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,
906 [IORING_OP_RECVMSG] = {
908 .unbound_nonreg_file = 1,
911 .needs_async_data = 1,
912 .async_size = sizeof(struct io_async_msghdr),
913 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
915 [IORING_OP_TIMEOUT] = {
916 .needs_async_data = 1,
917 .async_size = sizeof(struct io_timeout_data),
918 .work_flags = IO_WQ_WORK_MM,
920 [IORING_OP_TIMEOUT_REMOVE] = {
921 /* used by timeout updates' prep() */
922 .work_flags = IO_WQ_WORK_MM,
924 [IORING_OP_ACCEPT] = {
926 .unbound_nonreg_file = 1,
928 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
930 [IORING_OP_ASYNC_CANCEL] = {},
931 [IORING_OP_LINK_TIMEOUT] = {
932 .needs_async_data = 1,
933 .async_size = sizeof(struct io_timeout_data),
934 .work_flags = IO_WQ_WORK_MM,
936 [IORING_OP_CONNECT] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_data = 1,
941 .async_size = sizeof(struct io_async_connect),
942 .work_flags = IO_WQ_WORK_MM,
944 [IORING_OP_FALLOCATE] = {
946 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
948 [IORING_OP_OPENAT] = {
949 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
950 IO_WQ_WORK_FS | IO_WQ_WORK_MM,
952 [IORING_OP_CLOSE] = {
953 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
955 [IORING_OP_FILES_UPDATE] = {
956 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
958 [IORING_OP_STATX] = {
959 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
960 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
964 .unbound_nonreg_file = 1,
968 .async_size = sizeof(struct io_async_rw),
969 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
971 [IORING_OP_WRITE] = {
973 .unbound_nonreg_file = 1,
976 .async_size = sizeof(struct io_async_rw),
977 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
980 [IORING_OP_FADVISE] = {
982 .work_flags = IO_WQ_WORK_BLKCG,
984 [IORING_OP_MADVISE] = {
985 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
989 .unbound_nonreg_file = 1,
991 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
995 .unbound_nonreg_file = 1,
998 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
1000 [IORING_OP_OPENAT2] = {
1001 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
1002 IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
1004 [IORING_OP_EPOLL_CTL] = {
1005 .unbound_nonreg_file = 1,
1006 .work_flags = IO_WQ_WORK_FILES,
1008 [IORING_OP_SPLICE] = {
1011 .unbound_nonreg_file = 1,
1012 .work_flags = IO_WQ_WORK_BLKCG,
1014 [IORING_OP_PROVIDE_BUFFERS] = {},
1015 [IORING_OP_REMOVE_BUFFERS] = {},
1019 .unbound_nonreg_file = 1,
1021 [IORING_OP_SHUTDOWN] = {
1024 [IORING_OP_RENAMEAT] = {
1025 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
1026 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
1028 [IORING_OP_UNLINKAT] = {
1029 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
1030 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
1034 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1035 struct task_struct *task,
1036 struct files_struct *files);
1037 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1038 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1039 struct io_ring_ctx *ctx);
1040 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
1041 struct fixed_rsrc_ref_node *ref_node);
1043 static bool io_rw_reissue(struct io_kiocb *req);
1044 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1045 static void io_put_req(struct io_kiocb *req);
1046 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1047 static void io_double_put_req(struct io_kiocb *req);
1048 static void io_dismantle_req(struct io_kiocb *req);
1049 static void io_put_task(struct task_struct *task, int nr);
1050 static void io_queue_next(struct io_kiocb *req);
1051 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1052 static void __io_queue_linked_timeout(struct io_kiocb *req);
1053 static void io_queue_linked_timeout(struct io_kiocb *req);
1054 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1055 struct io_uring_rsrc_update *ip,
1057 static void __io_clean_op(struct io_kiocb *req);
1058 static struct file *io_file_get(struct io_submit_state *state,
1059 struct io_kiocb *req, int fd, bool fixed);
1060 static void __io_queue_sqe(struct io_kiocb *req);
1061 static void io_rsrc_put_work(struct work_struct *work);
1063 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1064 struct iov_iter *iter, bool needs_lock);
1065 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1066 const struct iovec *fast_iov,
1067 struct iov_iter *iter, bool force);
1068 static void io_req_task_queue(struct io_kiocb *req);
1069 static void io_submit_flush_completions(struct io_comp_state *cs,
1070 struct io_ring_ctx *ctx);
1072 static struct kmem_cache *req_cachep;
1074 static const struct file_operations io_uring_fops;
1076 struct sock *io_uring_get_socket(struct file *file)
1078 #if defined(CONFIG_UNIX)
1079 if (file->f_op == &io_uring_fops) {
1080 struct io_ring_ctx *ctx = file->private_data;
1082 return ctx->ring_sock->sk;
1087 EXPORT_SYMBOL(io_uring_get_socket);
1089 #define io_for_each_link(pos, head) \
1090 for (pos = (head); pos; pos = pos->link)
1092 static inline void io_clean_op(struct io_kiocb *req)
1094 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1098 static inline void io_set_resource_node(struct io_kiocb *req)
1100 struct io_ring_ctx *ctx = req->ctx;
1102 if (!req->fixed_rsrc_refs) {
1103 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1104 percpu_ref_get(req->fixed_rsrc_refs);
1108 static bool io_match_task(struct io_kiocb *head,
1109 struct task_struct *task,
1110 struct files_struct *files)
1112 struct io_kiocb *req;
1114 if (task && head->task != task) {
1115 /* in terms of cancelation, always match if req task is dead */
1116 if (head->task->flags & PF_EXITING)
1123 io_for_each_link(req, head) {
1124 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1126 if (req->file && req->file->f_op == &io_uring_fops)
1128 if ((req->work.flags & IO_WQ_WORK_FILES) &&
1129 req->work.identity->files == files)
1135 static void io_sq_thread_drop_mm_files(void)
1137 struct files_struct *files = current->files;
1138 struct mm_struct *mm = current->mm;
1141 kthread_unuse_mm(mm);
1146 struct nsproxy *nsproxy = current->nsproxy;
1149 current->files = NULL;
1150 current->nsproxy = NULL;
1151 task_unlock(current);
1152 put_files_struct(files);
1153 put_nsproxy(nsproxy);
1157 static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1159 if (!current->files) {
1160 struct files_struct *files;
1161 struct nsproxy *nsproxy;
1163 task_lock(ctx->sqo_task);
1164 files = ctx->sqo_task->files;
1166 task_unlock(ctx->sqo_task);
1169 atomic_inc(&files->count);
1170 get_nsproxy(ctx->sqo_task->nsproxy);
1171 nsproxy = ctx->sqo_task->nsproxy;
1172 task_unlock(ctx->sqo_task);
1175 current->files = files;
1176 current->nsproxy = nsproxy;
1177 task_unlock(current);
1182 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1184 struct mm_struct *mm;
1189 task_lock(ctx->sqo_task);
1190 mm = ctx->sqo_task->mm;
1191 if (unlikely(!mm || !mmget_not_zero(mm)))
1193 task_unlock(ctx->sqo_task);
1203 static int __io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1204 struct io_kiocb *req)
1206 const struct io_op_def *def = &io_op_defs[req->opcode];
1209 if (def->work_flags & IO_WQ_WORK_MM) {
1210 ret = __io_sq_thread_acquire_mm(ctx);
1215 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) {
1216 ret = __io_sq_thread_acquire_files(ctx);
1224 static inline int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1225 struct io_kiocb *req)
1227 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1229 return __io_sq_thread_acquire_mm_files(ctx, req);
1232 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1233 struct cgroup_subsys_state **cur_css)
1236 #ifdef CONFIG_BLK_CGROUP
1237 /* puts the old one when swapping */
1238 if (*cur_css != ctx->sqo_blkcg_css) {
1239 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1240 *cur_css = ctx->sqo_blkcg_css;
1245 static void io_sq_thread_unassociate_blkcg(void)
1247 #ifdef CONFIG_BLK_CGROUP
1248 kthread_associate_blkcg(NULL);
1252 static inline void req_set_fail_links(struct io_kiocb *req)
1254 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1255 req->flags |= REQ_F_FAIL_LINK;
1259 * None of these are dereferenced, they are simply used to check if any of
1260 * them have changed. If we're under current and check they are still the
1261 * same, we're fine to grab references to them for actual out-of-line use.
1263 static void io_init_identity(struct io_identity *id)
1265 id->files = current->files;
1266 id->mm = current->mm;
1267 #ifdef CONFIG_BLK_CGROUP
1269 id->blkcg_css = blkcg_css();
1272 id->creds = current_cred();
1273 id->nsproxy = current->nsproxy;
1274 id->fs = current->fs;
1275 id->fsize = rlimit(RLIMIT_FSIZE);
1277 id->loginuid = current->loginuid;
1278 id->sessionid = current->sessionid;
1280 refcount_set(&id->count, 1);
1283 static inline void __io_req_init_async(struct io_kiocb *req)
1285 memset(&req->work, 0, sizeof(req->work));
1286 req->flags |= REQ_F_WORK_INITIALIZED;
1290 * Note: must call io_req_init_async() for the first time you
1291 * touch any members of io_wq_work.
1293 static inline void io_req_init_async(struct io_kiocb *req)
1295 struct io_uring_task *tctx = current->io_uring;
1297 if (req->flags & REQ_F_WORK_INITIALIZED)
1300 __io_req_init_async(req);
1302 /* Grab a ref if this isn't our static identity */
1303 req->work.identity = tctx->identity;
1304 if (tctx->identity != &tctx->__identity)
1305 refcount_inc(&req->work.identity->count);
1308 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1310 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1312 complete(&ctx->ref_comp);
1315 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1317 return !req->timeout.off;
1320 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1322 struct io_ring_ctx *ctx;
1325 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1330 * Use 5 bits less than the max cq entries, that should give us around
1331 * 32 entries per hash list if totally full and uniformly spread.
1333 hash_bits = ilog2(p->cq_entries);
1337 ctx->cancel_hash_bits = hash_bits;
1338 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1340 if (!ctx->cancel_hash)
1342 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1344 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1345 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1348 ctx->flags = p->flags;
1349 init_waitqueue_head(&ctx->sqo_sq_wait);
1350 INIT_LIST_HEAD(&ctx->sqd_list);
1351 init_waitqueue_head(&ctx->cq_wait);
1352 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1353 init_completion(&ctx->ref_comp);
1354 init_completion(&ctx->sq_thread_comp);
1355 idr_init(&ctx->io_buffer_idr);
1356 idr_init(&ctx->personality_idr);
1357 mutex_init(&ctx->uring_lock);
1358 init_waitqueue_head(&ctx->wait);
1359 spin_lock_init(&ctx->completion_lock);
1360 INIT_LIST_HEAD(&ctx->iopoll_list);
1361 INIT_LIST_HEAD(&ctx->defer_list);
1362 INIT_LIST_HEAD(&ctx->timeout_list);
1363 spin_lock_init(&ctx->inflight_lock);
1364 INIT_LIST_HEAD(&ctx->inflight_list);
1365 spin_lock_init(&ctx->rsrc_ref_lock);
1366 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1367 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1368 init_llist_head(&ctx->rsrc_put_llist);
1369 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1370 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1373 kfree(ctx->cancel_hash);
1378 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1380 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1381 struct io_ring_ctx *ctx = req->ctx;
1383 return seq != ctx->cached_cq_tail
1384 + READ_ONCE(ctx->cached_cq_overflow);
1390 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1392 if (req->work.identity == &tctx->__identity)
1394 if (refcount_dec_and_test(&req->work.identity->count))
1395 kfree(req->work.identity);
1398 static void io_req_clean_work(struct io_kiocb *req)
1400 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1403 if (req->work.flags & IO_WQ_WORK_MM)
1404 mmdrop(req->work.identity->mm);
1405 #ifdef CONFIG_BLK_CGROUP
1406 if (req->work.flags & IO_WQ_WORK_BLKCG)
1407 css_put(req->work.identity->blkcg_css);
1409 if (req->work.flags & IO_WQ_WORK_CREDS)
1410 put_cred(req->work.identity->creds);
1411 if (req->work.flags & IO_WQ_WORK_FS) {
1412 struct fs_struct *fs = req->work.identity->fs;
1414 spin_lock(&req->work.identity->fs->lock);
1417 spin_unlock(&req->work.identity->fs->lock);
1421 if (req->work.flags & IO_WQ_WORK_FILES) {
1422 put_files_struct(req->work.identity->files);
1423 put_nsproxy(req->work.identity->nsproxy);
1425 if (req->flags & REQ_F_INFLIGHT) {
1426 struct io_ring_ctx *ctx = req->ctx;
1427 struct io_uring_task *tctx = req->task->io_uring;
1428 unsigned long flags;
1430 spin_lock_irqsave(&ctx->inflight_lock, flags);
1431 list_del(&req->inflight_entry);
1432 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1433 req->flags &= ~REQ_F_INFLIGHT;
1434 if (atomic_read(&tctx->in_idle))
1435 wake_up(&tctx->wait);
1438 req->flags &= ~REQ_F_WORK_INITIALIZED;
1439 req->work.flags &= ~(IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG | IO_WQ_WORK_FS |
1440 IO_WQ_WORK_CREDS | IO_WQ_WORK_FILES);
1441 io_put_identity(req->task->io_uring, req);
1445 * Create a private copy of io_identity, since some fields don't match
1446 * the current context.
1448 static bool io_identity_cow(struct io_kiocb *req)
1450 struct io_uring_task *tctx = current->io_uring;
1451 const struct cred *creds = NULL;
1452 struct io_identity *id;
1454 if (req->work.flags & IO_WQ_WORK_CREDS)
1455 creds = req->work.identity->creds;
1457 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1458 if (unlikely(!id)) {
1459 req->work.flags |= IO_WQ_WORK_CANCEL;
1464 * We can safely just re-init the creds we copied Either the field
1465 * matches the current one, or we haven't grabbed it yet. The only
1466 * exception is ->creds, through registered personalities, so handle
1467 * that one separately.
1469 io_init_identity(id);
1473 /* add one for this request */
1474 refcount_inc(&id->count);
1476 /* drop tctx and req identity references, if needed */
1477 if (tctx->identity != &tctx->__identity &&
1478 refcount_dec_and_test(&tctx->identity->count))
1479 kfree(tctx->identity);
1480 if (req->work.identity != &tctx->__identity &&
1481 refcount_dec_and_test(&req->work.identity->count))
1482 kfree(req->work.identity);
1484 req->work.identity = id;
1485 tctx->identity = id;
1489 static void io_req_track_inflight(struct io_kiocb *req)
1491 struct io_ring_ctx *ctx = req->ctx;
1493 if (!(req->flags & REQ_F_INFLIGHT)) {
1494 io_req_init_async(req);
1495 req->flags |= REQ_F_INFLIGHT;
1497 spin_lock_irq(&ctx->inflight_lock);
1498 list_add(&req->inflight_entry, &ctx->inflight_list);
1499 spin_unlock_irq(&ctx->inflight_lock);
1503 static bool io_grab_identity(struct io_kiocb *req)
1505 const struct io_op_def *def = &io_op_defs[req->opcode];
1506 struct io_identity *id = req->work.identity;
1508 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1509 if (id->fsize != rlimit(RLIMIT_FSIZE))
1511 req->work.flags |= IO_WQ_WORK_FSIZE;
1513 #ifdef CONFIG_BLK_CGROUP
1514 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1515 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1517 if (id->blkcg_css != blkcg_css()) {
1522 * This should be rare, either the cgroup is dying or the task
1523 * is moving cgroups. Just punt to root for the handful of ios.
1525 if (css_tryget_online(id->blkcg_css))
1526 req->work.flags |= IO_WQ_WORK_BLKCG;
1530 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1531 if (id->creds != current_cred())
1533 get_cred(id->creds);
1534 req->work.flags |= IO_WQ_WORK_CREDS;
1537 if (!uid_eq(current->loginuid, id->loginuid) ||
1538 current->sessionid != id->sessionid)
1541 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1542 (def->work_flags & IO_WQ_WORK_FS)) {
1543 if (current->fs != id->fs)
1545 spin_lock(&id->fs->lock);
1546 if (!id->fs->in_exec) {
1548 req->work.flags |= IO_WQ_WORK_FS;
1550 req->work.flags |= IO_WQ_WORK_CANCEL;
1552 spin_unlock(¤t->fs->lock);
1554 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1555 (def->work_flags & IO_WQ_WORK_FILES) &&
1556 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1557 if (id->files != current->files ||
1558 id->nsproxy != current->nsproxy)
1560 atomic_inc(&id->files->count);
1561 get_nsproxy(id->nsproxy);
1562 req->work.flags |= IO_WQ_WORK_FILES;
1563 io_req_track_inflight(req);
1565 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1566 (def->work_flags & IO_WQ_WORK_MM)) {
1567 if (id->mm != current->mm)
1570 req->work.flags |= IO_WQ_WORK_MM;
1576 static void io_prep_async_work(struct io_kiocb *req)
1578 const struct io_op_def *def = &io_op_defs[req->opcode];
1579 struct io_ring_ctx *ctx = req->ctx;
1581 io_req_init_async(req);
1583 if (req->flags & REQ_F_FORCE_ASYNC)
1584 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1586 if (req->flags & REQ_F_ISREG) {
1587 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1588 io_wq_hash_work(&req->work, file_inode(req->file));
1590 if (def->unbound_nonreg_file)
1591 req->work.flags |= IO_WQ_WORK_UNBOUND;
1594 /* if we fail grabbing identity, we must COW, regrab, and retry */
1595 if (io_grab_identity(req))
1598 if (!io_identity_cow(req))
1601 /* can't fail at this point */
1602 if (!io_grab_identity(req))
1606 static void io_prep_async_link(struct io_kiocb *req)
1608 struct io_kiocb *cur;
1610 io_for_each_link(cur, req)
1611 io_prep_async_work(cur);
1614 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1616 struct io_ring_ctx *ctx = req->ctx;
1617 struct io_kiocb *link = io_prep_linked_timeout(req);
1619 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1620 &req->work, req->flags);
1621 io_wq_enqueue(ctx->io_wq, &req->work);
1625 static void io_queue_async_work(struct io_kiocb *req)
1627 struct io_kiocb *link;
1629 /* init ->work of the whole link before punting */
1630 io_prep_async_link(req);
1631 link = __io_queue_async_work(req);
1634 io_queue_linked_timeout(link);
1637 static void io_kill_timeout(struct io_kiocb *req)
1639 struct io_timeout_data *io = req->async_data;
1642 ret = hrtimer_try_to_cancel(&io->timer);
1644 atomic_set(&req->ctx->cq_timeouts,
1645 atomic_read(&req->ctx->cq_timeouts) + 1);
1646 list_del_init(&req->timeout.list);
1647 io_cqring_fill_event(req, 0);
1648 io_put_req_deferred(req, 1);
1653 * Returns true if we found and killed one or more timeouts
1655 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1656 struct files_struct *files)
1658 struct io_kiocb *req, *tmp;
1661 spin_lock_irq(&ctx->completion_lock);
1662 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1663 if (io_match_task(req, tsk, files)) {
1664 io_kill_timeout(req);
1668 spin_unlock_irq(&ctx->completion_lock);
1669 return canceled != 0;
1672 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1675 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1676 struct io_defer_entry, list);
1678 if (req_need_defer(de->req, de->seq))
1680 list_del_init(&de->list);
1681 io_req_task_queue(de->req);
1683 } while (!list_empty(&ctx->defer_list));
1686 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1690 if (list_empty(&ctx->timeout_list))
1693 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1696 u32 events_needed, events_got;
1697 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1698 struct io_kiocb, timeout.list);
1700 if (io_is_timeout_noseq(req))
1704 * Since seq can easily wrap around over time, subtract
1705 * the last seq at which timeouts were flushed before comparing.
1706 * Assuming not more than 2^31-1 events have happened since,
1707 * these subtractions won't have wrapped, so we can check if
1708 * target is in [last_seq, current_seq] by comparing the two.
1710 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1711 events_got = seq - ctx->cq_last_tm_flush;
1712 if (events_got < events_needed)
1715 list_del_init(&req->timeout.list);
1716 io_kill_timeout(req);
1717 } while (!list_empty(&ctx->timeout_list));
1719 ctx->cq_last_tm_flush = seq;
1722 static void io_commit_cqring(struct io_ring_ctx *ctx)
1724 io_flush_timeouts(ctx);
1726 /* order cqe stores with ring update */
1727 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1729 if (unlikely(!list_empty(&ctx->defer_list)))
1730 __io_queue_deferred(ctx);
1733 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1735 struct io_rings *r = ctx->rings;
1737 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1740 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1742 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1745 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1747 struct io_rings *rings = ctx->rings;
1751 * writes to the cq entry need to come after reading head; the
1752 * control dependency is enough as we're using WRITE_ONCE to
1755 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1758 tail = ctx->cached_cq_tail++;
1759 return &rings->cqes[tail & ctx->cq_mask];
1762 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1766 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1768 if (!ctx->eventfd_async)
1770 return io_wq_current_is_worker();
1773 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1775 /* see waitqueue_active() comment */
1778 if (waitqueue_active(&ctx->wait))
1779 wake_up(&ctx->wait);
1780 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1781 wake_up(&ctx->sq_data->wait);
1782 if (io_should_trigger_evfd(ctx))
1783 eventfd_signal(ctx->cq_ev_fd, 1);
1784 if (waitqueue_active(&ctx->cq_wait)) {
1785 wake_up_interruptible(&ctx->cq_wait);
1786 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1790 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1792 /* see waitqueue_active() comment */
1795 if (ctx->flags & IORING_SETUP_SQPOLL) {
1796 if (waitqueue_active(&ctx->wait))
1797 wake_up(&ctx->wait);
1799 if (io_should_trigger_evfd(ctx))
1800 eventfd_signal(ctx->cq_ev_fd, 1);
1801 if (waitqueue_active(&ctx->cq_wait)) {
1802 wake_up_interruptible(&ctx->cq_wait);
1803 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1807 /* Returns true if there are no backlogged entries after the flush */
1808 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1809 struct task_struct *tsk,
1810 struct files_struct *files)
1812 struct io_rings *rings = ctx->rings;
1813 struct io_kiocb *req, *tmp;
1814 struct io_uring_cqe *cqe;
1815 unsigned long flags;
1816 bool all_flushed, posted;
1819 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1823 spin_lock_irqsave(&ctx->completion_lock, flags);
1824 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1825 if (!io_match_task(req, tsk, files))
1828 cqe = io_get_cqring(ctx);
1832 list_move(&req->compl.list, &list);
1834 WRITE_ONCE(cqe->user_data, req->user_data);
1835 WRITE_ONCE(cqe->res, req->result);
1836 WRITE_ONCE(cqe->flags, req->compl.cflags);
1838 ctx->cached_cq_overflow++;
1839 WRITE_ONCE(ctx->rings->cq_overflow,
1840 ctx->cached_cq_overflow);
1845 all_flushed = list_empty(&ctx->cq_overflow_list);
1847 clear_bit(0, &ctx->sq_check_overflow);
1848 clear_bit(0, &ctx->cq_check_overflow);
1849 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1853 io_commit_cqring(ctx);
1854 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1856 io_cqring_ev_posted(ctx);
1858 while (!list_empty(&list)) {
1859 req = list_first_entry(&list, struct io_kiocb, compl.list);
1860 list_del(&req->compl.list);
1867 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1868 struct task_struct *tsk,
1869 struct files_struct *files)
1871 if (test_bit(0, &ctx->cq_check_overflow)) {
1872 /* iopoll syncs against uring_lock, not completion_lock */
1873 if (ctx->flags & IORING_SETUP_IOPOLL)
1874 mutex_lock(&ctx->uring_lock);
1875 __io_cqring_overflow_flush(ctx, force, tsk, files);
1876 if (ctx->flags & IORING_SETUP_IOPOLL)
1877 mutex_unlock(&ctx->uring_lock);
1881 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1883 struct io_ring_ctx *ctx = req->ctx;
1884 struct io_uring_cqe *cqe;
1886 trace_io_uring_complete(ctx, req->user_data, res);
1889 * If we can't get a cq entry, userspace overflowed the
1890 * submission (by quite a lot). Increment the overflow count in
1893 cqe = io_get_cqring(ctx);
1895 WRITE_ONCE(cqe->user_data, req->user_data);
1896 WRITE_ONCE(cqe->res, res);
1897 WRITE_ONCE(cqe->flags, cflags);
1898 } else if (ctx->cq_overflow_flushed ||
1899 atomic_read(&req->task->io_uring->in_idle)) {
1901 * If we're in ring overflow flush mode, or in task cancel mode,
1902 * then we cannot store the request for later flushing, we need
1903 * to drop it on the floor.
1905 ctx->cached_cq_overflow++;
1906 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1908 if (list_empty(&ctx->cq_overflow_list)) {
1909 set_bit(0, &ctx->sq_check_overflow);
1910 set_bit(0, &ctx->cq_check_overflow);
1911 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1915 req->compl.cflags = cflags;
1916 refcount_inc(&req->refs);
1917 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1921 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1923 __io_cqring_fill_event(req, res, 0);
1926 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1927 unsigned int cflags)
1929 struct io_ring_ctx *ctx = req->ctx;
1930 unsigned long flags;
1932 spin_lock_irqsave(&ctx->completion_lock, flags);
1933 __io_cqring_fill_event(req, res, cflags);
1934 io_commit_cqring(ctx);
1936 * If we're the last reference to this request, add to our locked
1939 if (refcount_dec_and_test(&req->refs)) {
1940 struct io_comp_state *cs = &ctx->submit_state.comp;
1942 io_dismantle_req(req);
1943 io_put_task(req->task, 1);
1944 list_add(&req->compl.list, &cs->locked_free_list);
1945 cs->locked_free_nr++;
1948 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1950 io_cqring_ev_posted(ctx);
1953 percpu_ref_put(&ctx->refs);
1957 static void io_req_complete_state(struct io_kiocb *req, long res,
1958 unsigned int cflags)
1962 req->compl.cflags = cflags;
1963 req->flags |= REQ_F_COMPLETE_INLINE;
1966 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1967 long res, unsigned cflags)
1969 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1970 io_req_complete_state(req, res, cflags);
1972 io_req_complete_post(req, res, cflags);
1975 static inline void io_req_complete(struct io_kiocb *req, long res)
1977 __io_req_complete(req, 0, res, 0);
1980 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1982 struct io_submit_state *state = &ctx->submit_state;
1983 struct io_comp_state *cs = &state->comp;
1984 struct io_kiocb *req = NULL;
1987 * If we have more than a batch's worth of requests in our IRQ side
1988 * locked cache, grab the lock and move them over to our submission
1991 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1992 spin_lock_irq(&ctx->completion_lock);
1993 list_splice_init(&cs->locked_free_list, &cs->free_list);
1994 cs->locked_free_nr = 0;
1995 spin_unlock_irq(&ctx->completion_lock);
1998 while (!list_empty(&cs->free_list)) {
1999 req = list_first_entry(&cs->free_list, struct io_kiocb,
2001 list_del(&req->compl.list);
2002 state->reqs[state->free_reqs++] = req;
2003 if (state->free_reqs == ARRAY_SIZE(state->reqs))
2010 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
2012 struct io_submit_state *state = &ctx->submit_state;
2014 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
2016 if (!state->free_reqs) {
2017 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
2020 if (io_flush_cached_reqs(ctx))
2023 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
2027 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2028 * retry single alloc to be on the safe side.
2030 if (unlikely(ret <= 0)) {
2031 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
2032 if (!state->reqs[0])
2036 state->free_reqs = ret;
2040 return state->reqs[state->free_reqs];
2043 static inline void io_put_file(struct io_kiocb *req, struct file *file,
2050 static void io_dismantle_req(struct io_kiocb *req)
2054 if (req->async_data)
2055 kfree(req->async_data);
2057 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
2058 if (req->fixed_rsrc_refs)
2059 percpu_ref_put(req->fixed_rsrc_refs);
2060 io_req_clean_work(req);
2063 static inline void io_put_task(struct task_struct *task, int nr)
2065 struct io_uring_task *tctx = task->io_uring;
2067 percpu_counter_sub(&tctx->inflight, nr);
2068 if (unlikely(atomic_read(&tctx->in_idle)))
2069 wake_up(&tctx->wait);
2070 put_task_struct_many(task, nr);
2073 static void __io_free_req(struct io_kiocb *req)
2075 struct io_ring_ctx *ctx = req->ctx;
2077 io_dismantle_req(req);
2078 io_put_task(req->task, 1);
2080 kmem_cache_free(req_cachep, req);
2081 percpu_ref_put(&ctx->refs);
2084 static inline void io_remove_next_linked(struct io_kiocb *req)
2086 struct io_kiocb *nxt = req->link;
2088 req->link = nxt->link;
2092 static void io_kill_linked_timeout(struct io_kiocb *req)
2094 struct io_ring_ctx *ctx = req->ctx;
2095 struct io_kiocb *link;
2096 bool cancelled = false;
2097 unsigned long flags;
2099 spin_lock_irqsave(&ctx->completion_lock, flags);
2103 * Can happen if a linked timeout fired and link had been like
2104 * req -> link t-out -> link t-out [-> ...]
2106 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
2107 struct io_timeout_data *io = link->async_data;
2110 io_remove_next_linked(req);
2111 link->timeout.head = NULL;
2112 ret = hrtimer_try_to_cancel(&io->timer);
2114 io_cqring_fill_event(link, -ECANCELED);
2115 io_commit_cqring(ctx);
2119 req->flags &= ~REQ_F_LINK_TIMEOUT;
2120 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2123 io_cqring_ev_posted(ctx);
2129 static void io_fail_links(struct io_kiocb *req)
2131 struct io_kiocb *link, *nxt;
2132 struct io_ring_ctx *ctx = req->ctx;
2133 unsigned long flags;
2135 spin_lock_irqsave(&ctx->completion_lock, flags);
2143 trace_io_uring_fail_link(req, link);
2144 io_cqring_fill_event(link, -ECANCELED);
2147 * It's ok to free under spinlock as they're not linked anymore,
2148 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2151 if (link->flags & REQ_F_WORK_INITIALIZED)
2152 io_put_req_deferred(link, 2);
2154 io_double_put_req(link);
2157 io_commit_cqring(ctx);
2158 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2160 io_cqring_ev_posted(ctx);
2163 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2165 if (req->flags & REQ_F_LINK_TIMEOUT)
2166 io_kill_linked_timeout(req);
2169 * If LINK is set, we have dependent requests in this chain. If we
2170 * didn't fail this request, queue the first one up, moving any other
2171 * dependencies to the next request. In case of failure, fail the rest
2174 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2175 struct io_kiocb *nxt = req->link;
2184 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2186 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2188 return __io_req_find_next(req);
2191 static bool __tctx_task_work(struct io_uring_task *tctx)
2193 struct io_ring_ctx *ctx = NULL;
2194 struct io_wq_work_list list;
2195 struct io_wq_work_node *node;
2197 if (wq_list_empty(&tctx->task_list))
2200 spin_lock_irq(&tctx->task_lock);
2201 list = tctx->task_list;
2202 INIT_WQ_LIST(&tctx->task_list);
2203 spin_unlock_irq(&tctx->task_lock);
2207 struct io_wq_work_node *next = node->next;
2208 struct io_ring_ctx *this_ctx;
2209 struct io_kiocb *req;
2211 req = container_of(node, struct io_kiocb, io_task_work.node);
2212 this_ctx = req->ctx;
2213 req->task_work.func(&req->task_work);
2218 } else if (ctx != this_ctx) {
2219 mutex_lock(&ctx->uring_lock);
2220 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
2221 mutex_unlock(&ctx->uring_lock);
2226 if (ctx && ctx->submit_state.comp.nr) {
2227 mutex_lock(&ctx->uring_lock);
2228 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
2229 mutex_unlock(&ctx->uring_lock);
2232 return list.first != NULL;
2235 static void tctx_task_work(struct callback_head *cb)
2237 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
2239 while (__tctx_task_work(tctx))
2242 clear_bit(0, &tctx->task_state);
2245 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
2246 enum task_work_notify_mode notify)
2248 struct io_uring_task *tctx = tsk->io_uring;
2249 struct io_wq_work_node *node, *prev;
2250 unsigned long flags;
2253 WARN_ON_ONCE(!tctx);
2255 spin_lock_irqsave(&tctx->task_lock, flags);
2256 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2257 spin_unlock_irqrestore(&tctx->task_lock, flags);
2259 /* task_work already pending, we're done */
2260 if (test_bit(0, &tctx->task_state) ||
2261 test_and_set_bit(0, &tctx->task_state))
2264 if (!task_work_add(tsk, &tctx->task_work, notify))
2268 * Slow path - we failed, find and delete work. if the work is not
2269 * in the list, it got run and we're fine.
2272 spin_lock_irqsave(&tctx->task_lock, flags);
2273 wq_list_for_each(node, prev, &tctx->task_list) {
2274 if (&req->io_task_work.node == node) {
2275 wq_list_del(&tctx->task_list, node, prev);
2280 spin_unlock_irqrestore(&tctx->task_lock, flags);
2281 clear_bit(0, &tctx->task_state);
2285 static int io_req_task_work_add(struct io_kiocb *req)
2287 struct task_struct *tsk = req->task;
2288 struct io_ring_ctx *ctx = req->ctx;
2289 enum task_work_notify_mode notify;
2292 if (tsk->flags & PF_EXITING)
2296 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2297 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2298 * processing task_work. There's no reliable way to tell if TWA_RESUME
2302 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2303 notify = TWA_SIGNAL;
2305 ret = io_task_work_add(tsk, req, notify);
2307 wake_up_process(tsk);
2312 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2313 task_work_func_t cb)
2315 struct task_struct *tsk = io_wq_get_task(req->ctx->io_wq);
2317 init_task_work(&req->task_work, cb);
2318 task_work_add(tsk, &req->task_work, TWA_NONE);
2319 wake_up_process(tsk);
2322 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2324 struct io_ring_ctx *ctx = req->ctx;
2326 spin_lock_irq(&ctx->completion_lock);
2327 io_cqring_fill_event(req, error);
2328 io_commit_cqring(ctx);
2329 spin_unlock_irq(&ctx->completion_lock);
2331 io_cqring_ev_posted(ctx);
2332 req_set_fail_links(req);
2333 io_double_put_req(req);
2336 static void io_req_task_cancel(struct callback_head *cb)
2338 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2339 struct io_ring_ctx *ctx = req->ctx;
2341 mutex_lock(&ctx->uring_lock);
2342 __io_req_task_cancel(req, req->result);
2343 mutex_unlock(&ctx->uring_lock);
2344 percpu_ref_put(&ctx->refs);
2347 static void __io_req_task_submit(struct io_kiocb *req)
2349 struct io_ring_ctx *ctx = req->ctx;
2351 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2352 mutex_lock(&ctx->uring_lock);
2353 if (!ctx->sqo_dead && !(current->flags & PF_EXITING) &&
2354 !io_sq_thread_acquire_mm_files(ctx, req))
2355 __io_queue_sqe(req);
2357 __io_req_task_cancel(req, -EFAULT);
2358 mutex_unlock(&ctx->uring_lock);
2361 static void io_req_task_submit(struct callback_head *cb)
2363 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2365 __io_req_task_submit(req);
2368 static void io_req_task_queue(struct io_kiocb *req)
2372 req->task_work.func = io_req_task_submit;
2373 ret = io_req_task_work_add(req);
2374 if (unlikely(ret)) {
2375 req->result = -ECANCELED;
2376 percpu_ref_get(&req->ctx->refs);
2377 io_req_task_work_add_fallback(req, io_req_task_cancel);
2381 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2383 percpu_ref_get(&req->ctx->refs);
2385 req->task_work.func = io_req_task_cancel;
2387 if (unlikely(io_req_task_work_add(req)))
2388 io_req_task_work_add_fallback(req, io_req_task_cancel);
2391 static inline void io_queue_next(struct io_kiocb *req)
2393 struct io_kiocb *nxt = io_req_find_next(req);
2396 io_req_task_queue(nxt);
2399 static void io_free_req(struct io_kiocb *req)
2406 struct task_struct *task;
2411 static inline void io_init_req_batch(struct req_batch *rb)
2418 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2419 struct req_batch *rb)
2422 io_put_task(rb->task, rb->task_refs);
2424 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2427 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2428 struct io_submit_state *state)
2432 if (req->task != rb->task) {
2434 io_put_task(rb->task, rb->task_refs);
2435 rb->task = req->task;
2441 io_dismantle_req(req);
2442 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2443 state->reqs[state->free_reqs++] = req;
2445 list_add(&req->compl.list, &state->comp.free_list);
2448 static void io_submit_flush_completions(struct io_comp_state *cs,
2449 struct io_ring_ctx *ctx)
2452 struct io_kiocb *req;
2453 struct req_batch rb;
2455 io_init_req_batch(&rb);
2456 spin_lock_irq(&ctx->completion_lock);
2457 for (i = 0; i < nr; i++) {
2459 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2461 io_commit_cqring(ctx);
2462 spin_unlock_irq(&ctx->completion_lock);
2464 io_cqring_ev_posted(ctx);
2465 for (i = 0; i < nr; i++) {
2468 /* submission and completion refs */
2469 if (refcount_sub_and_test(2, &req->refs))
2470 io_req_free_batch(&rb, req, &ctx->submit_state);
2473 io_req_free_batch_finish(ctx, &rb);
2478 * Drop reference to request, return next in chain (if there is one) if this
2479 * was the last reference to this request.
2481 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2483 struct io_kiocb *nxt = NULL;
2485 if (refcount_dec_and_test(&req->refs)) {
2486 nxt = io_req_find_next(req);
2492 static void io_put_req(struct io_kiocb *req)
2494 if (refcount_dec_and_test(&req->refs))
2498 static void io_put_req_deferred_cb(struct callback_head *cb)
2500 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2505 static void io_free_req_deferred(struct io_kiocb *req)
2509 req->task_work.func = io_put_req_deferred_cb;
2510 ret = io_req_task_work_add(req);
2512 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2515 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2517 if (refcount_sub_and_test(refs, &req->refs))
2518 io_free_req_deferred(req);
2521 static void io_double_put_req(struct io_kiocb *req)
2523 /* drop both submit and complete references */
2524 if (refcount_sub_and_test(2, &req->refs))
2528 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2530 /* See comment at the top of this file */
2532 return __io_cqring_events(ctx);
2535 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2537 struct io_rings *rings = ctx->rings;
2539 /* make sure SQ entry isn't read before tail */
2540 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2543 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2545 unsigned int cflags;
2547 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2548 cflags |= IORING_CQE_F_BUFFER;
2549 req->flags &= ~REQ_F_BUFFER_SELECTED;
2554 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2556 struct io_buffer *kbuf;
2558 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2559 return io_put_kbuf(req, kbuf);
2562 static inline bool io_run_task_work(void)
2565 * Not safe to run on exiting task, and the task_work handling will
2566 * not add work to such a task.
2568 if (unlikely(current->flags & PF_EXITING))
2570 if (current->task_works) {
2571 __set_current_state(TASK_RUNNING);
2580 * Find and free completed poll iocbs
2582 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2583 struct list_head *done)
2585 struct req_batch rb;
2586 struct io_kiocb *req;
2588 /* order with ->result store in io_complete_rw_iopoll() */
2591 io_init_req_batch(&rb);
2592 while (!list_empty(done)) {
2595 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2596 list_del(&req->inflight_entry);
2598 if (READ_ONCE(req->result) == -EAGAIN) {
2599 req->iopoll_completed = 0;
2600 if (io_rw_reissue(req))
2604 if (req->flags & REQ_F_BUFFER_SELECTED)
2605 cflags = io_put_rw_kbuf(req);
2607 __io_cqring_fill_event(req, req->result, cflags);
2610 if (refcount_dec_and_test(&req->refs))
2611 io_req_free_batch(&rb, req, &ctx->submit_state);
2614 io_commit_cqring(ctx);
2615 io_cqring_ev_posted_iopoll(ctx);
2616 io_req_free_batch_finish(ctx, &rb);
2619 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2622 struct io_kiocb *req, *tmp;
2628 * Only spin for completions if we don't have multiple devices hanging
2629 * off our complete list, and we're under the requested amount.
2631 spin = !ctx->poll_multi_file && *nr_events < min;
2634 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2635 struct kiocb *kiocb = &req->rw.kiocb;
2638 * Move completed and retryable entries to our local lists.
2639 * If we find a request that requires polling, break out
2640 * and complete those lists first, if we have entries there.
2642 if (READ_ONCE(req->iopoll_completed)) {
2643 list_move_tail(&req->inflight_entry, &done);
2646 if (!list_empty(&done))
2649 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2653 /* iopoll may have completed current req */
2654 if (READ_ONCE(req->iopoll_completed))
2655 list_move_tail(&req->inflight_entry, &done);
2662 if (!list_empty(&done))
2663 io_iopoll_complete(ctx, nr_events, &done);
2669 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2670 * non-spinning poll check - we'll still enter the driver poll loop, but only
2671 * as a non-spinning completion check.
2673 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2676 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2679 ret = io_do_iopoll(ctx, nr_events, min);
2682 if (*nr_events >= min)
2690 * We can't just wait for polled events to come to us, we have to actively
2691 * find and complete them.
2693 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2695 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2698 mutex_lock(&ctx->uring_lock);
2699 while (!list_empty(&ctx->iopoll_list)) {
2700 unsigned int nr_events = 0;
2702 io_do_iopoll(ctx, &nr_events, 0);
2704 /* let it sleep and repeat later if can't complete a request */
2708 * Ensure we allow local-to-the-cpu processing to take place,
2709 * in this case we need to ensure that we reap all events.
2710 * Also let task_work, etc. to progress by releasing the mutex
2712 if (need_resched()) {
2713 mutex_unlock(&ctx->uring_lock);
2715 mutex_lock(&ctx->uring_lock);
2718 mutex_unlock(&ctx->uring_lock);
2721 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2723 unsigned int nr_events = 0;
2724 int iters = 0, ret = 0;
2727 * We disallow the app entering submit/complete with polling, but we
2728 * still need to lock the ring to prevent racing with polled issue
2729 * that got punted to a workqueue.
2731 mutex_lock(&ctx->uring_lock);
2734 * Don't enter poll loop if we already have events pending.
2735 * If we do, we can potentially be spinning for commands that
2736 * already triggered a CQE (eg in error).
2738 if (test_bit(0, &ctx->cq_check_overflow))
2739 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2740 if (io_cqring_events(ctx))
2744 * If a submit got punted to a workqueue, we can have the
2745 * application entering polling for a command before it gets
2746 * issued. That app will hold the uring_lock for the duration
2747 * of the poll right here, so we need to take a breather every
2748 * now and then to ensure that the issue has a chance to add
2749 * the poll to the issued list. Otherwise we can spin here
2750 * forever, while the workqueue is stuck trying to acquire the
2753 if (!(++iters & 7)) {
2754 mutex_unlock(&ctx->uring_lock);
2756 mutex_lock(&ctx->uring_lock);
2759 ret = io_iopoll_getevents(ctx, &nr_events, min);
2763 } while (min && !nr_events && !need_resched());
2765 mutex_unlock(&ctx->uring_lock);
2769 static void kiocb_end_write(struct io_kiocb *req)
2772 * Tell lockdep we inherited freeze protection from submission
2775 if (req->flags & REQ_F_ISREG) {
2776 struct inode *inode = file_inode(req->file);
2778 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2780 file_end_write(req->file);
2784 static bool io_resubmit_prep(struct io_kiocb *req)
2786 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2788 struct iov_iter iter;
2790 /* already prepared */
2791 if (req->async_data)
2794 switch (req->opcode) {
2795 case IORING_OP_READV:
2796 case IORING_OP_READ_FIXED:
2797 case IORING_OP_READ:
2800 case IORING_OP_WRITEV:
2801 case IORING_OP_WRITE_FIXED:
2802 case IORING_OP_WRITE:
2806 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2811 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2814 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2818 static bool io_rw_reissue(struct io_kiocb *req)
2821 umode_t mode = file_inode(req->file)->i_mode;
2824 if (!S_ISBLK(mode) && !S_ISREG(mode))
2826 if ((req->flags & REQ_F_NOWAIT) || io_wq_current_is_worker())
2829 lockdep_assert_held(&req->ctx->uring_lock);
2831 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2833 if (!ret && io_resubmit_prep(req)) {
2834 refcount_inc(&req->refs);
2835 io_queue_async_work(req);
2838 req_set_fail_links(req);
2843 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2844 unsigned int issue_flags)
2848 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2850 if (res != req->result)
2851 req_set_fail_links(req);
2853 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2854 kiocb_end_write(req);
2855 if (req->flags & REQ_F_BUFFER_SELECTED)
2856 cflags = io_put_rw_kbuf(req);
2857 __io_req_complete(req, issue_flags, res, cflags);
2860 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2862 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2864 __io_complete_rw(req, res, res2, 0);
2867 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2869 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2871 if (kiocb->ki_flags & IOCB_WRITE)
2872 kiocb_end_write(req);
2874 if (res != -EAGAIN && res != req->result)
2875 req_set_fail_links(req);
2877 WRITE_ONCE(req->result, res);
2878 /* order with io_poll_complete() checking ->result */
2880 WRITE_ONCE(req->iopoll_completed, 1);
2884 * After the iocb has been issued, it's safe to be found on the poll list.
2885 * Adding the kiocb to the list AFTER submission ensures that we don't
2886 * find it from a io_iopoll_getevents() thread before the issuer is done
2887 * accessing the kiocb cookie.
2889 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2891 struct io_ring_ctx *ctx = req->ctx;
2894 * Track whether we have multiple files in our lists. This will impact
2895 * how we do polling eventually, not spinning if we're on potentially
2896 * different devices.
2898 if (list_empty(&ctx->iopoll_list)) {
2899 ctx->poll_multi_file = false;
2900 } else if (!ctx->poll_multi_file) {
2901 struct io_kiocb *list_req;
2903 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2905 if (list_req->file != req->file)
2906 ctx->poll_multi_file = true;
2910 * For fast devices, IO may have already completed. If it has, add
2911 * it to the front so we find it first.
2913 if (READ_ONCE(req->iopoll_completed))
2914 list_add(&req->inflight_entry, &ctx->iopoll_list);
2916 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2919 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2920 * task context or in io worker task context. If current task context is
2921 * sq thread, we don't need to check whether should wake up sq thread.
2923 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2924 wq_has_sleeper(&ctx->sq_data->wait))
2925 wake_up(&ctx->sq_data->wait);
2928 static inline void io_state_file_put(struct io_submit_state *state)
2930 if (state->file_refs) {
2931 fput_many(state->file, state->file_refs);
2932 state->file_refs = 0;
2937 * Get as many references to a file as we have IOs left in this submission,
2938 * assuming most submissions are for one file, or at least that each file
2939 * has more than one submission.
2941 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2946 if (state->file_refs) {
2947 if (state->fd == fd) {
2951 io_state_file_put(state);
2953 state->file = fget_many(fd, state->ios_left);
2954 if (unlikely(!state->file))
2958 state->file_refs = state->ios_left - 1;
2962 static bool io_bdev_nowait(struct block_device *bdev)
2964 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2968 * If we tracked the file through the SCM inflight mechanism, we could support
2969 * any file. For now, just ensure that anything potentially problematic is done
2972 static bool io_file_supports_async(struct file *file, int rw)
2974 umode_t mode = file_inode(file)->i_mode;
2976 if (S_ISBLK(mode)) {
2977 if (IS_ENABLED(CONFIG_BLOCK) &&
2978 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2982 if (S_ISCHR(mode) || S_ISSOCK(mode))
2984 if (S_ISREG(mode)) {
2985 if (IS_ENABLED(CONFIG_BLOCK) &&
2986 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2987 file->f_op != &io_uring_fops)
2992 /* any ->read/write should understand O_NONBLOCK */
2993 if (file->f_flags & O_NONBLOCK)
2996 if (!(file->f_mode & FMODE_NOWAIT))
3000 return file->f_op->read_iter != NULL;
3002 return file->f_op->write_iter != NULL;
3005 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3007 struct io_ring_ctx *ctx = req->ctx;
3008 struct kiocb *kiocb = &req->rw.kiocb;
3009 struct file *file = req->file;
3013 if (S_ISREG(file_inode(file)->i_mode))
3014 req->flags |= REQ_F_ISREG;
3016 kiocb->ki_pos = READ_ONCE(sqe->off);
3017 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
3018 req->flags |= REQ_F_CUR_POS;
3019 kiocb->ki_pos = file->f_pos;
3021 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
3022 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
3023 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
3027 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
3028 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
3029 req->flags |= REQ_F_NOWAIT;
3031 ioprio = READ_ONCE(sqe->ioprio);
3033 ret = ioprio_check_cap(ioprio);
3037 kiocb->ki_ioprio = ioprio;
3039 kiocb->ki_ioprio = get_current_ioprio();
3041 if (ctx->flags & IORING_SETUP_IOPOLL) {
3042 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
3043 !kiocb->ki_filp->f_op->iopoll)
3046 kiocb->ki_flags |= IOCB_HIPRI;
3047 kiocb->ki_complete = io_complete_rw_iopoll;
3048 req->iopoll_completed = 0;
3050 if (kiocb->ki_flags & IOCB_HIPRI)
3052 kiocb->ki_complete = io_complete_rw;
3055 req->rw.addr = READ_ONCE(sqe->addr);
3056 req->rw.len = READ_ONCE(sqe->len);
3057 req->buf_index = READ_ONCE(sqe->buf_index);
3061 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
3067 case -ERESTARTNOINTR:
3068 case -ERESTARTNOHAND:
3069 case -ERESTART_RESTARTBLOCK:
3071 * We can't just restart the syscall, since previously
3072 * submitted sqes may already be in progress. Just fail this
3078 kiocb->ki_complete(kiocb, ret, 0);
3082 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
3083 unsigned int issue_flags)
3085 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
3086 struct io_async_rw *io = req->async_data;
3088 /* add previously done IO, if any */
3089 if (io && io->bytes_done > 0) {
3091 ret = io->bytes_done;
3093 ret += io->bytes_done;
3096 if (req->flags & REQ_F_CUR_POS)
3097 req->file->f_pos = kiocb->ki_pos;
3098 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
3099 __io_complete_rw(req, ret, 0, issue_flags);
3101 io_rw_done(kiocb, ret);
3104 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3106 struct io_ring_ctx *ctx = req->ctx;
3107 size_t len = req->rw.len;
3108 struct io_mapped_ubuf *imu;
3109 u16 index, buf_index = req->buf_index;
3113 if (unlikely(buf_index >= ctx->nr_user_bufs))
3115 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3116 imu = &ctx->user_bufs[index];
3117 buf_addr = req->rw.addr;
3120 if (buf_addr + len < buf_addr)
3122 /* not inside the mapped region */
3123 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
3127 * May not be a start of buffer, set size appropriately
3128 * and advance us to the beginning.
3130 offset = buf_addr - imu->ubuf;
3131 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3135 * Don't use iov_iter_advance() here, as it's really slow for
3136 * using the latter parts of a big fixed buffer - it iterates
3137 * over each segment manually. We can cheat a bit here, because
3140 * 1) it's a BVEC iter, we set it up
3141 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3142 * first and last bvec
3144 * So just find our index, and adjust the iterator afterwards.
3145 * If the offset is within the first bvec (or the whole first
3146 * bvec, just use iov_iter_advance(). This makes it easier
3147 * since we can just skip the first segment, which may not
3148 * be PAGE_SIZE aligned.
3150 const struct bio_vec *bvec = imu->bvec;
3152 if (offset <= bvec->bv_len) {
3153 iov_iter_advance(iter, offset);
3155 unsigned long seg_skip;
3157 /* skip first vec */
3158 offset -= bvec->bv_len;
3159 seg_skip = 1 + (offset >> PAGE_SHIFT);
3161 iter->bvec = bvec + seg_skip;
3162 iter->nr_segs -= seg_skip;
3163 iter->count -= bvec->bv_len + offset;
3164 iter->iov_offset = offset & ~PAGE_MASK;
3171 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3174 mutex_unlock(&ctx->uring_lock);
3177 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3180 * "Normal" inline submissions always hold the uring_lock, since we
3181 * grab it from the system call. Same is true for the SQPOLL offload.
3182 * The only exception is when we've detached the request and issue it
3183 * from an async worker thread, grab the lock for that case.
3186 mutex_lock(&ctx->uring_lock);
3189 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3190 int bgid, struct io_buffer *kbuf,
3193 struct io_buffer *head;
3195 if (req->flags & REQ_F_BUFFER_SELECTED)
3198 io_ring_submit_lock(req->ctx, needs_lock);
3200 lockdep_assert_held(&req->ctx->uring_lock);
3202 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3204 if (!list_empty(&head->list)) {
3205 kbuf = list_last_entry(&head->list, struct io_buffer,
3207 list_del(&kbuf->list);
3210 idr_remove(&req->ctx->io_buffer_idr, bgid);
3212 if (*len > kbuf->len)
3215 kbuf = ERR_PTR(-ENOBUFS);
3218 io_ring_submit_unlock(req->ctx, needs_lock);
3223 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3226 struct io_buffer *kbuf;
3229 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3230 bgid = req->buf_index;
3231 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3234 req->rw.addr = (u64) (unsigned long) kbuf;
3235 req->flags |= REQ_F_BUFFER_SELECTED;
3236 return u64_to_user_ptr(kbuf->addr);
3239 #ifdef CONFIG_COMPAT
3240 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3243 struct compat_iovec __user *uiov;
3244 compat_ssize_t clen;
3248 uiov = u64_to_user_ptr(req->rw.addr);
3249 if (!access_ok(uiov, sizeof(*uiov)))
3251 if (__get_user(clen, &uiov->iov_len))
3257 buf = io_rw_buffer_select(req, &len, needs_lock);
3259 return PTR_ERR(buf);
3260 iov[0].iov_base = buf;
3261 iov[0].iov_len = (compat_size_t) len;
3266 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3269 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3273 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3276 len = iov[0].iov_len;
3279 buf = io_rw_buffer_select(req, &len, needs_lock);
3281 return PTR_ERR(buf);
3282 iov[0].iov_base = buf;
3283 iov[0].iov_len = len;
3287 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3290 if (req->flags & REQ_F_BUFFER_SELECTED) {
3291 struct io_buffer *kbuf;
3293 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3294 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3295 iov[0].iov_len = kbuf->len;
3298 if (req->rw.len != 1)
3301 #ifdef CONFIG_COMPAT
3302 if (req->ctx->compat)
3303 return io_compat_import(req, iov, needs_lock);
3306 return __io_iov_buffer_select(req, iov, needs_lock);
3309 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3310 struct iov_iter *iter, bool needs_lock)
3312 void __user *buf = u64_to_user_ptr(req->rw.addr);
3313 size_t sqe_len = req->rw.len;
3314 u8 opcode = req->opcode;
3317 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3319 return io_import_fixed(req, rw, iter);
3322 /* buffer index only valid with fixed read/write, or buffer select */
3323 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3326 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3327 if (req->flags & REQ_F_BUFFER_SELECT) {
3328 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3330 return PTR_ERR(buf);
3331 req->rw.len = sqe_len;
3334 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3339 if (req->flags & REQ_F_BUFFER_SELECT) {
3340 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3342 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3347 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3351 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3353 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3357 * For files that don't have ->read_iter() and ->write_iter(), handle them
3358 * by looping over ->read() or ->write() manually.
3360 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3362 struct kiocb *kiocb = &req->rw.kiocb;
3363 struct file *file = req->file;
3367 * Don't support polled IO through this interface, and we can't
3368 * support non-blocking either. For the latter, this just causes
3369 * the kiocb to be handled from an async context.
3371 if (kiocb->ki_flags & IOCB_HIPRI)
3373 if (kiocb->ki_flags & IOCB_NOWAIT)
3376 while (iov_iter_count(iter)) {
3380 if (!iov_iter_is_bvec(iter)) {
3381 iovec = iov_iter_iovec(iter);
3383 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3384 iovec.iov_len = req->rw.len;
3388 nr = file->f_op->read(file, iovec.iov_base,
3389 iovec.iov_len, io_kiocb_ppos(kiocb));
3391 nr = file->f_op->write(file, iovec.iov_base,
3392 iovec.iov_len, io_kiocb_ppos(kiocb));
3401 if (nr != iovec.iov_len)
3405 iov_iter_advance(iter, nr);
3411 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3412 const struct iovec *fast_iov, struct iov_iter *iter)
3414 struct io_async_rw *rw = req->async_data;
3416 memcpy(&rw->iter, iter, sizeof(*iter));
3417 rw->free_iovec = iovec;
3419 /* can only be fixed buffers, no need to do anything */
3420 if (iov_iter_is_bvec(iter))
3423 unsigned iov_off = 0;
3425 rw->iter.iov = rw->fast_iov;
3426 if (iter->iov != fast_iov) {
3427 iov_off = iter->iov - fast_iov;
3428 rw->iter.iov += iov_off;
3430 if (rw->fast_iov != fast_iov)
3431 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3432 sizeof(struct iovec) * iter->nr_segs);
3434 req->flags |= REQ_F_NEED_CLEANUP;
3438 static inline int __io_alloc_async_data(struct io_kiocb *req)
3440 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3441 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3442 return req->async_data == NULL;
3445 static int io_alloc_async_data(struct io_kiocb *req)
3447 if (!io_op_defs[req->opcode].needs_async_data)
3450 return __io_alloc_async_data(req);
3453 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3454 const struct iovec *fast_iov,
3455 struct iov_iter *iter, bool force)
3457 if (!force && !io_op_defs[req->opcode].needs_async_data)
3459 if (!req->async_data) {
3460 if (__io_alloc_async_data(req)) {
3465 io_req_map_rw(req, iovec, fast_iov, iter);
3470 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3472 struct io_async_rw *iorw = req->async_data;
3473 struct iovec *iov = iorw->fast_iov;
3476 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3477 if (unlikely(ret < 0))
3480 iorw->bytes_done = 0;
3481 iorw->free_iovec = iov;
3483 req->flags |= REQ_F_NEED_CLEANUP;
3487 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3489 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3491 return io_prep_rw(req, sqe);
3495 * This is our waitqueue callback handler, registered through lock_page_async()
3496 * when we initially tried to do the IO with the iocb armed our waitqueue.
3497 * This gets called when the page is unlocked, and we generally expect that to
3498 * happen when the page IO is completed and the page is now uptodate. This will
3499 * queue a task_work based retry of the operation, attempting to copy the data
3500 * again. If the latter fails because the page was NOT uptodate, then we will
3501 * do a thread based blocking retry of the operation. That's the unexpected
3504 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3505 int sync, void *arg)
3507 struct wait_page_queue *wpq;
3508 struct io_kiocb *req = wait->private;
3509 struct wait_page_key *key = arg;
3511 wpq = container_of(wait, struct wait_page_queue, wait);
3513 if (!wake_page_match(wpq, key))
3516 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3517 list_del_init(&wait->entry);
3519 /* submit ref gets dropped, acquire a new one */
3520 refcount_inc(&req->refs);
3521 io_req_task_queue(req);
3526 * This controls whether a given IO request should be armed for async page
3527 * based retry. If we return false here, the request is handed to the async
3528 * worker threads for retry. If we're doing buffered reads on a regular file,
3529 * we prepare a private wait_page_queue entry and retry the operation. This
3530 * will either succeed because the page is now uptodate and unlocked, or it
3531 * will register a callback when the page is unlocked at IO completion. Through
3532 * that callback, io_uring uses task_work to setup a retry of the operation.
3533 * That retry will attempt the buffered read again. The retry will generally
3534 * succeed, or in rare cases where it fails, we then fall back to using the
3535 * async worker threads for a blocking retry.
3537 static bool io_rw_should_retry(struct io_kiocb *req)
3539 struct io_async_rw *rw = req->async_data;
3540 struct wait_page_queue *wait = &rw->wpq;
3541 struct kiocb *kiocb = &req->rw.kiocb;
3543 /* never retry for NOWAIT, we just complete with -EAGAIN */
3544 if (req->flags & REQ_F_NOWAIT)
3547 /* Only for buffered IO */
3548 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3552 * just use poll if we can, and don't attempt if the fs doesn't
3553 * support callback based unlocks
3555 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3558 wait->wait.func = io_async_buf_func;
3559 wait->wait.private = req;
3560 wait->wait.flags = 0;
3561 INIT_LIST_HEAD(&wait->wait.entry);
3562 kiocb->ki_flags |= IOCB_WAITQ;
3563 kiocb->ki_flags &= ~IOCB_NOWAIT;
3564 kiocb->ki_waitq = wait;
3568 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3570 if (req->file->f_op->read_iter)
3571 return call_read_iter(req->file, &req->rw.kiocb, iter);
3572 else if (req->file->f_op->read)
3573 return loop_rw_iter(READ, req, iter);
3578 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3580 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3581 struct kiocb *kiocb = &req->rw.kiocb;
3582 struct iov_iter __iter, *iter = &__iter;
3583 struct io_async_rw *rw = req->async_data;
3584 ssize_t io_size, ret, ret2;
3585 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3591 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3595 io_size = iov_iter_count(iter);
3596 req->result = io_size;
3598 /* Ensure we clear previously set non-block flag */
3599 if (!force_nonblock)
3600 kiocb->ki_flags &= ~IOCB_NOWAIT;
3602 kiocb->ki_flags |= IOCB_NOWAIT;
3604 /* If the file doesn't support async, just async punt */
3605 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3606 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3607 return ret ?: -EAGAIN;
3610 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3611 if (unlikely(ret)) {
3616 ret = io_iter_do_read(req, iter);
3618 if (ret == -EIOCBQUEUED) {
3620 } else if (ret == -EAGAIN) {
3621 /* IOPOLL retry should happen for io-wq threads */
3622 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3624 /* no retry on NONBLOCK nor RWF_NOWAIT */
3625 if (req->flags & REQ_F_NOWAIT)
3627 /* some cases will consume bytes even on error returns */
3628 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3630 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3631 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3632 /* read all, failed, already did sync or don't want to retry */
3636 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3641 rw = req->async_data;
3642 /* now use our persistent iterator, if we aren't already */
3647 rw->bytes_done += ret;
3648 /* if we can retry, do so with the callbacks armed */
3649 if (!io_rw_should_retry(req)) {
3650 kiocb->ki_flags &= ~IOCB_WAITQ;
3655 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3656 * we get -EIOCBQUEUED, then we'll get a notification when the
3657 * desired page gets unlocked. We can also get a partial read
3658 * here, and if we do, then just retry at the new offset.
3660 ret = io_iter_do_read(req, iter);
3661 if (ret == -EIOCBQUEUED)
3663 /* we got some bytes, but not all. retry. */
3664 } while (ret > 0 && ret < io_size);
3666 kiocb_done(kiocb, ret, issue_flags);
3668 /* it's faster to check here then delegate to kfree */
3674 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3676 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3678 return io_prep_rw(req, sqe);
3681 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3683 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3684 struct kiocb *kiocb = &req->rw.kiocb;
3685 struct iov_iter __iter, *iter = &__iter;
3686 struct io_async_rw *rw = req->async_data;
3687 ssize_t ret, ret2, io_size;
3688 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3694 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3698 io_size = iov_iter_count(iter);
3699 req->result = io_size;
3701 /* Ensure we clear previously set non-block flag */
3702 if (!force_nonblock)
3703 kiocb->ki_flags &= ~IOCB_NOWAIT;
3705 kiocb->ki_flags |= IOCB_NOWAIT;
3707 /* If the file doesn't support async, just async punt */
3708 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3711 /* file path doesn't support NOWAIT for non-direct_IO */
3712 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3713 (req->flags & REQ_F_ISREG))
3716 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3721 * Open-code file_start_write here to grab freeze protection,
3722 * which will be released by another thread in
3723 * io_complete_rw(). Fool lockdep by telling it the lock got
3724 * released so that it doesn't complain about the held lock when
3725 * we return to userspace.
3727 if (req->flags & REQ_F_ISREG) {
3728 sb_start_write(file_inode(req->file)->i_sb);
3729 __sb_writers_release(file_inode(req->file)->i_sb,
3732 kiocb->ki_flags |= IOCB_WRITE;
3734 if (req->file->f_op->write_iter)
3735 ret2 = call_write_iter(req->file, kiocb, iter);
3736 else if (req->file->f_op->write)
3737 ret2 = loop_rw_iter(WRITE, req, iter);
3742 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3743 * retry them without IOCB_NOWAIT.
3745 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3747 /* no retry on NONBLOCK nor RWF_NOWAIT */
3748 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3750 if (!force_nonblock || ret2 != -EAGAIN) {
3751 /* IOPOLL retry should happen for io-wq threads */
3752 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3755 kiocb_done(kiocb, ret2, issue_flags);
3758 /* some cases will consume bytes even on error returns */
3759 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3760 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3761 return ret ?: -EAGAIN;
3764 /* it's reportedly faster than delegating the null check to kfree() */
3770 static int io_renameat_prep(struct io_kiocb *req,
3771 const struct io_uring_sqe *sqe)
3773 struct io_rename *ren = &req->rename;
3774 const char __user *oldf, *newf;
3776 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3779 ren->old_dfd = READ_ONCE(sqe->fd);
3780 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3781 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3782 ren->new_dfd = READ_ONCE(sqe->len);
3783 ren->flags = READ_ONCE(sqe->rename_flags);
3785 ren->oldpath = getname(oldf);
3786 if (IS_ERR(ren->oldpath))
3787 return PTR_ERR(ren->oldpath);
3789 ren->newpath = getname(newf);
3790 if (IS_ERR(ren->newpath)) {
3791 putname(ren->oldpath);
3792 return PTR_ERR(ren->newpath);
3795 req->flags |= REQ_F_NEED_CLEANUP;
3799 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3801 struct io_rename *ren = &req->rename;
3804 if (issue_flags & IO_URING_F_NONBLOCK)
3807 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3808 ren->newpath, ren->flags);
3810 req->flags &= ~REQ_F_NEED_CLEANUP;
3812 req_set_fail_links(req);
3813 io_req_complete(req, ret);
3817 static int io_unlinkat_prep(struct io_kiocb *req,
3818 const struct io_uring_sqe *sqe)
3820 struct io_unlink *un = &req->unlink;
3821 const char __user *fname;
3823 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3826 un->dfd = READ_ONCE(sqe->fd);
3828 un->flags = READ_ONCE(sqe->unlink_flags);
3829 if (un->flags & ~AT_REMOVEDIR)
3832 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3833 un->filename = getname(fname);
3834 if (IS_ERR(un->filename))
3835 return PTR_ERR(un->filename);
3837 req->flags |= REQ_F_NEED_CLEANUP;
3841 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3843 struct io_unlink *un = &req->unlink;
3846 if (issue_flags & IO_URING_F_NONBLOCK)
3849 if (un->flags & AT_REMOVEDIR)
3850 ret = do_rmdir(un->dfd, un->filename);
3852 ret = do_unlinkat(un->dfd, un->filename);
3854 req->flags &= ~REQ_F_NEED_CLEANUP;
3856 req_set_fail_links(req);
3857 io_req_complete(req, ret);
3861 static int io_shutdown_prep(struct io_kiocb *req,
3862 const struct io_uring_sqe *sqe)
3864 #if defined(CONFIG_NET)
3865 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3867 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3871 req->shutdown.how = READ_ONCE(sqe->len);
3878 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3880 #if defined(CONFIG_NET)
3881 struct socket *sock;
3884 if (issue_flags & IO_URING_F_NONBLOCK)
3887 sock = sock_from_file(req->file);
3888 if (unlikely(!sock))
3891 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3893 req_set_fail_links(req);
3894 io_req_complete(req, ret);
3901 static int __io_splice_prep(struct io_kiocb *req,
3902 const struct io_uring_sqe *sqe)
3904 struct io_splice* sp = &req->splice;
3905 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3907 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3911 sp->len = READ_ONCE(sqe->len);
3912 sp->flags = READ_ONCE(sqe->splice_flags);
3914 if (unlikely(sp->flags & ~valid_flags))
3917 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3918 (sp->flags & SPLICE_F_FD_IN_FIXED));
3921 req->flags |= REQ_F_NEED_CLEANUP;
3923 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3925 * Splice operation will be punted aync, and here need to
3926 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3928 io_req_init_async(req);
3929 req->work.flags |= IO_WQ_WORK_UNBOUND;
3935 static int io_tee_prep(struct io_kiocb *req,
3936 const struct io_uring_sqe *sqe)
3938 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3940 return __io_splice_prep(req, sqe);
3943 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3945 struct io_splice *sp = &req->splice;
3946 struct file *in = sp->file_in;
3947 struct file *out = sp->file_out;
3948 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3951 if (issue_flags & IO_URING_F_NONBLOCK)
3954 ret = do_tee(in, out, sp->len, flags);
3956 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3957 req->flags &= ~REQ_F_NEED_CLEANUP;
3960 req_set_fail_links(req);
3961 io_req_complete(req, ret);
3965 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3967 struct io_splice* sp = &req->splice;
3969 sp->off_in = READ_ONCE(sqe->splice_off_in);
3970 sp->off_out = READ_ONCE(sqe->off);
3971 return __io_splice_prep(req, sqe);
3974 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3976 struct io_splice *sp = &req->splice;
3977 struct file *in = sp->file_in;
3978 struct file *out = sp->file_out;
3979 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3980 loff_t *poff_in, *poff_out;
3983 if (issue_flags & IO_URING_F_NONBLOCK)
3986 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3987 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3990 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3992 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3993 req->flags &= ~REQ_F_NEED_CLEANUP;
3996 req_set_fail_links(req);
3997 io_req_complete(req, ret);
4002 * IORING_OP_NOP just posts a completion event, nothing else.
4004 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4006 struct io_ring_ctx *ctx = req->ctx;
4008 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4011 __io_req_complete(req, issue_flags, 0, 0);
4015 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4017 struct io_ring_ctx *ctx = req->ctx;
4022 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4024 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4027 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4028 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4031 req->sync.off = READ_ONCE(sqe->off);
4032 req->sync.len = READ_ONCE(sqe->len);
4036 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4038 loff_t end = req->sync.off + req->sync.len;
4041 /* fsync always requires a blocking context */
4042 if (issue_flags & IO_URING_F_NONBLOCK)
4045 ret = vfs_fsync_range(req->file, req->sync.off,
4046 end > 0 ? end : LLONG_MAX,
4047 req->sync.flags & IORING_FSYNC_DATASYNC);
4049 req_set_fail_links(req);
4050 io_req_complete(req, ret);
4054 static int io_fallocate_prep(struct io_kiocb *req,
4055 const struct io_uring_sqe *sqe)
4057 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
4059 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4062 req->sync.off = READ_ONCE(sqe->off);
4063 req->sync.len = READ_ONCE(sqe->addr);
4064 req->sync.mode = READ_ONCE(sqe->len);
4068 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4072 /* fallocate always requiring blocking context */
4073 if (issue_flags & IO_URING_F_NONBLOCK)
4075 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4078 req_set_fail_links(req);
4079 io_req_complete(req, ret);
4083 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4085 const char __user *fname;
4088 if (unlikely(sqe->ioprio || sqe->buf_index))
4090 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4093 /* open.how should be already initialised */
4094 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4095 req->open.how.flags |= O_LARGEFILE;
4097 req->open.dfd = READ_ONCE(sqe->fd);
4098 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4099 req->open.filename = getname(fname);
4100 if (IS_ERR(req->open.filename)) {
4101 ret = PTR_ERR(req->open.filename);
4102 req->open.filename = NULL;
4105 req->open.nofile = rlimit(RLIMIT_NOFILE);
4106 req->flags |= REQ_F_NEED_CLEANUP;
4110 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4114 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4116 mode = READ_ONCE(sqe->len);
4117 flags = READ_ONCE(sqe->open_flags);
4118 req->open.how = build_open_how(flags, mode);
4119 return __io_openat_prep(req, sqe);
4122 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4124 struct open_how __user *how;
4128 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4130 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4131 len = READ_ONCE(sqe->len);
4132 if (len < OPEN_HOW_SIZE_VER0)
4135 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4140 return __io_openat_prep(req, sqe);
4143 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4145 struct open_flags op;
4148 bool resolve_nonblock;
4151 ret = build_open_flags(&req->open.how, &op);
4154 nonblock_set = op.open_flag & O_NONBLOCK;
4155 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4156 if (issue_flags & IO_URING_F_NONBLOCK) {
4158 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4159 * it'll always -EAGAIN
4161 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4163 op.lookup_flags |= LOOKUP_CACHED;
4164 op.open_flag |= O_NONBLOCK;
4167 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4171 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4172 /* only retry if RESOLVE_CACHED wasn't already set by application */
4173 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
4174 file == ERR_PTR(-EAGAIN)) {
4176 * We could hang on to this 'fd', but seems like marginal
4177 * gain for something that is now known to be a slower path.
4178 * So just put it, and we'll get a new one when we retry.
4186 ret = PTR_ERR(file);
4188 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4189 file->f_flags &= ~O_NONBLOCK;
4190 fsnotify_open(file);
4191 fd_install(ret, file);
4194 putname(req->open.filename);
4195 req->flags &= ~REQ_F_NEED_CLEANUP;
4197 req_set_fail_links(req);
4198 io_req_complete(req, ret);
4202 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4204 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
4207 static int io_remove_buffers_prep(struct io_kiocb *req,
4208 const struct io_uring_sqe *sqe)
4210 struct io_provide_buf *p = &req->pbuf;
4213 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4216 tmp = READ_ONCE(sqe->fd);
4217 if (!tmp || tmp > USHRT_MAX)
4220 memset(p, 0, sizeof(*p));
4222 p->bgid = READ_ONCE(sqe->buf_group);
4226 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4227 int bgid, unsigned nbufs)
4231 /* shouldn't happen */
4235 /* the head kbuf is the list itself */
4236 while (!list_empty(&buf->list)) {
4237 struct io_buffer *nxt;
4239 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4240 list_del(&nxt->list);
4247 idr_remove(&ctx->io_buffer_idr, bgid);
4252 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4254 struct io_provide_buf *p = &req->pbuf;
4255 struct io_ring_ctx *ctx = req->ctx;
4256 struct io_buffer *head;
4258 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4260 io_ring_submit_lock(ctx, !force_nonblock);
4262 lockdep_assert_held(&ctx->uring_lock);
4265 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4267 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4269 req_set_fail_links(req);
4271 /* need to hold the lock to complete IOPOLL requests */
4272 if (ctx->flags & IORING_SETUP_IOPOLL) {
4273 __io_req_complete(req, issue_flags, ret, 0);
4274 io_ring_submit_unlock(ctx, !force_nonblock);
4276 io_ring_submit_unlock(ctx, !force_nonblock);
4277 __io_req_complete(req, issue_flags, ret, 0);
4282 static int io_provide_buffers_prep(struct io_kiocb *req,
4283 const struct io_uring_sqe *sqe)
4285 struct io_provide_buf *p = &req->pbuf;
4288 if (sqe->ioprio || sqe->rw_flags)
4291 tmp = READ_ONCE(sqe->fd);
4292 if (!tmp || tmp > USHRT_MAX)
4295 p->addr = READ_ONCE(sqe->addr);
4296 p->len = READ_ONCE(sqe->len);
4298 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4301 p->bgid = READ_ONCE(sqe->buf_group);
4302 tmp = READ_ONCE(sqe->off);
4303 if (tmp > USHRT_MAX)
4309 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4311 struct io_buffer *buf;
4312 u64 addr = pbuf->addr;
4313 int i, bid = pbuf->bid;
4315 for (i = 0; i < pbuf->nbufs; i++) {
4316 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4321 buf->len = pbuf->len;
4326 INIT_LIST_HEAD(&buf->list);
4329 list_add_tail(&buf->list, &(*head)->list);
4333 return i ? i : -ENOMEM;
4336 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4338 struct io_provide_buf *p = &req->pbuf;
4339 struct io_ring_ctx *ctx = req->ctx;
4340 struct io_buffer *head, *list;
4342 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4344 io_ring_submit_lock(ctx, !force_nonblock);
4346 lockdep_assert_held(&ctx->uring_lock);
4348 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4350 ret = io_add_buffers(p, &head);
4355 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4358 __io_remove_buffers(ctx, head, p->bgid, -1U);
4364 req_set_fail_links(req);
4366 /* need to hold the lock to complete IOPOLL requests */
4367 if (ctx->flags & IORING_SETUP_IOPOLL) {
4368 __io_req_complete(req, issue_flags, ret, 0);
4369 io_ring_submit_unlock(ctx, !force_nonblock);
4371 io_ring_submit_unlock(ctx, !force_nonblock);
4372 __io_req_complete(req, issue_flags, ret, 0);
4377 static int io_epoll_ctl_prep(struct io_kiocb *req,
4378 const struct io_uring_sqe *sqe)
4380 #if defined(CONFIG_EPOLL)
4381 if (sqe->ioprio || sqe->buf_index)
4383 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4386 req->epoll.epfd = READ_ONCE(sqe->fd);
4387 req->epoll.op = READ_ONCE(sqe->len);
4388 req->epoll.fd = READ_ONCE(sqe->off);
4390 if (ep_op_has_event(req->epoll.op)) {
4391 struct epoll_event __user *ev;
4393 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4394 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4404 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4406 #if defined(CONFIG_EPOLL)
4407 struct io_epoll *ie = &req->epoll;
4409 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4411 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4412 if (force_nonblock && ret == -EAGAIN)
4416 req_set_fail_links(req);
4417 __io_req_complete(req, issue_flags, ret, 0);
4424 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4426 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4427 if (sqe->ioprio || sqe->buf_index || sqe->off)
4429 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4432 req->madvise.addr = READ_ONCE(sqe->addr);
4433 req->madvise.len = READ_ONCE(sqe->len);
4434 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4441 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4443 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4444 struct io_madvise *ma = &req->madvise;
4447 if (issue_flags & IO_URING_F_NONBLOCK)
4450 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4452 req_set_fail_links(req);
4453 io_req_complete(req, ret);
4460 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4462 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4464 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4467 req->fadvise.offset = READ_ONCE(sqe->off);
4468 req->fadvise.len = READ_ONCE(sqe->len);
4469 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4473 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4475 struct io_fadvise *fa = &req->fadvise;
4478 if (issue_flags & IO_URING_F_NONBLOCK) {
4479 switch (fa->advice) {
4480 case POSIX_FADV_NORMAL:
4481 case POSIX_FADV_RANDOM:
4482 case POSIX_FADV_SEQUENTIAL:
4489 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4491 req_set_fail_links(req);
4492 io_req_complete(req, ret);
4496 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4498 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4500 if (sqe->ioprio || sqe->buf_index)
4502 if (req->flags & REQ_F_FIXED_FILE)
4505 req->statx.dfd = READ_ONCE(sqe->fd);
4506 req->statx.mask = READ_ONCE(sqe->len);
4507 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4508 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4509 req->statx.flags = READ_ONCE(sqe->statx_flags);
4514 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4516 struct io_statx *ctx = &req->statx;
4519 if (issue_flags & IO_URING_F_NONBLOCK) {
4520 /* only need file table for an actual valid fd */
4521 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4522 req->flags |= REQ_F_NO_FILE_TABLE;
4526 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4530 req_set_fail_links(req);
4531 io_req_complete(req, ret);
4535 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4537 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4539 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4540 sqe->rw_flags || sqe->buf_index)
4542 if (req->flags & REQ_F_FIXED_FILE)
4545 req->close.fd = READ_ONCE(sqe->fd);
4549 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4551 struct files_struct *files = current->files;
4552 struct io_close *close = &req->close;
4553 struct fdtable *fdt;
4559 spin_lock(&files->file_lock);
4560 fdt = files_fdtable(files);
4561 if (close->fd >= fdt->max_fds) {
4562 spin_unlock(&files->file_lock);
4565 file = fdt->fd[close->fd];
4567 spin_unlock(&files->file_lock);
4571 if (file->f_op == &io_uring_fops) {
4572 spin_unlock(&files->file_lock);
4577 /* if the file has a flush method, be safe and punt to async */
4578 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4579 spin_unlock(&files->file_lock);
4583 ret = __close_fd_get_file(close->fd, &file);
4584 spin_unlock(&files->file_lock);
4591 /* No ->flush() or already async, safely close from here */
4592 ret = filp_close(file, current->files);
4595 req_set_fail_links(req);
4598 __io_req_complete(req, issue_flags, ret, 0);
4602 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4604 struct io_ring_ctx *ctx = req->ctx;
4606 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4608 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4611 req->sync.off = READ_ONCE(sqe->off);
4612 req->sync.len = READ_ONCE(sqe->len);
4613 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4617 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4621 /* sync_file_range always requires a blocking context */
4622 if (issue_flags & IO_URING_F_NONBLOCK)
4625 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4628 req_set_fail_links(req);
4629 io_req_complete(req, ret);
4633 #if defined(CONFIG_NET)
4634 static int io_setup_async_msg(struct io_kiocb *req,
4635 struct io_async_msghdr *kmsg)
4637 struct io_async_msghdr *async_msg = req->async_data;
4641 if (io_alloc_async_data(req)) {
4642 kfree(kmsg->free_iov);
4645 async_msg = req->async_data;
4646 req->flags |= REQ_F_NEED_CLEANUP;
4647 memcpy(async_msg, kmsg, sizeof(*kmsg));
4648 async_msg->msg.msg_name = &async_msg->addr;
4649 /* if were using fast_iov, set it to the new one */
4650 if (!async_msg->free_iov)
4651 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4656 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4657 struct io_async_msghdr *iomsg)
4659 iomsg->msg.msg_name = &iomsg->addr;
4660 iomsg->free_iov = iomsg->fast_iov;
4661 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4662 req->sr_msg.msg_flags, &iomsg->free_iov);
4665 static int io_sendmsg_prep_async(struct io_kiocb *req)
4669 if (!io_op_defs[req->opcode].needs_async_data)
4671 ret = io_sendmsg_copy_hdr(req, req->async_data);
4673 req->flags |= REQ_F_NEED_CLEANUP;
4677 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4679 struct io_sr_msg *sr = &req->sr_msg;
4681 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4684 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4685 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4686 sr->len = READ_ONCE(sqe->len);
4688 #ifdef CONFIG_COMPAT
4689 if (req->ctx->compat)
4690 sr->msg_flags |= MSG_CMSG_COMPAT;
4695 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4697 struct io_async_msghdr iomsg, *kmsg;
4698 struct socket *sock;
4702 sock = sock_from_file(req->file);
4703 if (unlikely(!sock))
4706 kmsg = req->async_data;
4708 ret = io_sendmsg_copy_hdr(req, &iomsg);
4714 flags = req->sr_msg.msg_flags;
4715 if (flags & MSG_DONTWAIT)
4716 req->flags |= REQ_F_NOWAIT;
4717 else if (issue_flags & IO_URING_F_NONBLOCK)
4718 flags |= MSG_DONTWAIT;
4720 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4721 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4722 return io_setup_async_msg(req, kmsg);
4723 if (ret == -ERESTARTSYS)
4726 /* fast path, check for non-NULL to avoid function call */
4728 kfree(kmsg->free_iov);
4729 req->flags &= ~REQ_F_NEED_CLEANUP;
4731 req_set_fail_links(req);
4732 __io_req_complete(req, issue_flags, ret, 0);
4736 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4738 struct io_sr_msg *sr = &req->sr_msg;
4741 struct socket *sock;
4745 sock = sock_from_file(req->file);
4746 if (unlikely(!sock))
4749 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4753 msg.msg_name = NULL;
4754 msg.msg_control = NULL;
4755 msg.msg_controllen = 0;
4756 msg.msg_namelen = 0;
4758 flags = req->sr_msg.msg_flags;
4759 if (flags & MSG_DONTWAIT)
4760 req->flags |= REQ_F_NOWAIT;
4761 else if (issue_flags & IO_URING_F_NONBLOCK)
4762 flags |= MSG_DONTWAIT;
4764 msg.msg_flags = flags;
4765 ret = sock_sendmsg(sock, &msg);
4766 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4768 if (ret == -ERESTARTSYS)
4772 req_set_fail_links(req);
4773 __io_req_complete(req, issue_flags, ret, 0);
4777 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4778 struct io_async_msghdr *iomsg)
4780 struct io_sr_msg *sr = &req->sr_msg;
4781 struct iovec __user *uiov;
4785 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4786 &iomsg->uaddr, &uiov, &iov_len);
4790 if (req->flags & REQ_F_BUFFER_SELECT) {
4793 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4795 sr->len = iomsg->fast_iov[0].iov_len;
4796 iomsg->free_iov = NULL;
4798 iomsg->free_iov = iomsg->fast_iov;
4799 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4800 &iomsg->free_iov, &iomsg->msg.msg_iter,
4809 #ifdef CONFIG_COMPAT
4810 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4811 struct io_async_msghdr *iomsg)
4813 struct compat_msghdr __user *msg_compat;
4814 struct io_sr_msg *sr = &req->sr_msg;
4815 struct compat_iovec __user *uiov;
4820 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4821 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4826 uiov = compat_ptr(ptr);
4827 if (req->flags & REQ_F_BUFFER_SELECT) {
4828 compat_ssize_t clen;
4832 if (!access_ok(uiov, sizeof(*uiov)))
4834 if (__get_user(clen, &uiov->iov_len))
4839 iomsg->free_iov = NULL;
4841 iomsg->free_iov = iomsg->fast_iov;
4842 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4843 UIO_FASTIOV, &iomsg->free_iov,
4844 &iomsg->msg.msg_iter, true);
4853 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4854 struct io_async_msghdr *iomsg)
4856 iomsg->msg.msg_name = &iomsg->addr;
4858 #ifdef CONFIG_COMPAT
4859 if (req->ctx->compat)
4860 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4863 return __io_recvmsg_copy_hdr(req, iomsg);
4866 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4869 struct io_sr_msg *sr = &req->sr_msg;
4870 struct io_buffer *kbuf;
4872 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4877 req->flags |= REQ_F_BUFFER_SELECTED;
4881 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4883 return io_put_kbuf(req, req->sr_msg.kbuf);
4886 static int io_recvmsg_prep_async(struct io_kiocb *req)
4890 if (!io_op_defs[req->opcode].needs_async_data)
4892 ret = io_recvmsg_copy_hdr(req, req->async_data);
4894 req->flags |= REQ_F_NEED_CLEANUP;
4898 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4900 struct io_sr_msg *sr = &req->sr_msg;
4902 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4905 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4906 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4907 sr->len = READ_ONCE(sqe->len);
4908 sr->bgid = READ_ONCE(sqe->buf_group);
4910 #ifdef CONFIG_COMPAT
4911 if (req->ctx->compat)
4912 sr->msg_flags |= MSG_CMSG_COMPAT;
4917 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4919 struct io_async_msghdr iomsg, *kmsg;
4920 struct socket *sock;
4921 struct io_buffer *kbuf;
4923 int ret, cflags = 0;
4924 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4926 sock = sock_from_file(req->file);
4927 if (unlikely(!sock))
4930 kmsg = req->async_data;
4932 ret = io_recvmsg_copy_hdr(req, &iomsg);
4938 if (req->flags & REQ_F_BUFFER_SELECT) {
4939 kbuf = io_recv_buffer_select(req, !force_nonblock);
4941 return PTR_ERR(kbuf);
4942 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4943 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4944 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4945 1, req->sr_msg.len);
4948 flags = req->sr_msg.msg_flags;
4949 if (flags & MSG_DONTWAIT)
4950 req->flags |= REQ_F_NOWAIT;
4951 else if (force_nonblock)
4952 flags |= MSG_DONTWAIT;
4954 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4955 kmsg->uaddr, flags);
4956 if (force_nonblock && ret == -EAGAIN)
4957 return io_setup_async_msg(req, kmsg);
4958 if (ret == -ERESTARTSYS)
4961 if (req->flags & REQ_F_BUFFER_SELECTED)
4962 cflags = io_put_recv_kbuf(req);
4963 /* fast path, check for non-NULL to avoid function call */
4965 kfree(kmsg->free_iov);
4966 req->flags &= ~REQ_F_NEED_CLEANUP;
4968 req_set_fail_links(req);
4969 __io_req_complete(req, issue_flags, ret, cflags);
4973 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4975 struct io_buffer *kbuf;
4976 struct io_sr_msg *sr = &req->sr_msg;
4978 void __user *buf = sr->buf;
4979 struct socket *sock;
4982 int ret, cflags = 0;
4983 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4985 sock = sock_from_file(req->file);
4986 if (unlikely(!sock))
4989 if (req->flags & REQ_F_BUFFER_SELECT) {
4990 kbuf = io_recv_buffer_select(req, !force_nonblock);
4992 return PTR_ERR(kbuf);
4993 buf = u64_to_user_ptr(kbuf->addr);
4996 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5000 msg.msg_name = NULL;
5001 msg.msg_control = NULL;
5002 msg.msg_controllen = 0;
5003 msg.msg_namelen = 0;
5004 msg.msg_iocb = NULL;
5007 flags = req->sr_msg.msg_flags;
5008 if (flags & MSG_DONTWAIT)
5009 req->flags |= REQ_F_NOWAIT;
5010 else if (force_nonblock)
5011 flags |= MSG_DONTWAIT;
5013 ret = sock_recvmsg(sock, &msg, flags);
5014 if (force_nonblock && ret == -EAGAIN)
5016 if (ret == -ERESTARTSYS)
5019 if (req->flags & REQ_F_BUFFER_SELECTED)
5020 cflags = io_put_recv_kbuf(req);
5022 req_set_fail_links(req);
5023 __io_req_complete(req, issue_flags, ret, cflags);
5027 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5029 struct io_accept *accept = &req->accept;
5031 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5033 if (sqe->ioprio || sqe->len || sqe->buf_index)
5036 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5037 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5038 accept->flags = READ_ONCE(sqe->accept_flags);
5039 accept->nofile = rlimit(RLIMIT_NOFILE);
5043 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5045 struct io_accept *accept = &req->accept;
5046 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5047 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5050 if (req->file->f_flags & O_NONBLOCK)
5051 req->flags |= REQ_F_NOWAIT;
5053 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
5054 accept->addr_len, accept->flags,
5056 if (ret == -EAGAIN && force_nonblock)
5059 if (ret == -ERESTARTSYS)
5061 req_set_fail_links(req);
5063 __io_req_complete(req, issue_flags, ret, 0);
5067 static int io_connect_prep_async(struct io_kiocb *req)
5069 struct io_async_connect *io = req->async_data;
5070 struct io_connect *conn = &req->connect;
5072 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5075 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5077 struct io_connect *conn = &req->connect;
5079 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5081 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
5084 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5085 conn->addr_len = READ_ONCE(sqe->addr2);
5089 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5091 struct io_async_connect __io, *io;
5092 unsigned file_flags;
5094 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5096 if (req->async_data) {
5097 io = req->async_data;
5099 ret = move_addr_to_kernel(req->connect.addr,
5100 req->connect.addr_len,
5107 file_flags = force_nonblock ? O_NONBLOCK : 0;
5109 ret = __sys_connect_file(req->file, &io->address,
5110 req->connect.addr_len, file_flags);
5111 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5112 if (req->async_data)
5114 if (io_alloc_async_data(req)) {
5118 io = req->async_data;
5119 memcpy(req->async_data, &__io, sizeof(__io));
5122 if (ret == -ERESTARTSYS)
5126 req_set_fail_links(req);
5127 __io_req_complete(req, issue_flags, ret, 0);
5130 #else /* !CONFIG_NET */
5131 #define IO_NETOP_FN(op) \
5132 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5134 return -EOPNOTSUPP; \
5137 #define IO_NETOP_PREP(op) \
5139 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5141 return -EOPNOTSUPP; \
5144 #define IO_NETOP_PREP_ASYNC(op) \
5146 static int io_##op##_prep_async(struct io_kiocb *req) \
5148 return -EOPNOTSUPP; \
5151 IO_NETOP_PREP_ASYNC(sendmsg);
5152 IO_NETOP_PREP_ASYNC(recvmsg);
5153 IO_NETOP_PREP_ASYNC(connect);
5154 IO_NETOP_PREP(accept);
5157 #endif /* CONFIG_NET */
5159 struct io_poll_table {
5160 struct poll_table_struct pt;
5161 struct io_kiocb *req;
5165 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5166 __poll_t mask, task_work_func_t func)
5170 /* for instances that support it check for an event match first: */
5171 if (mask && !(mask & poll->events))
5174 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5176 list_del_init(&poll->wait.entry);
5179 req->task_work.func = func;
5180 percpu_ref_get(&req->ctx->refs);
5183 * If this fails, then the task is exiting. When a task exits, the
5184 * work gets canceled, so just cancel this request as well instead
5185 * of executing it. We can't safely execute it anyway, as we may not
5186 * have the needed state needed for it anyway.
5188 ret = io_req_task_work_add(req);
5189 if (unlikely(ret)) {
5190 WRITE_ONCE(poll->canceled, true);
5191 io_req_task_work_add_fallback(req, func);
5196 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5197 __acquires(&req->ctx->completion_lock)
5199 struct io_ring_ctx *ctx = req->ctx;
5201 if (!req->result && !READ_ONCE(poll->canceled)) {
5202 struct poll_table_struct pt = { ._key = poll->events };
5204 req->result = vfs_poll(req->file, &pt) & poll->events;
5207 spin_lock_irq(&ctx->completion_lock);
5208 if (!req->result && !READ_ONCE(poll->canceled)) {
5209 add_wait_queue(poll->head, &poll->wait);
5216 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5218 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5219 if (req->opcode == IORING_OP_POLL_ADD)
5220 return req->async_data;
5221 return req->apoll->double_poll;
5224 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5226 if (req->opcode == IORING_OP_POLL_ADD)
5228 return &req->apoll->poll;
5231 static void io_poll_remove_double(struct io_kiocb *req)
5233 struct io_poll_iocb *poll = io_poll_get_double(req);
5235 lockdep_assert_held(&req->ctx->completion_lock);
5237 if (poll && poll->head) {
5238 struct wait_queue_head *head = poll->head;
5240 spin_lock(&head->lock);
5241 list_del_init(&poll->wait.entry);
5242 if (poll->wait.private)
5243 refcount_dec(&req->refs);
5245 spin_unlock(&head->lock);
5249 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5251 struct io_ring_ctx *ctx = req->ctx;
5253 io_poll_remove_double(req);
5254 req->poll.done = true;
5255 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5256 io_commit_cqring(ctx);
5259 static void io_poll_task_func(struct callback_head *cb)
5261 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5262 struct io_ring_ctx *ctx = req->ctx;
5263 struct io_kiocb *nxt;
5265 if (io_poll_rewait(req, &req->poll)) {
5266 spin_unlock_irq(&ctx->completion_lock);
5268 hash_del(&req->hash_node);
5269 io_poll_complete(req, req->result, 0);
5270 spin_unlock_irq(&ctx->completion_lock);
5272 nxt = io_put_req_find_next(req);
5273 io_cqring_ev_posted(ctx);
5275 __io_req_task_submit(nxt);
5278 percpu_ref_put(&ctx->refs);
5281 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5282 int sync, void *key)
5284 struct io_kiocb *req = wait->private;
5285 struct io_poll_iocb *poll = io_poll_get_single(req);
5286 __poll_t mask = key_to_poll(key);
5288 /* for instances that support it check for an event match first: */
5289 if (mask && !(mask & poll->events))
5292 list_del_init(&wait->entry);
5294 if (poll && poll->head) {
5297 spin_lock(&poll->head->lock);
5298 done = list_empty(&poll->wait.entry);
5300 list_del_init(&poll->wait.entry);
5301 /* make sure double remove sees this as being gone */
5302 wait->private = NULL;
5303 spin_unlock(&poll->head->lock);
5305 /* use wait func handler, so it matches the rq type */
5306 poll->wait.func(&poll->wait, mode, sync, key);
5309 refcount_dec(&req->refs);
5313 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5314 wait_queue_func_t wake_func)
5318 poll->canceled = false;
5319 poll->events = events;
5320 INIT_LIST_HEAD(&poll->wait.entry);
5321 init_waitqueue_func_entry(&poll->wait, wake_func);
5324 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5325 struct wait_queue_head *head,
5326 struct io_poll_iocb **poll_ptr)
5328 struct io_kiocb *req = pt->req;
5331 * If poll->head is already set, it's because the file being polled
5332 * uses multiple waitqueues for poll handling (eg one for read, one
5333 * for write). Setup a separate io_poll_iocb if this happens.
5335 if (unlikely(poll->head)) {
5336 struct io_poll_iocb *poll_one = poll;
5338 /* already have a 2nd entry, fail a third attempt */
5340 pt->error = -EINVAL;
5343 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5345 pt->error = -ENOMEM;
5348 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5349 refcount_inc(&req->refs);
5350 poll->wait.private = req;
5357 if (poll->events & EPOLLEXCLUSIVE)
5358 add_wait_queue_exclusive(head, &poll->wait);
5360 add_wait_queue(head, &poll->wait);
5363 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5364 struct poll_table_struct *p)
5366 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5367 struct async_poll *apoll = pt->req->apoll;
5369 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5372 static void io_async_task_func(struct callback_head *cb)
5374 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5375 struct async_poll *apoll = req->apoll;
5376 struct io_ring_ctx *ctx = req->ctx;
5378 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5380 if (io_poll_rewait(req, &apoll->poll)) {
5381 spin_unlock_irq(&ctx->completion_lock);
5382 percpu_ref_put(&ctx->refs);
5386 /* If req is still hashed, it cannot have been canceled. Don't check. */
5387 if (hash_hashed(&req->hash_node))
5388 hash_del(&req->hash_node);
5390 io_poll_remove_double(req);
5391 spin_unlock_irq(&ctx->completion_lock);
5393 if (!READ_ONCE(apoll->poll.canceled))
5394 __io_req_task_submit(req);
5396 __io_req_task_cancel(req, -ECANCELED);
5398 percpu_ref_put(&ctx->refs);
5399 kfree(apoll->double_poll);
5403 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5406 struct io_kiocb *req = wait->private;
5407 struct io_poll_iocb *poll = &req->apoll->poll;
5409 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5412 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5415 static void io_poll_req_insert(struct io_kiocb *req)
5417 struct io_ring_ctx *ctx = req->ctx;
5418 struct hlist_head *list;
5420 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5421 hlist_add_head(&req->hash_node, list);
5424 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5425 struct io_poll_iocb *poll,
5426 struct io_poll_table *ipt, __poll_t mask,
5427 wait_queue_func_t wake_func)
5428 __acquires(&ctx->completion_lock)
5430 struct io_ring_ctx *ctx = req->ctx;
5431 bool cancel = false;
5433 INIT_HLIST_NODE(&req->hash_node);
5434 io_init_poll_iocb(poll, mask, wake_func);
5435 poll->file = req->file;
5436 poll->wait.private = req;
5438 ipt->pt._key = mask;
5440 ipt->error = -EINVAL;
5442 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5444 spin_lock_irq(&ctx->completion_lock);
5445 if (likely(poll->head)) {
5446 spin_lock(&poll->head->lock);
5447 if (unlikely(list_empty(&poll->wait.entry))) {
5453 if (mask || ipt->error)
5454 list_del_init(&poll->wait.entry);
5456 WRITE_ONCE(poll->canceled, true);
5457 else if (!poll->done) /* actually waiting for an event */
5458 io_poll_req_insert(req);
5459 spin_unlock(&poll->head->lock);
5465 static bool io_arm_poll_handler(struct io_kiocb *req)
5467 const struct io_op_def *def = &io_op_defs[req->opcode];
5468 struct io_ring_ctx *ctx = req->ctx;
5469 struct async_poll *apoll;
5470 struct io_poll_table ipt;
5474 if (!req->file || !file_can_poll(req->file))
5476 if (req->flags & REQ_F_POLLED)
5480 else if (def->pollout)
5484 /* if we can't nonblock try, then no point in arming a poll handler */
5485 if (!io_file_supports_async(req->file, rw))
5488 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5489 if (unlikely(!apoll))
5491 apoll->double_poll = NULL;
5493 req->flags |= REQ_F_POLLED;
5498 mask |= POLLIN | POLLRDNORM;
5500 mask |= POLLOUT | POLLWRNORM;
5502 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5503 if ((req->opcode == IORING_OP_RECVMSG) &&
5504 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5507 mask |= POLLERR | POLLPRI;
5509 ipt.pt._qproc = io_async_queue_proc;
5511 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5513 if (ret || ipt.error) {
5514 io_poll_remove_double(req);
5515 spin_unlock_irq(&ctx->completion_lock);
5516 kfree(apoll->double_poll);
5520 spin_unlock_irq(&ctx->completion_lock);
5521 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5522 apoll->poll.events);
5526 static bool __io_poll_remove_one(struct io_kiocb *req,
5527 struct io_poll_iocb *poll)
5529 bool do_complete = false;
5531 spin_lock(&poll->head->lock);
5532 WRITE_ONCE(poll->canceled, true);
5533 if (!list_empty(&poll->wait.entry)) {
5534 list_del_init(&poll->wait.entry);
5537 spin_unlock(&poll->head->lock);
5538 hash_del(&req->hash_node);
5542 static bool io_poll_remove_one(struct io_kiocb *req)
5546 io_poll_remove_double(req);
5548 if (req->opcode == IORING_OP_POLL_ADD) {
5549 do_complete = __io_poll_remove_one(req, &req->poll);
5551 struct async_poll *apoll = req->apoll;
5553 /* non-poll requests have submit ref still */
5554 do_complete = __io_poll_remove_one(req, &apoll->poll);
5557 kfree(apoll->double_poll);
5563 io_cqring_fill_event(req, -ECANCELED);
5564 io_commit_cqring(req->ctx);
5565 req_set_fail_links(req);
5566 io_put_req_deferred(req, 1);
5573 * Returns true if we found and killed one or more poll requests
5575 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5576 struct files_struct *files)
5578 struct hlist_node *tmp;
5579 struct io_kiocb *req;
5582 spin_lock_irq(&ctx->completion_lock);
5583 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5584 struct hlist_head *list;
5586 list = &ctx->cancel_hash[i];
5587 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5588 if (io_match_task(req, tsk, files))
5589 posted += io_poll_remove_one(req);
5592 spin_unlock_irq(&ctx->completion_lock);
5595 io_cqring_ev_posted(ctx);
5600 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5602 struct hlist_head *list;
5603 struct io_kiocb *req;
5605 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5606 hlist_for_each_entry(req, list, hash_node) {
5607 if (sqe_addr != req->user_data)
5609 if (io_poll_remove_one(req))
5617 static int io_poll_remove_prep(struct io_kiocb *req,
5618 const struct io_uring_sqe *sqe)
5620 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5622 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5626 req->poll_remove.addr = READ_ONCE(sqe->addr);
5631 * Find a running poll command that matches one specified in sqe->addr,
5632 * and remove it if found.
5634 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5636 struct io_ring_ctx *ctx = req->ctx;
5639 spin_lock_irq(&ctx->completion_lock);
5640 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5641 spin_unlock_irq(&ctx->completion_lock);
5644 req_set_fail_links(req);
5645 io_req_complete(req, ret);
5649 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5652 struct io_kiocb *req = wait->private;
5653 struct io_poll_iocb *poll = &req->poll;
5655 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5658 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5659 struct poll_table_struct *p)
5661 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5663 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5666 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5668 struct io_poll_iocb *poll = &req->poll;
5671 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5673 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5676 events = READ_ONCE(sqe->poll32_events);
5678 events = swahw32(events);
5680 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5681 (events & EPOLLEXCLUSIVE);
5685 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5687 struct io_poll_iocb *poll = &req->poll;
5688 struct io_ring_ctx *ctx = req->ctx;
5689 struct io_poll_table ipt;
5692 ipt.pt._qproc = io_poll_queue_proc;
5694 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5697 if (mask) { /* no async, we'd stolen it */
5699 io_poll_complete(req, mask, 0);
5701 spin_unlock_irq(&ctx->completion_lock);
5704 io_cqring_ev_posted(ctx);
5710 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5712 struct io_timeout_data *data = container_of(timer,
5713 struct io_timeout_data, timer);
5714 struct io_kiocb *req = data->req;
5715 struct io_ring_ctx *ctx = req->ctx;
5716 unsigned long flags;
5718 spin_lock_irqsave(&ctx->completion_lock, flags);
5719 list_del_init(&req->timeout.list);
5720 atomic_set(&req->ctx->cq_timeouts,
5721 atomic_read(&req->ctx->cq_timeouts) + 1);
5723 io_cqring_fill_event(req, -ETIME);
5724 io_commit_cqring(ctx);
5725 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5727 io_cqring_ev_posted(ctx);
5728 req_set_fail_links(req);
5730 return HRTIMER_NORESTART;
5733 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5736 struct io_timeout_data *io;
5737 struct io_kiocb *req;
5740 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5741 if (user_data == req->user_data) {
5748 return ERR_PTR(ret);
5750 io = req->async_data;
5751 ret = hrtimer_try_to_cancel(&io->timer);
5753 return ERR_PTR(-EALREADY);
5754 list_del_init(&req->timeout.list);
5758 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5760 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5763 return PTR_ERR(req);
5765 req_set_fail_links(req);
5766 io_cqring_fill_event(req, -ECANCELED);
5767 io_put_req_deferred(req, 1);
5771 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5772 struct timespec64 *ts, enum hrtimer_mode mode)
5774 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5775 struct io_timeout_data *data;
5778 return PTR_ERR(req);
5780 req->timeout.off = 0; /* noseq */
5781 data = req->async_data;
5782 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5783 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5784 data->timer.function = io_timeout_fn;
5785 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5789 static int io_timeout_remove_prep(struct io_kiocb *req,
5790 const struct io_uring_sqe *sqe)
5792 struct io_timeout_rem *tr = &req->timeout_rem;
5794 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5796 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5798 if (sqe->ioprio || sqe->buf_index || sqe->len)
5801 tr->addr = READ_ONCE(sqe->addr);
5802 tr->flags = READ_ONCE(sqe->timeout_flags);
5803 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5804 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5806 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5808 } else if (tr->flags) {
5809 /* timeout removal doesn't support flags */
5816 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5818 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5823 * Remove or update an existing timeout command
5825 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5827 struct io_timeout_rem *tr = &req->timeout_rem;
5828 struct io_ring_ctx *ctx = req->ctx;
5831 spin_lock_irq(&ctx->completion_lock);
5832 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5833 ret = io_timeout_cancel(ctx, tr->addr);
5835 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5836 io_translate_timeout_mode(tr->flags));
5838 io_cqring_fill_event(req, ret);
5839 io_commit_cqring(ctx);
5840 spin_unlock_irq(&ctx->completion_lock);
5841 io_cqring_ev_posted(ctx);
5843 req_set_fail_links(req);
5848 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5849 bool is_timeout_link)
5851 struct io_timeout_data *data;
5853 u32 off = READ_ONCE(sqe->off);
5855 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5857 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5859 if (off && is_timeout_link)
5861 flags = READ_ONCE(sqe->timeout_flags);
5862 if (flags & ~IORING_TIMEOUT_ABS)
5865 req->timeout.off = off;
5867 if (!req->async_data && io_alloc_async_data(req))
5870 data = req->async_data;
5873 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5876 data->mode = io_translate_timeout_mode(flags);
5877 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5881 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5883 struct io_ring_ctx *ctx = req->ctx;
5884 struct io_timeout_data *data = req->async_data;
5885 struct list_head *entry;
5886 u32 tail, off = req->timeout.off;
5888 spin_lock_irq(&ctx->completion_lock);
5891 * sqe->off holds how many events that need to occur for this
5892 * timeout event to be satisfied. If it isn't set, then this is
5893 * a pure timeout request, sequence isn't used.
5895 if (io_is_timeout_noseq(req)) {
5896 entry = ctx->timeout_list.prev;
5900 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5901 req->timeout.target_seq = tail + off;
5903 /* Update the last seq here in case io_flush_timeouts() hasn't.
5904 * This is safe because ->completion_lock is held, and submissions
5905 * and completions are never mixed in the same ->completion_lock section.
5907 ctx->cq_last_tm_flush = tail;
5910 * Insertion sort, ensuring the first entry in the list is always
5911 * the one we need first.
5913 list_for_each_prev(entry, &ctx->timeout_list) {
5914 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5917 if (io_is_timeout_noseq(nxt))
5919 /* nxt.seq is behind @tail, otherwise would've been completed */
5920 if (off >= nxt->timeout.target_seq - tail)
5924 list_add(&req->timeout.list, entry);
5925 data->timer.function = io_timeout_fn;
5926 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5927 spin_unlock_irq(&ctx->completion_lock);
5931 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5933 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5935 return req->user_data == (unsigned long) data;
5938 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5940 enum io_wq_cancel cancel_ret;
5943 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5944 switch (cancel_ret) {
5945 case IO_WQ_CANCEL_OK:
5948 case IO_WQ_CANCEL_RUNNING:
5951 case IO_WQ_CANCEL_NOTFOUND:
5959 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5960 struct io_kiocb *req, __u64 sqe_addr,
5963 unsigned long flags;
5966 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5967 if (ret != -ENOENT) {
5968 spin_lock_irqsave(&ctx->completion_lock, flags);
5972 spin_lock_irqsave(&ctx->completion_lock, flags);
5973 ret = io_timeout_cancel(ctx, sqe_addr);
5976 ret = io_poll_cancel(ctx, sqe_addr);
5980 io_cqring_fill_event(req, ret);
5981 io_commit_cqring(ctx);
5982 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5983 io_cqring_ev_posted(ctx);
5986 req_set_fail_links(req);
5990 static int io_async_cancel_prep(struct io_kiocb *req,
5991 const struct io_uring_sqe *sqe)
5993 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5995 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5997 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
6000 req->cancel.addr = READ_ONCE(sqe->addr);
6004 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6006 struct io_ring_ctx *ctx = req->ctx;
6008 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
6012 static int io_rsrc_update_prep(struct io_kiocb *req,
6013 const struct io_uring_sqe *sqe)
6015 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
6017 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6019 if (sqe->ioprio || sqe->rw_flags)
6022 req->rsrc_update.offset = READ_ONCE(sqe->off);
6023 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6024 if (!req->rsrc_update.nr_args)
6026 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6030 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6032 struct io_ring_ctx *ctx = req->ctx;
6033 struct io_uring_rsrc_update up;
6036 if (issue_flags & IO_URING_F_NONBLOCK)
6039 up.offset = req->rsrc_update.offset;
6040 up.data = req->rsrc_update.arg;
6042 mutex_lock(&ctx->uring_lock);
6043 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
6044 mutex_unlock(&ctx->uring_lock);
6047 req_set_fail_links(req);
6048 __io_req_complete(req, issue_flags, ret, 0);
6052 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6054 switch (req->opcode) {
6057 case IORING_OP_READV:
6058 case IORING_OP_READ_FIXED:
6059 case IORING_OP_READ:
6060 return io_read_prep(req, sqe);
6061 case IORING_OP_WRITEV:
6062 case IORING_OP_WRITE_FIXED:
6063 case IORING_OP_WRITE:
6064 return io_write_prep(req, sqe);
6065 case IORING_OP_POLL_ADD:
6066 return io_poll_add_prep(req, sqe);
6067 case IORING_OP_POLL_REMOVE:
6068 return io_poll_remove_prep(req, sqe);
6069 case IORING_OP_FSYNC:
6070 return io_fsync_prep(req, sqe);
6071 case IORING_OP_SYNC_FILE_RANGE:
6072 return io_sfr_prep(req, sqe);
6073 case IORING_OP_SENDMSG:
6074 case IORING_OP_SEND:
6075 return io_sendmsg_prep(req, sqe);
6076 case IORING_OP_RECVMSG:
6077 case IORING_OP_RECV:
6078 return io_recvmsg_prep(req, sqe);
6079 case IORING_OP_CONNECT:
6080 return io_connect_prep(req, sqe);
6081 case IORING_OP_TIMEOUT:
6082 return io_timeout_prep(req, sqe, false);
6083 case IORING_OP_TIMEOUT_REMOVE:
6084 return io_timeout_remove_prep(req, sqe);
6085 case IORING_OP_ASYNC_CANCEL:
6086 return io_async_cancel_prep(req, sqe);
6087 case IORING_OP_LINK_TIMEOUT:
6088 return io_timeout_prep(req, sqe, true);
6089 case IORING_OP_ACCEPT:
6090 return io_accept_prep(req, sqe);
6091 case IORING_OP_FALLOCATE:
6092 return io_fallocate_prep(req, sqe);
6093 case IORING_OP_OPENAT:
6094 return io_openat_prep(req, sqe);
6095 case IORING_OP_CLOSE:
6096 return io_close_prep(req, sqe);
6097 case IORING_OP_FILES_UPDATE:
6098 return io_rsrc_update_prep(req, sqe);
6099 case IORING_OP_STATX:
6100 return io_statx_prep(req, sqe);
6101 case IORING_OP_FADVISE:
6102 return io_fadvise_prep(req, sqe);
6103 case IORING_OP_MADVISE:
6104 return io_madvise_prep(req, sqe);
6105 case IORING_OP_OPENAT2:
6106 return io_openat2_prep(req, sqe);
6107 case IORING_OP_EPOLL_CTL:
6108 return io_epoll_ctl_prep(req, sqe);
6109 case IORING_OP_SPLICE:
6110 return io_splice_prep(req, sqe);
6111 case IORING_OP_PROVIDE_BUFFERS:
6112 return io_provide_buffers_prep(req, sqe);
6113 case IORING_OP_REMOVE_BUFFERS:
6114 return io_remove_buffers_prep(req, sqe);
6116 return io_tee_prep(req, sqe);
6117 case IORING_OP_SHUTDOWN:
6118 return io_shutdown_prep(req, sqe);
6119 case IORING_OP_RENAMEAT:
6120 return io_renameat_prep(req, sqe);
6121 case IORING_OP_UNLINKAT:
6122 return io_unlinkat_prep(req, sqe);
6125 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6130 static int io_req_prep_async(struct io_kiocb *req)
6132 switch (req->opcode) {
6133 case IORING_OP_READV:
6134 case IORING_OP_READ_FIXED:
6135 case IORING_OP_READ:
6136 return io_rw_prep_async(req, READ);
6137 case IORING_OP_WRITEV:
6138 case IORING_OP_WRITE_FIXED:
6139 case IORING_OP_WRITE:
6140 return io_rw_prep_async(req, WRITE);
6141 case IORING_OP_SENDMSG:
6142 case IORING_OP_SEND:
6143 return io_sendmsg_prep_async(req);
6144 case IORING_OP_RECVMSG:
6145 case IORING_OP_RECV:
6146 return io_recvmsg_prep_async(req);
6147 case IORING_OP_CONNECT:
6148 return io_connect_prep_async(req);
6153 static int io_req_defer_prep(struct io_kiocb *req)
6155 if (!io_op_defs[req->opcode].needs_async_data)
6157 /* some opcodes init it during the inital prep */
6158 if (req->async_data)
6160 if (__io_alloc_async_data(req))
6162 return io_req_prep_async(req);
6165 static u32 io_get_sequence(struct io_kiocb *req)
6167 struct io_kiocb *pos;
6168 struct io_ring_ctx *ctx = req->ctx;
6169 u32 total_submitted, nr_reqs = 0;
6171 io_for_each_link(pos, req)
6174 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6175 return total_submitted - nr_reqs;
6178 static int io_req_defer(struct io_kiocb *req)
6180 struct io_ring_ctx *ctx = req->ctx;
6181 struct io_defer_entry *de;
6185 /* Still need defer if there is pending req in defer list. */
6186 if (likely(list_empty_careful(&ctx->defer_list) &&
6187 !(req->flags & REQ_F_IO_DRAIN)))
6190 seq = io_get_sequence(req);
6191 /* Still a chance to pass the sequence check */
6192 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6195 ret = io_req_defer_prep(req);
6198 io_prep_async_link(req);
6199 de = kmalloc(sizeof(*de), GFP_KERNEL);
6203 spin_lock_irq(&ctx->completion_lock);
6204 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6205 spin_unlock_irq(&ctx->completion_lock);
6207 io_queue_async_work(req);
6208 return -EIOCBQUEUED;
6211 trace_io_uring_defer(ctx, req, req->user_data);
6214 list_add_tail(&de->list, &ctx->defer_list);
6215 spin_unlock_irq(&ctx->completion_lock);
6216 return -EIOCBQUEUED;
6219 static void __io_clean_op(struct io_kiocb *req)
6221 if (req->flags & REQ_F_BUFFER_SELECTED) {
6222 switch (req->opcode) {
6223 case IORING_OP_READV:
6224 case IORING_OP_READ_FIXED:
6225 case IORING_OP_READ:
6226 kfree((void *)(unsigned long)req->rw.addr);
6228 case IORING_OP_RECVMSG:
6229 case IORING_OP_RECV:
6230 kfree(req->sr_msg.kbuf);
6233 req->flags &= ~REQ_F_BUFFER_SELECTED;
6236 if (req->flags & REQ_F_NEED_CLEANUP) {
6237 switch (req->opcode) {
6238 case IORING_OP_READV:
6239 case IORING_OP_READ_FIXED:
6240 case IORING_OP_READ:
6241 case IORING_OP_WRITEV:
6242 case IORING_OP_WRITE_FIXED:
6243 case IORING_OP_WRITE: {
6244 struct io_async_rw *io = req->async_data;
6246 kfree(io->free_iovec);
6249 case IORING_OP_RECVMSG:
6250 case IORING_OP_SENDMSG: {
6251 struct io_async_msghdr *io = req->async_data;
6253 kfree(io->free_iov);
6256 case IORING_OP_SPLICE:
6258 io_put_file(req, req->splice.file_in,
6259 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6261 case IORING_OP_OPENAT:
6262 case IORING_OP_OPENAT2:
6263 if (req->open.filename)
6264 putname(req->open.filename);
6266 case IORING_OP_RENAMEAT:
6267 putname(req->rename.oldpath);
6268 putname(req->rename.newpath);
6270 case IORING_OP_UNLINKAT:
6271 putname(req->unlink.filename);
6274 req->flags &= ~REQ_F_NEED_CLEANUP;
6278 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6280 struct io_ring_ctx *ctx = req->ctx;
6283 switch (req->opcode) {
6285 ret = io_nop(req, issue_flags);
6287 case IORING_OP_READV:
6288 case IORING_OP_READ_FIXED:
6289 case IORING_OP_READ:
6290 ret = io_read(req, issue_flags);
6292 case IORING_OP_WRITEV:
6293 case IORING_OP_WRITE_FIXED:
6294 case IORING_OP_WRITE:
6295 ret = io_write(req, issue_flags);
6297 case IORING_OP_FSYNC:
6298 ret = io_fsync(req, issue_flags);
6300 case IORING_OP_POLL_ADD:
6301 ret = io_poll_add(req, issue_flags);
6303 case IORING_OP_POLL_REMOVE:
6304 ret = io_poll_remove(req, issue_flags);
6306 case IORING_OP_SYNC_FILE_RANGE:
6307 ret = io_sync_file_range(req, issue_flags);
6309 case IORING_OP_SENDMSG:
6310 ret = io_sendmsg(req, issue_flags);
6312 case IORING_OP_SEND:
6313 ret = io_send(req, issue_flags);
6315 case IORING_OP_RECVMSG:
6316 ret = io_recvmsg(req, issue_flags);
6318 case IORING_OP_RECV:
6319 ret = io_recv(req, issue_flags);
6321 case IORING_OP_TIMEOUT:
6322 ret = io_timeout(req, issue_flags);
6324 case IORING_OP_TIMEOUT_REMOVE:
6325 ret = io_timeout_remove(req, issue_flags);
6327 case IORING_OP_ACCEPT:
6328 ret = io_accept(req, issue_flags);
6330 case IORING_OP_CONNECT:
6331 ret = io_connect(req, issue_flags);
6333 case IORING_OP_ASYNC_CANCEL:
6334 ret = io_async_cancel(req, issue_flags);
6336 case IORING_OP_FALLOCATE:
6337 ret = io_fallocate(req, issue_flags);
6339 case IORING_OP_OPENAT:
6340 ret = io_openat(req, issue_flags);
6342 case IORING_OP_CLOSE:
6343 ret = io_close(req, issue_flags);
6345 case IORING_OP_FILES_UPDATE:
6346 ret = io_files_update(req, issue_flags);
6348 case IORING_OP_STATX:
6349 ret = io_statx(req, issue_flags);
6351 case IORING_OP_FADVISE:
6352 ret = io_fadvise(req, issue_flags);
6354 case IORING_OP_MADVISE:
6355 ret = io_madvise(req, issue_flags);
6357 case IORING_OP_OPENAT2:
6358 ret = io_openat2(req, issue_flags);
6360 case IORING_OP_EPOLL_CTL:
6361 ret = io_epoll_ctl(req, issue_flags);
6363 case IORING_OP_SPLICE:
6364 ret = io_splice(req, issue_flags);
6366 case IORING_OP_PROVIDE_BUFFERS:
6367 ret = io_provide_buffers(req, issue_flags);
6369 case IORING_OP_REMOVE_BUFFERS:
6370 ret = io_remove_buffers(req, issue_flags);
6373 ret = io_tee(req, issue_flags);
6375 case IORING_OP_SHUTDOWN:
6376 ret = io_shutdown(req, issue_flags);
6378 case IORING_OP_RENAMEAT:
6379 ret = io_renameat(req, issue_flags);
6381 case IORING_OP_UNLINKAT:
6382 ret = io_unlinkat(req, issue_flags);
6392 /* If the op doesn't have a file, we're not polling for it */
6393 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6394 const bool in_async = io_wq_current_is_worker();
6396 /* workqueue context doesn't hold uring_lock, grab it now */
6398 mutex_lock(&ctx->uring_lock);
6400 io_iopoll_req_issued(req, in_async);
6403 mutex_unlock(&ctx->uring_lock);
6409 static void io_wq_submit_work(struct io_wq_work *work)
6411 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6412 struct io_kiocb *timeout;
6415 timeout = io_prep_linked_timeout(req);
6417 io_queue_linked_timeout(timeout);
6419 if (work->flags & IO_WQ_WORK_CANCEL)
6424 ret = io_issue_sqe(req, 0);
6426 * We can get EAGAIN for polled IO even though we're
6427 * forcing a sync submission from here, since we can't
6428 * wait for request slots on the block side.
6436 /* avoid locking problems by failing it from a clean context */
6438 /* io-wq is going to take one down */
6439 refcount_inc(&req->refs);
6440 io_req_task_queue_fail(req, ret);
6444 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6447 struct fixed_rsrc_table *table;
6449 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6450 return table->files[index & IORING_FILE_TABLE_MASK];
6453 static struct file *io_file_get(struct io_submit_state *state,
6454 struct io_kiocb *req, int fd, bool fixed)
6456 struct io_ring_ctx *ctx = req->ctx;
6460 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6462 fd = array_index_nospec(fd, ctx->nr_user_files);
6463 file = io_file_from_index(ctx, fd);
6464 io_set_resource_node(req);
6466 trace_io_uring_file_get(ctx, fd);
6467 file = __io_file_get(state, fd);
6470 if (file && unlikely(file->f_op == &io_uring_fops))
6471 io_req_track_inflight(req);
6475 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6477 struct io_timeout_data *data = container_of(timer,
6478 struct io_timeout_data, timer);
6479 struct io_kiocb *prev, *req = data->req;
6480 struct io_ring_ctx *ctx = req->ctx;
6481 unsigned long flags;
6483 spin_lock_irqsave(&ctx->completion_lock, flags);
6484 prev = req->timeout.head;
6485 req->timeout.head = NULL;
6488 * We don't expect the list to be empty, that will only happen if we
6489 * race with the completion of the linked work.
6491 if (prev && refcount_inc_not_zero(&prev->refs))
6492 io_remove_next_linked(prev);
6495 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6498 req_set_fail_links(prev);
6499 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6500 io_put_req_deferred(prev, 1);
6502 io_req_complete_post(req, -ETIME, 0);
6503 io_put_req_deferred(req, 1);
6505 return HRTIMER_NORESTART;
6508 static void __io_queue_linked_timeout(struct io_kiocb *req)
6511 * If the back reference is NULL, then our linked request finished
6512 * before we got a chance to setup the timer
6514 if (req->timeout.head) {
6515 struct io_timeout_data *data = req->async_data;
6517 data->timer.function = io_link_timeout_fn;
6518 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6523 static void io_queue_linked_timeout(struct io_kiocb *req)
6525 struct io_ring_ctx *ctx = req->ctx;
6527 spin_lock_irq(&ctx->completion_lock);
6528 __io_queue_linked_timeout(req);
6529 spin_unlock_irq(&ctx->completion_lock);
6531 /* drop submission reference */
6535 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6537 struct io_kiocb *nxt = req->link;
6539 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6540 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6543 nxt->timeout.head = req;
6544 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6545 req->flags |= REQ_F_LINK_TIMEOUT;
6549 static void __io_queue_sqe(struct io_kiocb *req)
6551 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6552 const struct cred *old_creds = NULL;
6555 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6556 (req->work.flags & IO_WQ_WORK_CREDS) &&
6557 req->work.identity->creds != current_cred())
6558 old_creds = override_creds(req->work.identity->creds);
6560 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6563 revert_creds(old_creds);
6566 * We async punt it if the file wasn't marked NOWAIT, or if the file
6567 * doesn't support non-blocking read/write attempts
6569 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6570 if (!io_arm_poll_handler(req)) {
6572 * Queued up for async execution, worker will release
6573 * submit reference when the iocb is actually submitted.
6575 io_queue_async_work(req);
6577 } else if (likely(!ret)) {
6578 /* drop submission reference */
6579 if (req->flags & REQ_F_COMPLETE_INLINE) {
6580 struct io_ring_ctx *ctx = req->ctx;
6581 struct io_comp_state *cs = &ctx->submit_state.comp;
6583 cs->reqs[cs->nr++] = req;
6584 if (cs->nr == ARRAY_SIZE(cs->reqs))
6585 io_submit_flush_completions(cs, ctx);
6590 req_set_fail_links(req);
6592 io_req_complete(req, ret);
6595 io_queue_linked_timeout(linked_timeout);
6598 static void io_queue_sqe(struct io_kiocb *req)
6602 ret = io_req_defer(req);
6604 if (ret != -EIOCBQUEUED) {
6606 req_set_fail_links(req);
6608 io_req_complete(req, ret);
6610 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6611 ret = io_req_defer_prep(req);
6614 io_queue_async_work(req);
6616 __io_queue_sqe(req);
6621 * Check SQE restrictions (opcode and flags).
6623 * Returns 'true' if SQE is allowed, 'false' otherwise.
6625 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6626 struct io_kiocb *req,
6627 unsigned int sqe_flags)
6629 if (!ctx->restricted)
6632 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6635 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6636 ctx->restrictions.sqe_flags_required)
6639 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6640 ctx->restrictions.sqe_flags_required))
6646 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6647 const struct io_uring_sqe *sqe)
6649 struct io_submit_state *state;
6650 unsigned int sqe_flags;
6653 req->opcode = READ_ONCE(sqe->opcode);
6654 /* same numerical values with corresponding REQ_F_*, safe to copy */
6655 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6656 req->user_data = READ_ONCE(sqe->user_data);
6657 req->async_data = NULL;
6661 req->fixed_rsrc_refs = NULL;
6662 /* one is dropped after submission, the other at completion */
6663 refcount_set(&req->refs, 2);
6664 req->task = current;
6667 /* enforce forwards compatibility on users */
6668 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6671 if (unlikely(req->opcode >= IORING_OP_LAST))
6674 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6677 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6680 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6681 !io_op_defs[req->opcode].buffer_select)
6684 id = READ_ONCE(sqe->personality);
6686 struct io_identity *iod;
6688 iod = idr_find(&ctx->personality_idr, id);
6691 refcount_inc(&iod->count);
6693 __io_req_init_async(req);
6694 get_cred(iod->creds);
6695 req->work.identity = iod;
6696 req->work.flags |= IO_WQ_WORK_CREDS;
6699 state = &ctx->submit_state;
6702 * Plug now if we have more than 1 IO left after this, and the target
6703 * is potentially a read/write to block based storage.
6705 if (!state->plug_started && state->ios_left > 1 &&
6706 io_op_defs[req->opcode].plug) {
6707 blk_start_plug(&state->plug);
6708 state->plug_started = true;
6711 if (io_op_defs[req->opcode].needs_file) {
6712 bool fixed = req->flags & REQ_F_FIXED_FILE;
6714 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6715 if (unlikely(!req->file))
6723 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6724 const struct io_uring_sqe *sqe)
6726 struct io_submit_link *link = &ctx->submit_state.link;
6729 ret = io_init_req(ctx, req, sqe);
6730 if (unlikely(ret)) {
6733 io_req_complete(req, ret);
6735 /* fail even hard links since we don't submit */
6736 link->head->flags |= REQ_F_FAIL_LINK;
6737 io_put_req(link->head);
6738 io_req_complete(link->head, -ECANCELED);
6743 ret = io_req_prep(req, sqe);
6747 /* don't need @sqe from now on */
6748 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6749 true, ctx->flags & IORING_SETUP_SQPOLL);
6752 * If we already have a head request, queue this one for async
6753 * submittal once the head completes. If we don't have a head but
6754 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6755 * submitted sync once the chain is complete. If none of those
6756 * conditions are true (normal request), then just queue it.
6759 struct io_kiocb *head = link->head;
6762 * Taking sequential execution of a link, draining both sides
6763 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6764 * requests in the link. So, it drains the head and the
6765 * next after the link request. The last one is done via
6766 * drain_next flag to persist the effect across calls.
6768 if (req->flags & REQ_F_IO_DRAIN) {
6769 head->flags |= REQ_F_IO_DRAIN;
6770 ctx->drain_next = 1;
6772 ret = io_req_defer_prep(req);
6775 trace_io_uring_link(ctx, req, head);
6776 link->last->link = req;
6779 /* last request of a link, enqueue the link */
6780 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6785 if (unlikely(ctx->drain_next)) {
6786 req->flags |= REQ_F_IO_DRAIN;
6787 ctx->drain_next = 0;
6789 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6801 * Batched submission is done, ensure local IO is flushed out.
6803 static void io_submit_state_end(struct io_submit_state *state,
6804 struct io_ring_ctx *ctx)
6806 if (state->link.head)
6807 io_queue_sqe(state->link.head);
6809 io_submit_flush_completions(&state->comp, ctx);
6810 if (state->plug_started)
6811 blk_finish_plug(&state->plug);
6812 io_state_file_put(state);
6816 * Start submission side cache.
6818 static void io_submit_state_start(struct io_submit_state *state,
6819 unsigned int max_ios)
6821 state->plug_started = false;
6822 state->ios_left = max_ios;
6823 /* set only head, no need to init link_last in advance */
6824 state->link.head = NULL;
6827 static void io_commit_sqring(struct io_ring_ctx *ctx)
6829 struct io_rings *rings = ctx->rings;
6832 * Ensure any loads from the SQEs are done at this point,
6833 * since once we write the new head, the application could
6834 * write new data to them.
6836 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6840 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6841 * that is mapped by userspace. This means that care needs to be taken to
6842 * ensure that reads are stable, as we cannot rely on userspace always
6843 * being a good citizen. If members of the sqe are validated and then later
6844 * used, it's important that those reads are done through READ_ONCE() to
6845 * prevent a re-load down the line.
6847 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6849 u32 *sq_array = ctx->sq_array;
6853 * The cached sq head (or cq tail) serves two purposes:
6855 * 1) allows us to batch the cost of updating the user visible
6857 * 2) allows the kernel side to track the head on its own, even
6858 * though the application is the one updating it.
6860 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6861 if (likely(head < ctx->sq_entries))
6862 return &ctx->sq_sqes[head];
6864 /* drop invalid entries */
6865 ctx->cached_sq_dropped++;
6866 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6870 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6874 /* if we have a backlog and couldn't flush it all, return BUSY */
6875 if (test_bit(0, &ctx->sq_check_overflow)) {
6876 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6880 /* make sure SQ entry isn't read before tail */
6881 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6883 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6886 percpu_counter_add(¤t->io_uring->inflight, nr);
6887 refcount_add(nr, ¤t->usage);
6888 io_submit_state_start(&ctx->submit_state, nr);
6890 while (submitted < nr) {
6891 const struct io_uring_sqe *sqe;
6892 struct io_kiocb *req;
6894 req = io_alloc_req(ctx);
6895 if (unlikely(!req)) {
6897 submitted = -EAGAIN;
6900 sqe = io_get_sqe(ctx);
6901 if (unlikely(!sqe)) {
6902 kmem_cache_free(req_cachep, req);
6905 /* will complete beyond this point, count as submitted */
6907 if (io_submit_sqe(ctx, req, sqe))
6911 if (unlikely(submitted != nr)) {
6912 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6913 struct io_uring_task *tctx = current->io_uring;
6914 int unused = nr - ref_used;
6916 percpu_ref_put_many(&ctx->refs, unused);
6917 percpu_counter_sub(&tctx->inflight, unused);
6918 put_task_struct_many(current, unused);
6921 io_submit_state_end(&ctx->submit_state, ctx);
6922 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6923 io_commit_sqring(ctx);
6928 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6930 /* Tell userspace we may need a wakeup call */
6931 spin_lock_irq(&ctx->completion_lock);
6932 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6933 spin_unlock_irq(&ctx->completion_lock);
6936 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6938 spin_lock_irq(&ctx->completion_lock);
6939 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6940 spin_unlock_irq(&ctx->completion_lock);
6943 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6945 unsigned int to_submit;
6948 to_submit = io_sqring_entries(ctx);
6949 /* if we're handling multiple rings, cap submit size for fairness */
6950 if (cap_entries && to_submit > 8)
6953 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6954 unsigned nr_events = 0;
6956 mutex_lock(&ctx->uring_lock);
6957 if (!list_empty(&ctx->iopoll_list))
6958 io_do_iopoll(ctx, &nr_events, 0);
6960 if (to_submit && !ctx->sqo_dead &&
6961 likely(!percpu_ref_is_dying(&ctx->refs)))
6962 ret = io_submit_sqes(ctx, to_submit);
6963 mutex_unlock(&ctx->uring_lock);
6966 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6967 wake_up(&ctx->sqo_sq_wait);
6972 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6974 struct io_ring_ctx *ctx;
6975 unsigned sq_thread_idle = 0;
6977 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6978 if (sq_thread_idle < ctx->sq_thread_idle)
6979 sq_thread_idle = ctx->sq_thread_idle;
6982 sqd->sq_thread_idle = sq_thread_idle;
6985 static void io_sqd_init_new(struct io_sq_data *sqd)
6987 struct io_ring_ctx *ctx;
6989 while (!list_empty(&sqd->ctx_new_list)) {
6990 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6991 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6992 complete(&ctx->sq_thread_comp);
6995 io_sqd_update_thread_idle(sqd);
6998 static int io_sq_thread(void *data)
7000 struct cgroup_subsys_state *cur_css = NULL;
7001 struct files_struct *old_files = current->files;
7002 struct nsproxy *old_nsproxy = current->nsproxy;
7003 const struct cred *old_cred = NULL;
7004 struct io_sq_data *sqd = data;
7005 struct io_ring_ctx *ctx;
7006 unsigned long timeout = 0;
7010 current->files = NULL;
7011 current->nsproxy = NULL;
7012 task_unlock(current);
7014 while (!kthread_should_stop()) {
7016 bool cap_entries, sqt_spin, needs_sched;
7019 * Any changes to the sqd lists are synchronized through the
7020 * kthread parking. This synchronizes the thread vs users,
7021 * the users are synchronized on the sqd->ctx_lock.
7023 if (kthread_should_park()) {
7026 * When sq thread is unparked, in case the previous park operation
7027 * comes from io_put_sq_data(), which means that sq thread is going
7028 * to be stopped, so here needs to have a check.
7030 if (kthread_should_stop())
7034 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7035 io_sqd_init_new(sqd);
7036 timeout = jiffies + sqd->sq_thread_idle;
7040 cap_entries = !list_is_singular(&sqd->ctx_list);
7041 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7042 if (current->cred != ctx->creds) {
7044 revert_creds(old_cred);
7045 old_cred = override_creds(ctx->creds);
7047 io_sq_thread_associate_blkcg(ctx, &cur_css);
7049 current->loginuid = ctx->loginuid;
7050 current->sessionid = ctx->sessionid;
7053 ret = __io_sq_thread(ctx, cap_entries);
7054 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7057 io_sq_thread_drop_mm_files();
7060 if (sqt_spin || !time_after(jiffies, timeout)) {
7062 io_sq_thread_drop_mm_files();
7065 timeout = jiffies + sqd->sq_thread_idle;
7070 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7071 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7072 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7073 !list_empty_careful(&ctx->iopoll_list)) {
7074 needs_sched = false;
7077 if (io_sqring_entries(ctx)) {
7078 needs_sched = false;
7083 if (needs_sched && !kthread_should_park()) {
7084 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7085 io_ring_set_wakeup_flag(ctx);
7088 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7089 io_ring_clear_wakeup_flag(ctx);
7092 finish_wait(&sqd->wait, &wait);
7093 timeout = jiffies + sqd->sq_thread_idle;
7097 io_sq_thread_drop_mm_files();
7100 io_sq_thread_unassociate_blkcg();
7102 revert_creds(old_cred);
7105 current->files = old_files;
7106 current->nsproxy = old_nsproxy;
7107 task_unlock(current);
7114 struct io_wait_queue {
7115 struct wait_queue_entry wq;
7116 struct io_ring_ctx *ctx;
7118 unsigned nr_timeouts;
7121 static inline bool io_should_wake(struct io_wait_queue *iowq)
7123 struct io_ring_ctx *ctx = iowq->ctx;
7126 * Wake up if we have enough events, or if a timeout occurred since we
7127 * started waiting. For timeouts, we always want to return to userspace,
7128 * regardless of event count.
7130 return io_cqring_events(ctx) >= iowq->to_wait ||
7131 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7134 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7135 int wake_flags, void *key)
7137 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7141 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7142 * the task, and the next invocation will do it.
7144 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
7145 return autoremove_wake_function(curr, mode, wake_flags, key);
7149 static int io_run_task_work_sig(void)
7151 if (io_run_task_work())
7153 if (!signal_pending(current))
7155 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7156 return -ERESTARTSYS;
7160 /* when returns >0, the caller should retry */
7161 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7162 struct io_wait_queue *iowq,
7163 signed long *timeout)
7167 /* make sure we run task_work before checking for signals */
7168 ret = io_run_task_work_sig();
7169 if (ret || io_should_wake(iowq))
7171 /* let the caller flush overflows, retry */
7172 if (test_bit(0, &ctx->cq_check_overflow))
7175 *timeout = schedule_timeout(*timeout);
7176 return !*timeout ? -ETIME : 1;
7180 * Wait until events become available, if we don't already have some. The
7181 * application must reap them itself, as they reside on the shared cq ring.
7183 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7184 const sigset_t __user *sig, size_t sigsz,
7185 struct __kernel_timespec __user *uts)
7187 struct io_wait_queue iowq = {
7190 .func = io_wake_function,
7191 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7194 .to_wait = min_events,
7196 struct io_rings *rings = ctx->rings;
7197 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7201 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7202 if (io_cqring_events(ctx) >= min_events)
7204 if (!io_run_task_work())
7209 #ifdef CONFIG_COMPAT
7210 if (in_compat_syscall())
7211 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7215 ret = set_user_sigmask(sig, sigsz);
7222 struct timespec64 ts;
7224 if (get_timespec64(&ts, uts))
7226 timeout = timespec64_to_jiffies(&ts);
7229 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7230 trace_io_uring_cqring_wait(ctx, min_events);
7232 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7233 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7234 TASK_INTERRUPTIBLE);
7235 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7236 finish_wait(&ctx->wait, &iowq.wq);
7239 restore_saved_sigmask_unless(ret == -EINTR);
7241 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7244 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7246 #if defined(CONFIG_UNIX)
7247 if (ctx->ring_sock) {
7248 struct sock *sock = ctx->ring_sock->sk;
7249 struct sk_buff *skb;
7251 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7257 for (i = 0; i < ctx->nr_user_files; i++) {
7260 file = io_file_from_index(ctx, i);
7267 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
7269 struct fixed_rsrc_data *data;
7271 data = container_of(ref, struct fixed_rsrc_data, refs);
7272 complete(&data->done);
7275 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7277 spin_lock_bh(&ctx->rsrc_ref_lock);
7280 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7282 spin_unlock_bh(&ctx->rsrc_ref_lock);
7285 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7286 struct fixed_rsrc_data *rsrc_data,
7287 struct fixed_rsrc_ref_node *ref_node)
7289 io_rsrc_ref_lock(ctx);
7290 rsrc_data->node = ref_node;
7291 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7292 io_rsrc_ref_unlock(ctx);
7293 percpu_ref_get(&rsrc_data->refs);
7296 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7298 struct fixed_rsrc_ref_node *ref_node = NULL;
7300 io_rsrc_ref_lock(ctx);
7301 ref_node = data->node;
7303 io_rsrc_ref_unlock(ctx);
7305 percpu_ref_kill(&ref_node->refs);
7308 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7309 struct io_ring_ctx *ctx,
7310 struct fixed_rsrc_ref_node *backup_node)
7317 data->quiesce = true;
7319 io_sqe_rsrc_kill_node(ctx, data);
7320 percpu_ref_kill(&data->refs);
7321 flush_delayed_work(&ctx->rsrc_put_work);
7323 ret = wait_for_completion_interruptible(&data->done);
7327 percpu_ref_resurrect(&data->refs);
7328 io_sqe_rsrc_set_node(ctx, data, backup_node);
7330 reinit_completion(&data->done);
7331 mutex_unlock(&ctx->uring_lock);
7332 ret = io_run_task_work_sig();
7333 mutex_lock(&ctx->uring_lock);
7337 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7341 init_fixed_file_ref_node(ctx, backup_node);
7343 data->quiesce = false;
7346 destroy_fixed_rsrc_ref_node(backup_node);
7350 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7352 struct fixed_rsrc_data *data;
7354 data = kzalloc(sizeof(*data), GFP_KERNEL);
7358 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7359 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7364 init_completion(&data->done);
7368 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7370 percpu_ref_exit(&data->refs);
7375 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7377 struct fixed_rsrc_data *data = ctx->file_data;
7378 struct fixed_rsrc_ref_node *backup_node;
7379 unsigned nr_tables, i;
7383 * percpu_ref_is_dying() is to stop parallel files unregister
7384 * Since we possibly drop uring lock later in this function to
7387 if (!data || percpu_ref_is_dying(&data->refs))
7389 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7392 init_fixed_file_ref_node(ctx, backup_node);
7394 ret = io_rsrc_ref_quiesce(data, ctx, backup_node);
7398 __io_sqe_files_unregister(ctx);
7399 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7400 for (i = 0; i < nr_tables; i++)
7401 kfree(data->table[i].files);
7402 free_fixed_rsrc_data(data);
7403 ctx->file_data = NULL;
7404 ctx->nr_user_files = 0;
7408 static void io_put_sq_data(struct io_sq_data *sqd)
7410 if (refcount_dec_and_test(&sqd->refs)) {
7412 * The park is a bit of a work-around, without it we get
7413 * warning spews on shutdown with SQPOLL set and affinity
7414 * set to a single CPU.
7417 kthread_park(sqd->thread);
7418 kthread_stop(sqd->thread);
7425 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7427 struct io_ring_ctx *ctx_attach;
7428 struct io_sq_data *sqd;
7431 f = fdget(p->wq_fd);
7433 return ERR_PTR(-ENXIO);
7434 if (f.file->f_op != &io_uring_fops) {
7436 return ERR_PTR(-EINVAL);
7439 ctx_attach = f.file->private_data;
7440 sqd = ctx_attach->sq_data;
7443 return ERR_PTR(-EINVAL);
7446 refcount_inc(&sqd->refs);
7451 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7453 struct io_sq_data *sqd;
7455 if (p->flags & IORING_SETUP_ATTACH_WQ)
7456 return io_attach_sq_data(p);
7458 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7460 return ERR_PTR(-ENOMEM);
7462 refcount_set(&sqd->refs, 1);
7463 INIT_LIST_HEAD(&sqd->ctx_list);
7464 INIT_LIST_HEAD(&sqd->ctx_new_list);
7465 mutex_init(&sqd->ctx_lock);
7466 mutex_init(&sqd->lock);
7467 init_waitqueue_head(&sqd->wait);
7471 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7472 __releases(&sqd->lock)
7476 kthread_unpark(sqd->thread);
7477 mutex_unlock(&sqd->lock);
7480 static void io_sq_thread_park(struct io_sq_data *sqd)
7481 __acquires(&sqd->lock)
7485 mutex_lock(&sqd->lock);
7486 kthread_park(sqd->thread);
7489 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7491 struct io_sq_data *sqd = ctx->sq_data;
7496 * We may arrive here from the error branch in
7497 * io_sq_offload_create() where the kthread is created
7498 * without being waked up, thus wake it up now to make
7499 * sure the wait will complete.
7501 wake_up_process(sqd->thread);
7502 wait_for_completion(&ctx->sq_thread_comp);
7504 io_sq_thread_park(sqd);
7507 mutex_lock(&sqd->ctx_lock);
7508 list_del(&ctx->sqd_list);
7509 io_sqd_update_thread_idle(sqd);
7510 mutex_unlock(&sqd->ctx_lock);
7513 io_sq_thread_unpark(sqd);
7515 io_put_sq_data(sqd);
7516 ctx->sq_data = NULL;
7520 static void io_finish_async(struct io_ring_ctx *ctx)
7522 io_sq_thread_stop(ctx);
7525 io_wq_destroy(ctx->io_wq);
7530 #if defined(CONFIG_UNIX)
7532 * Ensure the UNIX gc is aware of our file set, so we are certain that
7533 * the io_uring can be safely unregistered on process exit, even if we have
7534 * loops in the file referencing.
7536 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7538 struct sock *sk = ctx->ring_sock->sk;
7539 struct scm_fp_list *fpl;
7540 struct sk_buff *skb;
7543 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7547 skb = alloc_skb(0, GFP_KERNEL);
7556 fpl->user = get_uid(ctx->user);
7557 for (i = 0; i < nr; i++) {
7558 struct file *file = io_file_from_index(ctx, i + offset);
7562 fpl->fp[nr_files] = get_file(file);
7563 unix_inflight(fpl->user, fpl->fp[nr_files]);
7568 fpl->max = SCM_MAX_FD;
7569 fpl->count = nr_files;
7570 UNIXCB(skb).fp = fpl;
7571 skb->destructor = unix_destruct_scm;
7572 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7573 skb_queue_head(&sk->sk_receive_queue, skb);
7575 for (i = 0; i < nr_files; i++)
7586 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7587 * causes regular reference counting to break down. We rely on the UNIX
7588 * garbage collection to take care of this problem for us.
7590 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7592 unsigned left, total;
7596 left = ctx->nr_user_files;
7598 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7600 ret = __io_sqe_files_scm(ctx, this_files, total);
7604 total += this_files;
7610 while (total < ctx->nr_user_files) {
7611 struct file *file = io_file_from_index(ctx, total);
7621 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7627 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7628 unsigned nr_tables, unsigned nr_files)
7632 for (i = 0; i < nr_tables; i++) {
7633 struct fixed_rsrc_table *table = &file_data->table[i];
7634 unsigned this_files;
7636 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7637 table->files = kcalloc(this_files, sizeof(struct file *),
7641 nr_files -= this_files;
7647 for (i = 0; i < nr_tables; i++) {
7648 struct fixed_rsrc_table *table = &file_data->table[i];
7649 kfree(table->files);
7654 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7656 struct file *file = prsrc->file;
7657 #if defined(CONFIG_UNIX)
7658 struct sock *sock = ctx->ring_sock->sk;
7659 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7660 struct sk_buff *skb;
7663 __skb_queue_head_init(&list);
7666 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7667 * remove this entry and rearrange the file array.
7669 skb = skb_dequeue(head);
7671 struct scm_fp_list *fp;
7673 fp = UNIXCB(skb).fp;
7674 for (i = 0; i < fp->count; i++) {
7677 if (fp->fp[i] != file)
7680 unix_notinflight(fp->user, fp->fp[i]);
7681 left = fp->count - 1 - i;
7683 memmove(&fp->fp[i], &fp->fp[i + 1],
7684 left * sizeof(struct file *));
7691 __skb_queue_tail(&list, skb);
7701 __skb_queue_tail(&list, skb);
7703 skb = skb_dequeue(head);
7706 if (skb_peek(&list)) {
7707 spin_lock_irq(&head->lock);
7708 while ((skb = __skb_dequeue(&list)) != NULL)
7709 __skb_queue_tail(head, skb);
7710 spin_unlock_irq(&head->lock);
7717 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7719 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7720 struct io_ring_ctx *ctx = rsrc_data->ctx;
7721 struct io_rsrc_put *prsrc, *tmp;
7723 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7724 list_del(&prsrc->list);
7725 ref_node->rsrc_put(ctx, prsrc);
7729 percpu_ref_exit(&ref_node->refs);
7731 percpu_ref_put(&rsrc_data->refs);
7734 static void io_rsrc_put_work(struct work_struct *work)
7736 struct io_ring_ctx *ctx;
7737 struct llist_node *node;
7739 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7740 node = llist_del_all(&ctx->rsrc_put_llist);
7743 struct fixed_rsrc_ref_node *ref_node;
7744 struct llist_node *next = node->next;
7746 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7747 __io_rsrc_put_work(ref_node);
7752 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7755 struct fixed_rsrc_table *table;
7757 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7758 return &table->files[i & IORING_FILE_TABLE_MASK];
7761 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7763 struct fixed_rsrc_ref_node *ref_node;
7764 struct fixed_rsrc_data *data;
7765 struct io_ring_ctx *ctx;
7766 bool first_add = false;
7769 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7770 data = ref_node->rsrc_data;
7773 io_rsrc_ref_lock(ctx);
7774 ref_node->done = true;
7776 while (!list_empty(&ctx->rsrc_ref_list)) {
7777 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7778 struct fixed_rsrc_ref_node, node);
7779 /* recycle ref nodes in order */
7780 if (!ref_node->done)
7782 list_del(&ref_node->node);
7783 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7785 io_rsrc_ref_unlock(ctx);
7787 if (percpu_ref_is_dying(&data->refs))
7791 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7793 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7796 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7797 struct io_ring_ctx *ctx)
7799 struct fixed_rsrc_ref_node *ref_node;
7801 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7805 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7810 INIT_LIST_HEAD(&ref_node->node);
7811 INIT_LIST_HEAD(&ref_node->rsrc_list);
7812 ref_node->done = false;
7816 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7817 struct fixed_rsrc_ref_node *ref_node)
7819 ref_node->rsrc_data = ctx->file_data;
7820 ref_node->rsrc_put = io_ring_file_put;
7823 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7825 percpu_ref_exit(&ref_node->refs);
7830 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7833 __s32 __user *fds = (__s32 __user *) arg;
7834 unsigned nr_tables, i;
7836 int fd, ret = -ENOMEM;
7837 struct fixed_rsrc_ref_node *ref_node;
7838 struct fixed_rsrc_data *file_data;
7844 if (nr_args > IORING_MAX_FIXED_FILES)
7847 file_data = alloc_fixed_rsrc_data(ctx);
7850 ctx->file_data = file_data;
7852 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7853 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7855 if (!file_data->table)
7858 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7861 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7862 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7866 /* allow sparse sets */
7876 * Don't allow io_uring instances to be registered. If UNIX
7877 * isn't enabled, then this causes a reference cycle and this
7878 * instance can never get freed. If UNIX is enabled we'll
7879 * handle it just fine, but there's still no point in allowing
7880 * a ring fd as it doesn't support regular read/write anyway.
7882 if (file->f_op == &io_uring_fops) {
7886 *io_fixed_file_slot(file_data, i) = file;
7889 ret = io_sqe_files_scm(ctx);
7891 io_sqe_files_unregister(ctx);
7895 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7897 io_sqe_files_unregister(ctx);
7900 init_fixed_file_ref_node(ctx, ref_node);
7902 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7905 for (i = 0; i < ctx->nr_user_files; i++) {
7906 file = io_file_from_index(ctx, i);
7910 for (i = 0; i < nr_tables; i++)
7911 kfree(file_data->table[i].files);
7912 ctx->nr_user_files = 0;
7914 free_fixed_rsrc_data(ctx->file_data);
7915 ctx->file_data = NULL;
7919 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7922 #if defined(CONFIG_UNIX)
7923 struct sock *sock = ctx->ring_sock->sk;
7924 struct sk_buff_head *head = &sock->sk_receive_queue;
7925 struct sk_buff *skb;
7928 * See if we can merge this file into an existing skb SCM_RIGHTS
7929 * file set. If there's no room, fall back to allocating a new skb
7930 * and filling it in.
7932 spin_lock_irq(&head->lock);
7933 skb = skb_peek(head);
7935 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7937 if (fpl->count < SCM_MAX_FD) {
7938 __skb_unlink(skb, head);
7939 spin_unlock_irq(&head->lock);
7940 fpl->fp[fpl->count] = get_file(file);
7941 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7943 spin_lock_irq(&head->lock);
7944 __skb_queue_head(head, skb);
7949 spin_unlock_irq(&head->lock);
7956 return __io_sqe_files_scm(ctx, 1, index);
7962 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7964 struct io_rsrc_put *prsrc;
7965 struct fixed_rsrc_ref_node *ref_node = data->node;
7967 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7972 list_add(&prsrc->list, &ref_node->rsrc_list);
7977 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7980 return io_queue_rsrc_removal(data, (void *)file);
7983 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7984 struct io_uring_rsrc_update *up,
7987 struct fixed_rsrc_data *data = ctx->file_data;
7988 struct fixed_rsrc_ref_node *ref_node;
7989 struct file *file, **file_slot;
7993 bool needs_switch = false;
7995 if (check_add_overflow(up->offset, nr_args, &done))
7997 if (done > ctx->nr_user_files)
8000 ref_node = alloc_fixed_rsrc_ref_node(ctx);
8003 init_fixed_file_ref_node(ctx, ref_node);
8005 fds = u64_to_user_ptr(up->data);
8006 for (done = 0; done < nr_args; done++) {
8008 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
8012 if (fd == IORING_REGISTER_FILES_SKIP)
8015 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8016 file_slot = io_fixed_file_slot(ctx->file_data, i);
8019 err = io_queue_file_removal(data, *file_slot);
8023 needs_switch = true;
8032 * Don't allow io_uring instances to be registered. If
8033 * UNIX isn't enabled, then this causes a reference
8034 * cycle and this instance can never get freed. If UNIX
8035 * is enabled we'll handle it just fine, but there's
8036 * still no point in allowing a ring fd as it doesn't
8037 * support regular read/write anyway.
8039 if (file->f_op == &io_uring_fops) {
8045 err = io_sqe_file_register(ctx, file, i);
8055 percpu_ref_kill(&data->node->refs);
8056 io_sqe_rsrc_set_node(ctx, data, ref_node);
8058 destroy_fixed_rsrc_ref_node(ref_node);
8060 return done ? done : err;
8063 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
8066 struct io_uring_rsrc_update up;
8068 if (!ctx->file_data)
8072 if (copy_from_user(&up, arg, sizeof(up)))
8077 return __io_sqe_files_update(ctx, &up, nr_args);
8080 static struct io_wq_work *io_free_work(struct io_wq_work *work)
8082 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8084 req = io_put_req_find_next(req);
8085 return req ? &req->work : NULL;
8088 static int io_init_wq_offload(struct io_ring_ctx *ctx,
8089 struct io_uring_params *p)
8091 struct io_wq_data data;
8093 struct io_ring_ctx *ctx_attach;
8094 unsigned int concurrency;
8097 data.user = ctx->user;
8098 data.free_work = io_free_work;
8099 data.do_work = io_wq_submit_work;
8101 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
8102 /* Do QD, or 4 * CPUS, whatever is smallest */
8103 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8105 ctx->io_wq = io_wq_create(concurrency, &data);
8106 if (IS_ERR(ctx->io_wq)) {
8107 ret = PTR_ERR(ctx->io_wq);
8113 f = fdget(p->wq_fd);
8117 if (f.file->f_op != &io_uring_fops) {
8122 ctx_attach = f.file->private_data;
8123 /* @io_wq is protected by holding the fd */
8124 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8129 ctx->io_wq = ctx_attach->io_wq;
8135 static int io_uring_alloc_task_context(struct task_struct *task)
8137 struct io_uring_task *tctx;
8140 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8141 if (unlikely(!tctx))
8144 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8145 if (unlikely(ret)) {
8151 init_waitqueue_head(&tctx->wait);
8153 atomic_set(&tctx->in_idle, 0);
8154 tctx->sqpoll = false;
8155 io_init_identity(&tctx->__identity);
8156 tctx->identity = &tctx->__identity;
8157 task->io_uring = tctx;
8158 spin_lock_init(&tctx->task_lock);
8159 INIT_WQ_LIST(&tctx->task_list);
8160 tctx->task_state = 0;
8161 init_task_work(&tctx->task_work, tctx_task_work);
8165 void __io_uring_free(struct task_struct *tsk)
8167 struct io_uring_task *tctx = tsk->io_uring;
8169 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8170 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8171 if (tctx->identity != &tctx->__identity)
8172 kfree(tctx->identity);
8173 percpu_counter_destroy(&tctx->inflight);
8175 tsk->io_uring = NULL;
8178 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8179 struct io_uring_params *p)
8183 if (ctx->flags & IORING_SETUP_SQPOLL) {
8184 struct io_sq_data *sqd;
8187 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8190 sqd = io_get_sq_data(p);
8197 io_sq_thread_park(sqd);
8198 mutex_lock(&sqd->ctx_lock);
8199 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8200 mutex_unlock(&sqd->ctx_lock);
8201 io_sq_thread_unpark(sqd);
8203 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8204 if (!ctx->sq_thread_idle)
8205 ctx->sq_thread_idle = HZ;
8210 if (p->flags & IORING_SETUP_SQ_AFF) {
8211 int cpu = p->sq_thread_cpu;
8214 if (cpu >= nr_cpu_ids)
8216 if (!cpu_online(cpu))
8219 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8220 cpu, "io_uring-sq");
8222 sqd->thread = kthread_create(io_sq_thread, sqd,
8225 if (IS_ERR(sqd->thread)) {
8226 ret = PTR_ERR(sqd->thread);
8230 ret = io_uring_alloc_task_context(sqd->thread);
8233 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8234 /* Can't have SQ_AFF without SQPOLL */
8240 ret = io_init_wq_offload(ctx, p);
8246 io_finish_async(ctx);
8250 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8252 struct io_sq_data *sqd = ctx->sq_data;
8254 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8255 wake_up_process(sqd->thread);
8258 static inline void __io_unaccount_mem(struct user_struct *user,
8259 unsigned long nr_pages)
8261 atomic_long_sub(nr_pages, &user->locked_vm);
8264 static inline int __io_account_mem(struct user_struct *user,
8265 unsigned long nr_pages)
8267 unsigned long page_limit, cur_pages, new_pages;
8269 /* Don't allow more pages than we can safely lock */
8270 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8273 cur_pages = atomic_long_read(&user->locked_vm);
8274 new_pages = cur_pages + nr_pages;
8275 if (new_pages > page_limit)
8277 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8278 new_pages) != cur_pages);
8283 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8286 __io_unaccount_mem(ctx->user, nr_pages);
8288 if (ctx->mm_account)
8289 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8292 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8296 if (ctx->limit_mem) {
8297 ret = __io_account_mem(ctx->user, nr_pages);
8302 if (ctx->mm_account)
8303 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8308 static void io_mem_free(void *ptr)
8315 page = virt_to_head_page(ptr);
8316 if (put_page_testzero(page))
8317 free_compound_page(page);
8320 static void *io_mem_alloc(size_t size)
8322 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8323 __GFP_NORETRY | __GFP_ACCOUNT;
8325 return (void *) __get_free_pages(gfp_flags, get_order(size));
8328 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8331 struct io_rings *rings;
8332 size_t off, sq_array_size;
8334 off = struct_size(rings, cqes, cq_entries);
8335 if (off == SIZE_MAX)
8339 off = ALIGN(off, SMP_CACHE_BYTES);
8347 sq_array_size = array_size(sizeof(u32), sq_entries);
8348 if (sq_array_size == SIZE_MAX)
8351 if (check_add_overflow(off, sq_array_size, &off))
8357 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8361 if (!ctx->user_bufs)
8364 for (i = 0; i < ctx->nr_user_bufs; i++) {
8365 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8367 for (j = 0; j < imu->nr_bvecs; j++)
8368 unpin_user_page(imu->bvec[j].bv_page);
8370 if (imu->acct_pages)
8371 io_unaccount_mem(ctx, imu->acct_pages);
8376 kfree(ctx->user_bufs);
8377 ctx->user_bufs = NULL;
8378 ctx->nr_user_bufs = 0;
8382 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8383 void __user *arg, unsigned index)
8385 struct iovec __user *src;
8387 #ifdef CONFIG_COMPAT
8389 struct compat_iovec __user *ciovs;
8390 struct compat_iovec ciov;
8392 ciovs = (struct compat_iovec __user *) arg;
8393 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8396 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8397 dst->iov_len = ciov.iov_len;
8401 src = (struct iovec __user *) arg;
8402 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8408 * Not super efficient, but this is just a registration time. And we do cache
8409 * the last compound head, so generally we'll only do a full search if we don't
8412 * We check if the given compound head page has already been accounted, to
8413 * avoid double accounting it. This allows us to account the full size of the
8414 * page, not just the constituent pages of a huge page.
8416 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8417 int nr_pages, struct page *hpage)
8421 /* check current page array */
8422 for (i = 0; i < nr_pages; i++) {
8423 if (!PageCompound(pages[i]))
8425 if (compound_head(pages[i]) == hpage)
8429 /* check previously registered pages */
8430 for (i = 0; i < ctx->nr_user_bufs; i++) {
8431 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8433 for (j = 0; j < imu->nr_bvecs; j++) {
8434 if (!PageCompound(imu->bvec[j].bv_page))
8436 if (compound_head(imu->bvec[j].bv_page) == hpage)
8444 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8445 int nr_pages, struct io_mapped_ubuf *imu,
8446 struct page **last_hpage)
8450 for (i = 0; i < nr_pages; i++) {
8451 if (!PageCompound(pages[i])) {
8456 hpage = compound_head(pages[i]);
8457 if (hpage == *last_hpage)
8459 *last_hpage = hpage;
8460 if (headpage_already_acct(ctx, pages, i, hpage))
8462 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8466 if (!imu->acct_pages)
8469 ret = io_account_mem(ctx, imu->acct_pages);
8471 imu->acct_pages = 0;
8475 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8476 struct io_mapped_ubuf *imu,
8477 struct page **last_hpage)
8479 struct vm_area_struct **vmas = NULL;
8480 struct page **pages = NULL;
8481 unsigned long off, start, end, ubuf;
8483 int ret, pret, nr_pages, i;
8485 ubuf = (unsigned long) iov->iov_base;
8486 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8487 start = ubuf >> PAGE_SHIFT;
8488 nr_pages = end - start;
8492 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8496 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8501 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8507 mmap_read_lock(current->mm);
8508 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8510 if (pret == nr_pages) {
8511 /* don't support file backed memory */
8512 for (i = 0; i < nr_pages; i++) {
8513 struct vm_area_struct *vma = vmas[i];
8516 !is_file_hugepages(vma->vm_file)) {
8522 ret = pret < 0 ? pret : -EFAULT;
8524 mmap_read_unlock(current->mm);
8527 * if we did partial map, or found file backed vmas,
8528 * release any pages we did get
8531 unpin_user_pages(pages, pret);
8536 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8538 unpin_user_pages(pages, pret);
8543 off = ubuf & ~PAGE_MASK;
8544 size = iov->iov_len;
8545 for (i = 0; i < nr_pages; i++) {
8548 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8549 imu->bvec[i].bv_page = pages[i];
8550 imu->bvec[i].bv_len = vec_len;
8551 imu->bvec[i].bv_offset = off;
8555 /* store original address for later verification */
8557 imu->len = iov->iov_len;
8558 imu->nr_bvecs = nr_pages;
8566 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8570 if (!nr_args || nr_args > UIO_MAXIOV)
8573 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8575 if (!ctx->user_bufs)
8581 static int io_buffer_validate(struct iovec *iov)
8584 * Don't impose further limits on the size and buffer
8585 * constraints here, we'll -EINVAL later when IO is
8586 * submitted if they are wrong.
8588 if (!iov->iov_base || !iov->iov_len)
8591 /* arbitrary limit, but we need something */
8592 if (iov->iov_len > SZ_1G)
8598 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8599 unsigned int nr_args)
8603 struct page *last_hpage = NULL;
8605 ret = io_buffers_map_alloc(ctx, nr_args);
8609 for (i = 0; i < nr_args; i++) {
8610 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8612 ret = io_copy_iov(ctx, &iov, arg, i);
8616 ret = io_buffer_validate(&iov);
8620 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8624 ctx->nr_user_bufs++;
8628 io_sqe_buffers_unregister(ctx);
8633 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8635 __s32 __user *fds = arg;
8641 if (copy_from_user(&fd, fds, sizeof(*fds)))
8644 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8645 if (IS_ERR(ctx->cq_ev_fd)) {
8646 int ret = PTR_ERR(ctx->cq_ev_fd);
8647 ctx->cq_ev_fd = NULL;
8654 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8656 if (ctx->cq_ev_fd) {
8657 eventfd_ctx_put(ctx->cq_ev_fd);
8658 ctx->cq_ev_fd = NULL;
8665 static int __io_destroy_buffers(int id, void *p, void *data)
8667 struct io_ring_ctx *ctx = data;
8668 struct io_buffer *buf = p;
8670 __io_remove_buffers(ctx, buf, id, -1U);
8674 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8676 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8677 idr_destroy(&ctx->io_buffer_idr);
8680 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8682 struct io_kiocb *req, *nxt;
8684 list_for_each_entry_safe(req, nxt, list, compl.list) {
8685 if (tsk && req->task != tsk)
8687 list_del(&req->compl.list);
8688 kmem_cache_free(req_cachep, req);
8692 static void io_req_caches_free(struct io_ring_ctx *ctx, struct task_struct *tsk)
8694 struct io_submit_state *submit_state = &ctx->submit_state;
8696 mutex_lock(&ctx->uring_lock);
8698 if (submit_state->free_reqs)
8699 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8700 submit_state->reqs);
8702 io_req_cache_free(&submit_state->comp.free_list, NULL);
8704 spin_lock_irq(&ctx->completion_lock);
8705 io_req_cache_free(&submit_state->comp.locked_free_list, NULL);
8706 spin_unlock_irq(&ctx->completion_lock);
8708 mutex_unlock(&ctx->uring_lock);
8711 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8714 * Some may use context even when all refs and requests have been put,
8715 * and they are free to do so while still holding uring_lock, see
8716 * __io_req_task_submit(). Wait for them to finish.
8718 mutex_lock(&ctx->uring_lock);
8719 mutex_unlock(&ctx->uring_lock);
8721 io_finish_async(ctx);
8722 io_sqe_buffers_unregister(ctx);
8724 if (ctx->sqo_task) {
8725 put_task_struct(ctx->sqo_task);
8726 ctx->sqo_task = NULL;
8727 mmdrop(ctx->mm_account);
8728 ctx->mm_account = NULL;
8731 #ifdef CONFIG_BLK_CGROUP
8732 if (ctx->sqo_blkcg_css)
8733 css_put(ctx->sqo_blkcg_css);
8736 mutex_lock(&ctx->uring_lock);
8737 io_sqe_files_unregister(ctx);
8738 mutex_unlock(&ctx->uring_lock);
8739 io_eventfd_unregister(ctx);
8740 io_destroy_buffers(ctx);
8741 idr_destroy(&ctx->personality_idr);
8743 #if defined(CONFIG_UNIX)
8744 if (ctx->ring_sock) {
8745 ctx->ring_sock->file = NULL; /* so that iput() is called */
8746 sock_release(ctx->ring_sock);
8750 io_mem_free(ctx->rings);
8751 io_mem_free(ctx->sq_sqes);
8753 percpu_ref_exit(&ctx->refs);
8754 free_uid(ctx->user);
8755 put_cred(ctx->creds);
8756 io_req_caches_free(ctx, NULL);
8757 kfree(ctx->cancel_hash);
8761 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8763 struct io_ring_ctx *ctx = file->private_data;
8766 poll_wait(file, &ctx->cq_wait, wait);
8768 * synchronizes with barrier from wq_has_sleeper call in
8772 if (!io_sqring_full(ctx))
8773 mask |= EPOLLOUT | EPOLLWRNORM;
8776 * Don't flush cqring overflow list here, just do a simple check.
8777 * Otherwise there could possible be ABBA deadlock:
8780 * lock(&ctx->uring_lock);
8782 * lock(&ctx->uring_lock);
8785 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8786 * pushs them to do the flush.
8788 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8789 mask |= EPOLLIN | EPOLLRDNORM;
8794 static int io_uring_fasync(int fd, struct file *file, int on)
8796 struct io_ring_ctx *ctx = file->private_data;
8798 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8801 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8803 struct io_identity *iod;
8805 iod = idr_remove(&ctx->personality_idr, id);
8807 put_cred(iod->creds);
8808 if (refcount_dec_and_test(&iod->count))
8816 static int io_remove_personalities(int id, void *p, void *data)
8818 struct io_ring_ctx *ctx = data;
8820 io_unregister_personality(ctx, id);
8824 static void io_ring_exit_work(struct work_struct *work)
8826 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8830 * If we're doing polled IO and end up having requests being
8831 * submitted async (out-of-line), then completions can come in while
8832 * we're waiting for refs to drop. We need to reap these manually,
8833 * as nobody else will be looking for them.
8836 io_uring_try_cancel_requests(ctx, NULL, NULL);
8837 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8838 io_ring_ctx_free(ctx);
8841 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8843 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8845 return req->ctx == data;
8848 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8850 mutex_lock(&ctx->uring_lock);
8851 percpu_ref_kill(&ctx->refs);
8853 if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead))
8856 /* if force is set, the ring is going away. always drop after that */
8857 ctx->cq_overflow_flushed = 1;
8859 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8860 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8861 mutex_unlock(&ctx->uring_lock);
8863 io_kill_timeouts(ctx, NULL, NULL);
8864 io_poll_remove_all(ctx, NULL, NULL);
8867 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8869 /* if we failed setting up the ctx, we might not have any rings */
8870 io_iopoll_try_reap_events(ctx);
8872 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8874 * Use system_unbound_wq to avoid spawning tons of event kworkers
8875 * if we're exiting a ton of rings at the same time. It just adds
8876 * noise and overhead, there's no discernable change in runtime
8877 * over using system_wq.
8879 queue_work(system_unbound_wq, &ctx->exit_work);
8882 static int io_uring_release(struct inode *inode, struct file *file)
8884 struct io_ring_ctx *ctx = file->private_data;
8886 file->private_data = NULL;
8887 io_ring_ctx_wait_and_kill(ctx);
8891 struct io_task_cancel {
8892 struct task_struct *task;
8893 struct files_struct *files;
8896 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8898 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8899 struct io_task_cancel *cancel = data;
8902 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8903 unsigned long flags;
8904 struct io_ring_ctx *ctx = req->ctx;
8906 /* protect against races with linked timeouts */
8907 spin_lock_irqsave(&ctx->completion_lock, flags);
8908 ret = io_match_task(req, cancel->task, cancel->files);
8909 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8911 ret = io_match_task(req, cancel->task, cancel->files);
8916 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8917 struct task_struct *task,
8918 struct files_struct *files)
8920 struct io_defer_entry *de = NULL;
8923 spin_lock_irq(&ctx->completion_lock);
8924 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8925 if (io_match_task(de->req, task, files)) {
8926 list_cut_position(&list, &ctx->defer_list, &de->list);
8930 spin_unlock_irq(&ctx->completion_lock);
8932 while (!list_empty(&list)) {
8933 de = list_first_entry(&list, struct io_defer_entry, list);
8934 list_del_init(&de->list);
8935 req_set_fail_links(de->req);
8936 io_put_req(de->req);
8937 io_req_complete(de->req, -ECANCELED);
8942 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8943 struct task_struct *task,
8944 struct files_struct *files)
8946 struct io_task_cancel cancel = { .task = task, .files = files, };
8949 enum io_wq_cancel cret;
8953 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb,
8955 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8958 /* SQPOLL thread does its own polling */
8959 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8960 while (!list_empty_careful(&ctx->iopoll_list)) {
8961 io_iopoll_try_reap_events(ctx);
8966 ret |= io_poll_remove_all(ctx, task, files);
8967 ret |= io_kill_timeouts(ctx, task, files);
8968 ret |= io_run_task_work();
8969 io_cqring_overflow_flush(ctx, true, task, files);
8976 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8977 struct task_struct *task,
8978 struct files_struct *files)
8980 struct io_kiocb *req;
8983 spin_lock_irq(&ctx->inflight_lock);
8984 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8985 cnt += io_match_task(req, task, files);
8986 spin_unlock_irq(&ctx->inflight_lock);
8990 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8991 struct task_struct *task,
8992 struct files_struct *files)
8994 while (!list_empty_careful(&ctx->inflight_list)) {
8998 inflight = io_uring_count_inflight(ctx, task, files);
9002 io_uring_try_cancel_requests(ctx, task, files);
9005 io_sq_thread_unpark(ctx->sq_data);
9006 prepare_to_wait(&task->io_uring->wait, &wait,
9007 TASK_UNINTERRUPTIBLE);
9008 if (inflight == io_uring_count_inflight(ctx, task, files))
9010 finish_wait(&task->io_uring->wait, &wait);
9012 io_sq_thread_park(ctx->sq_data);
9016 static void io_disable_sqo_submit(struct io_ring_ctx *ctx)
9018 mutex_lock(&ctx->uring_lock);
9020 mutex_unlock(&ctx->uring_lock);
9022 /* make sure callers enter the ring to get error */
9024 io_ring_set_wakeup_flag(ctx);
9028 * We need to iteratively cancel requests, in case a request has dependent
9029 * hard links. These persist even for failure of cancelations, hence keep
9030 * looping until none are found.
9032 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
9033 struct files_struct *files)
9035 struct task_struct *task = current;
9037 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
9038 io_disable_sqo_submit(ctx);
9039 task = ctx->sq_data->thread;
9040 atomic_inc(&task->io_uring->in_idle);
9041 io_sq_thread_park(ctx->sq_data);
9044 io_cancel_defer_files(ctx, task, files);
9046 io_uring_cancel_files(ctx, task, files);
9048 io_uring_try_cancel_requests(ctx, task, NULL);
9050 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
9051 atomic_dec(&task->io_uring->in_idle);
9053 * If the files that are going away are the ones in the thread
9054 * identity, clear them out.
9056 if (task->io_uring->identity->files == files)
9057 task->io_uring->identity->files = NULL;
9058 io_sq_thread_unpark(ctx->sq_data);
9063 * Note that this task has used io_uring. We use it for cancelation purposes.
9065 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
9067 struct io_uring_task *tctx = current->io_uring;
9070 if (unlikely(!tctx)) {
9071 ret = io_uring_alloc_task_context(current);
9074 tctx = current->io_uring;
9076 if (tctx->last != file) {
9077 void *old = xa_load(&tctx->xa, (unsigned long)file);
9081 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
9088 /* one and only SQPOLL file note, held by sqo_task */
9089 WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) &&
9090 current != ctx->sqo_task);
9096 * This is race safe in that the task itself is doing this, hence it
9097 * cannot be going through the exit/cancel paths at the same time.
9098 * This cannot be modified while exit/cancel is running.
9100 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
9101 tctx->sqpoll = true;
9107 * Remove this io_uring_file -> task mapping.
9109 static void io_uring_del_task_file(struct file *file)
9111 struct io_uring_task *tctx = current->io_uring;
9113 if (tctx->last == file)
9115 file = xa_erase(&tctx->xa, (unsigned long)file);
9120 static void io_uring_remove_task_files(struct io_uring_task *tctx)
9123 unsigned long index;
9125 xa_for_each(&tctx->xa, index, file)
9126 io_uring_del_task_file(file);
9129 void __io_uring_files_cancel(struct files_struct *files)
9131 struct io_uring_task *tctx = current->io_uring;
9133 unsigned long index;
9135 /* make sure overflow events are dropped */
9136 atomic_inc(&tctx->in_idle);
9137 xa_for_each(&tctx->xa, index, file)
9138 io_uring_cancel_task_requests(file->private_data, files);
9139 atomic_dec(&tctx->in_idle);
9142 io_uring_remove_task_files(tctx);
9145 static s64 tctx_inflight(struct io_uring_task *tctx)
9147 return percpu_counter_sum(&tctx->inflight);
9150 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
9152 struct io_uring_task *tctx;
9158 tctx = ctx->sq_data->thread->io_uring;
9159 io_disable_sqo_submit(ctx);
9161 atomic_inc(&tctx->in_idle);
9163 /* read completions before cancelations */
9164 inflight = tctx_inflight(tctx);
9167 io_uring_cancel_task_requests(ctx, NULL);
9169 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9171 * If we've seen completions, retry without waiting. This
9172 * avoids a race where a completion comes in before we did
9173 * prepare_to_wait().
9175 if (inflight == tctx_inflight(tctx))
9177 finish_wait(&tctx->wait, &wait);
9179 atomic_dec(&tctx->in_idle);
9183 * Find any io_uring fd that this task has registered or done IO on, and cancel
9186 void __io_uring_task_cancel(void)
9188 struct io_uring_task *tctx = current->io_uring;
9192 /* make sure overflow events are dropped */
9193 atomic_inc(&tctx->in_idle);
9195 /* trigger io_disable_sqo_submit() */
9198 unsigned long index;
9200 xa_for_each(&tctx->xa, index, file)
9201 io_uring_cancel_sqpoll(file->private_data);
9205 /* read completions before cancelations */
9206 inflight = tctx_inflight(tctx);
9209 __io_uring_files_cancel(NULL);
9211 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9214 * If we've seen completions, retry without waiting. This
9215 * avoids a race where a completion comes in before we did
9216 * prepare_to_wait().
9218 if (inflight == tctx_inflight(tctx))
9220 finish_wait(&tctx->wait, &wait);
9223 atomic_dec(&tctx->in_idle);
9225 io_uring_remove_task_files(tctx);
9228 static int io_uring_flush(struct file *file, void *data)
9230 struct io_uring_task *tctx = current->io_uring;
9231 struct io_ring_ctx *ctx = file->private_data;
9233 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
9234 io_uring_cancel_task_requests(ctx, NULL);
9235 io_req_caches_free(ctx, current);
9241 /* we should have cancelled and erased it before PF_EXITING */
9242 WARN_ON_ONCE((current->flags & PF_EXITING) &&
9243 xa_load(&tctx->xa, (unsigned long)file));
9246 * fput() is pending, will be 2 if the only other ref is our potential
9247 * task file note. If the task is exiting, drop regardless of count.
9249 if (atomic_long_read(&file->f_count) != 2)
9252 if (ctx->flags & IORING_SETUP_SQPOLL) {
9253 /* there is only one file note, which is owned by sqo_task */
9254 WARN_ON_ONCE(ctx->sqo_task != current &&
9255 xa_load(&tctx->xa, (unsigned long)file));
9256 /* sqo_dead check is for when this happens after cancellation */
9257 WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead &&
9258 !xa_load(&tctx->xa, (unsigned long)file));
9260 io_disable_sqo_submit(ctx);
9263 if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current)
9264 io_uring_del_task_file(file);
9268 static void *io_uring_validate_mmap_request(struct file *file,
9269 loff_t pgoff, size_t sz)
9271 struct io_ring_ctx *ctx = file->private_data;
9272 loff_t offset = pgoff << PAGE_SHIFT;
9277 case IORING_OFF_SQ_RING:
9278 case IORING_OFF_CQ_RING:
9281 case IORING_OFF_SQES:
9285 return ERR_PTR(-EINVAL);
9288 page = virt_to_head_page(ptr);
9289 if (sz > page_size(page))
9290 return ERR_PTR(-EINVAL);
9297 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9299 size_t sz = vma->vm_end - vma->vm_start;
9303 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9305 return PTR_ERR(ptr);
9307 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9308 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9311 #else /* !CONFIG_MMU */
9313 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9315 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9318 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9320 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9323 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9324 unsigned long addr, unsigned long len,
9325 unsigned long pgoff, unsigned long flags)
9329 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9331 return PTR_ERR(ptr);
9333 return (unsigned long) ptr;
9336 #endif /* !CONFIG_MMU */
9338 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9344 if (!io_sqring_full(ctx))
9347 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9349 if (unlikely(ctx->sqo_dead)) {
9354 if (!io_sqring_full(ctx))
9358 } while (!signal_pending(current));
9360 finish_wait(&ctx->sqo_sq_wait, &wait);
9365 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9366 struct __kernel_timespec __user **ts,
9367 const sigset_t __user **sig)
9369 struct io_uring_getevents_arg arg;
9372 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9373 * is just a pointer to the sigset_t.
9375 if (!(flags & IORING_ENTER_EXT_ARG)) {
9376 *sig = (const sigset_t __user *) argp;
9382 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9383 * timespec and sigset_t pointers if good.
9385 if (*argsz != sizeof(arg))
9387 if (copy_from_user(&arg, argp, sizeof(arg)))
9389 *sig = u64_to_user_ptr(arg.sigmask);
9390 *argsz = arg.sigmask_sz;
9391 *ts = u64_to_user_ptr(arg.ts);
9395 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9396 u32, min_complete, u32, flags, const void __user *, argp,
9399 struct io_ring_ctx *ctx;
9406 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9407 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9415 if (f.file->f_op != &io_uring_fops)
9419 ctx = f.file->private_data;
9420 if (!percpu_ref_tryget(&ctx->refs))
9424 if (ctx->flags & IORING_SETUP_R_DISABLED)
9428 * For SQ polling, the thread will do all submissions and completions.
9429 * Just return the requested submit count, and wake the thread if
9433 if (ctx->flags & IORING_SETUP_SQPOLL) {
9434 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9437 if (unlikely(ctx->sqo_dead))
9439 if (flags & IORING_ENTER_SQ_WAKEUP)
9440 wake_up(&ctx->sq_data->wait);
9441 if (flags & IORING_ENTER_SQ_WAIT) {
9442 ret = io_sqpoll_wait_sq(ctx);
9446 submitted = to_submit;
9447 } else if (to_submit) {
9448 ret = io_uring_add_task_file(ctx, f.file);
9451 mutex_lock(&ctx->uring_lock);
9452 submitted = io_submit_sqes(ctx, to_submit);
9453 mutex_unlock(&ctx->uring_lock);
9455 if (submitted != to_submit)
9458 if (flags & IORING_ENTER_GETEVENTS) {
9459 const sigset_t __user *sig;
9460 struct __kernel_timespec __user *ts;
9462 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9466 min_complete = min(min_complete, ctx->cq_entries);
9469 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9470 * space applications don't need to do io completion events
9471 * polling again, they can rely on io_sq_thread to do polling
9472 * work, which can reduce cpu usage and uring_lock contention.
9474 if (ctx->flags & IORING_SETUP_IOPOLL &&
9475 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9476 ret = io_iopoll_check(ctx, min_complete);
9478 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9483 percpu_ref_put(&ctx->refs);
9486 return submitted ? submitted : ret;
9489 #ifdef CONFIG_PROC_FS
9490 static int io_uring_show_cred(int id, void *p, void *data)
9492 struct io_identity *iod = p;
9493 const struct cred *cred = iod->creds;
9494 struct seq_file *m = data;
9495 struct user_namespace *uns = seq_user_ns(m);
9496 struct group_info *gi;
9501 seq_printf(m, "%5d\n", id);
9502 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9503 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9504 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9505 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9506 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9507 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9508 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9509 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9510 seq_puts(m, "\n\tGroups:\t");
9511 gi = cred->group_info;
9512 for (g = 0; g < gi->ngroups; g++) {
9513 seq_put_decimal_ull(m, g ? " " : "",
9514 from_kgid_munged(uns, gi->gid[g]));
9516 seq_puts(m, "\n\tCapEff:\t");
9517 cap = cred->cap_effective;
9518 CAP_FOR_EACH_U32(__capi)
9519 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9524 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9526 struct io_sq_data *sq = NULL;
9531 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9532 * since fdinfo case grabs it in the opposite direction of normal use
9533 * cases. If we fail to get the lock, we just don't iterate any
9534 * structures that could be going away outside the io_uring mutex.
9536 has_lock = mutex_trylock(&ctx->uring_lock);
9538 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9541 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9542 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9543 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9544 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9545 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9548 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9550 seq_printf(m, "%5u: <none>\n", i);
9552 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9553 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9554 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9556 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9557 (unsigned int) buf->len);
9559 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9560 seq_printf(m, "Personalities:\n");
9561 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9563 seq_printf(m, "PollList:\n");
9564 spin_lock_irq(&ctx->completion_lock);
9565 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9566 struct hlist_head *list = &ctx->cancel_hash[i];
9567 struct io_kiocb *req;
9569 hlist_for_each_entry(req, list, hash_node)
9570 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9571 req->task->task_works != NULL);
9573 spin_unlock_irq(&ctx->completion_lock);
9575 mutex_unlock(&ctx->uring_lock);
9578 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9580 struct io_ring_ctx *ctx = f->private_data;
9582 if (percpu_ref_tryget(&ctx->refs)) {
9583 __io_uring_show_fdinfo(ctx, m);
9584 percpu_ref_put(&ctx->refs);
9589 static const struct file_operations io_uring_fops = {
9590 .release = io_uring_release,
9591 .flush = io_uring_flush,
9592 .mmap = io_uring_mmap,
9594 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9595 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9597 .poll = io_uring_poll,
9598 .fasync = io_uring_fasync,
9599 #ifdef CONFIG_PROC_FS
9600 .show_fdinfo = io_uring_show_fdinfo,
9604 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9605 struct io_uring_params *p)
9607 struct io_rings *rings;
9608 size_t size, sq_array_offset;
9610 /* make sure these are sane, as we already accounted them */
9611 ctx->sq_entries = p->sq_entries;
9612 ctx->cq_entries = p->cq_entries;
9614 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9615 if (size == SIZE_MAX)
9618 rings = io_mem_alloc(size);
9623 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9624 rings->sq_ring_mask = p->sq_entries - 1;
9625 rings->cq_ring_mask = p->cq_entries - 1;
9626 rings->sq_ring_entries = p->sq_entries;
9627 rings->cq_ring_entries = p->cq_entries;
9628 ctx->sq_mask = rings->sq_ring_mask;
9629 ctx->cq_mask = rings->cq_ring_mask;
9631 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9632 if (size == SIZE_MAX) {
9633 io_mem_free(ctx->rings);
9638 ctx->sq_sqes = io_mem_alloc(size);
9639 if (!ctx->sq_sqes) {
9640 io_mem_free(ctx->rings);
9648 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9652 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9656 ret = io_uring_add_task_file(ctx, file);
9661 fd_install(fd, file);
9666 * Allocate an anonymous fd, this is what constitutes the application
9667 * visible backing of an io_uring instance. The application mmaps this
9668 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9669 * we have to tie this fd to a socket for file garbage collection purposes.
9671 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9674 #if defined(CONFIG_UNIX)
9677 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9680 return ERR_PTR(ret);
9683 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9684 O_RDWR | O_CLOEXEC);
9685 #if defined(CONFIG_UNIX)
9687 sock_release(ctx->ring_sock);
9688 ctx->ring_sock = NULL;
9690 ctx->ring_sock->file = file;
9696 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9697 struct io_uring_params __user *params)
9699 struct user_struct *user = NULL;
9700 struct io_ring_ctx *ctx;
9706 if (entries > IORING_MAX_ENTRIES) {
9707 if (!(p->flags & IORING_SETUP_CLAMP))
9709 entries = IORING_MAX_ENTRIES;
9713 * Use twice as many entries for the CQ ring. It's possible for the
9714 * application to drive a higher depth than the size of the SQ ring,
9715 * since the sqes are only used at submission time. This allows for
9716 * some flexibility in overcommitting a bit. If the application has
9717 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9718 * of CQ ring entries manually.
9720 p->sq_entries = roundup_pow_of_two(entries);
9721 if (p->flags & IORING_SETUP_CQSIZE) {
9723 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9724 * to a power-of-two, if it isn't already. We do NOT impose
9725 * any cq vs sq ring sizing.
9729 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9730 if (!(p->flags & IORING_SETUP_CLAMP))
9732 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9734 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9735 if (p->cq_entries < p->sq_entries)
9738 p->cq_entries = 2 * p->sq_entries;
9741 user = get_uid(current_user());
9743 ctx = io_ring_ctx_alloc(p);
9748 ctx->compat = in_compat_syscall();
9749 ctx->limit_mem = !capable(CAP_IPC_LOCK);
9751 ctx->creds = get_current_cred();
9753 ctx->loginuid = current->loginuid;
9754 ctx->sessionid = current->sessionid;
9756 ctx->sqo_task = get_task_struct(current);
9759 * This is just grabbed for accounting purposes. When a process exits,
9760 * the mm is exited and dropped before the files, hence we need to hang
9761 * on to this mm purely for the purposes of being able to unaccount
9762 * memory (locked/pinned vm). It's not used for anything else.
9764 mmgrab(current->mm);
9765 ctx->mm_account = current->mm;
9767 #ifdef CONFIG_BLK_CGROUP
9769 * The sq thread will belong to the original cgroup it was inited in.
9770 * If the cgroup goes offline (e.g. disabling the io controller), then
9771 * issued bios will be associated with the closest cgroup later in the
9775 ctx->sqo_blkcg_css = blkcg_css();
9776 ret = css_tryget_online(ctx->sqo_blkcg_css);
9779 /* don't init against a dying cgroup, have the user try again */
9780 ctx->sqo_blkcg_css = NULL;
9785 ret = io_allocate_scq_urings(ctx, p);
9789 ret = io_sq_offload_create(ctx, p);
9793 if (!(p->flags & IORING_SETUP_R_DISABLED))
9794 io_sq_offload_start(ctx);
9796 memset(&p->sq_off, 0, sizeof(p->sq_off));
9797 p->sq_off.head = offsetof(struct io_rings, sq.head);
9798 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9799 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9800 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9801 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9802 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9803 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9805 memset(&p->cq_off, 0, sizeof(p->cq_off));
9806 p->cq_off.head = offsetof(struct io_rings, cq.head);
9807 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9808 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9809 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9810 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9811 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9812 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9814 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9815 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9816 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9817 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9818 IORING_FEAT_EXT_ARG;
9820 if (copy_to_user(params, p, sizeof(*p))) {
9825 file = io_uring_get_file(ctx);
9827 ret = PTR_ERR(file);
9832 * Install ring fd as the very last thing, so we don't risk someone
9833 * having closed it before we finish setup
9835 ret = io_uring_install_fd(ctx, file);
9837 io_disable_sqo_submit(ctx);
9838 /* fput will clean it up */
9843 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9846 io_disable_sqo_submit(ctx);
9847 io_ring_ctx_wait_and_kill(ctx);
9852 * Sets up an aio uring context, and returns the fd. Applications asks for a
9853 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9854 * params structure passed in.
9856 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9858 struct io_uring_params p;
9861 if (copy_from_user(&p, params, sizeof(p)))
9863 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9868 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9869 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9870 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9871 IORING_SETUP_R_DISABLED))
9874 return io_uring_create(entries, &p, params);
9877 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9878 struct io_uring_params __user *, params)
9880 return io_uring_setup(entries, params);
9883 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9885 struct io_uring_probe *p;
9889 size = struct_size(p, ops, nr_args);
9890 if (size == SIZE_MAX)
9892 p = kzalloc(size, GFP_KERNEL);
9897 if (copy_from_user(p, arg, size))
9900 if (memchr_inv(p, 0, size))
9903 p->last_op = IORING_OP_LAST - 1;
9904 if (nr_args > IORING_OP_LAST)
9905 nr_args = IORING_OP_LAST;
9907 for (i = 0; i < nr_args; i++) {
9909 if (!io_op_defs[i].not_supported)
9910 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9915 if (copy_to_user(arg, p, size))
9922 static int io_register_personality(struct io_ring_ctx *ctx)
9924 struct io_identity *id;
9927 id = kmalloc(sizeof(*id), GFP_KERNEL);
9931 io_init_identity(id);
9932 id->creds = get_current_cred();
9934 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9936 put_cred(id->creds);
9942 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9943 unsigned int nr_args)
9945 struct io_uring_restriction *res;
9949 /* Restrictions allowed only if rings started disabled */
9950 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9953 /* We allow only a single restrictions registration */
9954 if (ctx->restrictions.registered)
9957 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9960 size = array_size(nr_args, sizeof(*res));
9961 if (size == SIZE_MAX)
9964 res = memdup_user(arg, size);
9966 return PTR_ERR(res);
9970 for (i = 0; i < nr_args; i++) {
9971 switch (res[i].opcode) {
9972 case IORING_RESTRICTION_REGISTER_OP:
9973 if (res[i].register_op >= IORING_REGISTER_LAST) {
9978 __set_bit(res[i].register_op,
9979 ctx->restrictions.register_op);
9981 case IORING_RESTRICTION_SQE_OP:
9982 if (res[i].sqe_op >= IORING_OP_LAST) {
9987 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9989 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9990 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9992 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9993 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10002 /* Reset all restrictions if an error happened */
10004 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10006 ctx->restrictions.registered = true;
10012 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10014 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10017 if (ctx->restrictions.registered)
10018 ctx->restricted = 1;
10020 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10022 io_sq_offload_start(ctx);
10027 static bool io_register_op_must_quiesce(int op)
10030 case IORING_UNREGISTER_FILES:
10031 case IORING_REGISTER_FILES_UPDATE:
10032 case IORING_REGISTER_PROBE:
10033 case IORING_REGISTER_PERSONALITY:
10034 case IORING_UNREGISTER_PERSONALITY:
10041 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10042 void __user *arg, unsigned nr_args)
10043 __releases(ctx->uring_lock)
10044 __acquires(ctx->uring_lock)
10049 * We're inside the ring mutex, if the ref is already dying, then
10050 * someone else killed the ctx or is already going through
10051 * io_uring_register().
10053 if (percpu_ref_is_dying(&ctx->refs))
10056 if (io_register_op_must_quiesce(opcode)) {
10057 percpu_ref_kill(&ctx->refs);
10060 * Drop uring mutex before waiting for references to exit. If
10061 * another thread is currently inside io_uring_enter() it might
10062 * need to grab the uring_lock to make progress. If we hold it
10063 * here across the drain wait, then we can deadlock. It's safe
10064 * to drop the mutex here, since no new references will come in
10065 * after we've killed the percpu ref.
10067 mutex_unlock(&ctx->uring_lock);
10069 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10072 ret = io_run_task_work_sig();
10077 mutex_lock(&ctx->uring_lock);
10080 percpu_ref_resurrect(&ctx->refs);
10085 if (ctx->restricted) {
10086 if (opcode >= IORING_REGISTER_LAST) {
10091 if (!test_bit(opcode, ctx->restrictions.register_op)) {
10098 case IORING_REGISTER_BUFFERS:
10099 ret = io_sqe_buffers_register(ctx, arg, nr_args);
10101 case IORING_UNREGISTER_BUFFERS:
10103 if (arg || nr_args)
10105 ret = io_sqe_buffers_unregister(ctx);
10107 case IORING_REGISTER_FILES:
10108 ret = io_sqe_files_register(ctx, arg, nr_args);
10110 case IORING_UNREGISTER_FILES:
10112 if (arg || nr_args)
10114 ret = io_sqe_files_unregister(ctx);
10116 case IORING_REGISTER_FILES_UPDATE:
10117 ret = io_sqe_files_update(ctx, arg, nr_args);
10119 case IORING_REGISTER_EVENTFD:
10120 case IORING_REGISTER_EVENTFD_ASYNC:
10124 ret = io_eventfd_register(ctx, arg);
10127 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10128 ctx->eventfd_async = 1;
10130 ctx->eventfd_async = 0;
10132 case IORING_UNREGISTER_EVENTFD:
10134 if (arg || nr_args)
10136 ret = io_eventfd_unregister(ctx);
10138 case IORING_REGISTER_PROBE:
10140 if (!arg || nr_args > 256)
10142 ret = io_probe(ctx, arg, nr_args);
10144 case IORING_REGISTER_PERSONALITY:
10146 if (arg || nr_args)
10148 ret = io_register_personality(ctx);
10150 case IORING_UNREGISTER_PERSONALITY:
10154 ret = io_unregister_personality(ctx, nr_args);
10156 case IORING_REGISTER_ENABLE_RINGS:
10158 if (arg || nr_args)
10160 ret = io_register_enable_rings(ctx);
10162 case IORING_REGISTER_RESTRICTIONS:
10163 ret = io_register_restrictions(ctx, arg, nr_args);
10171 if (io_register_op_must_quiesce(opcode)) {
10172 /* bring the ctx back to life */
10173 percpu_ref_reinit(&ctx->refs);
10175 reinit_completion(&ctx->ref_comp);
10180 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10181 void __user *, arg, unsigned int, nr_args)
10183 struct io_ring_ctx *ctx;
10192 if (f.file->f_op != &io_uring_fops)
10195 ctx = f.file->private_data;
10197 mutex_lock(&ctx->uring_lock);
10198 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10199 mutex_unlock(&ctx->uring_lock);
10200 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10201 ctx->cq_ev_fd != NULL, ret);
10207 static int __init io_uring_init(void)
10209 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10210 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10211 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10214 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10215 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10216 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10217 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10218 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10219 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10220 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10221 BUILD_BUG_SQE_ELEM(8, __u64, off);
10222 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10223 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10224 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10225 BUILD_BUG_SQE_ELEM(24, __u32, len);
10226 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10227 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10228 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10229 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10230 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10231 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10232 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10233 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10234 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10235 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10236 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10237 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10238 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10239 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10240 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10241 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10242 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10243 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10244 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10246 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10247 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10248 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10252 __initcall(io_uring_init);