io_openat_prep(req, sqe);
4154 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4156 struct open_flags op;
4159 bool resolve_nonblock;
4162 ret = build_open_flags(&req->open.how, &op);
4165 nonblock_set = op.open_flag & O_NONBLOCK;
4166 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4167 if (issue_flags & IO_URING_F_NONBLOCK) {
4169 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4170 * it'll always -EAGAIN
4172 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4174 op.lookup_flags |= LOOKUP_CACHED;
4175 op.open_flag |= O_NONBLOCK;
4178 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4182 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4183 /* only retry if RESOLVE_CACHED wasn't already set by application */
4184 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
4185 file == ERR_PTR(-EAGAIN)) {
4187 * We could hang on to this 'fd', but seems like marginal
4188 * gain for something that is now known to be a slower path.
4189 * So just put it, and we'll get a new one when we retry.
4197 ret = PTR_ERR(file);
4199 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4200 file->f_flags &= ~O_NONBLOCK;
4201 fsnotify_open(file);
4202 fd_install(ret, file);
4205 putname(req->open.filename);
4206 req->flags &= ~REQ_F_NEED_CLEANUP;
4208 req_set_fail_links(req);
4209 io_req_complete(req, ret);
4213 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4215 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
4218 static int io_remove_buffers_prep(struct io_kiocb *req,
4219 const struct io_uring_sqe *sqe)
4221 struct io_provide_buf *p = &req->pbuf;
4224 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4227 tmp = READ_ONCE(sqe->fd);
4228 if (!tmp || tmp > USHRT_MAX)
4231 memset(p, 0, sizeof(*p));
4233 p->bgid = READ_ONCE(sqe->buf_group);
4237 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4238 int bgid, unsigned nbufs)
4242 /* shouldn't happen */
4246 /* the head kbuf is the list itself */
4247 while (!list_empty(&buf->list)) {
4248 struct io_buffer *nxt;
4250 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4251 list_del(&nxt->list);
4258 idr_remove(&ctx->io_buffer_idr, bgid);
4263 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4265 struct io_provide_buf *p = &req->pbuf;
4266 struct io_ring_ctx *ctx = req->ctx;
4267 struct io_buffer *head;
4269 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4271 io_ring_submit_lock(ctx, !force_nonblock);
4273 lockdep_assert_held(&ctx->uring_lock);
4276 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4278 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4280 req_set_fail_links(req);
4282 /* need to hold the lock to complete IOPOLL requests */
4283 if (ctx->flags & IORING_SETUP_IOPOLL) {
4284 __io_req_complete(req, issue_flags, ret, 0);
4285 io_ring_submit_unlock(ctx, !force_nonblock);
4287 io_ring_submit_unlock(ctx, !force_nonblock);
4288 __io_req_complete(req, issue_flags, ret, 0);
4293 static int io_provide_buffers_prep(struct io_kiocb *req,
4294 const struct io_uring_sqe *sqe)
4296 struct io_provide_buf *p = &req->pbuf;
4299 if (sqe->ioprio || sqe->rw_flags)
4302 tmp = READ_ONCE(sqe->fd);
4303 if (!tmp || tmp > USHRT_MAX)
4306 p->addr = READ_ONCE(sqe->addr);
4307 p->len = READ_ONCE(sqe->len);
4309 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4312 p->bgid = READ_ONCE(sqe->buf_group);
4313 tmp = READ_ONCE(sqe->off);
4314 if (tmp > USHRT_MAX)
4320 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4322 struct io_buffer *buf;
4323 u64 addr = pbuf->addr;
4324 int i, bid = pbuf->bid;
4326 for (i = 0; i < pbuf->nbufs; i++) {
4327 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4332 buf->len = pbuf->len;
4337 INIT_LIST_HEAD(&buf->list);
4340 list_add_tail(&buf->list, &(*head)->list);
4344 return i ? i : -ENOMEM;
4347 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4349 struct io_provide_buf *p = &req->pbuf;
4350 struct io_ring_ctx *ctx = req->ctx;
4351 struct io_buffer *head, *list;
4353 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4355 io_ring_submit_lock(ctx, !force_nonblock);
4357 lockdep_assert_held(&ctx->uring_lock);
4359 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4361 ret = io_add_buffers(p, &head);
4366 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4369 __io_remove_buffers(ctx, head, p->bgid, -1U);
4375 req_set_fail_links(req);
4377 /* need to hold the lock to complete IOPOLL requests */
4378 if (ctx->flags & IORING_SETUP_IOPOLL) {
4379 __io_req_complete(req, issue_flags, ret, 0);
4380 io_ring_submit_unlock(ctx, !force_nonblock);
4382 io_ring_submit_unlock(ctx, !force_nonblock);
4383 __io_req_complete(req, issue_flags, ret, 0);
4388 static int io_epoll_ctl_prep(struct io_kiocb *req,
4389 const struct io_uring_sqe *sqe)
4391 #if defined(CONFIG_EPOLL)
4392 if (sqe->ioprio || sqe->buf_index)
4394 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4397 req->epoll.epfd = READ_ONCE(sqe->fd);
4398 req->epoll.op = READ_ONCE(sqe->len);
4399 req->epoll.fd = READ_ONCE(sqe->off);
4401 if (ep_op_has_event(req->epoll.op)) {
4402 struct epoll_event __user *ev;
4404 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4405 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4415 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4417 #if defined(CONFIG_EPOLL)
4418 struct io_epoll *ie = &req->epoll;
4420 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4422 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4423 if (force_nonblock && ret == -EAGAIN)
4427 req_set_fail_links(req);
4428 __io_req_complete(req, issue_flags, ret, 0);
4435 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4437 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4438 if (sqe->ioprio || sqe->buf_index || sqe->off)
4440 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4443 req->madvise.addr = READ_ONCE(sqe->addr);
4444 req->madvise.len = READ_ONCE(sqe->len);
4445 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4452 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4454 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4455 struct io_madvise *ma = &req->madvise;
4458 if (issue_flags & IO_URING_F_NONBLOCK)
4461 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4463 req_set_fail_links(req);
4464 io_req_complete(req, ret);
4471 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4473 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4475 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4478 req->fadvise.offset = READ_ONCE(sqe->off);
4479 req->fadvise.len = READ_ONCE(sqe->len);
4480 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4484 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4486 struct io_fadvise *fa = &req->fadvise;
4489 if (issue_flags & IO_URING_F_NONBLOCK) {
4490 switch (fa->advice) {
4491 case POSIX_FADV_NORMAL:
4492 case POSIX_FADV_RANDOM:
4493 case POSIX_FADV_SEQUENTIAL:
4500 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4502 req_set_fail_links(req);
4503 io_req_complete(req, ret);
4507 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4509 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4511 if (sqe->ioprio || sqe->buf_index)
4513 if (req->flags & REQ_F_FIXED_FILE)
4516 req->statx.dfd = READ_ONCE(sqe->fd);
4517 req->statx.mask = READ_ONCE(sqe->len);
4518 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4519 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4520 req->statx.flags = READ_ONCE(sqe->statx_flags);
4525 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4527 struct io_statx *ctx = &req->statx;
4530 if (issue_flags & IO_URING_F_NONBLOCK) {
4531 /* only need file table for an actual valid fd */
4532 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4533 req->flags |= REQ_F_NO_FILE_TABLE;
4537 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4541 req_set_fail_links(req);
4542 io_req_complete(req, ret);
4546 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4548 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4550 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4551 sqe->rw_flags || sqe->buf_index)
4553 if (req->flags & REQ_F_FIXED_FILE)
4556 req->close.fd = READ_ONCE(sqe->fd);
4560 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4562 struct files_struct *files = current->files;
4563 struct io_close *close = &req->close;
4564 struct fdtable *fdt;
4570 spin_lock(&files->file_lock);
4571 fdt = files_fdtable(files);
4572 if (close->fd >= fdt->max_fds) {
4573 spin_unlock(&files->file_lock);
4576 file = fdt->fd[close->fd];
4578 spin_unlock(&files->file_lock);
4582 if (file->f_op == &io_uring_fops) {
4583 spin_unlock(&files->file_lock);
4588 /* if the file has a flush method, be safe and punt to async */
4589 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4590 spin_unlock(&files->file_lock);
4594 ret = __close_fd_get_file(close->fd, &file);
4595 spin_unlock(&files->file_lock);
4602 /* No ->flush() or already async, safely close from here */
4603 ret = filp_close(file, current->files);
4606 req_set_fail_links(req);
4609 __io_req_complete(req, issue_flags, ret, 0);
4613 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4615 struct io_ring_ctx *ctx = req->ctx;
4617 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4619 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4622 req->sync.off = READ_ONCE(sqe->off);
4623 req->sync.len = READ_ONCE(sqe->len);
4624 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4628 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4632 /* sync_file_range always requires a blocking context */
4633 if (issue_flags & IO_URING_F_NONBLOCK)
4636 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4639 req_set_fail_links(req);
4640 io_req_complete(req, ret);
4644 #if defined(CONFIG_NET)
4645 static int io_setup_async_msg(struct io_kiocb *req,
4646 struct io_async_msghdr *kmsg)
4648 struct io_async_msghdr *async_msg = req->async_data;
4652 if (io_alloc_async_data(req)) {
4653 kfree(kmsg->free_iov);
4656 async_msg = req->async_data;
4657 req->flags |= REQ_F_NEED_CLEANUP;
4658 memcpy(async_msg, kmsg, sizeof(*kmsg));
4659 async_msg->msg.msg_name = &async_msg->addr;
4660 /* if were using fast_iov, set it to the new one */
4661 if (!async_msg->free_iov)
4662 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4667 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4668 struct io_async_msghdr *iomsg)
4670 iomsg->msg.msg_name = &iomsg->addr;
4671 iomsg->free_iov = iomsg->fast_iov;
4672 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4673 req->sr_msg.msg_flags, &iomsg->free_iov);
4676 static int io_sendmsg_prep_async(struct io_kiocb *req)
4680 if (!io_op_defs[req->opcode].needs_async_data)
4682 ret = io_sendmsg_copy_hdr(req, req->async_data);
4684 req->flags |= REQ_F_NEED_CLEANUP;
4688 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4690 struct io_sr_msg *sr = &req->sr_msg;
4692 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4695 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4696 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4697 sr->len = READ_ONCE(sqe->len);
4699 #ifdef CONFIG_COMPAT
4700 if (req->ctx->compat)
4701 sr->msg_flags |= MSG_CMSG_COMPAT;
4706 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4708 struct io_async_msghdr iomsg, *kmsg;
4709 struct socket *sock;
4713 sock = sock_from_file(req->file);
4714 if (unlikely(!sock))
4717 kmsg = req->async_data;
4719 ret = io_sendmsg_copy_hdr(req, &iomsg);
4725 flags = req->sr_msg.msg_flags;
4726 if (flags & MSG_DONTWAIT)
4727 req->flags |= REQ_F_NOWAIT;
4728 else if (issue_flags & IO_URING_F_NONBLOCK)
4729 flags |= MSG_DONTWAIT;
4731 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4732 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4733 return io_setup_async_msg(req, kmsg);
4734 if (ret == -ERESTARTSYS)
4737 /* fast path, check for non-NULL to avoid function call */
4739 kfree(kmsg->free_iov);
4740 req->flags &= ~REQ_F_NEED_CLEANUP;
4742 req_set_fail_links(req);
4743 __io_req_complete(req, issue_flags, ret, 0);
4747 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4749 struct io_sr_msg *sr = &req->sr_msg;
4752 struct socket *sock;
4756 sock = sock_from_file(req->file);
4757 if (unlikely(!sock))
4760 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4764 msg.msg_name = NULL;
4765 msg.msg_control = NULL;
4766 msg.msg_controllen = 0;
4767 msg.msg_namelen = 0;
4769 flags = req->sr_msg.msg_flags;
4770 if (flags & MSG_DONTWAIT)
4771 req->flags |= REQ_F_NOWAIT;
4772 else if (issue_flags & IO_URING_F_NONBLOCK)
4773 flags |= MSG_DONTWAIT;
4775 msg.msg_flags = flags;
4776 ret = sock_sendmsg(sock, &msg);
4777 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4779 if (ret == -ERESTARTSYS)
4783 req_set_fail_links(req);
4784 __io_req_complete(req, issue_flags, ret, 0);
4788 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4789 struct io_async_msghdr *iomsg)
4791 struct io_sr_msg *sr = &req->sr_msg;
4792 struct iovec __user *uiov;
4796 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4797 &iomsg->uaddr, &uiov, &iov_len);
4801 if (req->flags & REQ_F_BUFFER_SELECT) {
4804 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4806 sr->len = iomsg->fast_iov[0].iov_len;
4807 iomsg->free_iov = NULL;
4809 iomsg->free_iov = iomsg->fast_iov;
4810 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4811 &iomsg->free_iov, &iomsg->msg.msg_iter,
4820 #ifdef CONFIG_COMPAT
4821 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4822 struct io_async_msghdr *iomsg)
4824 struct compat_msghdr __user *msg_compat;
4825 struct io_sr_msg *sr = &req->sr_msg;
4826 struct compat_iovec __user *uiov;
4831 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4832 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4837 uiov = compat_ptr(ptr);
4838 if (req->flags & REQ_F_BUFFER_SELECT) {
4839 compat_ssize_t clen;
4843 if (!access_ok(uiov, sizeof(*uiov)))
4845 if (__get_user(clen, &uiov->iov_len))
4850 iomsg->free_iov = NULL;
4852 iomsg->free_iov = iomsg->fast_iov;
4853 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4854 UIO_FASTIOV, &iomsg->free_iov,
4855 &iomsg->msg.msg_iter, true);
4864 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4865 struct io_async_msghdr *iomsg)
4867 iomsg->msg.msg_name = &iomsg->addr;
4869 #ifdef CONFIG_COMPAT
4870 if (req->ctx->compat)
4871 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4874 return __io_recvmsg_copy_hdr(req, iomsg);
4877 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4880 struct io_sr_msg *sr = &req->sr_msg;
4881 struct io_buffer *kbuf;
4883 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4888 req->flags |= REQ_F_BUFFER_SELECTED;
4892 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4894 return io_put_kbuf(req, req->sr_msg.kbuf);
4897 static int io_recvmsg_prep_async(struct io_kiocb *req)
4901 if (!io_op_defs[req->opcode].needs_async_data)
4903 ret = io_recvmsg_copy_hdr(req, req->async_data);
4905 req->flags |= REQ_F_NEED_CLEANUP;
4909 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4911 struct io_sr_msg *sr = &req->sr_msg;
4913 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4916 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4917 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4918 sr->len = READ_ONCE(sqe->len);
4919 sr->bgid = READ_ONCE(sqe->buf_group);
4921 #ifdef CONFIG_COMPAT
4922 if (req->ctx->compat)
4923 sr->msg_flags |= MSG_CMSG_COMPAT;
4928 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4930 struct io_async_msghdr iomsg, *kmsg;
4931 struct socket *sock;
4932 struct io_buffer *kbuf;
4934 int ret, cflags = 0;
4935 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4937 sock = sock_from_file(req->file);
4938 if (unlikely(!sock))
4941 kmsg = req->async_data;
4943 ret = io_recvmsg_copy_hdr(req, &iomsg);
4949 if (req->flags & REQ_F_BUFFER_SELECT) {
4950 kbuf = io_recv_buffer_select(req, !force_nonblock);
4952 return PTR_ERR(kbuf);
4953 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4954 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4955 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4956 1, req->sr_msg.len);
4959 flags = req->sr_msg.msg_flags;
4960 if (flags & MSG_DONTWAIT)
4961 req->flags |= REQ_F_NOWAIT;
4962 else if (force_nonblock)
4963 flags |= MSG_DONTWAIT;
4965 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4966 kmsg->uaddr, flags);
4967 if (force_nonblock && ret == -EAGAIN)
4968 return io_setup_async_msg(req, kmsg);
4969 if (ret == -ERESTARTSYS)
4972 if (req->flags & REQ_F_BUFFER_SELECTED)
4973 cflags = io_put_recv_kbuf(req);
4974 /* fast path, check for non-NULL to avoid function call */
4976 kfree(kmsg->free_iov);
4977 req->flags &= ~REQ_F_NEED_CLEANUP;
4979 req_set_fail_links(req);
4980 __io_req_complete(req, issue_flags, ret, cflags);
4984 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4986 struct io_buffer *kbuf;
4987 struct io_sr_msg *sr = &req->sr_msg;
4989 void __user *buf = sr->buf;
4990 struct socket *sock;
4993 int ret, cflags = 0;
4994 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4996 sock = sock_from_file(req->file);
4997 if (unlikely(!sock))
5000 if (req->flags & REQ_F_BUFFER_SELECT) {
5001 kbuf = io_recv_buffer_select(req, !force_nonblock);
5003 return PTR_ERR(kbuf);
5004 buf = u64_to_user_ptr(kbuf->addr);
5007 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5011 msg.msg_name = NULL;
5012 msg.msg_control = NULL;
5013 msg.msg_controllen = 0;
5014 msg.msg_namelen = 0;
5015 msg.msg_iocb = NULL;
5018 flags = req->sr_msg.msg_flags;
5019 if (flags & MSG_DONTWAIT)
5020 req->flags |= REQ_F_NOWAIT;
5021 else if (force_nonblock)
5022 flags |= MSG_DONTWAIT;
5024 ret = sock_recvmsg(sock, &msg, flags);
5025 if (force_nonblock && ret == -EAGAIN)
5027 if (ret == -ERESTARTSYS)
5030 if (req->flags & REQ_F_BUFFER_SELECTED)
5031 cflags = io_put_recv_kbuf(req);
5033 req_set_fail_links(req);
5034 __io_req_complete(req, issue_flags, ret, cflags);
5038 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5040 struct io_accept *accept = &req->accept;
5042 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5044 if (sqe->ioprio || sqe->len || sqe->buf_index)
5047 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5048 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5049 accept->flags = READ_ONCE(sqe->accept_flags);
5050 accept->nofile = rlimit(RLIMIT_NOFILE);
5054 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5056 struct io_accept *accept = &req->accept;
5057 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5058 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5061 if (req->file->f_flags & O_NONBLOCK)
5062 req->flags |= REQ_F_NOWAIT;
5064 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
5065 accept->addr_len, accept->flags,
5067 if (ret == -EAGAIN && force_nonblock)
5070 if (ret == -ERESTARTSYS)
5072 req_set_fail_links(req);
5074 __io_req_complete(req, issue_flags, ret, 0);
5078 static int io_connect_prep_async(struct io_kiocb *req)
5080 struct io_async_connect *io = req->async_data;
5081 struct io_connect *conn = &req->connect;
5083 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5086 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5088 struct io_connect *conn = &req->connect;
5090 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5092 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
5095 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5096 conn->addr_len = READ_ONCE(sqe->addr2);
5100 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5102 struct io_async_connect __io, *io;
5103 unsigned file_flags;
5105 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5107 if (req->async_data) {
5108 io = req->async_data;
5110 ret = move_addr_to_kernel(req->connect.addr,
5111 req->connect.addr_len,
5118 file_flags = force_nonblock ? O_NONBLOCK : 0;
5120 ret = __sys_connect_file(req->file, &io->address,
5121 req->connect.addr_len, file_flags);
5122 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5123 if (req->async_data)
5125 if (io_alloc_async_data(req)) {
5129 io = req->async_data;
5130 memcpy(req->async_data, &__io, sizeof(__io));
5133 if (ret == -ERESTARTSYS)
5137 req_set_fail_links(req);
5138 __io_req_complete(req, issue_flags, ret, 0);
5141 #else /* !CONFIG_NET */
5142 #define IO_NETOP_FN(op) \
5143 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5145 return -EOPNOTSUPP; \
5148 #define IO_NETOP_PREP(op) \
5150 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5152 return -EOPNOTSUPP; \
5155 #define IO_NETOP_PREP_ASYNC(op) \
5157 static int io_##op##_prep_async(struct io_kiocb *req) \
5159 return -EOPNOTSUPP; \
5162 IO_NETOP_PREP_ASYNC(sendmsg);
5163 IO_NETOP_PREP_ASYNC(recvmsg);
5164 IO_NETOP_PREP_ASYNC(connect);
5165 IO_NETOP_PREP(accept);
5168 #endif /* CONFIG_NET */
5170 struct io_poll_table {
5171 struct poll_table_struct pt;
5172 struct io_kiocb *req;
5176 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5177 __poll_t mask, task_work_func_t func)
5181 /* for instances that support it check for an event match first: */
5182 if (mask && !(mask & poll->events))
5185 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5187 list_del_init(&poll->wait.entry);
5190 req->task_work.func = func;
5191 percpu_ref_get(&req->ctx->refs);
5194 * If this fails, then the task is exiting. When a task exits, the
5195 * work gets canceled, so just cancel this request as well instead
5196 * of executing it. We can't safely execute it anyway, as we may not
5197 * have the needed state needed for it anyway.
5199 ret = io_req_task_work_add(req);
5200 if (unlikely(ret)) {
5201 WRITE_ONCE(poll->canceled, true);
5202 io_req_task_work_add_fallback(req, func);
5207 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5208 __acquires(&req->ctx->completion_lock)
5210 struct io_ring_ctx *ctx = req->ctx;
5212 if (!req->result && !READ_ONCE(poll->canceled)) {
5213 struct poll_table_struct pt = { ._key = poll->events };
5215 req->result = vfs_poll(req->file, &pt) & poll->events;
5218 spin_lock_irq(&ctx->completion_lock);
5219 if (!req->result && !READ_ONCE(poll->canceled)) {
5220 add_wait_queue(poll->head, &poll->wait);
5227 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5229 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5230 if (req->opcode == IORING_OP_POLL_ADD)
5231 return req->async_data;
5232 return req->apoll->double_poll;
5235 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5237 if (req->opcode == IORING_OP_POLL_ADD)
5239 return &req->apoll->poll;
5242 static void io_poll_remove_double(struct io_kiocb *req)
5244 struct io_poll_iocb *poll = io_poll_get_double(req);
5246 lockdep_assert_held(&req->ctx->completion_lock);
5248 if (poll && poll->head) {
5249 struct wait_queue_head *head = poll->head;
5251 spin_lock(&head->lock);
5252 list_del_init(&poll->wait.entry);
5253 if (poll->wait.private)
5254 refcount_dec(&req->refs);
5256 spin_unlock(&head->lock);
5260 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5262 struct io_ring_ctx *ctx = req->ctx;
5264 io_poll_remove_double(req);
5265 req->poll.done = true;
5266 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5267 io_commit_cqring(ctx);
5270 static void io_poll_task_func(struct callback_head *cb)
5272 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5273 struct io_ring_ctx *ctx = req->ctx;
5274 struct io_kiocb *nxt;
5276 if (io_poll_rewait(req, &req->poll)) {
5277 spin_unlock_irq(&ctx->completion_lock);
5279 hash_del(&req->hash_node);
5280 io_poll_complete(req, req->result, 0);
5281 spin_unlock_irq(&ctx->completion_lock);
5283 nxt = io_put_req_find_next(req);
5284 io_cqring_ev_posted(ctx);
5286 __io_req_task_submit(nxt);
5289 percpu_ref_put(&ctx->refs);
5292 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5293 int sync, void *key)
5295 struct io_kiocb *req = wait->private;
5296 struct io_poll_iocb *poll = io_poll_get_single(req);
5297 __poll_t mask = key_to_poll(key);
5299 /* for instances that support it check for an event match first: */
5300 if (mask && !(mask & poll->events))
5303 list_del_init(&wait->entry);
5305 if (poll && poll->head) {
5308 spin_lock(&poll->head->lock);
5309 done = list_empty(&poll->wait.entry);
5311 list_del_init(&poll->wait.entry);
5312 /* make sure double remove sees this as being gone */
5313 wait->private = NULL;
5314 spin_unlock(&poll->head->lock);
5316 /* use wait func handler, so it matches the rq type */
5317 poll->wait.func(&poll->wait, mode, sync, key);
5320 refcount_dec(&req->refs);
5324 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5325 wait_queue_func_t wake_func)
5329 poll->canceled = false;
5330 poll->events = events;
5331 INIT_LIST_HEAD(&poll->wait.entry);
5332 init_waitqueue_func_entry(&poll->wait, wake_func);
5335 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5336 struct wait_queue_head *head,
5337 struct io_poll_iocb **poll_ptr)
5339 struct io_kiocb *req = pt->req;
5342 * If poll->head is already set, it's because the file being polled
5343 * uses multiple waitqueues for poll handling (eg one for read, one
5344 * for write). Setup a separate io_poll_iocb if this happens.
5346 if (unlikely(poll->head)) {
5347 struct io_poll_iocb *poll_one = poll;
5349 /* already have a 2nd entry, fail a third attempt */
5351 pt->error = -EINVAL;
5354 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5356 pt->error = -ENOMEM;
5359 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5360 refcount_inc(&req->refs);
5361 poll->wait.private = req;
5368 if (poll->events & EPOLLEXCLUSIVE)
5369 add_wait_queue_exclusive(head, &poll->wait);
5371 add_wait_queue(head, &poll->wait);
5374 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5375 struct poll_table_struct *p)
5377 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5378 struct async_poll *apoll = pt->req->apoll;
5380 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5383 static void io_async_task_func(struct callback_head *cb)
5385 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5386 struct async_poll *apoll = req->apoll;
5387 struct io_ring_ctx *ctx = req->ctx;
5389 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5391 if (io_poll_rewait(req, &apoll->poll)) {
5392 spin_unlock_irq(&ctx->completion_lock);
5393 percpu_ref_put(&ctx->refs);
5397 /* If req is still hashed, it cannot have been canceled. Don't check. */
5398 if (hash_hashed(&req->hash_node))
5399 hash_del(&req->hash_node);
5401 io_poll_remove_double(req);
5402 spin_unlock_irq(&ctx->completion_lock);
5404 if (!READ_ONCE(apoll->poll.canceled))
5405 __io_req_task_submit(req);
5407 __io_req_task_cancel(req, -ECANCELED);
5409 percpu_ref_put(&ctx->refs);
5410 kfree(apoll->double_poll);
5414 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5417 struct io_kiocb *req = wait->private;
5418 struct io_poll_iocb *poll = &req->apoll->poll;
5420 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5423 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5426 static void io_poll_req_insert(struct io_kiocb *req)
5428 struct io_ring_ctx *ctx = req->ctx;
5429 struct hlist_head *list;
5431 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5432 hlist_add_head(&req->hash_node, list);
5435 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5436 struct io_poll_iocb *poll,
5437 struct io_poll_table *ipt, __poll_t mask,
5438 wait_queue_func_t wake_func)
5439 __acquires(&ctx->completion_lock)
5441 struct io_ring_ctx *ctx = req->ctx;
5442 bool cancel = false;
5444 INIT_HLIST_NODE(&req->hash_node);
5445 io_init_poll_iocb(poll, mask, wake_func);
5446 poll->file = req->file;
5447 poll->wait.private = req;
5449 ipt->pt._key = mask;
5451 ipt->error = -EINVAL;
5453 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5455 spin_lock_irq(&ctx->completion_lock);
5456 if (likely(poll->head)) {
5457 spin_lock(&poll->head->lock);
5458 if (unlikely(list_empty(&poll->wait.entry))) {
5464 if (mask || ipt->error)
5465 list_del_init(&poll->wait.entry);
5467 WRITE_ONCE(poll->canceled, true);
5468 else if (!poll->done) /* actually waiting for an event */
5469 io_poll_req_insert(req);
5470 spin_unlock(&poll->head->lock);
5476 static bool io_arm_poll_handler(struct io_kiocb *req)
5478 const struct io_op_def *def = &io_op_defs[req->opcode];
5479 struct io_ring_ctx *ctx = req->ctx;
5480 struct async_poll *apoll;
5481 struct io_poll_table ipt;
5485 if (!req->file || !file_can_poll(req->file))
5487 if (req->flags & REQ_F_POLLED)
5491 else if (def->pollout)
5495 /* if we can't nonblock try, then no point in arming a poll handler */
5496 if (!io_file_supports_async(req->file, rw))
5499 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5500 if (unlikely(!apoll))
5502 apoll->double_poll = NULL;
5504 req->flags |= REQ_F_POLLED;
5509 mask |= POLLIN | POLLRDNORM;
5511 mask |= POLLOUT | POLLWRNORM;
5513 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5514 if ((req->opcode == IORING_OP_RECVMSG) &&
5515 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5518 mask |= POLLERR | POLLPRI;
5520 ipt.pt._qproc = io_async_queue_proc;
5522 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5524 if (ret || ipt.error) {
5525 io_poll_remove_double(req);
5526 spin_unlock_irq(&ctx->completion_lock);
5527 kfree(apoll->double_poll);
5531 spin_unlock_irq(&ctx->completion_lock);
5532 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5533 apoll->poll.events);
5537 static bool __io_poll_remove_one(struct io_kiocb *req,
5538 struct io_poll_iocb *poll)
5540 bool do_complete = false;
5542 spin_lock(&poll->head->lock);
5543 WRITE_ONCE(poll->canceled, true);
5544 if (!list_empty(&poll->wait.entry)) {
5545 list_del_init(&poll->wait.entry);
5548 spin_unlock(&poll->head->lock);
5549 hash_del(&req->hash_node);
5553 static bool io_poll_remove_one(struct io_kiocb *req)
5557 io_poll_remove_double(req);
5559 if (req->opcode == IORING_OP_POLL_ADD) {
5560 do_complete = __io_poll_remove_one(req, &req->poll);
5562 struct async_poll *apoll = req->apoll;
5564 /* non-poll requests have submit ref still */
5565 do_complete = __io_poll_remove_one(req, &apoll->poll);
5568 kfree(apoll->double_poll);
5574 io_cqring_fill_event(req, -ECANCELED);
5575 io_commit_cqring(req->ctx);
5576 req_set_fail_links(req);
5577 io_put_req_deferred(req, 1);
5584 * Returns true if we found and killed one or more poll requests
5586 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5587 struct files_struct *files)
5589 struct hlist_node *tmp;
5590 struct io_kiocb *req;
5593 spin_lock_irq(&ctx->completion_lock);
5594 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5595 struct hlist_head *list;
5597 list = &ctx->cancel_hash[i];
5598 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5599 if (io_match_task(req, tsk, files))
5600 posted += io_poll_remove_one(req);
5603 spin_unlock_irq(&ctx->completion_lock);
5606 io_cqring_ev_posted(ctx);
5611 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5613 struct hlist_head *list;
5614 struct io_kiocb *req;
5616 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5617 hlist_for_each_entry(req, list, hash_node) {
5618 if (sqe_addr != req->user_data)
5620 if (io_poll_remove_one(req))
5628 static int io_poll_remove_prep(struct io_kiocb *req,
5629 const struct io_uring_sqe *sqe)
5631 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5633 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5637 req->poll_remove.addr = READ_ONCE(sqe->addr);
5642 * Find a running poll command that matches one specified in sqe->addr,
5643 * and remove it if found.
5645 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5647 struct io_ring_ctx *ctx = req->ctx;
5650 spin_lock_irq(&ctx->completion_lock);
5651 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5652 spin_unlock_irq(&ctx->completion_lock);
5655 req_set_fail_links(req);
5656 io_req_complete(req, ret);
5660 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5663 struct io_kiocb *req = wait->private;
5664 struct io_poll_iocb *poll = &req->poll;
5666 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5669 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5670 struct poll_table_struct *p)
5672 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5674 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5677 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5679 struct io_poll_iocb *poll = &req->poll;
5682 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5684 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5687 events = READ_ONCE(sqe->poll32_events);
5689 events = swahw32(events);
5691 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5692 (events & EPOLLEXCLUSIVE);
5696 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5698 struct io_poll_iocb *poll = &req->poll;
5699 struct io_ring_ctx *ctx = req->ctx;
5700 struct io_poll_table ipt;
5703 ipt.pt._qproc = io_poll_queue_proc;
5705 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5708 if (mask) { /* no async, we'd stolen it */
5710 io_poll_complete(req, mask, 0);
5712 spin_unlock_irq(&ctx->completion_lock);
5715 io_cqring_ev_posted(ctx);
5721 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5723 struct io_timeout_data *data = container_of(timer,
5724 struct io_timeout_data, timer);
5725 struct io_kiocb *req = data->req;
5726 struct io_ring_ctx *ctx = req->ctx;
5727 unsigned long flags;
5729 spin_lock_irqsave(&ctx->completion_lock, flags);
5730 list_del_init(&req->timeout.list);
5731 atomic_set(&req->ctx->cq_timeouts,
5732 atomic_read(&req->ctx->cq_timeouts) + 1);
5734 io_cqring_fill_event(req, -ETIME);
5735 io_commit_cqring(ctx);
5736 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5738 io_cqring_ev_posted(ctx);
5739 req_set_fail_links(req);
5741 return HRTIMER_NORESTART;
5744 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5747 struct io_timeout_data *io;
5748 struct io_kiocb *req;
5751 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5752 if (user_data == req->user_data) {
5759 return ERR_PTR(ret);
5761 io = req->async_data;
5762 ret = hrtimer_try_to_cancel(&io->timer);
5764 return ERR_PTR(-EALREADY);
5765 list_del_init(&req->timeout.list);
5769 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5771 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5774 return PTR_ERR(req);
5776 req_set_fail_links(req);
5777 io_cqring_fill_event(req, -ECANCELED);
5778 io_put_req_deferred(req, 1);
5782 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5783 struct timespec64 *ts, enum hrtimer_mode mode)
5785 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5786 struct io_timeout_data *data;
5789 return PTR_ERR(req);
5791 req->timeout.off = 0; /* noseq */
5792 data = req->async_data;
5793 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5794 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5795 data->timer.function = io_timeout_fn;
5796 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5800 static int io_timeout_remove_prep(struct io_kiocb *req,
5801 const struct io_uring_sqe *sqe)
5803 struct io_timeout_rem *tr = &req->timeout_rem;
5805 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5807 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5809 if (sqe->ioprio || sqe->buf_index || sqe->len)
5812 tr->addr = READ_ONCE(sqe->addr);
5813 tr->flags = READ_ONCE(sqe->timeout_flags);
5814 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5815 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5817 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5819 } else if (tr->flags) {
5820 /* timeout removal doesn't support flags */
5827 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5829 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5834 * Remove or update an existing timeout command
5836 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5838 struct io_timeout_rem *tr = &req->timeout_rem;
5839 struct io_ring_ctx *ctx = req->ctx;
5842 spin_lock_irq(&ctx->completion_lock);
5843 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5844 ret = io_timeout_cancel(ctx, tr->addr);
5846 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5847 io_translate_timeout_mode(tr->flags));
5849 io_cqring_fill_event(req, ret);
5850 io_commit_cqring(ctx);
5851 spin_unlock_irq(&ctx->completion_lock);
5852 io_cqring_ev_posted(ctx);
5854 req_set_fail_links(req);
5859 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5860 bool is_timeout_link)
5862 struct io_timeout_data *data;
5864 u32 off = READ_ONCE(sqe->off);
5866 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5868 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5870 if (off && is_timeout_link)
5872 flags = READ_ONCE(sqe->timeout_flags);
5873 if (flags & ~IORING_TIMEOUT_ABS)
5876 req->timeout.off = off;
5878 if (!req->async_data && io_alloc_async_data(req))
5881 data = req->async_data;
5884 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5887 data->mode = io_translate_timeout_mode(flags);
5888 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5892 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5894 struct io_ring_ctx *ctx = req->ctx;
5895 struct io_timeout_data *data = req->async_data;
5896 struct list_head *entry;
5897 u32 tail, off = req->timeout.off;
5899 spin_lock_irq(&ctx->completion_lock);
5902 * sqe->off holds how many events that need to occur for this
5903 * timeout event to be satisfied. If it isn't set, then this is
5904 * a pure timeout request, sequence isn't used.
5906 if (io_is_timeout_noseq(req)) {
5907 entry = ctx->timeout_list.prev;
5911 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5912 req->timeout.target_seq = tail + off;
5914 /* Update the last seq here in case io_flush_timeouts() hasn't.
5915 * This is safe because ->completion_lock is held, and submissions
5916 * and completions are never mixed in the same ->completion_lock section.
5918 ctx->cq_last_tm_flush = tail;
5921 * Insertion sort, ensuring the first entry in the list is always
5922 * the one we need first.
5924 list_for_each_prev(entry, &ctx->timeout_list) {
5925 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5928 if (io_is_timeout_noseq(nxt))
5930 /* nxt.seq is behind @tail, otherwise would've been completed */
5931 if (off >= nxt->timeout.target_seq - tail)
5935 list_add(&req->timeout.list, entry);
5936 data->timer.function = io_timeout_fn;
5937 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5938 spin_unlock_irq(&ctx->completion_lock);
5942 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5944 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5946 return req->user_data == (unsigned long) data;
5949 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5951 enum io_wq_cancel cancel_ret;
5954 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5955 switch (cancel_ret) {
5956 case IO_WQ_CANCEL_OK:
5959 case IO_WQ_CANCEL_RUNNING:
5962 case IO_WQ_CANCEL_NOTFOUND:
5970 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5971 struct io_kiocb *req, __u64 sqe_addr,
5974 unsigned long flags;
5977 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5978 if (ret != -ENOENT) {
5979 spin_lock_irqsave(&ctx->completion_lock, flags);
5983 spin_lock_irqsave(&ctx->completion_lock, flags);
5984 ret = io_timeout_cancel(ctx, sqe_addr);
5987 ret = io_poll_cancel(ctx, sqe_addr);
5991 io_cqring_fill_event(req, ret);
5992 io_commit_cqring(ctx);
5993 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5994 io_cqring_ev_posted(ctx);
5997 req_set_fail_links(req);
6001 static int io_async_cancel_prep(struct io_kiocb *req,
6002 const struct io_uring_sqe *sqe)
6004 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6006 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6008 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
6011 req->cancel.addr = READ_ONCE(sqe->addr);
6015 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6017 struct io_ring_ctx *ctx = req->ctx;
6019 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
6023 static int io_rsrc_update_prep(struct io_kiocb *req,
6024 const struct io_uring_sqe *sqe)
6026 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
6028 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6030 if (sqe->ioprio || sqe->rw_flags)
6033 req->rsrc_update.offset = READ_ONCE(sqe->off);
6034 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6035 if (!req->rsrc_update.nr_args)
6037 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6041 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6043 struct io_ring_ctx *ctx = req->ctx;
6044 struct io_uring_rsrc_update up;
6047 if (issue_flags & IO_URING_F_NONBLOCK)
6050 up.offset = req->rsrc_update.offset;
6051 up.data = req->rsrc_update.arg;
6053 mutex_lock(&ctx->uring_lock);
6054 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
6055 mutex_unlock(&ctx->uring_lock);
6058 req_set_fail_links(req);
6059 __io_req_complete(req, issue_flags, ret, 0);
6063 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6065 switch (req->opcode) {
6068 case IORING_OP_READV:
6069 case IORING_OP_READ_FIXED:
6070 case IORING_OP_READ:
6071 return io_read_prep(req, sqe);
6072 case IORING_OP_WRITEV:
6073 case IORING_OP_WRITE_FIXED:
6074 case IORING_OP_WRITE:
6075 return io_write_prep(req, sqe);
6076 case IORING_OP_POLL_ADD:
6077 return io_poll_add_prep(req, sqe);
6078 case IORING_OP_POLL_REMOVE:
6079 return io_poll_remove_prep(req, sqe);
6080 case IORING_OP_FSYNC:
6081 return io_fsync_prep(req, sqe);
6082 case IORING_OP_SYNC_FILE_RANGE:
6083 return io_sfr_prep(req, sqe);
6084 case IORING_OP_SENDMSG:
6085 case IORING_OP_SEND:
6086 return io_sendmsg_prep(req, sqe);
6087 case IORING_OP_RECVMSG:
6088 case IORING_OP_RECV:
6089 return io_recvmsg_prep(req, sqe);
6090 case IORING_OP_CONNECT:
6091 return io_connect_prep(req, sqe);
6092 case IORING_OP_TIMEOUT:
6093 return io_timeout_prep(req, sqe, false);
6094 case IORING_OP_TIMEOUT_REMOVE:
6095 return io_timeout_remove_prep(req, sqe);
6096 case IORING_OP_ASYNC_CANCEL:
6097 return io_async_cancel_prep(req, sqe);
6098 case IORING_OP_LINK_TIMEOUT:
6099 return io_timeout_prep(req, sqe, true);
6100 case IORING_OP_ACCEPT:
6101 return io_accept_prep(req, sqe);
6102 case IORING_OP_FALLOCATE:
6103 return io_fallocate_prep(req, sqe);
6104 case IORING_OP_OPENAT:
6105 return io_openat_prep(req, sqe);
6106 case IORING_OP_CLOSE:
6107 return io_close_prep(req, sqe);
6108 case IORING_OP_FILES_UPDATE:
6109 return io_rsrc_update_prep(req, sqe);
6110 case IORING_OP_STATX:
6111 return io_statx_prep(req, sqe);
6112 case IORING_OP_FADVISE:
6113 return io_fadvise_prep(req, sqe);
6114 case IORING_OP_MADVISE:
6115 return io_madvise_prep(req, sqe);
6116 case IORING_OP_OPENAT2:
6117 return io_openat2_prep(req, sqe);
6118 case IORING_OP_EPOLL_CTL:
6119 return io_epoll_ctl_prep(req, sqe);
6120 case IORING_OP_SPLICE:
6121 return io_splice_prep(req, sqe);
6122 case IORING_OP_PROVIDE_BUFFERS:
6123 return io_provide_buffers_prep(req, sqe);
6124 case IORING_OP_REMOVE_BUFFERS:
6125 return io_remove_buffers_prep(req, sqe);
6127 return io_tee_prep(req, sqe);
6128 case IORING_OP_SHUTDOWN:
6129 return io_shutdown_prep(req, sqe);
6130 case IORING_OP_RENAMEAT:
6131 return io_renameat_prep(req, sqe);
6132 case IORING_OP_UNLINKAT:
6133 return io_unlinkat_prep(req, sqe);
6136 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6141 static int io_req_prep_async(struct io_kiocb *req)
6143 switch (req->opcode) {
6144 case IORING_OP_READV:
6145 case IORING_OP_READ_FIXED:
6146 case IORING_OP_READ:
6147 return io_rw_prep_async(req, READ);
6148 case IORING_OP_WRITEV:
6149 case IORING_OP_WRITE_FIXED:
6150 case IORING_OP_WRITE:
6151 return io_rw_prep_async(req, WRITE);
6152 case IORING_OP_SENDMSG:
6153 case IORING_OP_SEND:
6154 return io_sendmsg_prep_async(req);
6155 case IORING_OP_RECVMSG:
6156 case IORING_OP_RECV:
6157 return io_recvmsg_prep_async(req);
6158 case IORING_OP_CONNECT:
6159 return io_connect_prep_async(req);
6164 static int io_req_defer_prep(struct io_kiocb *req)
6166 if (!io_op_defs[req->opcode].needs_async_data)
6168 /* some opcodes init it during the inital prep */
6169 if (req->async_data)
6171 if (__io_alloc_async_data(req))
6173 return io_req_prep_async(req);
6176 static u32 io_get_sequence(struct io_kiocb *req)
6178 struct io_kiocb *pos;
6179 struct io_ring_ctx *ctx = req->ctx;
6180 u32 total_submitted, nr_reqs = 0;
6182 io_for_each_link(pos, req)
6185 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6186 return total_submitted - nr_reqs;
6189 static int io_req_defer(struct io_kiocb *req)
6191 struct io_ring_ctx *ctx = req->ctx;
6192 struct io_defer_entry *de;
6196 /* Still need defer if there is pending req in defer list. */
6197 if (likely(list_empty_careful(&ctx->defer_list) &&
6198 !(req->flags & REQ_F_IO_DRAIN)))
6201 seq = io_get_sequence(req);
6202 /* Still a chance to pass the sequence check */
6203 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6206 ret = io_req_defer_prep(req);
6209 io_prep_async_link(req);
6210 de = kmalloc(sizeof(*de), GFP_KERNEL);
6214 spin_lock_irq(&ctx->completion_lock);
6215 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6216 spin_unlock_irq(&ctx->completion_lock);
6218 io_queue_async_work(req);
6219 return -EIOCBQUEUED;
6222 trace_io_uring_defer(ctx, req, req->user_data);
6225 list_add_tail(&de->list, &ctx->defer_list);
6226 spin_unlock_irq(&ctx->completion_lock);
6227 return -EIOCBQUEUED;
6230 static void __io_clean_op(struct io_kiocb *req)
6232 if (req->flags & REQ_F_BUFFER_SELECTED) {
6233 switch (req->opcode) {
6234 case IORING_OP_READV:
6235 case IORING_OP_READ_FIXED:
6236 case IORING_OP_READ:
6237 kfree((void *)(unsigned long)req->rw.addr);
6239 case IORING_OP_RECVMSG:
6240 case IORING_OP_RECV:
6241 kfree(req->sr_msg.kbuf);
6244 req->flags &= ~REQ_F_BUFFER_SELECTED;
6247 if (req->flags & REQ_F_NEED_CLEANUP) {
6248 switch (req->opcode) {
6249 case IORING_OP_READV:
6250 case IORING_OP_READ_FIXED:
6251 case IORING_OP_READ:
6252 case IORING_OP_WRITEV:
6253 case IORING_OP_WRITE_FIXED:
6254 case IORING_OP_WRITE: {
6255 struct io_async_rw *io = req->async_data;
6257 kfree(io->free_iovec);
6260 case IORING_OP_RECVMSG:
6261 case IORING_OP_SENDMSG: {
6262 struct io_async_msghdr *io = req->async_data;
6264 kfree(io->free_iov);
6267 case IORING_OP_SPLICE:
6269 io_put_file(req, req->splice.file_in,
6270 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6272 case IORING_OP_OPENAT:
6273 case IORING_OP_OPENAT2:
6274 if (req->open.filename)
6275 putname(req->open.filename);
6277 case IORING_OP_RENAMEAT:
6278 putname(req->rename.oldpath);
6279 putname(req->rename.newpath);
6281 case IORING_OP_UNLINKAT:
6282 putname(req->unlink.filename);
6285 req->flags &= ~REQ_F_NEED_CLEANUP;
6289 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6291 struct io_ring_ctx *ctx = req->ctx;
6294 switch (req->opcode) {
6296 ret = io_nop(req, issue_flags);
6298 case IORING_OP_READV:
6299 case IORING_OP_READ_FIXED:
6300 case IORING_OP_READ:
6301 ret = io_read(req, issue_flags);
6303 case IORING_OP_WRITEV:
6304 case IORING_OP_WRITE_FIXED:
6305 case IORING_OP_WRITE:
6306 ret = io_write(req, issue_flags);
6308 case IORING_OP_FSYNC:
6309 ret = io_fsync(req, issue_flags);
6311 case IORING_OP_POLL_ADD:
6312 ret = io_poll_add(req, issue_flags);
6314 case IORING_OP_POLL_REMOVE:
6315 ret = io_poll_remove(req, issue_flags);
6317 case IORING_OP_SYNC_FILE_RANGE:
6318 ret = io_sync_file_range(req, issue_flags);
6320 case IORING_OP_SENDMSG:
6321 ret = io_sendmsg(req, issue_flags);
6323 case IORING_OP_SEND:
6324 ret = io_send(req, issue_flags);
6326 case IORING_OP_RECVMSG:
6327 ret = io_recvmsg(req, issue_flags);
6329 case IORING_OP_RECV:
6330 ret = io_recv(req, issue_flags);
6332 case IORING_OP_TIMEOUT:
6333 ret = io_timeout(req, issue_flags);
6335 case IORING_OP_TIMEOUT_REMOVE:
6336 ret = io_timeout_remove(req, issue_flags);
6338 case IORING_OP_ACCEPT:
6339 ret = io_accept(req, issue_flags);
6341 case IORING_OP_CONNECT:
6342 ret = io_connect(req, issue_flags);
6344 case IORING_OP_ASYNC_CANCEL:
6345 ret = io_async_cancel(req, issue_flags);
6347 case IORING_OP_FALLOCATE:
6348 ret = io_fallocate(req, issue_flags);
6350 case IORING_OP_OPENAT:
6351 ret = io_openat(req, issue_flags);
6353 case IORING_OP_CLOSE:
6354 ret = io_close(req, issue_flags);
6356 case IORING_OP_FILES_UPDATE:
6357 ret = io_files_update(req, issue_flags);
6359 case IORING_OP_STATX:
6360 ret = io_statx(req, issue_flags);
6362 case IORING_OP_FADVISE:
6363 ret = io_fadvise(req, issue_flags);
6365 case IORING_OP_MADVISE:
6366 ret = io_madvise(req, issue_flags);
6368 case IORING_OP_OPENAT2:
6369 ret = io_openat2(req, issue_flags);
6371 case IORING_OP_EPOLL_CTL:
6372 ret = io_epoll_ctl(req, issue_flags);
6374 case IORING_OP_SPLICE:
6375 ret = io_splice(req, issue_flags);
6377 case IORING_OP_PROVIDE_BUFFERS:
6378 ret = io_provide_buffers(req, issue_flags);
6380 case IORING_OP_REMOVE_BUFFERS:
6381 ret = io_remove_buffers(req, issue_flags);
6384 ret = io_tee(req, issue_flags);
6386 case IORING_OP_SHUTDOWN:
6387 ret = io_shutdown(req, issue_flags);
6389 case IORING_OP_RENAMEAT:
6390 ret = io_renameat(req, issue_flags);
6392 case IORING_OP_UNLINKAT:
6393 ret = io_unlinkat(req, issue_flags);
6403 /* If the op doesn't have a file, we're not polling for it */
6404 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6405 const bool in_async = io_wq_current_is_worker();
6407 /* workqueue context doesn't hold uring_lock, grab it now */
6409 mutex_lock(&ctx->uring_lock);
6411 io_iopoll_req_issued(req, in_async);
6414 mutex_unlock(&ctx->uring_lock);
6420 static void io_wq_submit_work(struct io_wq_work *work)
6422 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6423 struct io_kiocb *timeout;
6426 timeout = io_prep_linked_timeout(req);
6428 io_queue_linked_timeout(timeout);
6430 if (work->flags & IO_WQ_WORK_CANCEL)
6435 ret = io_issue_sqe(req, 0);
6437 * We can get EAGAIN for polled IO even though we're
6438 * forcing a sync submission from here, since we can't
6439 * wait for request slots on the block side.
6447 /* avoid locking problems by failing it from a clean context */
6449 /* io-wq is going to take one down */
6450 refcount_inc(&req->refs);
6451 io_req_task_queue_fail(req, ret);
6455 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6458 struct fixed_rsrc_table *table;
6460 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6461 return table->files[index & IORING_FILE_TABLE_MASK];
6464 static struct file *io_file_get(struct io_submit_state *state,
6465 struct io_kiocb *req, int fd, bool fixed)
6467 struct io_ring_ctx *ctx = req->ctx;
6471 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6473 fd = array_index_nospec(fd, ctx->nr_user_files);
6474 file = io_file_from_index(ctx, fd);
6475 io_set_resource_node(req);
6477 trace_io_uring_file_get(ctx, fd);
6478 file = __io_file_get(state, fd);
6481 if (file && unlikely(file->f_op == &io_uring_fops))
6482 io_req_track_inflight(req);
6486 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6488 struct io_timeout_data *data = container_of(timer,
6489 struct io_timeout_data, timer);
6490 struct io_kiocb *prev, *req = data->req;
6491 struct io_ring_ctx *ctx = req->ctx;
6492 unsigned long flags;
6494 spin_lock_irqsave(&ctx->completion_lock, flags);
6495 prev = req->timeout.head;
6496 req->timeout.head = NULL;
6499 * We don't expect the list to be empty, that will only happen if we
6500 * race with the completion of the linked work.
6502 if (prev && refcount_inc_not_zero(&prev->refs))
6503 io_remove_next_linked(prev);
6506 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6509 req_set_fail_links(prev);
6510 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6511 io_put_req_deferred(prev, 1);
6513 io_req_complete_post(req, -ETIME, 0);
6514 io_put_req_deferred(req, 1);
6516 return HRTIMER_NORESTART;
6519 static void __io_queue_linked_timeout(struct io_kiocb *req)
6522 * If the back reference is NULL, then our linked request finished
6523 * before we got a chance to setup the timer
6525 if (req->timeout.head) {
6526 struct io_timeout_data *data = req->async_data;
6528 data->timer.function = io_link_timeout_fn;
6529 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6534 static void io_queue_linked_timeout(struct io_kiocb *req)
6536 struct io_ring_ctx *ctx = req->ctx;
6538 spin_lock_irq(&ctx->completion_lock);
6539 __io_queue_linked_timeout(req);
6540 spin_unlock_irq(&ctx->completion_lock);
6542 /* drop submission reference */
6546 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6548 struct io_kiocb *nxt = req->link;
6550 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6551 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6554 nxt->timeout.head = req;
6555 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6556 req->flags |= REQ_F_LINK_TIMEOUT;
6560 static void __io_queue_sqe(struct io_kiocb *req)
6562 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6563 const struct cred *old_creds = NULL;
6566 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6567 (req->work.flags & IO_WQ_WORK_CREDS) &&
6568 req->work.identity->creds != current_cred())
6569 old_creds = override_creds(req->work.identity->creds);
6571 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6574 revert_creds(old_creds);
6577 * We async punt it if the file wasn't marked NOWAIT, or if the file
6578 * doesn't support non-blocking read/write attempts
6580 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6581 if (!io_arm_poll_handler(req)) {
6583 * Queued up for async execution, worker will release
6584 * submit reference when the iocb is actually submitted.
6586 io_queue_async_work(req);
6588 } else if (likely(!ret)) {
6589 /* drop submission reference */
6590 if (req->flags & REQ_F_COMPLETE_INLINE) {
6591 struct io_ring_ctx *ctx = req->ctx;
6592 struct io_comp_state *cs = &ctx->submit_state.comp;
6594 cs->reqs[cs->nr++] = req;
6595 if (cs->nr == ARRAY_SIZE(cs->reqs))
6596 io_submit_flush_completions(cs, ctx);
6601 req_set_fail_links(req);
6603 io_req_complete(req, ret);
6606 io_queue_linked_timeout(linked_timeout);
6609 static void io_queue_sqe(struct io_kiocb *req)
6613 ret = io_req_defer(req);
6615 if (ret != -EIOCBQUEUED) {
6617 req_set_fail_links(req);
6619 io_req_complete(req, ret);
6621 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6622 ret = io_req_defer_prep(req);
6625 io_queue_async_work(req);
6627 __io_queue_sqe(req);
6632 * Check SQE restrictions (opcode and flags).
6634 * Returns 'true' if SQE is allowed, 'false' otherwise.
6636 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6637 struct io_kiocb *req,
6638 unsigned int sqe_flags)
6640 if (!ctx->restricted)
6643 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6646 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6647 ctx->restrictions.sqe_flags_required)
6650 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6651 ctx->restrictions.sqe_flags_required))
6657 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6658 const struct io_uring_sqe *sqe)
6660 struct io_submit_state *state;
6661 unsigned int sqe_flags;
6664 req->opcode = READ_ONCE(sqe->opcode);
6665 /* same numerical values with corresponding REQ_F_*, safe to copy */
6666 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6667 req->user_data = READ_ONCE(sqe->user_data);
6668 req->async_data = NULL;
6672 req->fixed_rsrc_refs = NULL;
6673 /* one is dropped after submission, the other at completion */
6674 refcount_set(&req->refs, 2);
6675 req->task = current;
6678 /* enforce forwards compatibility on users */
6679 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6684 if (unlikely(req->opcode >= IORING_OP_LAST))
6687 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6690 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6693 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6694 !io_op_defs[req->opcode].buffer_select)
6697 id = READ_ONCE(sqe->personality);
6699 struct io_identity *iod;
6701 iod = idr_find(&ctx->personality_idr, id);
6704 refcount_inc(&iod->count);
6706 __io_req_init_async(req);
6707 get_cred(iod->creds);
6708 req->work.identity = iod;
6709 req->work.flags |= IO_WQ_WORK_CREDS;
6712 state = &ctx->submit_state;
6715 * Plug now if we have more than 1 IO left after this, and the target
6716 * is potentially a read/write to block based storage.
6718 if (!state->plug_started && state->ios_left > 1 &&
6719 io_op_defs[req->opcode].plug) {
6720 blk_start_plug(&state->plug);
6721 state->plug_started = true;
6724 if (io_op_defs[req->opcode].needs_file) {
6725 bool fixed = req->flags & REQ_F_FIXED_FILE;
6727 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6728 if (unlikely(!req->file))
6736 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6737 const struct io_uring_sqe *sqe)
6739 struct io_submit_link *link = &ctx->submit_state.link;
6742 ret = io_init_req(ctx, req, sqe);
6743 if (unlikely(ret)) {
6746 io_req_complete(req, ret);
6748 /* fail even hard links since we don't submit */
6749 link->head->flags |= REQ_F_FAIL_LINK;
6750 io_put_req(link->head);
6751 io_req_complete(link->head, -ECANCELED);
6756 ret = io_req_prep(req, sqe);
6760 /* don't need @sqe from now on */
6761 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6762 true, ctx->flags & IORING_SETUP_SQPOLL);
6765 * If we already have a head request, queue this one for async
6766 * submittal once the head completes. If we don't have a head but
6767 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6768 * submitted sync once the chain is complete. If none of those
6769 * conditions are true (normal request), then just queue it.
6772 struct io_kiocb *head = link->head;
6775 * Taking sequential execution of a link, draining both sides
6776 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6777 * requests in the link. So, it drains the head and the
6778 * next after the link request. The last one is done via
6779 * drain_next flag to persist the effect across calls.
6781 if (req->flags & REQ_F_IO_DRAIN) {
6782 head->flags |= REQ_F_IO_DRAIN;
6783 ctx->drain_next = 1;
6785 ret = io_req_defer_prep(req);
6788 trace_io_uring_link(ctx, req, head);
6789 link->last->link = req;
6792 /* last request of a link, enqueue the link */
6793 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6798 if (unlikely(ctx->drain_next)) {
6799 req->flags |= REQ_F_IO_DRAIN;
6800 ctx->drain_next = 0;
6802 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6814 * Batched submission is done, ensure local IO is flushed out.
6816 static void io_submit_state_end(struct io_submit_state *state,
6817 struct io_ring_ctx *ctx)
6819 if (state->link.head)
6820 io_queue_sqe(state->link.head);
6822 io_submit_flush_completions(&state->comp, ctx);
6823 if (state->plug_started)
6824 blk_finish_plug(&state->plug);
6825 io_state_file_put(state);
6829 * Start submission side cache.
6831 static void io_submit_state_start(struct io_submit_state *state,
6832 unsigned int max_ios)
6834 state->plug_started = false;
6835 state->ios_left = max_ios;
6836 /* set only head, no need to init link_last in advance */
6837 state->link.head = NULL;
6840 static void io_commit_sqring(struct io_ring_ctx *ctx)
6842 struct io_rings *rings = ctx->rings;
6845 * Ensure any loads from the SQEs are done at this point,
6846 * since once we write the new head, the application could
6847 * write new data to them.
6849 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6853 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6854 * that is mapped by userspace. This means that care needs to be taken to
6855 * ensure that reads are stable, as we cannot rely on userspace always
6856 * being a good citizen. If members of the sqe are validated and then later
6857 * used, it's important that those reads are done through READ_ONCE() to
6858 * prevent a re-load down the line.
6860 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6862 u32 *sq_array = ctx->sq_array;
6866 * The cached sq head (or cq tail) serves two purposes:
6868 * 1) allows us to batch the cost of updating the user visible
6870 * 2) allows the kernel side to track the head on its own, even
6871 * though the application is the one updating it.
6873 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6874 if (likely(head < ctx->sq_entries))
6875 return &ctx->sq_sqes[head];
6877 /* drop invalid entries */
6878 ctx->cached_sq_dropped++;
6879 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6883 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6887 /* if we have a backlog and couldn't flush it all, return BUSY */
6888 if (test_bit(0, &ctx->sq_check_overflow)) {
6889 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6893 /* make sure SQ entry isn't read before tail */
6894 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6896 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6899 percpu_counter_add(¤t->io_uring->inflight, nr);
6900 refcount_add(nr, ¤t->usage);
6901 io_submit_state_start(&ctx->submit_state, nr);
6903 while (submitted < nr) {
6904 const struct io_uring_sqe *sqe;
6905 struct io_kiocb *req;
6907 req = io_alloc_req(ctx);
6908 if (unlikely(!req)) {
6910 submitted = -EAGAIN;
6913 sqe = io_get_sqe(ctx);
6914 if (unlikely(!sqe)) {
6915 kmem_cache_free(req_cachep, req);
6918 /* will complete beyond this point, count as submitted */
6920 if (io_submit_sqe(ctx, req, sqe))
6924 if (unlikely(submitted != nr)) {
6925 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6926 struct io_uring_task *tctx = current->io_uring;
6927 int unused = nr - ref_used;
6929 percpu_ref_put_many(&ctx->refs, unused);
6930 percpu_counter_sub(&tctx->inflight, unused);
6931 put_task_struct_many(current, unused);
6934 io_submit_state_end(&ctx->submit_state, ctx);
6935 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6936 io_commit_sqring(ctx);
6941 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6943 /* Tell userspace we may need a wakeup call */
6944 spin_lock_irq(&ctx->completion_lock);
6945 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6946 spin_unlock_irq(&ctx->completion_lock);
6949 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6951 spin_lock_irq(&ctx->completion_lock);
6952 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6953 spin_unlock_irq(&ctx->completion_lock);
6956 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6958 unsigned int to_submit;
6961 to_submit = io_sqring_entries(ctx);
6962 /* if we're handling multiple rings, cap submit size for fairness */
6963 if (cap_entries && to_submit > 8)
6966 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6967 unsigned nr_events = 0;
6969 mutex_lock(&ctx->uring_lock);
6970 if (!list_empty(&ctx->iopoll_list))
6971 io_do_iopoll(ctx, &nr_events, 0);
6973 if (to_submit && !ctx->sqo_dead &&
6974 likely(!percpu_ref_is_dying(&ctx->refs)))
6975 ret = io_submit_sqes(ctx, to_submit);
6976 mutex_unlock(&ctx->uring_lock);
6979 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6980 wake_up(&ctx->sqo_sq_wait);
6985 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6987 struct io_ring_ctx *ctx;
6988 unsigned sq_thread_idle = 0;
6990 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6991 if (sq_thread_idle < ctx->sq_thread_idle)
6992 sq_thread_idle = ctx->sq_thread_idle;
6995 sqd->sq_thread_idle = sq_thread_idle;
6998 static void io_sqd_init_new(struct io_sq_data *sqd)
7000 struct io_ring_ctx *ctx;
7002 while (!list_empty(&sqd->ctx_new_list)) {
7003 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
7004 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
7005 complete(&ctx->sq_thread_comp);
7008 io_sqd_update_thread_idle(sqd);
7011 static int io_sq_thread(void *data)
7013 struct cgroup_subsys_state *cur_css = NULL;
7014 struct files_struct *old_files = current->files;
7015 struct nsproxy *old_nsproxy = current->nsproxy;
7016 const struct cred *old_cred = NULL;
7017 struct io_sq_data *sqd = data;
7018 struct io_ring_ctx *ctx;
7019 unsigned long timeout = 0;
7023 current->files = NULL;
7024 current->nsproxy = NULL;
7025 task_unlock(current);
7027 while (!kthread_should_stop()) {
7029 bool cap_entries, sqt_spin, needs_sched;
7032 * Any changes to the sqd lists are synchronized through the
7033 * kthread parking. This synchronizes the thread vs users,
7034 * the users are synchronized on the sqd->ctx_lock.
7036 if (kthread_should_park()) {
7039 * When sq thread is unparked, in case the previous park operation
7040 * comes from io_put_sq_data(), which means that sq thread is going
7041 * to be stopped, so here needs to have a check.
7043 if (kthread_should_stop())
7047 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7048 io_sqd_init_new(sqd);
7049 timeout = jiffies + sqd->sq_thread_idle;
7053 cap_entries = !list_is_singular(&sqd->ctx_list);
7054 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7055 if (current->cred != ctx->creds) {
7057 revert_creds(old_cred);
7058 old_cred = override_creds(ctx->creds);
7060 io_sq_thread_associate_blkcg(ctx, &cur_css);
7062 current->loginuid = ctx->loginuid;
7063 current->sessionid = ctx->sessionid;
7066 ret = __io_sq_thread(ctx, cap_entries);
7067 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7070 io_sq_thread_drop_mm_files();
7073 if (sqt_spin || !time_after(jiffies, timeout)) {
7075 io_sq_thread_drop_mm_files();
7078 timeout = jiffies + sqd->sq_thread_idle;
7083 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7084 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7085 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7086 !list_empty_careful(&ctx->iopoll_list)) {
7087 needs_sched = false;
7090 if (io_sqring_entries(ctx)) {
7091 needs_sched = false;
7096 if (needs_sched && !kthread_should_park()) {
7097 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7098 io_ring_set_wakeup_flag(ctx);
7101 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7102 io_ring_clear_wakeup_flag(ctx);
7105 finish_wait(&sqd->wait, &wait);
7106 timeout = jiffies + sqd->sq_thread_idle;
7110 io_sq_thread_drop_mm_files();
7113 io_sq_thread_unassociate_blkcg();
7115 revert_creds(old_cred);
7118 current->files = old_files;
7119 current->nsproxy = old_nsproxy;
7120 task_unlock(current);
7127 struct io_wait_queue {
7128 struct wait_queue_entry wq;
7129 struct io_ring_ctx *ctx;
7131 unsigned nr_timeouts;
7134 static inline bool io_should_wake(struct io_wait_queue *iowq)
7136 struct io_ring_ctx *ctx = iowq->ctx;
7139 * Wake up if we have enough events, or if a timeout occurred since we
7140 * started waiting. For timeouts, we always want to return to userspace,
7141 * regardless of event count.
7143 return io_cqring_events(ctx) >= iowq->to_wait ||
7144 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7147 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7148 int wake_flags, void *key)
7150 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7154 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7155 * the task, and the next invocation will do it.
7157 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
7158 return autoremove_wake_function(curr, mode, wake_flags, key);
7162 static int io_run_task_work_sig(void)
7164 if (io_run_task_work())
7166 if (!signal_pending(current))
7168 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7169 return -ERESTARTSYS;
7173 /* when returns >0, the caller should retry */
7174 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7175 struct io_wait_queue *iowq,
7176 signed long *timeout)
7180 /* make sure we run task_work before checking for signals */
7181 ret = io_run_task_work_sig();
7182 if (ret || io_should_wake(iowq))
7184 /* let the caller flush overflows, retry */
7185 if (test_bit(0, &ctx->cq_check_overflow))
7188 *timeout = schedule_timeout(*timeout);
7189 return !*timeout ? -ETIME : 1;
7193 * Wait until events become available, if we don't already have some. The
7194 * application must reap them itself, as they reside on the shared cq ring.
7196 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7197 const sigset_t __user *sig, size_t sigsz,
7198 struct __kernel_timespec __user *uts)
7200 struct io_wait_queue iowq = {
7203 .func = io_wake_function,
7204 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7207 .to_wait = min_events,
7209 struct io_rings *rings = ctx->rings;
7210 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7214 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7215 if (io_cqring_events(ctx) >= min_events)
7217 if (!io_run_task_work())
7222 #ifdef CONFIG_COMPAT
7223 if (in_compat_syscall())
7224 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7228 ret = set_user_sigmask(sig, sigsz);
7235 struct timespec64 ts;
7237 if (get_timespec64(&ts, uts))
7239 timeout = timespec64_to_jiffies(&ts);
7242 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7243 trace_io_uring_cqring_wait(ctx, min_events);
7245 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7246 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7247 TASK_INTERRUPTIBLE);
7248 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7249 finish_wait(&ctx->wait, &iowq.wq);
7252 restore_saved_sigmask_unless(ret == -EINTR);
7254 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7257 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7259 #if defined(CONFIG_UNIX)
7260 if (ctx->ring_sock) {
7261 struct sock *sock = ctx->ring_sock->sk;
7262 struct sk_buff *skb;
7264 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7270 for (i = 0; i < ctx->nr_user_files; i++) {
7273 file = io_file_from_index(ctx, i);
7280 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
7282 struct fixed_rsrc_data *data;
7284 data = container_of(ref, struct fixed_rsrc_data, refs);
7285 complete(&data->done);
7288 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7290 spin_lock_bh(&ctx->rsrc_ref_lock);
7293 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7295 spin_unlock_bh(&ctx->rsrc_ref_lock);
7298 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7299 struct fixed_rsrc_data *rsrc_data,
7300 struct fixed_rsrc_ref_node *ref_node)
7302 io_rsrc_ref_lock(ctx);
7303 rsrc_data->node = ref_node;
7304 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7305 io_rsrc_ref_unlock(ctx);
7306 percpu_ref_get(&rsrc_data->refs);
7309 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7311 struct fixed_rsrc_ref_node *ref_node = NULL;
7313 io_rsrc_ref_lock(ctx);
7314 ref_node = data->node;
7316 io_rsrc_ref_unlock(ctx);
7318 percpu_ref_kill(&ref_node->refs);
7321 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7322 struct io_ring_ctx *ctx,
7323 void (*rsrc_put)(struct io_ring_ctx *ctx,
7324 struct io_rsrc_put *prsrc))
7326 struct fixed_rsrc_ref_node *backup_node;
7332 data->quiesce = true;
7335 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7338 backup_node->rsrc_data = data;
7339 backup_node->rsrc_put = rsrc_put;
7341 io_sqe_rsrc_kill_node(ctx, data);
7342 percpu_ref_kill(&data->refs);
7343 flush_delayed_work(&ctx->rsrc_put_work);
7345 ret = wait_for_completion_interruptible(&data->done);
7349 percpu_ref_resurrect(&data->refs);
7350 io_sqe_rsrc_set_node(ctx, data, backup_node);
7352 reinit_completion(&data->done);
7353 mutex_unlock(&ctx->uring_lock);
7354 ret = io_run_task_work_sig();
7355 mutex_lock(&ctx->uring_lock);
7357 data->quiesce = false;
7360 destroy_fixed_rsrc_ref_node(backup_node);
7364 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7366 struct fixed_rsrc_data *data;
7368 data = kzalloc(sizeof(*data), GFP_KERNEL);
7372 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7373 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7378 init_completion(&data->done);
7382 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7384 percpu_ref_exit(&data->refs);
7389 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7391 struct fixed_rsrc_data *data = ctx->file_data;
7392 unsigned nr_tables, i;
7396 * percpu_ref_is_dying() is to stop parallel files unregister
7397 * Since we possibly drop uring lock later in this function to
7400 if (!data || percpu_ref_is_dying(&data->refs))
7402 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7406 __io_sqe_files_unregister(ctx);
7407 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7408 for (i = 0; i < nr_tables; i++)
7409 kfree(data->table[i].files);
7410 free_fixed_rsrc_data(data);
7411 ctx->file_data = NULL;
7412 ctx->nr_user_files = 0;
7416 static void io_put_sq_data(struct io_sq_data *sqd)
7418 if (refcount_dec_and_test(&sqd->refs)) {
7420 * The park is a bit of a work-around, without it we get
7421 * warning spews on shutdown with SQPOLL set and affinity
7422 * set to a single CPU.
7425 kthread_park(sqd->thread);
7426 kthread_stop(sqd->thread);
7433 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7435 struct io_ring_ctx *ctx_attach;
7436 struct io_sq_data *sqd;
7439 f = fdget(p->wq_fd);
7441 return ERR_PTR(-ENXIO);
7442 if (f.file->f_op != &io_uring_fops) {
7444 return ERR_PTR(-EINVAL);
7447 ctx_attach = f.file->private_data;
7448 sqd = ctx_attach->sq_data;
7451 return ERR_PTR(-EINVAL);
7454 refcount_inc(&sqd->refs);
7459 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7461 struct io_sq_data *sqd;
7463 if (p->flags & IORING_SETUP_ATTACH_WQ)
7464 return io_attach_sq_data(p);
7466 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7468 return ERR_PTR(-ENOMEM);
7470 refcount_set(&sqd->refs, 1);
7471 INIT_LIST_HEAD(&sqd->ctx_list);
7472 INIT_LIST_HEAD(&sqd->ctx_new_list);
7473 mutex_init(&sqd->ctx_lock);
7474 mutex_init(&sqd->lock);
7475 init_waitqueue_head(&sqd->wait);
7479 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7480 __releases(&sqd->lock)
7484 kthread_unpark(sqd->thread);
7485 mutex_unlock(&sqd->lock);
7488 static void io_sq_thread_park(struct io_sq_data *sqd)
7489 __acquires(&sqd->lock)
7493 mutex_lock(&sqd->lock);
7494 kthread_park(sqd->thread);
7497 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7499 struct io_sq_data *sqd = ctx->sq_data;
7504 * We may arrive here from the error branch in
7505 * io_sq_offload_create() where the kthread is created
7506 * without being waked up, thus wake it up now to make
7507 * sure the wait will complete.
7509 wake_up_process(sqd->thread);
7510 wait_for_completion(&ctx->sq_thread_comp);
7512 io_sq_thread_park(sqd);
7515 mutex_lock(&sqd->ctx_lock);
7516 list_del(&ctx->sqd_list);
7517 io_sqd_update_thread_idle(sqd);
7518 mutex_unlock(&sqd->ctx_lock);
7521 io_sq_thread_unpark(sqd);
7523 io_put_sq_data(sqd);
7524 ctx->sq_data = NULL;
7528 static void io_finish_async(struct io_ring_ctx *ctx)
7530 io_sq_thread_stop(ctx);
7533 io_wq_destroy(ctx->io_wq);
7538 #if defined(CONFIG_UNIX)
7540 * Ensure the UNIX gc is aware of our file set, so we are certain that
7541 * the io_uring can be safely unregistered on process exit, even if we have
7542 * loops in the file referencing.
7544 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7546 struct sock *sk = ctx->ring_sock->sk;
7547 struct scm_fp_list *fpl;
7548 struct sk_buff *skb;
7551 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7555 skb = alloc_skb(0, GFP_KERNEL);
7564 fpl->user = get_uid(ctx->user);
7565 for (i = 0; i < nr; i++) {
7566 struct file *file = io_file_from_index(ctx, i + offset);
7570 fpl->fp[nr_files] = get_file(file);
7571 unix_inflight(fpl->user, fpl->fp[nr_files]);
7576 fpl->max = SCM_MAX_FD;
7577 fpl->count = nr_files;
7578 UNIXCB(skb).fp = fpl;
7579 skb->destructor = unix_destruct_scm;
7580 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7581 skb_queue_head(&sk->sk_receive_queue, skb);
7583 for (i = 0; i < nr_files; i++)
7594 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7595 * causes regular reference counting to break down. We rely on the UNIX
7596 * garbage collection to take care of this problem for us.
7598 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7600 unsigned left, total;
7604 left = ctx->nr_user_files;
7606 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7608 ret = __io_sqe_files_scm(ctx, this_files, total);
7612 total += this_files;
7618 while (total < ctx->nr_user_files) {
7619 struct file *file = io_file_from_index(ctx, total);
7629 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7635 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7636 unsigned nr_tables, unsigned nr_files)
7640 for (i = 0; i < nr_tables; i++) {
7641 struct fixed_rsrc_table *table = &file_data->table[i];
7642 unsigned this_files;
7644 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7645 table->files = kcalloc(this_files, sizeof(struct file *),
7649 nr_files -= this_files;
7655 for (i = 0; i < nr_tables; i++) {
7656 struct fixed_rsrc_table *table = &file_data->table[i];
7657 kfree(table->files);
7662 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7664 struct file *file = prsrc->file;
7665 #if defined(CONFIG_UNIX)
7666 struct sock *sock = ctx->ring_sock->sk;
7667 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7668 struct sk_buff *skb;
7671 __skb_queue_head_init(&list);
7674 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7675 * remove this entry and rearrange the file array.
7677 skb = skb_dequeue(head);
7679 struct scm_fp_list *fp;
7681 fp = UNIXCB(skb).fp;
7682 for (i = 0; i < fp->count; i++) {
7685 if (fp->fp[i] != file)
7688 unix_notinflight(fp->user, fp->fp[i]);
7689 left = fp->count - 1 - i;
7691 memmove(&fp->fp[i], &fp->fp[i + 1],
7692 left * sizeof(struct file *));
7699 __skb_queue_tail(&list, skb);
7709 __skb_queue_tail(&list, skb);
7711 skb = skb_dequeue(head);
7714 if (skb_peek(&list)) {
7715 spin_lock_irq(&head->lock);
7716 while ((skb = __skb_dequeue(&list)) != NULL)
7717 __skb_queue_tail(head, skb);
7718 spin_unlock_irq(&head->lock);
7725 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7727 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7728 struct io_ring_ctx *ctx = rsrc_data->ctx;
7729 struct io_rsrc_put *prsrc, *tmp;
7731 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7732 list_del(&prsrc->list);
7733 ref_node->rsrc_put(ctx, prsrc);
7737 percpu_ref_exit(&ref_node->refs);
7739 percpu_ref_put(&rsrc_data->refs);
7742 static void io_rsrc_put_work(struct work_struct *work)
7744 struct io_ring_ctx *ctx;
7745 struct llist_node *node;
7747 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7748 node = llist_del_all(&ctx->rsrc_put_llist);
7751 struct fixed_rsrc_ref_node *ref_node;
7752 struct llist_node *next = node->next;
7754 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7755 __io_rsrc_put_work(ref_node);
7760 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7763 struct fixed_rsrc_table *table;
7765 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7766 return &table->files[i & IORING_FILE_TABLE_MASK];
7769 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7771 struct fixed_rsrc_ref_node *ref_node;
7772 struct fixed_rsrc_data *data;
7773 struct io_ring_ctx *ctx;
7774 bool first_add = false;
7777 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7778 data = ref_node->rsrc_data;
7781 io_rsrc_ref_lock(ctx);
7782 ref_node->done = true;
7784 while (!list_empty(&ctx->rsrc_ref_list)) {
7785 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7786 struct fixed_rsrc_ref_node, node);
7787 /* recycle ref nodes in order */
7788 if (!ref_node->done)
7790 list_del(&ref_node->node);
7791 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7793 io_rsrc_ref_unlock(ctx);
7795 if (percpu_ref_is_dying(&data->refs))
7799 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7801 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7804 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7805 struct io_ring_ctx *ctx)
7807 struct fixed_rsrc_ref_node *ref_node;
7809 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7813 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7818 INIT_LIST_HEAD(&ref_node->node);
7819 INIT_LIST_HEAD(&ref_node->rsrc_list);
7820 ref_node->done = false;
7824 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7825 struct fixed_rsrc_ref_node *ref_node)
7827 ref_node->rsrc_data = ctx->file_data;
7828 ref_node->rsrc_put = io_ring_file_put;
7831 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7833 percpu_ref_exit(&ref_node->refs);
7838 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7841 __s32 __user *fds = (__s32 __user *) arg;
7842 unsigned nr_tables, i;
7844 int fd, ret = -ENOMEM;
7845 struct fixed_rsrc_ref_node *ref_node;
7846 struct fixed_rsrc_data *file_data;
7852 if (nr_args > IORING_MAX_FIXED_FILES)
7855 file_data = alloc_fixed_rsrc_data(ctx);
7858 ctx->file_data = file_data;
7860 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7861 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7863 if (!file_data->table)
7866 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7869 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7870 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7874 /* allow sparse sets */
7884 * Don't allow io_uring instances to be registered. If UNIX
7885 * isn't enabled, then this causes a reference cycle and this
7886 * instance can never get freed. If UNIX is enabled we'll
7887 * handle it just fine, but there's still no point in allowing
7888 * a ring fd as it doesn't support regular read/write anyway.
7890 if (file->f_op == &io_uring_fops) {
7894 *io_fixed_file_slot(file_data, i) = file;
7897 ret = io_sqe_files_scm(ctx);
7899 io_sqe_files_unregister(ctx);
7903 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7905 io_sqe_files_unregister(ctx);
7908 init_fixed_file_ref_node(ctx, ref_node);
7910 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7913 for (i = 0; i < ctx->nr_user_files; i++) {
7914 file = io_file_from_index(ctx, i);
7918 for (i = 0; i < nr_tables; i++)
7919 kfree(file_data->table[i].files);
7920 ctx->nr_user_files = 0;
7922 free_fixed_rsrc_data(ctx->file_data);
7923 ctx->file_data = NULL;
7927 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7930 #if defined(CONFIG_UNIX)
7931 struct sock *sock = ctx->ring_sock->sk;
7932 struct sk_buff_head *head = &sock->sk_receive_queue;
7933 struct sk_buff *skb;
7936 * See if we can merge this file into an existing skb SCM_RIGHTS
7937 * file set. If there's no room, fall back to allocating a new skb
7938 * and filling it in.
7940 spin_lock_irq(&head->lock);
7941 skb = skb_peek(head);
7943 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7945 if (fpl->count < SCM_MAX_FD) {
7946 __skb_unlink(skb, head);
7947 spin_unlock_irq(&head->lock);
7948 fpl->fp[fpl->count] = get_file(file);
7949 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7951 spin_lock_irq(&head->lock);
7952 __skb_queue_head(head, skb);
7957 spin_unlock_irq(&head->lock);
7964 return __io_sqe_files_scm(ctx, 1, index);
7970 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7972 struct io_rsrc_put *prsrc;
7973 struct fixed_rsrc_ref_node *ref_node = data->node;
7975 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7980 list_add(&prsrc->list, &ref_node->rsrc_list);
7985 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7988 return io_queue_rsrc_removal(data, (void *)file);
7991 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7992 struct io_uring_rsrc_update *up,
7995 struct fixed_rsrc_data *data = ctx->file_data;
7996 struct fixed_rsrc_ref_node *ref_node;
7997 struct file *file, **file_slot;
8001 bool needs_switch = false;
8003 if (check_add_overflow(up->offset, nr_args, &done))
8005 if (done > ctx->nr_user_files)
8008 ref_node = alloc_fixed_rsrc_ref_node(ctx);
8011 init_fixed_file_ref_node(ctx, ref_node);
8013 fds = u64_to_user_ptr(up->data);
8014 for (done = 0; done < nr_args; done++) {
8016 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
8020 if (fd == IORING_REGISTER_FILES_SKIP)
8023 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8024 file_slot = io_fixed_file_slot(ctx->file_data, i);
8027 err = io_queue_file_removal(data, *file_slot);
8031 needs_switch = true;
8040 * Don't allow io_uring instances to be registered. If
8041 * UNIX isn't enabled, then this causes a reference
8042 * cycle and this instance can never get freed. If UNIX
8043 * is enabled we'll handle it just fine, but there's
8044 * still no point in allowing a ring fd as it doesn't
8045 * support regular read/write anyway.
8047 if (file->f_op == &io_uring_fops) {
8053 err = io_sqe_file_register(ctx, file, i);
8063 percpu_ref_kill(&data->node->refs);
8064 io_sqe_rsrc_set_node(ctx, data, ref_node);
8066 destroy_fixed_rsrc_ref_node(ref_node);
8068 return done ? done : err;
8071 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
8074 struct io_uring_rsrc_update up;
8076 if (!ctx->file_data)
8080 if (copy_from_user(&up, arg, sizeof(up)))
8085 return __io_sqe_files_update(ctx, &up, nr_args);
8088 static struct io_wq_work *io_free_work(struct io_wq_work *work)
8090 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8092 req = io_put_req_find_next(req);
8093 return req ? &req->work : NULL;
8096 static int io_init_wq_offload(struct io_ring_ctx *ctx,
8097 struct io_uring_params *p)
8099 struct io_wq_data data;
8101 struct io_ring_ctx *ctx_attach;
8102 unsigned int concurrency;
8105 data.user = ctx->user;
8106 data.free_work = io_free_work;
8107 data.do_work = io_wq_submit_work;
8109 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
8110 /* Do QD, or 4 * CPUS, whatever is smallest */
8111 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8113 ctx->io_wq = io_wq_create(concurrency, &data);
8114 if (IS_ERR(ctx->io_wq)) {
8115 ret = PTR_ERR(ctx->io_wq);
8121 f = fdget(p->wq_fd);
8125 if (f.file->f_op != &io_uring_fops) {
8130 ctx_attach = f.file->private_data;
8131 /* @io_wq is protected by holding the fd */
8132 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8137 ctx->io_wq = ctx_attach->io_wq;
8143 static int io_uring_alloc_task_context(struct task_struct *task)
8145 struct io_uring_task *tctx;
8148 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8149 if (unlikely(!tctx))
8152 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8153 if (unlikely(ret)) {
8159 init_waitqueue_head(&tctx->wait);
8161 atomic_set(&tctx->in_idle, 0);
8162 tctx->sqpoll = false;
8163 io_init_identity(&tctx->__identity);
8164 tctx->identity = &tctx->__identity;
8165 task->io_uring = tctx;
8166 spin_lock_init(&tctx->task_lock);
8167 INIT_WQ_LIST(&tctx->task_list);
8168 tctx->task_state = 0;
8169 init_task_work(&tctx->task_work, tctx_task_work);
8173 void __io_uring_free(struct task_struct *tsk)
8175 struct io_uring_task *tctx = tsk->io_uring;
8177 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8178 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8179 if (tctx->identity != &tctx->__identity)
8180 kfree(tctx->identity);
8181 percpu_counter_destroy(&tctx->inflight);
8183 tsk->io_uring = NULL;
8186 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8187 struct io_uring_params *p)
8191 if (ctx->flags & IORING_SETUP_SQPOLL) {
8192 struct io_sq_data *sqd;
8195 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8198 sqd = io_get_sq_data(p);
8205 io_sq_thread_park(sqd);
8206 mutex_lock(&sqd->ctx_lock);
8207 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8208 mutex_unlock(&sqd->ctx_lock);
8209 io_sq_thread_unpark(sqd);
8211 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8212 if (!ctx->sq_thread_idle)
8213 ctx->sq_thread_idle = HZ;
8218 if (p->flags & IORING_SETUP_SQ_AFF) {
8219 int cpu = p->sq_thread_cpu;
8222 if (cpu >= nr_cpu_ids)
8224 if (!cpu_online(cpu))
8227 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8228 cpu, "io_uring-sq");
8230 sqd->thread = kthread_create(io_sq_thread, sqd,
8233 if (IS_ERR(sqd->thread)) {
8234 ret = PTR_ERR(sqd->thread);
8238 ret = io_uring_alloc_task_context(sqd->thread);
8241 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8242 /* Can't have SQ_AFF without SQPOLL */
8248 ret = io_init_wq_offload(ctx, p);
8254 io_finish_async(ctx);
8258 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8260 struct io_sq_data *sqd = ctx->sq_data;
8262 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8263 wake_up_process(sqd->thread);
8266 static inline void __io_unaccount_mem(struct user_struct *user,
8267 unsigned long nr_pages)
8269 atomic_long_sub(nr_pages, &user->locked_vm);
8272 static inline int __io_account_mem(struct user_struct *user,
8273 unsigned long nr_pages)
8275 unsigned long page_limit, cur_pages, new_pages;
8277 /* Don't allow more pages than we can safely lock */
8278 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8281 cur_pages = atomic_long_read(&user->locked_vm);
8282 new_pages = cur_pages + nr_pages;
8283 if (new_pages > page_limit)
8285 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8286 new_pages) != cur_pages);
8291 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8294 __io_unaccount_mem(ctx->user, nr_pages);
8296 if (ctx->mm_account)
8297 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8300 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8304 if (ctx->limit_mem) {
8305 ret = __io_account_mem(ctx->user, nr_pages);
8310 if (ctx->mm_account)
8311 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8316 static void io_mem_free(void *ptr)
8323 page = virt_to_head_page(ptr);
8324 if (put_page_testzero(page))
8325 free_compound_page(page);
8328 static void *io_mem_alloc(size_t size)
8330 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8331 __GFP_NORETRY | __GFP_ACCOUNT;
8333 return (void *) __get_free_pages(gfp_flags, get_order(size));
8336 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8339 struct io_rings *rings;
8340 size_t off, sq_array_size;
8342 off = struct_size(rings, cqes, cq_entries);
8343 if (off == SIZE_MAX)
8347 off = ALIGN(off, SMP_CACHE_BYTES);
8355 sq_array_size = array_size(sizeof(u32), sq_entries);
8356 if (sq_array_size == SIZE_MAX)
8359 if (check_add_overflow(off, sq_array_size, &off))
8365 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8369 if (!ctx->user_bufs)
8372 for (i = 0; i < ctx->nr_user_bufs; i++) {
8373 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8375 for (j = 0; j < imu->nr_bvecs; j++)
8376 unpin_user_page(imu->bvec[j].bv_page);
8378 if (imu->acct_pages)
8379 io_unaccount_mem(ctx, imu->acct_pages);
8384 kfree(ctx->user_bufs);
8385 ctx->user_bufs = NULL;
8386 ctx->nr_user_bufs = 0;
8390 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8391 void __user *arg, unsigned index)
8393 struct iovec __user *src;
8395 #ifdef CONFIG_COMPAT
8397 struct compat_iovec __user *ciovs;
8398 struct compat_iovec ciov;
8400 ciovs = (struct compat_iovec __user *) arg;
8401 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8404 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8405 dst->iov_len = ciov.iov_len;
8409 src = (struct iovec __user *) arg;
8410 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8416 * Not super efficient, but this is just a registration time. And we do cache
8417 * the last compound head, so generally we'll only do a full search if we don't
8420 * We check if the given compound head page has already been accounted, to
8421 * avoid double accounting it. This allows us to account the full size of the
8422 * page, not just the constituent pages of a huge page.
8424 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8425 int nr_pages, struct page *hpage)
8429 /* check current page array */
8430 for (i = 0; i < nr_pages; i++) {
8431 if (!PageCompound(pages[i]))
8433 if (compound_head(pages[i]) == hpage)
8437 /* check previously registered pages */
8438 for (i = 0; i < ctx->nr_user_bufs; i++) {
8439 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8441 for (j = 0; j < imu->nr_bvecs; j++) {
8442 if (!PageCompound(imu->bvec[j].bv_page))
8444 if (compound_head(imu->bvec[j].bv_page) == hpage)
8452 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8453 int nr_pages, struct io_mapped_ubuf *imu,
8454 struct page **last_hpage)
8458 for (i = 0; i < nr_pages; i++) {
8459 if (!PageCompound(pages[i])) {
8464 hpage = compound_head(pages[i]);
8465 if (hpage == *last_hpage)
8467 *last_hpage = hpage;
8468 if (headpage_already_acct(ctx, pages, i, hpage))
8470 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8474 if (!imu->acct_pages)
8477 ret = io_account_mem(ctx, imu->acct_pages);
8479 imu->acct_pages = 0;
8483 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8484 struct io_mapped_ubuf *imu,
8485 struct page **last_hpage)
8487 struct vm_area_struct **vmas = NULL;
8488 struct page **pages = NULL;
8489 unsigned long off, start, end, ubuf;
8491 int ret, pret, nr_pages, i;
8493 ubuf = (unsigned long) iov->iov_base;
8494 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8495 start = ubuf >> PAGE_SHIFT;
8496 nr_pages = end - start;
8500 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8504 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8509 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8515 mmap_read_lock(current->mm);
8516 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8518 if (pret == nr_pages) {
8519 /* don't support file backed memory */
8520 for (i = 0; i < nr_pages; i++) {
8521 struct vm_area_struct *vma = vmas[i];
8524 !is_file_hugepages(vma->vm_file)) {
8530 ret = pret < 0 ? pret : -EFAULT;
8532 mmap_read_unlock(current->mm);
8535 * if we did partial map, or found file backed vmas,
8536 * release any pages we did get
8539 unpin_user_pages(pages, pret);
8544 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8546 unpin_user_pages(pages, pret);
8551 off = ubuf & ~PAGE_MASK;
8552 size = iov->iov_len;
8553 for (i = 0; i < nr_pages; i++) {
8556 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8557 imu->bvec[i].bv_page = pages[i];
8558 imu->bvec[i].bv_len = vec_len;
8559 imu->bvec[i].bv_offset = off;
8563 /* store original address for later verification */
8565 imu->len = iov->iov_len;
8566 imu->nr_bvecs = nr_pages;
8574 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8578 if (!nr_args || nr_args > UIO_MAXIOV)
8581 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8583 if (!ctx->user_bufs)
8589 static int io_buffer_validate(struct iovec *iov)
8592 * Don't impose further limits on the size and buffer
8593 * constraints here, we'll -EINVAL later when IO is
8594 * submitted if they are wrong.
8596 if (!iov->iov_base || !iov->iov_len)
8599 /* arbitrary limit, but we need something */
8600 if (iov->iov_len > SZ_1G)
8606 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8607 unsigned int nr_args)
8611 struct page *last_hpage = NULL;
8613 ret = io_buffers_map_alloc(ctx, nr_args);
8617 for (i = 0; i < nr_args; i++) {
8618 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8620 ret = io_copy_iov(ctx, &iov, arg, i);
8624 ret = io_buffer_validate(&iov);
8628 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8632 ctx->nr_user_bufs++;
8636 io_sqe_buffers_unregister(ctx);
8641 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8643 __s32 __user *fds = arg;
8649 if (copy_from_user(&fd, fds, sizeof(*fds)))
8652 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8653 if (IS_ERR(ctx->cq_ev_fd)) {
8654 int ret = PTR_ERR(ctx->cq_ev_fd);
8655 ctx->cq_ev_fd = NULL;
8662 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8664 if (ctx->cq_ev_fd) {
8665 eventfd_ctx_put(ctx->cq_ev_fd);
8666 ctx->cq_ev_fd = NULL;
8673 static int __io_destroy_buffers(int id, void *p, void *data)
8675 struct io_ring_ctx *ctx = data;
8676 struct io_buffer *buf = p;
8678 __io_remove_buffers(ctx, buf, id, -1U);
8682 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8684 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8685 idr_destroy(&ctx->io_buffer_idr);
8688 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8690 struct io_kiocb *req, *nxt;
8692 list_for_each_entry_safe(req, nxt, list, compl.list) {
8693 if (tsk && req->task != tsk)
8695 list_del(&req->compl.list);
8696 kmem_cache_free(req_cachep, req);
8700 static void io_req_caches_free(struct io_ring_ctx *ctx, struct task_struct *tsk)
8702 struct io_submit_state *submit_state = &ctx->submit_state;
8703 struct io_comp_state *cs = &ctx->submit_state.comp;
8705 mutex_lock(&ctx->uring_lock);
8707 if (submit_state->free_reqs) {
8708 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8709 submit_state->reqs);
8710 submit_state->free_reqs = 0;
8713 spin_lock_irq(&ctx->completion_lock);
8714 list_splice_init(&cs->locked_free_list, &cs->free_list);
8715 cs->locked_free_nr = 0;
8716 spin_unlock_irq(&ctx->completion_lock);
8718 io_req_cache_free(&cs->free_list, NULL);
8720 mutex_unlock(&ctx->uring_lock);
8723 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8726 * Some may use context even when all refs and requests have been put,
8727 * and they are free to do so while still holding uring_lock, see
8728 * __io_req_task_submit(). Wait for them to finish.
8730 mutex_lock(&ctx->uring_lock);
8731 mutex_unlock(&ctx->uring_lock);
8733 io_finish_async(ctx);
8734 io_sqe_buffers_unregister(ctx);
8736 if (ctx->sqo_task) {
8737 put_task_struct(ctx->sqo_task);
8738 ctx->sqo_task = NULL;
8739 mmdrop(ctx->mm_account);
8740 ctx->mm_account = NULL;
8743 #ifdef CONFIG_BLK_CGROUP
8744 if (ctx->sqo_blkcg_css)
8745 css_put(ctx->sqo_blkcg_css);
8748 mutex_lock(&ctx->uring_lock);
8749 io_sqe_files_unregister(ctx);
8750 mutex_unlock(&ctx->uring_lock);
8751 io_eventfd_unregister(ctx);
8752 io_destroy_buffers(ctx);
8753 idr_destroy(&ctx->personality_idr);
8755 #if defined(CONFIG_UNIX)
8756 if (ctx->ring_sock) {
8757 ctx->ring_sock->file = NULL; /* so that iput() is called */
8758 sock_release(ctx->ring_sock);
8762 io_mem_free(ctx->rings);
8763 io_mem_free(ctx->sq_sqes);
8765 percpu_ref_exit(&ctx->refs);
8766 free_uid(ctx->user);
8767 put_cred(ctx->creds);
8768 io_req_caches_free(ctx, NULL);
8769 kfree(ctx->cancel_hash);
8773 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8775 struct io_ring_ctx *ctx = file->private_data;
8778 poll_wait(file, &ctx->cq_wait, wait);
8780 * synchronizes with barrier from wq_has_sleeper call in
8784 if (!io_sqring_full(ctx))
8785 mask |= EPOLLOUT | EPOLLWRNORM;
8788 * Don't flush cqring overflow list here, just do a simple check.
8789 * Otherwise there could possible be ABBA deadlock:
8792 * lock(&ctx->uring_lock);
8794 * lock(&ctx->uring_lock);
8797 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8798 * pushs them to do the flush.
8800 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8801 mask |= EPOLLIN | EPOLLRDNORM;
8806 static int io_uring_fasync(int fd, struct file *file, int on)
8808 struct io_ring_ctx *ctx = file->private_data;
8810 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8813 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8815 struct io_identity *iod;
8817 iod = idr_remove(&ctx->personality_idr, id);
8819 put_cred(iod->creds);
8820 if (refcount_dec_and_test(&iod->count))
8828 static int io_remove_personalities(int id, void *p, void *data)
8830 struct io_ring_ctx *ctx = data;
8832 io_unregister_personality(ctx, id);
8836 static void io_ring_exit_work(struct work_struct *work)
8838 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8842 * If we're doing polled IO and end up having requests being
8843 * submitted async (out-of-line), then completions can come in while
8844 * we're waiting for refs to drop. We need to reap these manually,
8845 * as nobody else will be looking for them.
8848 io_uring_try_cancel_requests(ctx, NULL, NULL);
8849 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8850 io_ring_ctx_free(ctx);
8853 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8855 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8857 return req->ctx == data;
8860 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8862 mutex_lock(&ctx->uring_lock);
8863 percpu_ref_kill(&ctx->refs);
8865 if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead))
8868 /* if force is set, the ring is going away. always drop after that */
8869 ctx->cq_overflow_flushed = 1;
8871 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8872 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8873 mutex_unlock(&ctx->uring_lock);
8875 io_kill_timeouts(ctx, NULL, NULL);
8876 io_poll_remove_all(ctx, NULL, NULL);
8879 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8881 /* if we failed setting up the ctx, we might not have any rings */
8882 io_iopoll_try_reap_events(ctx);
8884 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8886 * Use system_unbound_wq to avoid spawning tons of event kworkers
8887 * if we're exiting a ton of rings at the same time. It just adds
8888 * noise and overhead, there's no discernable change in runtime
8889 * over using system_wq.
8891 queue_work(system_unbound_wq, &ctx->exit_work);
8894 static int io_uring_release(struct inode *inode, struct file *file)
8896 struct io_ring_ctx *ctx = file->private_data;
8898 file->private_data = NULL;
8899 io_ring_ctx_wait_and_kill(ctx);
8903 struct io_task_cancel {
8904 struct task_struct *task;
8905 struct files_struct *files;
8908 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8910 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8911 struct io_task_cancel *cancel = data;
8914 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8915 unsigned long flags;
8916 struct io_ring_ctx *ctx = req->ctx;
8918 /* protect against races with linked timeouts */
8919 spin_lock_irqsave(&ctx->completion_lock, flags);
8920 ret = io_match_task(req, cancel->task, cancel->files);
8921 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8923 ret = io_match_task(req, cancel->task, cancel->files);
8928 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8929 struct task_struct *task,
8930 struct files_struct *files)
8932 struct io_defer_entry *de = NULL;
8935 spin_lock_irq(&ctx->completion_lock);
8936 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8937 if (io_match_task(de->req, task, files)) {
8938 list_cut_position(&list, &ctx->defer_list, &de->list);
8942 spin_unlock_irq(&ctx->completion_lock);
8944 while (!list_empty(&list)) {
8945 de = list_first_entry(&list, struct io_defer_entry, list);
8946 list_del_init(&de->list);
8947 req_set_fail_links(de->req);
8948 io_put_req(de->req);
8949 io_req_complete(de->req, -ECANCELED);
8954 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8955 struct task_struct *task,
8956 struct files_struct *files)
8958 struct io_task_cancel cancel = { .task = task, .files = files, };
8961 enum io_wq_cancel cret;
8965 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb,
8967 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8970 /* SQPOLL thread does its own polling */
8971 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8972 while (!list_empty_careful(&ctx->iopoll_list)) {
8973 io_iopoll_try_reap_events(ctx);
8978 ret |= io_poll_remove_all(ctx, task, files);
8979 ret |= io_kill_timeouts(ctx, task, files);
8980 ret |= io_run_task_work();
8981 io_cqring_overflow_flush(ctx, true, task, files);
8988 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8989 struct task_struct *task,
8990 struct files_struct *files)
8992 struct io_kiocb *req;
8995 spin_lock_irq(&ctx->inflight_lock);
8996 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8997 cnt += io_match_task(req, task, files);
8998 spin_unlock_irq(&ctx->inflight_lock);
9002 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
9003 struct task_struct *task,
9004 struct files_struct *files)
9006 while (!list_empty_careful(&ctx->inflight_list)) {
9010 inflight = io_uring_count_inflight(ctx, task, files);
9014 io_uring_try_cancel_requests(ctx, task, files);
9017 io_sq_thread_unpark(ctx->sq_data);
9018 prepare_to_wait(&task->io_uring->wait, &wait,
9019 TASK_UNINTERRUPTIBLE);
9020 if (inflight == io_uring_count_inflight(ctx, task, files))
9022 finish_wait(&task->io_uring->wait, &wait);
9024 io_sq_thread_park(ctx->sq_data);
9028 static void io_disable_sqo_submit(struct io_ring_ctx *ctx)
9030 mutex_lock(&ctx->uring_lock);
9032 mutex_unlock(&ctx->uring_lock);
9034 /* make sure callers enter the ring to get error */
9036 io_ring_set_wakeup_flag(ctx);
9040 * We need to iteratively cancel requests, in case a request has dependent
9041 * hard links. These persist even for failure of cancelations, hence keep
9042 * looping until none are found.
9044 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
9045 struct files_struct *files)
9047 struct task_struct *task = current;
9049 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
9050 io_disable_sqo_submit(ctx);
9051 task = ctx->sq_data->thread;
9052 atomic_inc(&task->io_uring->in_idle);
9053 io_sq_thread_park(ctx->sq_data);
9056 io_cancel_defer_files(ctx, task, files);
9058 io_uring_cancel_files(ctx, task, files);
9060 io_uring_try_cancel_requests(ctx, task, NULL);
9062 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
9063 atomic_dec(&task->io_uring->in_idle);
9064 io_sq_thread_unpark(ctx->sq_data);
9069 * Note that this task has used io_uring. We use it for cancelation purposes.
9071 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
9073 struct io_uring_task *tctx = current->io_uring;
9076 if (unlikely(!tctx)) {
9077 ret = io_uring_alloc_task_context(current);
9080 tctx = current->io_uring;
9082 if (tctx->last != file) {
9083 void *old = xa_load(&tctx->xa, (unsigned long)file);
9087 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
9094 /* one and only SQPOLL file note, held by sqo_task */
9095 WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) &&
9096 current != ctx->sqo_task);
9102 * This is race safe in that the task itself is doing this, hence it
9103 * cannot be going through the exit/cancel paths at the same time.
9104 * This cannot be modified while exit/cancel is running.
9106 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
9107 tctx->sqpoll = true;
9113 * Remove this io_uring_file -> task mapping.
9115 static void io_uring_del_task_file(struct file *file)
9117 struct io_uring_task *tctx = current->io_uring;
9119 if (tctx->last == file)
9121 file = xa_erase(&tctx->xa, (unsigned long)file);
9126 static void io_uring_remove_task_files(struct io_uring_task *tctx)
9129 unsigned long index;
9131 xa_for_each(&tctx->xa, index, file)
9132 io_uring_del_task_file(file);
9135 void __io_uring_files_cancel(struct files_struct *files)
9137 struct io_uring_task *tctx = current->io_uring;
9139 unsigned long index;
9141 /* make sure overflow events are dropped */
9142 atomic_inc(&tctx->in_idle);
9143 xa_for_each(&tctx->xa, index, file)
9144 io_uring_cancel_task_requests(file->private_data, files);
9145 atomic_dec(&tctx->in_idle);
9148 io_uring_remove_task_files(tctx);
9151 static s64 tctx_inflight(struct io_uring_task *tctx)
9153 return percpu_counter_sum(&tctx->inflight);
9156 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
9158 struct io_uring_task *tctx;
9164 tctx = ctx->sq_data->thread->io_uring;
9165 io_disable_sqo_submit(ctx);
9167 atomic_inc(&tctx->in_idle);
9169 /* read completions before cancelations */
9170 inflight = tctx_inflight(tctx);
9173 io_uring_cancel_task_requests(ctx, NULL);
9175 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9177 * If we've seen completions, retry without waiting. This
9178 * avoids a race where a completion comes in before we did
9179 * prepare_to_wait().
9181 if (inflight == tctx_inflight(tctx))
9183 finish_wait(&tctx->wait, &wait);
9185 atomic_dec(&tctx->in_idle);
9189 * Find any io_uring fd that this task has registered or done IO on, and cancel
9192 void __io_uring_task_cancel(void)
9194 struct io_uring_task *tctx = current->io_uring;
9198 /* make sure overflow events are dropped */
9199 atomic_inc(&tctx->in_idle);
9201 /* trigger io_disable_sqo_submit() */
9204 unsigned long index;
9206 xa_for_each(&tctx->xa, index, file)
9207 io_uring_cancel_sqpoll(file->private_data);
9211 /* read completions before cancelations */
9212 inflight = tctx_inflight(tctx);
9215 __io_uring_files_cancel(NULL);
9217 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9220 * If we've seen completions, retry without waiting. This
9221 * avoids a race where a completion comes in before we did
9222 * prepare_to_wait().
9224 if (inflight == tctx_inflight(tctx))
9226 finish_wait(&tctx->wait, &wait);
9229 atomic_dec(&tctx->in_idle);
9231 io_uring_remove_task_files(tctx);
9234 static int io_uring_flush(struct file *file, void *data)
9236 struct io_uring_task *tctx = current->io_uring;
9237 struct io_ring_ctx *ctx = file->private_data;
9239 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
9240 io_uring_cancel_task_requests(ctx, NULL);
9241 io_req_caches_free(ctx, current);
9247 /* we should have cancelled and erased it before PF_EXITING */
9248 WARN_ON_ONCE((current->flags & PF_EXITING) &&
9249 xa_load(&tctx->xa, (unsigned long)file));
9252 * fput() is pending, will be 2 if the only other ref is our potential
9253 * task file note. If the task is exiting, drop regardless of count.
9255 if (atomic_long_read(&file->f_count) != 2)
9258 if (ctx->flags & IORING_SETUP_SQPOLL) {
9259 /* there is only one file note, which is owned by sqo_task */
9260 WARN_ON_ONCE(ctx->sqo_task != current &&
9261 xa_load(&tctx->xa, (unsigned long)file));
9262 /* sqo_dead check is for when this happens after cancellation */
9263 WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead &&
9264 !xa_load(&tctx->xa, (unsigned long)file));
9266 io_disable_sqo_submit(ctx);
9269 if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current)
9270 io_uring_del_task_file(file);
9274 static void *io_uring_validate_mmap_request(struct file *file,
9275 loff_t pgoff, size_t sz)
9277 struct io_ring_ctx *ctx = file->private_data;
9278 loff_t offset = pgoff << PAGE_SHIFT;
9283 case IORING_OFF_SQ_RING:
9284 case IORING_OFF_CQ_RING:
9287 case IORING_OFF_SQES:
9291 return ERR_PTR(-EINVAL);
9294 page = virt_to_head_page(ptr);
9295 if (sz > page_size(page))
9296 return ERR_PTR(-EINVAL);
9303 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9305 size_t sz = vma->vm_end - vma->vm_start;
9309 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9311 return PTR_ERR(ptr);
9313 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9314 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9317 #else /* !CONFIG_MMU */
9319 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9321 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9324 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9326 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9329 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9330 unsigned long addr, unsigned long len,
9331 unsigned long pgoff, unsigned long flags)
9335 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9337 return PTR_ERR(ptr);
9339 return (unsigned long) ptr;
9342 #endif /* !CONFIG_MMU */
9344 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9350 if (!io_sqring_full(ctx))
9353 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9355 if (unlikely(ctx->sqo_dead)) {
9360 if (!io_sqring_full(ctx))
9364 } while (!signal_pending(current));
9366 finish_wait(&ctx->sqo_sq_wait, &wait);
9371 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9372 struct __kernel_timespec __user **ts,
9373 const sigset_t __user **sig)
9375 struct io_uring_getevents_arg arg;
9378 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9379 * is just a pointer to the sigset_t.
9381 if (!(flags & IORING_ENTER_EXT_ARG)) {
9382 *sig = (const sigset_t __user *) argp;
9388 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9389 * timespec and sigset_t pointers if good.
9391 if (*argsz != sizeof(arg))
9393 if (copy_from_user(&arg, argp, sizeof(arg)))
9395 *sig = u64_to_user_ptr(arg.sigmask);
9396 *argsz = arg.sigmask_sz;
9397 *ts = u64_to_user_ptr(arg.ts);
9401 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9402 u32, min_complete, u32, flags, const void __user *, argp,
9405 struct io_ring_ctx *ctx;
9412 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9413 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9421 if (f.file->f_op != &io_uring_fops)
9425 ctx = f.file->private_data;
9426 if (!percpu_ref_tryget(&ctx->refs))
9430 if (ctx->flags & IORING_SETUP_R_DISABLED)
9434 * For SQ polling, the thread will do all submissions and completions.
9435 * Just return the requested submit count, and wake the thread if
9439 if (ctx->flags & IORING_SETUP_SQPOLL) {
9440 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9443 if (unlikely(ctx->sqo_dead))
9445 if (flags & IORING_ENTER_SQ_WAKEUP)
9446 wake_up(&ctx->sq_data->wait);
9447 if (flags & IORING_ENTER_SQ_WAIT) {
9448 ret = io_sqpoll_wait_sq(ctx);
9452 submitted = to_submit;
9453 } else if (to_submit) {
9454 ret = io_uring_add_task_file(ctx, f.file);
9457 mutex_lock(&ctx->uring_lock);
9458 submitted = io_submit_sqes(ctx, to_submit);
9459 mutex_unlock(&ctx->uring_lock);
9461 if (submitted != to_submit)
9464 if (flags & IORING_ENTER_GETEVENTS) {
9465 const sigset_t __user *sig;
9466 struct __kernel_timespec __user *ts;
9468 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9472 min_complete = min(min_complete, ctx->cq_entries);
9475 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9476 * space applications don't need to do io completion events
9477 * polling again, they can rely on io_sq_thread to do polling
9478 * work, which can reduce cpu usage and uring_lock contention.
9480 if (ctx->flags & IORING_SETUP_IOPOLL &&
9481 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9482 ret = io_iopoll_check(ctx, min_complete);
9484 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9489 percpu_ref_put(&ctx->refs);
9492 return submitted ? submitted : ret;
9495 #ifdef CONFIG_PROC_FS
9496 static int io_uring_show_cred(int id, void *p, void *data)
9498 struct io_identity *iod = p;
9499 const struct cred *cred = iod->creds;
9500 struct seq_file *m = data;
9501 struct user_namespace *uns = seq_user_ns(m);
9502 struct group_info *gi;
9507 seq_printf(m, "%5d\n", id);
9508 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9509 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9510 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9511 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9512 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9513 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9514 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9515 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9516 seq_puts(m, "\n\tGroups:\t");
9517 gi = cred->group_info;
9518 for (g = 0; g < gi->ngroups; g++) {
9519 seq_put_decimal_ull(m, g ? " " : "",
9520 from_kgid_munged(uns, gi->gid[g]));
9522 seq_puts(m, "\n\tCapEff:\t");
9523 cap = cred->cap_effective;
9524 CAP_FOR_EACH_U32(__capi)
9525 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9530 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9532 struct io_sq_data *sq = NULL;
9537 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9538 * since fdinfo case grabs it in the opposite direction of normal use
9539 * cases. If we fail to get the lock, we just don't iterate any
9540 * structures that could be going away outside the io_uring mutex.
9542 has_lock = mutex_trylock(&ctx->uring_lock);
9544 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9547 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9548 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9549 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9550 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9551 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9554 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9556 seq_printf(m, "%5u: <none>\n", i);
9558 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9559 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9560 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9562 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9563 (unsigned int) buf->len);
9565 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9566 seq_printf(m, "Personalities:\n");
9567 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9569 seq_printf(m, "PollList:\n");
9570 spin_lock_irq(&ctx->completion_lock);
9571 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9572 struct hlist_head *list = &ctx->cancel_hash[i];
9573 struct io_kiocb *req;
9575 hlist_for_each_entry(req, list, hash_node)
9576 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9577 req->task->task_works != NULL);
9579 spin_unlock_irq(&ctx->completion_lock);
9581 mutex_unlock(&ctx->uring_lock);
9584 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9586 struct io_ring_ctx *ctx = f->private_data;
9588 if (percpu_ref_tryget(&ctx->refs)) {
9589 __io_uring_show_fdinfo(ctx, m);
9590 percpu_ref_put(&ctx->refs);
9595 static const struct file_operations io_uring_fops = {
9596 .release = io_uring_release,
9597 .flush = io_uring_flush,
9598 .mmap = io_uring_mmap,
9600 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9601 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9603 .poll = io_uring_poll,
9604 .fasync = io_uring_fasync,
9605 #ifdef CONFIG_PROC_FS
9606 .show_fdinfo = io_uring_show_fdinfo,
9610 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9611 struct io_uring_params *p)
9613 struct io_rings *rings;
9614 size_t size, sq_array_offset;
9616 /* make sure these are sane, as we already accounted them */
9617 ctx->sq_entries = p->sq_entries;
9618 ctx->cq_entries = p->cq_entries;
9620 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9621 if (size == SIZE_MAX)
9624 rings = io_mem_alloc(size);
9629 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9630 rings->sq_ring_mask = p->sq_entries - 1;
9631 rings->cq_ring_mask = p->cq_entries - 1;
9632 rings->sq_ring_entries = p->sq_entries;
9633 rings->cq_ring_entries = p->cq_entries;
9634 ctx->sq_mask = rings->sq_ring_mask;
9635 ctx->cq_mask = rings->cq_ring_mask;
9637 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9638 if (size == SIZE_MAX) {
9639 io_mem_free(ctx->rings);
9644 ctx->sq_sqes = io_mem_alloc(size);
9645 if (!ctx->sq_sqes) {
9646 io_mem_free(ctx->rings);
9654 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9658 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9662 ret = io_uring_add_task_file(ctx, file);
9667 fd_install(fd, file);
9672 * Allocate an anonymous fd, this is what constitutes the application
9673 * visible backing of an io_uring instance. The application mmaps this
9674 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9675 * we have to tie this fd to a socket for file garbage collection purposes.
9677 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9680 #if defined(CONFIG_UNIX)
9683 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9686 return ERR_PTR(ret);
9689 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9690 O_RDWR | O_CLOEXEC);
9691 #if defined(CONFIG_UNIX)
9693 sock_release(ctx->ring_sock);
9694 ctx->ring_sock = NULL;
9696 ctx->ring_sock->file = file;
9702 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9703 struct io_uring_params __user *params)
9705 struct user_struct *user = NULL;
9706 struct io_ring_ctx *ctx;
9712 if (entries > IORING_MAX_ENTRIES) {
9713 if (!(p->flags & IORING_SETUP_CLAMP))
9715 entries = IORING_MAX_ENTRIES;
9719 * Use twice as many entries for the CQ ring. It's possible for the
9720 * application to drive a higher depth than the size of the SQ ring,
9721 * since the sqes are only used at submission time. This allows for
9722 * some flexibility in overcommitting a bit. If the application has
9723 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9724 * of CQ ring entries manually.
9726 p->sq_entries = roundup_pow_of_two(entries);
9727 if (p->flags & IORING_SETUP_CQSIZE) {
9729 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9730 * to a power-of-two, if it isn't already. We do NOT impose
9731 * any cq vs sq ring sizing.
9735 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9736 if (!(p->flags & IORING_SETUP_CLAMP))
9738 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9740 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9741 if (p->cq_entries < p->sq_entries)
9744 p->cq_entries = 2 * p->sq_entries;
9747 user = get_uid(current_user());
9749 ctx = io_ring_ctx_alloc(p);
9754 ctx->compat = in_compat_syscall();
9755 ctx->limit_mem = !capable(CAP_IPC_LOCK);
9757 ctx->creds = get_current_cred();
9759 ctx->loginuid = current->loginuid;
9760 ctx->sessionid = current->sessionid;
9762 ctx->sqo_task = get_task_struct(current);
9765 * This is just grabbed for accounting purposes. When a process exits,
9766 * the mm is exited and dropped before the files, hence we need to hang
9767 * on to this mm purely for the purposes of being able to unaccount
9768 * memory (locked/pinned vm). It's not used for anything else.
9770 mmgrab(current->mm);
9771 ctx->mm_account = current->mm;
9773 #ifdef CONFIG_BLK_CGROUP
9775 * The sq thread will belong to the original cgroup it was inited in.
9776 * If the cgroup goes offline (e.g. disabling the io controller), then
9777 * issued bios will be associated with the closest cgroup later in the
9781 ctx->sqo_blkcg_css = blkcg_css();
9782 ret = css_tryget_online(ctx->sqo_blkcg_css);
9785 /* don't init against a dying cgroup, have the user try again */
9786 ctx->sqo_blkcg_css = NULL;
9791 ret = io_allocate_scq_urings(ctx, p);
9795 ret = io_sq_offload_create(ctx, p);
9799 if (!(p->flags & IORING_SETUP_R_DISABLED))
9800 io_sq_offload_start(ctx);
9802 memset(&p->sq_off, 0, sizeof(p->sq_off));
9803 p->sq_off.head = offsetof(struct io_rings, sq.head);
9804 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9805 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9806 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9807 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9808 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9809 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9811 memset(&p->cq_off, 0, sizeof(p->cq_off));
9812 p->cq_off.head = offsetof(struct io_rings, cq.head);
9813 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9814 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9815 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9816 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9817 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9818 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9820 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9821 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9822 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9823 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9824 IORING_FEAT_EXT_ARG;
9826 if (copy_to_user(params, p, sizeof(*p))) {
9831 file = io_uring_get_file(ctx);
9833 ret = PTR_ERR(file);
9838 * Install ring fd as the very last thing, so we don't risk someone
9839 * having closed it before we finish setup
9841 ret = io_uring_install_fd(ctx, file);
9843 io_disable_sqo_submit(ctx);
9844 /* fput will clean it up */
9849 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9852 io_disable_sqo_submit(ctx);
9853 io_ring_ctx_wait_and_kill(ctx);
9858 * Sets up an aio uring context, and returns the fd. Applications asks for a
9859 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9860 * params structure passed in.
9862 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9864 struct io_uring_params p;
9867 if (copy_from_user(&p, params, sizeof(p)))
9869 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9874 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9875 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9876 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9877 IORING_SETUP_R_DISABLED))
9880 return io_uring_create(entries, &p, params);
9883 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9884 struct io_uring_params __user *, params)
9886 return io_uring_setup(entries, params);
9889 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9891 struct io_uring_probe *p;
9895 size = struct_size(p, ops, nr_args);
9896 if (size == SIZE_MAX)
9898 p = kzalloc(size, GFP_KERNEL);
9903 if (copy_from_user(p, arg, size))
9906 if (memchr_inv(p, 0, size))
9909 p->last_op = IORING_OP_LAST - 1;
9910 if (nr_args > IORING_OP_LAST)
9911 nr_args = IORING_OP_LAST;
9913 for (i = 0; i < nr_args; i++) {
9915 if (!io_op_defs[i].not_supported)
9916 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9921 if (copy_to_user(arg, p, size))
9928 static int io_register_personality(struct io_ring_ctx *ctx)
9930 struct io_identity *id;
9933 id = kmalloc(sizeof(*id), GFP_KERNEL);
9937 io_init_identity(id);
9938 id->creds = get_current_cred();
9940 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9942 put_cred(id->creds);
9948 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9949 unsigned int nr_args)
9951 struct io_uring_restriction *res;
9955 /* Restrictions allowed only if rings started disabled */
9956 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9959 /* We allow only a single restrictions registration */
9960 if (ctx->restrictions.registered)
9963 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9966 size = array_size(nr_args, sizeof(*res));
9967 if (size == SIZE_MAX)
9970 res = memdup_user(arg, size);
9972 return PTR_ERR(res);
9976 for (i = 0; i < nr_args; i++) {
9977 switch (res[i].opcode) {
9978 case IORING_RESTRICTION_REGISTER_OP:
9979 if (res[i].register_op >= IORING_REGISTER_LAST) {
9984 __set_bit(res[i].register_op,
9985 ctx->restrictions.register_op);
9987 case IORING_RESTRICTION_SQE_OP:
9988 if (res[i].sqe_op >= IORING_OP_LAST) {
9993 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9995 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9996 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9998 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9999 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10008 /* Reset all restrictions if an error happened */
10010 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10012 ctx->restrictions.registered = true;
10018 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10020 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10023 if (ctx->restrictions.registered)
10024 ctx->restricted = 1;
10026 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10028 io_sq_offload_start(ctx);
10033 static bool io_register_op_must_quiesce(int op)
10036 case IORING_UNREGISTER_FILES:
10037 case IORING_REGISTER_FILES_UPDATE:
10038 case IORING_REGISTER_PROBE:
10039 case IORING_REGISTER_PERSONALITY:
10040 case IORING_UNREGISTER_PERSONALITY:
10047 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10048 void __user *arg, unsigned nr_args)
10049 __releases(ctx->uring_lock)
10050 __acquires(ctx->uring_lock)
10055 * We're inside the ring mutex, if the ref is already dying, then
10056 * someone else killed the ctx or is already going through
10057 * io_uring_register().
10059 if (percpu_ref_is_dying(&ctx->refs))
10062 if (io_register_op_must_quiesce(opcode)) {
10063 percpu_ref_kill(&ctx->refs);
10066 * Drop uring mutex before waiting for references to exit. If
10067 * another thread is currently inside io_uring_enter() it might
10068 * need to grab the uring_lock to make progress. If we hold it
10069 * here across the drain wait, then we can deadlock. It's safe
10070 * to drop the mutex here, since no new references will come in
10071 * after we've killed the percpu ref.
10073 mutex_unlock(&ctx->uring_lock);
10075 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10078 ret = io_run_task_work_sig();
10083 mutex_lock(&ctx->uring_lock);
10086 percpu_ref_resurrect(&ctx->refs);
10091 if (ctx->restricted) {
10092 if (opcode >= IORING_REGISTER_LAST) {
10097 if (!test_bit(opcode, ctx->restrictions.register_op)) {
10104 case IORING_REGISTER_BUFFERS:
10105 ret = io_sqe_buffers_register(ctx, arg, nr_args);
10107 case IORING_UNREGISTER_BUFFERS:
10109 if (arg || nr_args)
10111 ret = io_sqe_buffers_unregister(ctx);
10113 case IORING_REGISTER_FILES:
10114 ret = io_sqe_files_register(ctx, arg, nr_args);
10116 case IORING_UNREGISTER_FILES:
10118 if (arg || nr_args)
10120 ret = io_sqe_files_unregister(ctx);
10122 case IORING_REGISTER_FILES_UPDATE:
10123 ret = io_sqe_files_update(ctx, arg, nr_args);
10125 case IORING_REGISTER_EVENTFD:
10126 case IORING_REGISTER_EVENTFD_ASYNC:
10130 ret = io_eventfd_register(ctx, arg);
10133 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10134 ctx->eventfd_async = 1;
10136 ctx->eventfd_async = 0;
10138 case IORING_UNREGISTER_EVENTFD:
10140 if (arg || nr_args)
10142 ret = io_eventfd_unregister(ctx);
10144 case IORING_REGISTER_PROBE:
10146 if (!arg || nr_args > 256)
10148 ret = io_probe(ctx, arg, nr_args);
10150 case IORING_REGISTER_PERSONALITY:
10152 if (arg || nr_args)
10154 ret = io_register_personality(ctx);
10156 case IORING_UNREGISTER_PERSONALITY:
10160 ret = io_unregister_personality(ctx, nr_args);
10162 case IORING_REGISTER_ENABLE_RINGS:
10164 if (arg || nr_args)
10166 ret = io_register_enable_rings(ctx);
10168 case IORING_REGISTER_RESTRICTIONS:
10169 ret = io_register_restrictions(ctx, arg, nr_args);
10177 if (io_register_op_must_quiesce(opcode)) {
10178 /* bring the ctx back to life */
10179 percpu_ref_reinit(&ctx->refs);
10181 reinit_completion(&ctx->ref_comp);
10186 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10187 void __user *, arg, unsigned int, nr_args)
10189 struct io_ring_ctx *ctx;
10198 if (f.file->f_op != &io_uring_fops)
10201 ctx = f.file->private_data;
10203 io_run_task_work();
10205 mutex_lock(&ctx->uring_lock);
10206 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10207 mutex_unlock(&ctx->uring_lock);
10208 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10209 ctx->cq_ev_fd != NULL, ret);
10215 static int __init io_uring_init(void)
10217 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10218 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10219 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10222 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10223 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10224 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10225 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10226 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10227 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10228 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10229 BUILD_BUG_SQE_ELEM(8, __u64, off);
10230 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10231 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10232 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10233 BUILD_BUG_SQE_ELEM(24, __u32, len);
10234 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10235 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10236 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10237 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10238 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10239 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10240 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10241 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10242 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10243 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10244 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10245 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10246 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10247 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10248 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10249 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10250 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10251 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10252 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10254 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10255 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10256 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10260 __initcall(io_uring_init);