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)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 struct io_mapped_ubuf {
193 struct bio_vec *bvec;
194 unsigned int nr_bvecs;
195 unsigned long acct_pages;
198 struct fixed_file_table {
202 struct fixed_file_ref_node {
203 struct percpu_ref refs;
204 struct list_head node;
205 struct list_head file_list;
206 struct fixed_file_data *file_data;
207 struct llist_node llist;
211 struct fixed_file_data {
212 struct fixed_file_table *table;
213 struct io_ring_ctx *ctx;
215 struct fixed_file_ref_node *node;
216 struct percpu_ref refs;
217 struct completion done;
218 struct list_head ref_list;
223 struct list_head list;
229 struct io_restriction {
230 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
231 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
232 u8 sqe_flags_allowed;
233 u8 sqe_flags_required;
241 /* ctx's that are using this sqd */
242 struct list_head ctx_list;
243 struct list_head ctx_new_list;
244 struct mutex ctx_lock;
246 struct task_struct *thread;
247 struct wait_queue_head wait;
249 unsigned sq_thread_idle;
254 struct percpu_ref refs;
255 } ____cacheline_aligned_in_smp;
259 unsigned int compat: 1;
260 unsigned int limit_mem: 1;
261 unsigned int cq_overflow_flushed: 1;
262 unsigned int drain_next: 1;
263 unsigned int eventfd_async: 1;
264 unsigned int restricted: 1;
265 unsigned int sqo_dead: 1;
268 * Ring buffer of indices into array of io_uring_sqe, which is
269 * mmapped by the application using the IORING_OFF_SQES offset.
271 * This indirection could e.g. be used to assign fixed
272 * io_uring_sqe entries to operations and only submit them to
273 * the queue when needed.
275 * The kernel modifies neither the indices array nor the entries
279 unsigned cached_sq_head;
282 unsigned sq_thread_idle;
283 unsigned cached_sq_dropped;
284 unsigned cached_cq_overflow;
285 unsigned long sq_check_overflow;
287 struct list_head defer_list;
288 struct list_head timeout_list;
289 struct list_head cq_overflow_list;
291 struct io_uring_sqe *sq_sqes;
292 } ____cacheline_aligned_in_smp;
294 struct io_rings *rings;
300 * For SQPOLL usage - we hold a reference to the parent task, so we
301 * have access to the ->files
303 struct task_struct *sqo_task;
305 /* Only used for accounting purposes */
306 struct mm_struct *mm_account;
308 #ifdef CONFIG_BLK_CGROUP
309 struct cgroup_subsys_state *sqo_blkcg_css;
312 struct io_sq_data *sq_data; /* if using sq thread polling */
314 struct wait_queue_head sqo_sq_wait;
315 struct list_head sqd_list;
318 * If used, fixed file set. Writers must ensure that ->refs is dead,
319 * readers must ensure that ->refs is alive as long as the file* is
320 * used. Only updated through io_uring_register(2).
322 struct fixed_file_data *file_data;
323 unsigned nr_user_files;
325 /* if used, fixed mapped user buffers */
326 unsigned nr_user_bufs;
327 struct io_mapped_ubuf *user_bufs;
329 struct user_struct *user;
331 const struct cred *creds;
335 unsigned int sessionid;
338 struct completion ref_comp;
339 struct completion sq_thread_comp;
341 /* if all else fails... */
342 struct io_kiocb *fallback_req;
344 #if defined(CONFIG_UNIX)
345 struct socket *ring_sock;
348 struct idr io_buffer_idr;
350 struct idr personality_idr;
353 unsigned cached_cq_tail;
356 atomic_t cq_timeouts;
357 unsigned cq_last_tm_flush;
358 unsigned long cq_check_overflow;
359 struct wait_queue_head cq_wait;
360 struct fasync_struct *cq_fasync;
361 struct eventfd_ctx *cq_ev_fd;
362 } ____cacheline_aligned_in_smp;
365 struct mutex uring_lock;
366 wait_queue_head_t wait;
367 } ____cacheline_aligned_in_smp;
370 spinlock_t completion_lock;
373 * ->iopoll_list is protected by the ctx->uring_lock for
374 * io_uring instances that don't use IORING_SETUP_SQPOLL.
375 * For SQPOLL, only the single threaded io_sq_thread() will
376 * manipulate the list, hence no extra locking is needed there.
378 struct list_head iopoll_list;
379 struct hlist_head *cancel_hash;
380 unsigned cancel_hash_bits;
381 bool poll_multi_file;
383 spinlock_t inflight_lock;
384 struct list_head inflight_list;
385 } ____cacheline_aligned_in_smp;
387 struct delayed_work file_put_work;
388 struct llist_head file_put_llist;
390 struct work_struct exit_work;
391 struct io_restriction restrictions;
395 * First field must be the file pointer in all the
396 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
398 struct io_poll_iocb {
400 struct wait_queue_head *head;
404 struct wait_queue_entry wait;
407 struct io_poll_remove {
414 struct file *put_file;
418 struct io_timeout_data {
419 struct io_kiocb *req;
420 struct hrtimer timer;
421 struct timespec64 ts;
422 enum hrtimer_mode mode;
427 struct sockaddr __user *addr;
428 int __user *addr_len;
430 unsigned long nofile;
450 struct list_head list;
451 /* head of the link, used by linked timeouts only */
452 struct io_kiocb *head;
455 struct io_timeout_rem {
460 struct timespec64 ts;
465 /* NOTE: kiocb has the file as the first member, so don't do it here */
473 struct sockaddr __user *addr;
480 struct user_msghdr __user *umsg;
486 struct io_buffer *kbuf;
492 bool ignore_nonblock;
493 struct filename *filename;
495 unsigned long nofile;
498 struct io_files_update {
524 struct epoll_event event;
528 struct file *file_out;
529 struct file *file_in;
536 struct io_provide_buf {
550 const char __user *filename;
551 struct statx __user *buffer;
563 struct filename *oldpath;
564 struct filename *newpath;
572 struct filename *filename;
575 struct io_completion {
577 struct list_head list;
581 struct io_async_connect {
582 struct sockaddr_storage address;
585 struct io_async_msghdr {
586 struct iovec fast_iov[UIO_FASTIOV];
588 struct sockaddr __user *uaddr;
590 struct sockaddr_storage addr;
594 struct iovec fast_iov[UIO_FASTIOV];
595 const struct iovec *free_iovec;
596 struct iov_iter iter;
598 struct wait_page_queue wpq;
602 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
603 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
604 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
605 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
606 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
607 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
613 REQ_F_LINK_TIMEOUT_BIT,
615 REQ_F_NEED_CLEANUP_BIT,
617 REQ_F_BUFFER_SELECTED_BIT,
618 REQ_F_NO_FILE_TABLE_BIT,
619 REQ_F_WORK_INITIALIZED_BIT,
620 REQ_F_LTIMEOUT_ACTIVE_BIT,
622 /* not a real bit, just to check we're not overflowing the space */
628 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
629 /* drain existing IO first */
630 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
632 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
633 /* doesn't sever on completion < 0 */
634 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
636 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
637 /* IOSQE_BUFFER_SELECT */
638 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
640 /* fail rest of links */
641 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
642 /* on inflight list */
643 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
644 /* read/write uses file position */
645 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
646 /* must not punt to workers */
647 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
648 /* has or had linked timeout */
649 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
651 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
653 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
654 /* already went through poll handler */
655 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
656 /* buffer already selected */
657 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
658 /* doesn't need file table for this request */
659 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
660 /* io_wq_work is initialized */
661 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
662 /* linked timeout is active, i.e. prepared by link's head */
663 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
667 struct io_poll_iocb poll;
668 struct io_poll_iocb *double_poll;
672 * NOTE! Each of the iocb union members has the file pointer
673 * as the first entry in their struct definition. So you can
674 * access the file pointer through any of the sub-structs,
675 * or directly as just 'ki_filp' in this struct.
681 struct io_poll_iocb poll;
682 struct io_poll_remove poll_remove;
683 struct io_accept accept;
685 struct io_cancel cancel;
686 struct io_timeout timeout;
687 struct io_timeout_rem timeout_rem;
688 struct io_connect connect;
689 struct io_sr_msg sr_msg;
691 struct io_close close;
692 struct io_files_update files_update;
693 struct io_fadvise fadvise;
694 struct io_madvise madvise;
695 struct io_epoll epoll;
696 struct io_splice splice;
697 struct io_provide_buf pbuf;
698 struct io_statx statx;
699 struct io_shutdown shutdown;
700 struct io_rename rename;
701 struct io_unlink unlink;
702 /* use only after cleaning per-op data, see io_clean_op() */
703 struct io_completion compl;
706 /* opcode allocated if it needs to store data for async defer */
709 /* polled IO has completed */
715 struct io_ring_ctx *ctx;
718 struct task_struct *task;
721 struct io_kiocb *link;
722 struct percpu_ref *fixed_file_refs;
725 * 1. used with ctx->iopoll_list with reads/writes
726 * 2. to track reqs with ->files (see io_op_def::file_table)
728 struct list_head inflight_entry;
729 struct callback_head task_work;
730 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
731 struct hlist_node hash_node;
732 struct async_poll *apoll;
733 struct io_wq_work work;
736 struct io_defer_entry {
737 struct list_head list;
738 struct io_kiocb *req;
742 #define IO_IOPOLL_BATCH 8
744 struct io_comp_state {
746 struct list_head list;
747 struct io_ring_ctx *ctx;
750 struct io_submit_state {
751 struct blk_plug plug;
754 * io_kiocb alloc cache
756 void *reqs[IO_IOPOLL_BATCH];
757 unsigned int free_reqs;
762 * Batch completion logic
764 struct io_comp_state comp;
767 * File reference cache
771 unsigned int file_refs;
772 unsigned int ios_left;
776 /* needs req->file assigned */
777 unsigned needs_file : 1;
778 /* don't fail if file grab fails */
779 unsigned needs_file_no_error : 1;
780 /* hash wq insertion if file is a regular file */
781 unsigned hash_reg_file : 1;
782 /* unbound wq insertion if file is a non-regular file */
783 unsigned unbound_nonreg_file : 1;
784 /* opcode is not supported by this kernel */
785 unsigned not_supported : 1;
786 /* set if opcode supports polled "wait" */
788 unsigned pollout : 1;
789 /* op supports buffer selection */
790 unsigned buffer_select : 1;
791 /* must always have async data allocated */
792 unsigned needs_async_data : 1;
793 /* should block plug */
795 /* size of async data needed, if any */
796 unsigned short async_size;
800 static const struct io_op_def io_op_defs[] = {
801 [IORING_OP_NOP] = {},
802 [IORING_OP_READV] = {
804 .unbound_nonreg_file = 1,
807 .needs_async_data = 1,
809 .async_size = sizeof(struct io_async_rw),
810 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
812 [IORING_OP_WRITEV] = {
815 .unbound_nonreg_file = 1,
817 .needs_async_data = 1,
819 .async_size = sizeof(struct io_async_rw),
820 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
823 [IORING_OP_FSYNC] = {
825 .work_flags = IO_WQ_WORK_BLKCG,
827 [IORING_OP_READ_FIXED] = {
829 .unbound_nonreg_file = 1,
832 .async_size = sizeof(struct io_async_rw),
833 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
835 [IORING_OP_WRITE_FIXED] = {
838 .unbound_nonreg_file = 1,
841 .async_size = sizeof(struct io_async_rw),
842 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
845 [IORING_OP_POLL_ADD] = {
847 .unbound_nonreg_file = 1,
849 [IORING_OP_POLL_REMOVE] = {},
850 [IORING_OP_SYNC_FILE_RANGE] = {
852 .work_flags = IO_WQ_WORK_BLKCG,
854 [IORING_OP_SENDMSG] = {
856 .unbound_nonreg_file = 1,
858 .needs_async_data = 1,
859 .async_size = sizeof(struct io_async_msghdr),
860 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
862 [IORING_OP_RECVMSG] = {
864 .unbound_nonreg_file = 1,
867 .needs_async_data = 1,
868 .async_size = sizeof(struct io_async_msghdr),
869 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
871 [IORING_OP_TIMEOUT] = {
872 .needs_async_data = 1,
873 .async_size = sizeof(struct io_timeout_data),
874 .work_flags = IO_WQ_WORK_MM,
876 [IORING_OP_TIMEOUT_REMOVE] = {
877 /* used by timeout updates' prep() */
878 .work_flags = IO_WQ_WORK_MM,
880 [IORING_OP_ACCEPT] = {
882 .unbound_nonreg_file = 1,
884 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
886 [IORING_OP_ASYNC_CANCEL] = {},
887 [IORING_OP_LINK_TIMEOUT] = {
888 .needs_async_data = 1,
889 .async_size = sizeof(struct io_timeout_data),
890 .work_flags = IO_WQ_WORK_MM,
892 [IORING_OP_CONNECT] = {
894 .unbound_nonreg_file = 1,
896 .needs_async_data = 1,
897 .async_size = sizeof(struct io_async_connect),
898 .work_flags = IO_WQ_WORK_MM,
900 [IORING_OP_FALLOCATE] = {
902 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
904 [IORING_OP_OPENAT] = {
905 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
906 IO_WQ_WORK_FS | IO_WQ_WORK_MM,
908 [IORING_OP_CLOSE] = {
910 .needs_file_no_error = 1,
911 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
913 [IORING_OP_FILES_UPDATE] = {
914 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
916 [IORING_OP_STATX] = {
917 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
918 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
922 .unbound_nonreg_file = 1,
926 .async_size = sizeof(struct io_async_rw),
927 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
929 [IORING_OP_WRITE] = {
931 .unbound_nonreg_file = 1,
934 .async_size = sizeof(struct io_async_rw),
935 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
938 [IORING_OP_FADVISE] = {
940 .work_flags = IO_WQ_WORK_BLKCG,
942 [IORING_OP_MADVISE] = {
943 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
947 .unbound_nonreg_file = 1,
949 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
953 .unbound_nonreg_file = 1,
956 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
958 [IORING_OP_OPENAT2] = {
959 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
960 IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
962 [IORING_OP_EPOLL_CTL] = {
963 .unbound_nonreg_file = 1,
964 .work_flags = IO_WQ_WORK_FILES,
966 [IORING_OP_SPLICE] = {
969 .unbound_nonreg_file = 1,
970 .work_flags = IO_WQ_WORK_BLKCG,
972 [IORING_OP_PROVIDE_BUFFERS] = {},
973 [IORING_OP_REMOVE_BUFFERS] = {},
977 .unbound_nonreg_file = 1,
979 [IORING_OP_SHUTDOWN] = {
982 [IORING_OP_RENAMEAT] = {
983 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
984 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
986 [IORING_OP_UNLINKAT] = {
987 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
988 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
992 enum io_mem_account {
997 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
998 struct task_struct *task);
1000 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node);
1001 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
1002 struct io_ring_ctx *ctx);
1004 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
1005 struct io_comp_state *cs);
1006 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1007 static void io_put_req(struct io_kiocb *req);
1008 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1009 static void io_double_put_req(struct io_kiocb *req);
1010 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1011 static void __io_queue_linked_timeout(struct io_kiocb *req);
1012 static void io_queue_linked_timeout(struct io_kiocb *req);
1013 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1014 struct io_uring_files_update *ip,
1016 static void __io_clean_op(struct io_kiocb *req);
1017 static struct file *io_file_get(struct io_submit_state *state,
1018 struct io_kiocb *req, int fd, bool fixed);
1019 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
1020 static void io_file_put_work(struct work_struct *work);
1022 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1023 struct iovec **iovec, struct iov_iter *iter,
1025 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1026 const struct iovec *fast_iov,
1027 struct iov_iter *iter, bool force);
1028 static void io_req_drop_files(struct io_kiocb *req);
1030 static struct kmem_cache *req_cachep;
1032 static const struct file_operations io_uring_fops;
1034 struct sock *io_uring_get_socket(struct file *file)
1036 #if defined(CONFIG_UNIX)
1037 if (file->f_op == &io_uring_fops) {
1038 struct io_ring_ctx *ctx = file->private_data;
1040 return ctx->ring_sock->sk;
1045 EXPORT_SYMBOL(io_uring_get_socket);
1047 #define io_for_each_link(pos, head) \
1048 for (pos = (head); pos; pos = pos->link)
1050 static inline void io_clean_op(struct io_kiocb *req)
1052 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1056 static inline void io_set_resource_node(struct io_kiocb *req)
1058 struct io_ring_ctx *ctx = req->ctx;
1060 if (!req->fixed_file_refs) {
1061 req->fixed_file_refs = &ctx->file_data->node->refs;
1062 percpu_ref_get(req->fixed_file_refs);
1066 static bool io_match_task(struct io_kiocb *head,
1067 struct task_struct *task,
1068 struct files_struct *files)
1070 struct io_kiocb *req;
1072 if (task && head->task != task)
1077 io_for_each_link(req, head) {
1078 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1080 if (req->file && req->file->f_op == &io_uring_fops)
1082 if ((req->work.flags & IO_WQ_WORK_FILES) &&
1083 req->work.identity->files == files)
1089 static void io_sq_thread_drop_mm_files(void)
1091 struct files_struct *files = current->files;
1092 struct mm_struct *mm = current->mm;
1095 kthread_unuse_mm(mm);
1100 struct nsproxy *nsproxy = current->nsproxy;
1103 current->files = NULL;
1104 current->nsproxy = NULL;
1105 task_unlock(current);
1106 put_files_struct(files);
1107 put_nsproxy(nsproxy);
1111 static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1113 if (current->flags & PF_EXITING)
1116 if (!current->files) {
1117 struct files_struct *files;
1118 struct nsproxy *nsproxy;
1120 task_lock(ctx->sqo_task);
1121 files = ctx->sqo_task->files;
1123 task_unlock(ctx->sqo_task);
1126 atomic_inc(&files->count);
1127 get_nsproxy(ctx->sqo_task->nsproxy);
1128 nsproxy = ctx->sqo_task->nsproxy;
1129 task_unlock(ctx->sqo_task);
1132 current->files = files;
1133 current->nsproxy = nsproxy;
1134 task_unlock(current);
1139 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1141 struct mm_struct *mm;
1143 if (current->flags & PF_EXITING)
1148 /* Should never happen */
1149 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1152 task_lock(ctx->sqo_task);
1153 mm = ctx->sqo_task->mm;
1154 if (unlikely(!mm || !mmget_not_zero(mm)))
1156 task_unlock(ctx->sqo_task);
1166 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1167 struct io_kiocb *req)
1169 const struct io_op_def *def = &io_op_defs[req->opcode];
1172 if (def->work_flags & IO_WQ_WORK_MM) {
1173 ret = __io_sq_thread_acquire_mm(ctx);
1178 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) {
1179 ret = __io_sq_thread_acquire_files(ctx);
1187 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1188 struct cgroup_subsys_state **cur_css)
1191 #ifdef CONFIG_BLK_CGROUP
1192 /* puts the old one when swapping */
1193 if (*cur_css != ctx->sqo_blkcg_css) {
1194 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1195 *cur_css = ctx->sqo_blkcg_css;
1200 static void io_sq_thread_unassociate_blkcg(void)
1202 #ifdef CONFIG_BLK_CGROUP
1203 kthread_associate_blkcg(NULL);
1207 static inline void req_set_fail_links(struct io_kiocb *req)
1209 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1210 req->flags |= REQ_F_FAIL_LINK;
1214 * None of these are dereferenced, they are simply used to check if any of
1215 * them have changed. If we're under current and check they are still the
1216 * same, we're fine to grab references to them for actual out-of-line use.
1218 static void io_init_identity(struct io_identity *id)
1220 id->files = current->files;
1221 id->mm = current->mm;
1222 #ifdef CONFIG_BLK_CGROUP
1224 id->blkcg_css = blkcg_css();
1227 id->creds = current_cred();
1228 id->nsproxy = current->nsproxy;
1229 id->fs = current->fs;
1230 id->fsize = rlimit(RLIMIT_FSIZE);
1232 id->loginuid = current->loginuid;
1233 id->sessionid = current->sessionid;
1235 refcount_set(&id->count, 1);
1238 static inline void __io_req_init_async(struct io_kiocb *req)
1240 memset(&req->work, 0, sizeof(req->work));
1241 req->flags |= REQ_F_WORK_INITIALIZED;
1245 * Note: must call io_req_init_async() for the first time you
1246 * touch any members of io_wq_work.
1248 static inline void io_req_init_async(struct io_kiocb *req)
1250 struct io_uring_task *tctx = current->io_uring;
1252 if (req->flags & REQ_F_WORK_INITIALIZED)
1255 __io_req_init_async(req);
1257 /* Grab a ref if this isn't our static identity */
1258 req->work.identity = tctx->identity;
1259 if (tctx->identity != &tctx->__identity)
1260 refcount_inc(&req->work.identity->count);
1263 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1265 return ctx->flags & IORING_SETUP_SQPOLL;
1268 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1270 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1272 complete(&ctx->ref_comp);
1275 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1277 return !req->timeout.off;
1280 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1282 struct io_ring_ctx *ctx;
1285 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1289 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1290 if (!ctx->fallback_req)
1294 * Use 5 bits less than the max cq entries, that should give us around
1295 * 32 entries per hash list if totally full and uniformly spread.
1297 hash_bits = ilog2(p->cq_entries);
1301 ctx->cancel_hash_bits = hash_bits;
1302 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1304 if (!ctx->cancel_hash)
1306 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1308 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1309 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1312 ctx->flags = p->flags;
1313 init_waitqueue_head(&ctx->sqo_sq_wait);
1314 INIT_LIST_HEAD(&ctx->sqd_list);
1315 init_waitqueue_head(&ctx->cq_wait);
1316 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1317 init_completion(&ctx->ref_comp);
1318 init_completion(&ctx->sq_thread_comp);
1319 idr_init(&ctx->io_buffer_idr);
1320 idr_init(&ctx->personality_idr);
1321 mutex_init(&ctx->uring_lock);
1322 init_waitqueue_head(&ctx->wait);
1323 spin_lock_init(&ctx->completion_lock);
1324 INIT_LIST_HEAD(&ctx->iopoll_list);
1325 INIT_LIST_HEAD(&ctx->defer_list);
1326 INIT_LIST_HEAD(&ctx->timeout_list);
1327 spin_lock_init(&ctx->inflight_lock);
1328 INIT_LIST_HEAD(&ctx->inflight_list);
1329 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1330 init_llist_head(&ctx->file_put_llist);
1333 if (ctx->fallback_req)
1334 kmem_cache_free(req_cachep, ctx->fallback_req);
1335 kfree(ctx->cancel_hash);
1340 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1342 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1343 struct io_ring_ctx *ctx = req->ctx;
1345 return seq != ctx->cached_cq_tail
1346 + READ_ONCE(ctx->cached_cq_overflow);
1352 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1354 struct io_rings *rings = ctx->rings;
1356 /* order cqe stores with ring update */
1357 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1360 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1362 if (req->work.identity == &tctx->__identity)
1364 if (refcount_dec_and_test(&req->work.identity->count))
1365 kfree(req->work.identity);
1368 static void io_req_clean_work(struct io_kiocb *req)
1370 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1373 req->flags &= ~REQ_F_WORK_INITIALIZED;
1375 if (req->work.flags & IO_WQ_WORK_MM) {
1376 mmdrop(req->work.identity->mm);
1377 req->work.flags &= ~IO_WQ_WORK_MM;
1379 #ifdef CONFIG_BLK_CGROUP
1380 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1381 css_put(req->work.identity->blkcg_css);
1382 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1385 if (req->work.flags & IO_WQ_WORK_CREDS) {
1386 put_cred(req->work.identity->creds);
1387 req->work.flags &= ~IO_WQ_WORK_CREDS;
1389 if (req->work.flags & IO_WQ_WORK_FS) {
1390 struct fs_struct *fs = req->work.identity->fs;
1392 spin_lock(&req->work.identity->fs->lock);
1395 spin_unlock(&req->work.identity->fs->lock);
1398 req->work.flags &= ~IO_WQ_WORK_FS;
1400 if (req->flags & REQ_F_INFLIGHT)
1401 io_req_drop_files(req);
1403 io_put_identity(req->task->io_uring, req);
1407 * Create a private copy of io_identity, since some fields don't match
1408 * the current context.
1410 static bool io_identity_cow(struct io_kiocb *req)
1412 struct io_uring_task *tctx = current->io_uring;
1413 const struct cred *creds = NULL;
1414 struct io_identity *id;
1416 if (req->work.flags & IO_WQ_WORK_CREDS)
1417 creds = req->work.identity->creds;
1419 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1420 if (unlikely(!id)) {
1421 req->work.flags |= IO_WQ_WORK_CANCEL;
1426 * We can safely just re-init the creds we copied Either the field
1427 * matches the current one, or we haven't grabbed it yet. The only
1428 * exception is ->creds, through registered personalities, so handle
1429 * that one separately.
1431 io_init_identity(id);
1435 /* add one for this request */
1436 refcount_inc(&id->count);
1438 /* drop tctx and req identity references, if needed */
1439 if (tctx->identity != &tctx->__identity &&
1440 refcount_dec_and_test(&tctx->identity->count))
1441 kfree(tctx->identity);
1442 if (req->work.identity != &tctx->__identity &&
1443 refcount_dec_and_test(&req->work.identity->count))
1444 kfree(req->work.identity);
1446 req->work.identity = id;
1447 tctx->identity = id;
1451 static bool io_grab_identity(struct io_kiocb *req)
1453 const struct io_op_def *def = &io_op_defs[req->opcode];
1454 struct io_identity *id = req->work.identity;
1455 struct io_ring_ctx *ctx = req->ctx;
1457 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1458 if (id->fsize != rlimit(RLIMIT_FSIZE))
1460 req->work.flags |= IO_WQ_WORK_FSIZE;
1462 #ifdef CONFIG_BLK_CGROUP
1463 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1464 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1466 if (id->blkcg_css != blkcg_css()) {
1471 * This should be rare, either the cgroup is dying or the task
1472 * is moving cgroups. Just punt to root for the handful of ios.
1474 if (css_tryget_online(id->blkcg_css))
1475 req->work.flags |= IO_WQ_WORK_BLKCG;
1479 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1480 if (id->creds != current_cred())
1482 get_cred(id->creds);
1483 req->work.flags |= IO_WQ_WORK_CREDS;
1486 if (!uid_eq(current->loginuid, id->loginuid) ||
1487 current->sessionid != id->sessionid)
1490 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1491 (def->work_flags & IO_WQ_WORK_FS)) {
1492 if (current->fs != id->fs)
1494 spin_lock(&id->fs->lock);
1495 if (!id->fs->in_exec) {
1497 req->work.flags |= IO_WQ_WORK_FS;
1499 req->work.flags |= IO_WQ_WORK_CANCEL;
1501 spin_unlock(¤t->fs->lock);
1503 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1504 (def->work_flags & IO_WQ_WORK_FILES) &&
1505 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1506 if (id->files != current->files ||
1507 id->nsproxy != current->nsproxy)
1509 atomic_inc(&id->files->count);
1510 get_nsproxy(id->nsproxy);
1512 if (!(req->flags & REQ_F_INFLIGHT)) {
1513 req->flags |= REQ_F_INFLIGHT;
1515 spin_lock_irq(&ctx->inflight_lock);
1516 list_add(&req->inflight_entry, &ctx->inflight_list);
1517 spin_unlock_irq(&ctx->inflight_lock);
1519 req->work.flags |= IO_WQ_WORK_FILES;
1521 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1522 (def->work_flags & IO_WQ_WORK_MM)) {
1523 if (id->mm != current->mm)
1526 req->work.flags |= IO_WQ_WORK_MM;
1532 static void io_prep_async_work(struct io_kiocb *req)
1534 const struct io_op_def *def = &io_op_defs[req->opcode];
1535 struct io_ring_ctx *ctx = req->ctx;
1537 io_req_init_async(req);
1539 if (req->flags & REQ_F_FORCE_ASYNC)
1540 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1542 if (req->flags & REQ_F_ISREG) {
1543 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1544 io_wq_hash_work(&req->work, file_inode(req->file));
1546 if (def->unbound_nonreg_file)
1547 req->work.flags |= IO_WQ_WORK_UNBOUND;
1550 /* if we fail grabbing identity, we must COW, regrab, and retry */
1551 if (io_grab_identity(req))
1554 if (!io_identity_cow(req))
1557 /* can't fail at this point */
1558 if (!io_grab_identity(req))
1562 static void io_prep_async_link(struct io_kiocb *req)
1564 struct io_kiocb *cur;
1566 io_for_each_link(cur, req)
1567 io_prep_async_work(cur);
1570 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1572 struct io_ring_ctx *ctx = req->ctx;
1573 struct io_kiocb *link = io_prep_linked_timeout(req);
1575 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1576 &req->work, req->flags);
1577 io_wq_enqueue(ctx->io_wq, &req->work);
1581 static void io_queue_async_work(struct io_kiocb *req)
1583 struct io_kiocb *link;
1585 /* init ->work of the whole link before punting */
1586 io_prep_async_link(req);
1587 link = __io_queue_async_work(req);
1590 io_queue_linked_timeout(link);
1593 static void io_kill_timeout(struct io_kiocb *req)
1595 struct io_timeout_data *io = req->async_data;
1598 ret = hrtimer_try_to_cancel(&io->timer);
1600 atomic_set(&req->ctx->cq_timeouts,
1601 atomic_read(&req->ctx->cq_timeouts) + 1);
1602 list_del_init(&req->timeout.list);
1603 io_cqring_fill_event(req, 0);
1604 io_put_req_deferred(req, 1);
1609 * Returns true if we found and killed one or more timeouts
1611 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1612 struct files_struct *files)
1614 struct io_kiocb *req, *tmp;
1617 spin_lock_irq(&ctx->completion_lock);
1618 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1619 if (io_match_task(req, tsk, files)) {
1620 io_kill_timeout(req);
1624 spin_unlock_irq(&ctx->completion_lock);
1625 return canceled != 0;
1628 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1631 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1632 struct io_defer_entry, list);
1633 struct io_kiocb *link;
1635 if (req_need_defer(de->req, de->seq))
1637 list_del_init(&de->list);
1638 /* punt-init is done before queueing for defer */
1639 link = __io_queue_async_work(de->req);
1641 __io_queue_linked_timeout(link);
1642 /* drop submission reference */
1643 io_put_req_deferred(link, 1);
1646 } while (!list_empty(&ctx->defer_list));
1649 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1653 if (list_empty(&ctx->timeout_list))
1656 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1659 u32 events_needed, events_got;
1660 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1661 struct io_kiocb, timeout.list);
1663 if (io_is_timeout_noseq(req))
1667 * Since seq can easily wrap around over time, subtract
1668 * the last seq at which timeouts were flushed before comparing.
1669 * Assuming not more than 2^31-1 events have happened since,
1670 * these subtractions won't have wrapped, so we can check if
1671 * target is in [last_seq, current_seq] by comparing the two.
1673 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1674 events_got = seq - ctx->cq_last_tm_flush;
1675 if (events_got < events_needed)
1678 list_del_init(&req->timeout.list);
1679 io_kill_timeout(req);
1680 } while (!list_empty(&ctx->timeout_list));
1682 ctx->cq_last_tm_flush = seq;
1685 static void io_commit_cqring(struct io_ring_ctx *ctx)
1687 io_flush_timeouts(ctx);
1688 __io_commit_cqring(ctx);
1690 if (unlikely(!list_empty(&ctx->defer_list)))
1691 __io_queue_deferred(ctx);
1694 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1696 struct io_rings *r = ctx->rings;
1698 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1701 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1703 struct io_rings *rings = ctx->rings;
1706 tail = ctx->cached_cq_tail;
1708 * writes to the cq entry need to come after reading head; the
1709 * control dependency is enough as we're using WRITE_ONCE to
1712 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1715 ctx->cached_cq_tail++;
1716 return &rings->cqes[tail & ctx->cq_mask];
1719 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1723 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1725 if (!ctx->eventfd_async)
1727 return io_wq_current_is_worker();
1730 static inline unsigned __io_cqring_events(struct io_ring_ctx *ctx)
1732 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1735 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1737 /* see waitqueue_active() comment */
1740 if (waitqueue_active(&ctx->wait))
1741 wake_up(&ctx->wait);
1742 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1743 wake_up(&ctx->sq_data->wait);
1744 if (io_should_trigger_evfd(ctx))
1745 eventfd_signal(ctx->cq_ev_fd, 1);
1746 if (waitqueue_active(&ctx->cq_wait)) {
1747 wake_up_interruptible(&ctx->cq_wait);
1748 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1752 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1754 /* see waitqueue_active() comment */
1757 if (ctx->flags & IORING_SETUP_SQPOLL) {
1758 if (waitqueue_active(&ctx->wait))
1759 wake_up(&ctx->wait);
1761 if (io_should_trigger_evfd(ctx))
1762 eventfd_signal(ctx->cq_ev_fd, 1);
1763 if (waitqueue_active(&ctx->cq_wait)) {
1764 wake_up_interruptible(&ctx->cq_wait);
1765 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1769 /* Returns true if there are no backlogged entries after the flush */
1770 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1771 struct task_struct *tsk,
1772 struct files_struct *files)
1774 struct io_rings *rings = ctx->rings;
1775 struct io_kiocb *req, *tmp;
1776 struct io_uring_cqe *cqe;
1777 unsigned long flags;
1781 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1784 spin_lock_irqsave(&ctx->completion_lock, flags);
1785 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1786 if (!io_match_task(req, tsk, files))
1789 cqe = io_get_cqring(ctx);
1793 list_move(&req->compl.list, &list);
1795 WRITE_ONCE(cqe->user_data, req->user_data);
1796 WRITE_ONCE(cqe->res, req->result);
1797 WRITE_ONCE(cqe->flags, req->compl.cflags);
1799 ctx->cached_cq_overflow++;
1800 WRITE_ONCE(ctx->rings->cq_overflow,
1801 ctx->cached_cq_overflow);
1805 all_flushed = list_empty(&ctx->cq_overflow_list);
1807 clear_bit(0, &ctx->sq_check_overflow);
1808 clear_bit(0, &ctx->cq_check_overflow);
1809 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1812 io_commit_cqring(ctx);
1813 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1814 io_cqring_ev_posted(ctx);
1816 while (!list_empty(&list)) {
1817 req = list_first_entry(&list, struct io_kiocb, compl.list);
1818 list_del(&req->compl.list);
1825 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1826 struct task_struct *tsk,
1827 struct files_struct *files)
1829 if (test_bit(0, &ctx->cq_check_overflow)) {
1830 /* iopoll syncs against uring_lock, not completion_lock */
1831 if (ctx->flags & IORING_SETUP_IOPOLL)
1832 mutex_lock(&ctx->uring_lock);
1833 __io_cqring_overflow_flush(ctx, force, tsk, files);
1834 if (ctx->flags & IORING_SETUP_IOPOLL)
1835 mutex_unlock(&ctx->uring_lock);
1839 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1841 struct io_ring_ctx *ctx = req->ctx;
1842 struct io_uring_cqe *cqe;
1844 trace_io_uring_complete(ctx, req->user_data, res);
1847 * If we can't get a cq entry, userspace overflowed the
1848 * submission (by quite a lot). Increment the overflow count in
1851 cqe = io_get_cqring(ctx);
1853 WRITE_ONCE(cqe->user_data, req->user_data);
1854 WRITE_ONCE(cqe->res, res);
1855 WRITE_ONCE(cqe->flags, cflags);
1856 } else if (ctx->cq_overflow_flushed ||
1857 atomic_read(&req->task->io_uring->in_idle)) {
1859 * If we're in ring overflow flush mode, or in task cancel mode,
1860 * then we cannot store the request for later flushing, we need
1861 * to drop it on the floor.
1863 ctx->cached_cq_overflow++;
1864 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1866 if (list_empty(&ctx->cq_overflow_list)) {
1867 set_bit(0, &ctx->sq_check_overflow);
1868 set_bit(0, &ctx->cq_check_overflow);
1869 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1873 req->compl.cflags = cflags;
1874 refcount_inc(&req->refs);
1875 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1879 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1881 __io_cqring_fill_event(req, res, 0);
1884 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1886 struct io_ring_ctx *ctx = req->ctx;
1887 unsigned long flags;
1889 spin_lock_irqsave(&ctx->completion_lock, flags);
1890 __io_cqring_fill_event(req, res, cflags);
1891 io_commit_cqring(ctx);
1892 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1894 io_cqring_ev_posted(ctx);
1897 static void io_submit_flush_completions(struct io_comp_state *cs)
1899 struct io_ring_ctx *ctx = cs->ctx;
1901 spin_lock_irq(&ctx->completion_lock);
1902 while (!list_empty(&cs->list)) {
1903 struct io_kiocb *req;
1905 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1906 list_del(&req->compl.list);
1907 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1910 * io_free_req() doesn't care about completion_lock unless one
1911 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1912 * because of a potential deadlock with req->work.fs->lock
1914 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1915 |REQ_F_WORK_INITIALIZED)) {
1916 spin_unlock_irq(&ctx->completion_lock);
1918 spin_lock_irq(&ctx->completion_lock);
1923 io_commit_cqring(ctx);
1924 spin_unlock_irq(&ctx->completion_lock);
1926 io_cqring_ev_posted(ctx);
1930 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1931 struct io_comp_state *cs)
1934 io_cqring_add_event(req, res, cflags);
1939 req->compl.cflags = cflags;
1940 list_add_tail(&req->compl.list, &cs->list);
1942 io_submit_flush_completions(cs);
1946 static void io_req_complete(struct io_kiocb *req, long res)
1948 __io_req_complete(req, res, 0, NULL);
1951 static inline bool io_is_fallback_req(struct io_kiocb *req)
1953 return req == (struct io_kiocb *)
1954 ((unsigned long) req->ctx->fallback_req & ~1UL);
1957 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1959 struct io_kiocb *req;
1961 req = ctx->fallback_req;
1962 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1968 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1969 struct io_submit_state *state)
1971 if (!state->free_reqs) {
1972 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1976 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1977 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1980 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1981 * retry single alloc to be on the safe side.
1983 if (unlikely(ret <= 0)) {
1984 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1985 if (!state->reqs[0])
1989 state->free_reqs = ret;
1993 return state->reqs[state->free_reqs];
1995 return io_get_fallback_req(ctx);
1998 static inline void io_put_file(struct io_kiocb *req, struct file *file,
2005 static void io_dismantle_req(struct io_kiocb *req)
2009 if (req->async_data)
2010 kfree(req->async_data);
2012 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
2013 if (req->fixed_file_refs)
2014 percpu_ref_put(req->fixed_file_refs);
2015 io_req_clean_work(req);
2018 static void __io_free_req(struct io_kiocb *req)
2020 struct io_uring_task *tctx = req->task->io_uring;
2021 struct io_ring_ctx *ctx = req->ctx;
2023 io_dismantle_req(req);
2025 percpu_counter_dec(&tctx->inflight);
2026 if (atomic_read(&tctx->in_idle))
2027 wake_up(&tctx->wait);
2028 put_task_struct(req->task);
2030 if (likely(!io_is_fallback_req(req)))
2031 kmem_cache_free(req_cachep, req);
2033 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
2034 percpu_ref_put(&ctx->refs);
2037 static inline void io_remove_next_linked(struct io_kiocb *req)
2039 struct io_kiocb *nxt = req->link;
2041 req->link = nxt->link;
2045 static void io_kill_linked_timeout(struct io_kiocb *req)
2047 struct io_ring_ctx *ctx = req->ctx;
2048 struct io_kiocb *link;
2049 bool cancelled = false;
2050 unsigned long flags;
2052 spin_lock_irqsave(&ctx->completion_lock, flags);
2056 * Can happen if a linked timeout fired and link had been like
2057 * req -> link t-out -> link t-out [-> ...]
2059 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
2060 struct io_timeout_data *io = link->async_data;
2063 io_remove_next_linked(req);
2064 link->timeout.head = NULL;
2065 ret = hrtimer_try_to_cancel(&io->timer);
2067 io_cqring_fill_event(link, -ECANCELED);
2068 io_commit_cqring(ctx);
2072 req->flags &= ~REQ_F_LINK_TIMEOUT;
2073 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2076 io_cqring_ev_posted(ctx);
2082 static void io_fail_links(struct io_kiocb *req)
2084 struct io_kiocb *link, *nxt;
2085 struct io_ring_ctx *ctx = req->ctx;
2086 unsigned long flags;
2088 spin_lock_irqsave(&ctx->completion_lock, flags);
2096 trace_io_uring_fail_link(req, link);
2097 io_cqring_fill_event(link, -ECANCELED);
2100 * It's ok to free under spinlock as they're not linked anymore,
2101 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2104 if (link->flags & REQ_F_WORK_INITIALIZED)
2105 io_put_req_deferred(link, 2);
2107 io_double_put_req(link);
2110 io_commit_cqring(ctx);
2111 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2113 io_cqring_ev_posted(ctx);
2116 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2118 if (req->flags & REQ_F_LINK_TIMEOUT)
2119 io_kill_linked_timeout(req);
2122 * If LINK is set, we have dependent requests in this chain. If we
2123 * didn't fail this request, queue the first one up, moving any other
2124 * dependencies to the next request. In case of failure, fail the rest
2127 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2128 struct io_kiocb *nxt = req->link;
2137 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2139 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2141 return __io_req_find_next(req);
2144 static int io_req_task_work_add(struct io_kiocb *req)
2146 struct task_struct *tsk = req->task;
2147 struct io_ring_ctx *ctx = req->ctx;
2148 enum task_work_notify_mode notify;
2151 if (tsk->flags & PF_EXITING)
2155 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2156 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2157 * processing task_work. There's no reliable way to tell if TWA_RESUME
2161 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2162 notify = TWA_SIGNAL;
2164 ret = task_work_add(tsk, &req->task_work, notify);
2166 wake_up_process(tsk);
2171 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2173 struct io_ring_ctx *ctx = req->ctx;
2175 spin_lock_irq(&ctx->completion_lock);
2176 io_cqring_fill_event(req, error);
2177 io_commit_cqring(ctx);
2178 spin_unlock_irq(&ctx->completion_lock);
2180 io_cqring_ev_posted(ctx);
2181 req_set_fail_links(req);
2182 io_double_put_req(req);
2185 static void io_req_task_cancel(struct callback_head *cb)
2187 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2188 struct io_ring_ctx *ctx = req->ctx;
2190 __io_req_task_cancel(req, -ECANCELED);
2191 percpu_ref_put(&ctx->refs);
2194 static void __io_req_task_submit(struct io_kiocb *req)
2196 struct io_ring_ctx *ctx = req->ctx;
2198 mutex_lock(&ctx->uring_lock);
2199 if (!ctx->sqo_dead &&
2200 !__io_sq_thread_acquire_mm(ctx) &&
2201 !__io_sq_thread_acquire_files(ctx))
2202 __io_queue_sqe(req, NULL);
2204 __io_req_task_cancel(req, -EFAULT);
2205 mutex_unlock(&ctx->uring_lock);
2208 static void io_req_task_submit(struct callback_head *cb)
2210 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2211 struct io_ring_ctx *ctx = req->ctx;
2213 __io_req_task_submit(req);
2214 percpu_ref_put(&ctx->refs);
2217 static void io_req_task_queue(struct io_kiocb *req)
2221 init_task_work(&req->task_work, io_req_task_submit);
2222 percpu_ref_get(&req->ctx->refs);
2224 ret = io_req_task_work_add(req);
2225 if (unlikely(ret)) {
2226 struct task_struct *tsk;
2228 init_task_work(&req->task_work, io_req_task_cancel);
2229 tsk = io_wq_get_task(req->ctx->io_wq);
2230 task_work_add(tsk, &req->task_work, TWA_NONE);
2231 wake_up_process(tsk);
2235 static inline void io_queue_next(struct io_kiocb *req)
2237 struct io_kiocb *nxt = io_req_find_next(req);
2240 io_req_task_queue(nxt);
2243 static void io_free_req(struct io_kiocb *req)
2250 void *reqs[IO_IOPOLL_BATCH];
2253 struct task_struct *task;
2257 static inline void io_init_req_batch(struct req_batch *rb)
2264 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2265 struct req_batch *rb)
2267 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2268 percpu_ref_put_many(&ctx->refs, rb->to_free);
2272 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2273 struct req_batch *rb)
2276 __io_req_free_batch_flush(ctx, rb);
2278 struct io_uring_task *tctx = rb->task->io_uring;
2280 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2281 if (atomic_read(&tctx->in_idle))
2282 wake_up(&tctx->wait);
2283 put_task_struct_many(rb->task, rb->task_refs);
2288 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2290 if (unlikely(io_is_fallback_req(req))) {
2296 if (req->task != rb->task) {
2298 struct io_uring_task *tctx = rb->task->io_uring;
2300 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2301 if (atomic_read(&tctx->in_idle))
2302 wake_up(&tctx->wait);
2303 put_task_struct_many(rb->task, rb->task_refs);
2305 rb->task = req->task;
2310 io_dismantle_req(req);
2311 rb->reqs[rb->to_free++] = req;
2312 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2313 __io_req_free_batch_flush(req->ctx, rb);
2317 * Drop reference to request, return next in chain (if there is one) if this
2318 * was the last reference to this request.
2320 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2322 struct io_kiocb *nxt = NULL;
2324 if (refcount_dec_and_test(&req->refs)) {
2325 nxt = io_req_find_next(req);
2331 static void io_put_req(struct io_kiocb *req)
2333 if (refcount_dec_and_test(&req->refs))
2337 static void io_put_req_deferred_cb(struct callback_head *cb)
2339 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2344 static void io_free_req_deferred(struct io_kiocb *req)
2348 init_task_work(&req->task_work, io_put_req_deferred_cb);
2349 ret = io_req_task_work_add(req);
2350 if (unlikely(ret)) {
2351 struct task_struct *tsk;
2353 tsk = io_wq_get_task(req->ctx->io_wq);
2354 task_work_add(tsk, &req->task_work, TWA_NONE);
2355 wake_up_process(tsk);
2359 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2361 if (refcount_sub_and_test(refs, &req->refs))
2362 io_free_req_deferred(req);
2365 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2367 struct io_kiocb *nxt;
2370 * A ref is owned by io-wq in which context we're. So, if that's the
2371 * last one, it's safe to steal next work. False negatives are Ok,
2372 * it just will be re-punted async in io_put_work()
2374 if (refcount_read(&req->refs) != 1)
2377 nxt = io_req_find_next(req);
2378 return nxt ? &nxt->work : NULL;
2381 static void io_double_put_req(struct io_kiocb *req)
2383 /* drop both submit and complete references */
2384 if (refcount_sub_and_test(2, &req->refs))
2388 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2390 /* See comment at the top of this file */
2392 return __io_cqring_events(ctx);
2395 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2397 struct io_rings *rings = ctx->rings;
2399 /* make sure SQ entry isn't read before tail */
2400 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2403 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2405 unsigned int cflags;
2407 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2408 cflags |= IORING_CQE_F_BUFFER;
2409 req->flags &= ~REQ_F_BUFFER_SELECTED;
2414 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2416 struct io_buffer *kbuf;
2418 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2419 return io_put_kbuf(req, kbuf);
2422 static inline bool io_run_task_work(void)
2425 * Not safe to run on exiting task, and the task_work handling will
2426 * not add work to such a task.
2428 if (unlikely(current->flags & PF_EXITING))
2430 if (current->task_works) {
2431 __set_current_state(TASK_RUNNING);
2439 static void io_iopoll_queue(struct list_head *again)
2441 struct io_kiocb *req;
2444 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2445 list_del(&req->inflight_entry);
2446 __io_complete_rw(req, -EAGAIN, 0, NULL);
2447 } while (!list_empty(again));
2451 * Find and free completed poll iocbs
2453 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2454 struct list_head *done)
2456 struct req_batch rb;
2457 struct io_kiocb *req;
2460 /* order with ->result store in io_complete_rw_iopoll() */
2463 io_init_req_batch(&rb);
2464 while (!list_empty(done)) {
2467 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2468 if (READ_ONCE(req->result) == -EAGAIN) {
2470 req->iopoll_completed = 0;
2471 list_move_tail(&req->inflight_entry, &again);
2474 list_del(&req->inflight_entry);
2476 if (req->flags & REQ_F_BUFFER_SELECTED)
2477 cflags = io_put_rw_kbuf(req);
2479 __io_cqring_fill_event(req, req->result, cflags);
2482 if (refcount_dec_and_test(&req->refs))
2483 io_req_free_batch(&rb, req);
2486 io_commit_cqring(ctx);
2487 io_cqring_ev_posted_iopoll(ctx);
2488 io_req_free_batch_finish(ctx, &rb);
2490 if (!list_empty(&again))
2491 io_iopoll_queue(&again);
2494 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2497 struct io_kiocb *req, *tmp;
2503 * Only spin for completions if we don't have multiple devices hanging
2504 * off our complete list, and we're under the requested amount.
2506 spin = !ctx->poll_multi_file && *nr_events < min;
2509 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2510 struct kiocb *kiocb = &req->rw.kiocb;
2513 * Move completed and retryable entries to our local lists.
2514 * If we find a request that requires polling, break out
2515 * and complete those lists first, if we have entries there.
2517 if (READ_ONCE(req->iopoll_completed)) {
2518 list_move_tail(&req->inflight_entry, &done);
2521 if (!list_empty(&done))
2524 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2528 /* iopoll may have completed current req */
2529 if (READ_ONCE(req->iopoll_completed))
2530 list_move_tail(&req->inflight_entry, &done);
2537 if (!list_empty(&done))
2538 io_iopoll_complete(ctx, nr_events, &done);
2544 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2545 * non-spinning poll check - we'll still enter the driver poll loop, but only
2546 * as a non-spinning completion check.
2548 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2551 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2554 ret = io_do_iopoll(ctx, nr_events, min);
2557 if (*nr_events >= min)
2565 * We can't just wait for polled events to come to us, we have to actively
2566 * find and complete them.
2568 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2570 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2573 mutex_lock(&ctx->uring_lock);
2574 while (!list_empty(&ctx->iopoll_list)) {
2575 unsigned int nr_events = 0;
2577 io_do_iopoll(ctx, &nr_events, 0);
2579 /* let it sleep and repeat later if can't complete a request */
2583 * Ensure we allow local-to-the-cpu processing to take place,
2584 * in this case we need to ensure that we reap all events.
2585 * Also let task_work, etc. to progress by releasing the mutex
2587 if (need_resched()) {
2588 mutex_unlock(&ctx->uring_lock);
2590 mutex_lock(&ctx->uring_lock);
2593 mutex_unlock(&ctx->uring_lock);
2596 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2598 unsigned int nr_events = 0;
2599 int iters = 0, ret = 0;
2602 * We disallow the app entering submit/complete with polling, but we
2603 * still need to lock the ring to prevent racing with polled issue
2604 * that got punted to a workqueue.
2606 mutex_lock(&ctx->uring_lock);
2609 * Don't enter poll loop if we already have events pending.
2610 * If we do, we can potentially be spinning for commands that
2611 * already triggered a CQE (eg in error).
2613 if (test_bit(0, &ctx->cq_check_overflow))
2614 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2615 if (io_cqring_events(ctx))
2619 * If a submit got punted to a workqueue, we can have the
2620 * application entering polling for a command before it gets
2621 * issued. That app will hold the uring_lock for the duration
2622 * of the poll right here, so we need to take a breather every
2623 * now and then to ensure that the issue has a chance to add
2624 * the poll to the issued list. Otherwise we can spin here
2625 * forever, while the workqueue is stuck trying to acquire the
2628 if (!(++iters & 7)) {
2629 mutex_unlock(&ctx->uring_lock);
2631 mutex_lock(&ctx->uring_lock);
2634 ret = io_iopoll_getevents(ctx, &nr_events, min);
2638 } while (min && !nr_events && !need_resched());
2640 mutex_unlock(&ctx->uring_lock);
2644 static void kiocb_end_write(struct io_kiocb *req)
2647 * Tell lockdep we inherited freeze protection from submission
2650 if (req->flags & REQ_F_ISREG) {
2651 struct inode *inode = file_inode(req->file);
2653 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2655 file_end_write(req->file);
2658 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2659 struct io_comp_state *cs)
2661 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2664 if (kiocb->ki_flags & IOCB_WRITE)
2665 kiocb_end_write(req);
2667 if (res != req->result)
2668 req_set_fail_links(req);
2669 if (req->flags & REQ_F_BUFFER_SELECTED)
2670 cflags = io_put_rw_kbuf(req);
2671 __io_req_complete(req, res, cflags, cs);
2675 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2677 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2678 ssize_t ret = -ECANCELED;
2679 struct iov_iter iter;
2687 switch (req->opcode) {
2688 case IORING_OP_READV:
2689 case IORING_OP_READ_FIXED:
2690 case IORING_OP_READ:
2693 case IORING_OP_WRITEV:
2694 case IORING_OP_WRITE_FIXED:
2695 case IORING_OP_WRITE:
2699 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2704 if (!req->async_data) {
2705 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2708 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2716 req_set_fail_links(req);
2721 static bool io_rw_reissue(struct io_kiocb *req, long res)
2724 umode_t mode = file_inode(req->file)->i_mode;
2727 if (!S_ISBLK(mode) && !S_ISREG(mode))
2729 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2732 lockdep_assert_held(&req->ctx->uring_lock);
2734 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2736 if (io_resubmit_prep(req, ret)) {
2737 refcount_inc(&req->refs);
2738 io_queue_async_work(req);
2746 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2747 struct io_comp_state *cs)
2749 if (!io_rw_reissue(req, res))
2750 io_complete_rw_common(&req->rw.kiocb, res, cs);
2753 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2755 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2757 __io_complete_rw(req, res, res2, NULL);
2760 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2762 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2764 if (kiocb->ki_flags & IOCB_WRITE)
2765 kiocb_end_write(req);
2767 if (res != -EAGAIN && res != req->result)
2768 req_set_fail_links(req);
2770 WRITE_ONCE(req->result, res);
2771 /* order with io_poll_complete() checking ->result */
2773 WRITE_ONCE(req->iopoll_completed, 1);
2777 * After the iocb has been issued, it's safe to be found on the poll list.
2778 * Adding the kiocb to the list AFTER submission ensures that we don't
2779 * find it from a io_iopoll_getevents() thread before the issuer is done
2780 * accessing the kiocb cookie.
2782 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2784 struct io_ring_ctx *ctx = req->ctx;
2787 * Track whether we have multiple files in our lists. This will impact
2788 * how we do polling eventually, not spinning if we're on potentially
2789 * different devices.
2791 if (list_empty(&ctx->iopoll_list)) {
2792 ctx->poll_multi_file = false;
2793 } else if (!ctx->poll_multi_file) {
2794 struct io_kiocb *list_req;
2796 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2798 if (list_req->file != req->file)
2799 ctx->poll_multi_file = true;
2803 * For fast devices, IO may have already completed. If it has, add
2804 * it to the front so we find it first.
2806 if (READ_ONCE(req->iopoll_completed))
2807 list_add(&req->inflight_entry, &ctx->iopoll_list);
2809 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2812 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2813 * task context or in io worker task context. If current task context is
2814 * sq thread, we don't need to check whether should wake up sq thread.
2816 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2817 wq_has_sleeper(&ctx->sq_data->wait))
2818 wake_up(&ctx->sq_data->wait);
2821 static inline void __io_state_file_put(struct io_submit_state *state)
2823 fput_many(state->file, state->file_refs);
2824 state->file_refs = 0;
2827 static inline void io_state_file_put(struct io_submit_state *state)
2829 if (state->file_refs)
2830 __io_state_file_put(state);
2834 * Get as many references to a file as we have IOs left in this submission,
2835 * assuming most submissions are for one file, or at least that each file
2836 * has more than one submission.
2838 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2843 if (state->file_refs) {
2844 if (state->fd == fd) {
2848 __io_state_file_put(state);
2850 state->file = fget_many(fd, state->ios_left);
2851 if (unlikely(!state->file))
2855 state->file_refs = state->ios_left - 1;
2859 static bool io_bdev_nowait(struct block_device *bdev)
2861 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2865 * If we tracked the file through the SCM inflight mechanism, we could support
2866 * any file. For now, just ensure that anything potentially problematic is done
2869 static bool io_file_supports_async(struct file *file, int rw)
2871 umode_t mode = file_inode(file)->i_mode;
2873 if (S_ISBLK(mode)) {
2874 if (IS_ENABLED(CONFIG_BLOCK) &&
2875 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2879 if (S_ISCHR(mode) || S_ISSOCK(mode))
2881 if (S_ISREG(mode)) {
2882 if (IS_ENABLED(CONFIG_BLOCK) &&
2883 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2884 file->f_op != &io_uring_fops)
2889 /* any ->read/write should understand O_NONBLOCK */
2890 if (file->f_flags & O_NONBLOCK)
2893 if (!(file->f_mode & FMODE_NOWAIT))
2897 return file->f_op->read_iter != NULL;
2899 return file->f_op->write_iter != NULL;
2902 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2904 struct io_ring_ctx *ctx = req->ctx;
2905 struct kiocb *kiocb = &req->rw.kiocb;
2909 if (S_ISREG(file_inode(req->file)->i_mode))
2910 req->flags |= REQ_F_ISREG;
2912 kiocb->ki_pos = READ_ONCE(sqe->off);
2913 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2914 req->flags |= REQ_F_CUR_POS;
2915 kiocb->ki_pos = req->file->f_pos;
2917 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2918 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2919 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2923 ioprio = READ_ONCE(sqe->ioprio);
2925 ret = ioprio_check_cap(ioprio);
2929 kiocb->ki_ioprio = ioprio;
2931 kiocb->ki_ioprio = get_current_ioprio();
2933 /* don't allow async punt if RWF_NOWAIT was requested */
2934 if (kiocb->ki_flags & IOCB_NOWAIT)
2935 req->flags |= REQ_F_NOWAIT;
2937 if (ctx->flags & IORING_SETUP_IOPOLL) {
2938 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2939 !kiocb->ki_filp->f_op->iopoll)
2942 kiocb->ki_flags |= IOCB_HIPRI;
2943 kiocb->ki_complete = io_complete_rw_iopoll;
2944 req->iopoll_completed = 0;
2946 if (kiocb->ki_flags & IOCB_HIPRI)
2948 kiocb->ki_complete = io_complete_rw;
2951 req->rw.addr = READ_ONCE(sqe->addr);
2952 req->rw.len = READ_ONCE(sqe->len);
2953 req->buf_index = READ_ONCE(sqe->buf_index);
2957 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2963 case -ERESTARTNOINTR:
2964 case -ERESTARTNOHAND:
2965 case -ERESTART_RESTARTBLOCK:
2967 * We can't just restart the syscall, since previously
2968 * submitted sqes may already be in progress. Just fail this
2974 kiocb->ki_complete(kiocb, ret, 0);
2978 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2979 struct io_comp_state *cs)
2981 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2982 struct io_async_rw *io = req->async_data;
2984 /* add previously done IO, if any */
2985 if (io && io->bytes_done > 0) {
2987 ret = io->bytes_done;
2989 ret += io->bytes_done;
2992 if (req->flags & REQ_F_CUR_POS)
2993 req->file->f_pos = kiocb->ki_pos;
2994 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2995 __io_complete_rw(req, ret, 0, cs);
2997 io_rw_done(kiocb, ret);
3000 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
3001 struct iov_iter *iter)
3003 struct io_ring_ctx *ctx = req->ctx;
3004 size_t len = req->rw.len;
3005 struct io_mapped_ubuf *imu;
3006 u16 index, buf_index = req->buf_index;
3010 if (unlikely(buf_index >= ctx->nr_user_bufs))
3012 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3013 imu = &ctx->user_bufs[index];
3014 buf_addr = req->rw.addr;
3017 if (buf_addr + len < buf_addr)
3019 /* not inside the mapped region */
3020 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
3024 * May not be a start of buffer, set size appropriately
3025 * and advance us to the beginning.
3027 offset = buf_addr - imu->ubuf;
3028 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3032 * Don't use iov_iter_advance() here, as it's really slow for
3033 * using the latter parts of a big fixed buffer - it iterates
3034 * over each segment manually. We can cheat a bit here, because
3037 * 1) it's a BVEC iter, we set it up
3038 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3039 * first and last bvec
3041 * So just find our index, and adjust the iterator afterwards.
3042 * If the offset is within the first bvec (or the whole first
3043 * bvec, just use iov_iter_advance(). This makes it easier
3044 * since we can just skip the first segment, which may not
3045 * be PAGE_SIZE aligned.
3047 const struct bio_vec *bvec = imu->bvec;
3049 if (offset <= bvec->bv_len) {
3050 iov_iter_advance(iter, offset);
3052 unsigned long seg_skip;
3054 /* skip first vec */
3055 offset -= bvec->bv_len;
3056 seg_skip = 1 + (offset >> PAGE_SHIFT);
3058 iter->bvec = bvec + seg_skip;
3059 iter->nr_segs -= seg_skip;
3060 iter->count -= bvec->bv_len + offset;
3061 iter->iov_offset = offset & ~PAGE_MASK;
3068 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3071 mutex_unlock(&ctx->uring_lock);
3074 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3077 * "Normal" inline submissions always hold the uring_lock, since we
3078 * grab it from the system call. Same is true for the SQPOLL offload.
3079 * The only exception is when we've detached the request and issue it
3080 * from an async worker thread, grab the lock for that case.
3083 mutex_lock(&ctx->uring_lock);
3086 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3087 int bgid, struct io_buffer *kbuf,
3090 struct io_buffer *head;
3092 if (req->flags & REQ_F_BUFFER_SELECTED)
3095 io_ring_submit_lock(req->ctx, needs_lock);
3097 lockdep_assert_held(&req->ctx->uring_lock);
3099 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3101 if (!list_empty(&head->list)) {
3102 kbuf = list_last_entry(&head->list, struct io_buffer,
3104 list_del(&kbuf->list);
3107 idr_remove(&req->ctx->io_buffer_idr, bgid);
3109 if (*len > kbuf->len)
3112 kbuf = ERR_PTR(-ENOBUFS);
3115 io_ring_submit_unlock(req->ctx, needs_lock);
3120 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3123 struct io_buffer *kbuf;
3126 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3127 bgid = req->buf_index;
3128 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3131 req->rw.addr = (u64) (unsigned long) kbuf;
3132 req->flags |= REQ_F_BUFFER_SELECTED;
3133 return u64_to_user_ptr(kbuf->addr);
3136 #ifdef CONFIG_COMPAT
3137 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3140 struct compat_iovec __user *uiov;
3141 compat_ssize_t clen;
3145 uiov = u64_to_user_ptr(req->rw.addr);
3146 if (!access_ok(uiov, sizeof(*uiov)))
3148 if (__get_user(clen, &uiov->iov_len))
3154 buf = io_rw_buffer_select(req, &len, needs_lock);
3156 return PTR_ERR(buf);
3157 iov[0].iov_base = buf;
3158 iov[0].iov_len = (compat_size_t) len;
3163 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3166 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3170 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3173 len = iov[0].iov_len;
3176 buf = io_rw_buffer_select(req, &len, needs_lock);
3178 return PTR_ERR(buf);
3179 iov[0].iov_base = buf;
3180 iov[0].iov_len = len;
3184 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3187 if (req->flags & REQ_F_BUFFER_SELECTED) {
3188 struct io_buffer *kbuf;
3190 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3191 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3192 iov[0].iov_len = kbuf->len;
3195 if (req->rw.len != 1)
3198 #ifdef CONFIG_COMPAT
3199 if (req->ctx->compat)
3200 return io_compat_import(req, iov, needs_lock);
3203 return __io_iov_buffer_select(req, iov, needs_lock);
3206 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3207 struct iovec **iovec, struct iov_iter *iter,
3210 void __user *buf = u64_to_user_ptr(req->rw.addr);
3211 size_t sqe_len = req->rw.len;
3215 opcode = req->opcode;
3216 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3218 return io_import_fixed(req, rw, iter);
3221 /* buffer index only valid with fixed read/write, or buffer select */
3222 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3225 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3226 if (req->flags & REQ_F_BUFFER_SELECT) {
3227 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3229 return PTR_ERR(buf);
3230 req->rw.len = sqe_len;
3233 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3238 if (req->flags & REQ_F_BUFFER_SELECT) {
3239 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3241 ret = (*iovec)->iov_len;
3242 iov_iter_init(iter, rw, *iovec, 1, ret);
3248 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3252 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3254 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3258 * For files that don't have ->read_iter() and ->write_iter(), handle them
3259 * by looping over ->read() or ->write() manually.
3261 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3263 struct kiocb *kiocb = &req->rw.kiocb;
3264 struct file *file = req->file;
3268 * Don't support polled IO through this interface, and we can't
3269 * support non-blocking either. For the latter, this just causes
3270 * the kiocb to be handled from an async context.
3272 if (kiocb->ki_flags & IOCB_HIPRI)
3274 if (kiocb->ki_flags & IOCB_NOWAIT)
3277 while (iov_iter_count(iter)) {
3281 if (!iov_iter_is_bvec(iter)) {
3282 iovec = iov_iter_iovec(iter);
3284 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3285 iovec.iov_len = req->rw.len;
3289 nr = file->f_op->read(file, iovec.iov_base,
3290 iovec.iov_len, io_kiocb_ppos(kiocb));
3292 nr = file->f_op->write(file, iovec.iov_base,
3293 iovec.iov_len, io_kiocb_ppos(kiocb));
3302 if (nr != iovec.iov_len)
3306 iov_iter_advance(iter, nr);
3312 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3313 const struct iovec *fast_iov, struct iov_iter *iter)
3315 struct io_async_rw *rw = req->async_data;
3317 memcpy(&rw->iter, iter, sizeof(*iter));
3318 rw->free_iovec = iovec;
3320 /* can only be fixed buffers, no need to do anything */
3321 if (iov_iter_is_bvec(iter))
3324 unsigned iov_off = 0;
3326 rw->iter.iov = rw->fast_iov;
3327 if (iter->iov != fast_iov) {
3328 iov_off = iter->iov - fast_iov;
3329 rw->iter.iov += iov_off;
3331 if (rw->fast_iov != fast_iov)
3332 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3333 sizeof(struct iovec) * iter->nr_segs);
3335 req->flags |= REQ_F_NEED_CLEANUP;
3339 static inline int __io_alloc_async_data(struct io_kiocb *req)
3341 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3342 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3343 return req->async_data == NULL;
3346 static int io_alloc_async_data(struct io_kiocb *req)
3348 if (!io_op_defs[req->opcode].needs_async_data)
3351 return __io_alloc_async_data(req);
3354 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3355 const struct iovec *fast_iov,
3356 struct iov_iter *iter, bool force)
3358 if (!force && !io_op_defs[req->opcode].needs_async_data)
3360 if (!req->async_data) {
3361 if (__io_alloc_async_data(req))
3364 io_req_map_rw(req, iovec, fast_iov, iter);
3369 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3371 struct io_async_rw *iorw = req->async_data;
3372 struct iovec *iov = iorw->fast_iov;
3375 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3376 if (unlikely(ret < 0))
3379 iorw->bytes_done = 0;
3380 iorw->free_iovec = iov;
3382 req->flags |= REQ_F_NEED_CLEANUP;
3386 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3390 ret = io_prep_rw(req, sqe);
3394 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3397 /* either don't need iovec imported or already have it */
3398 if (!req->async_data)
3400 return io_rw_prep_async(req, READ);
3404 * This is our waitqueue callback handler, registered through lock_page_async()
3405 * when we initially tried to do the IO with the iocb armed our waitqueue.
3406 * This gets called when the page is unlocked, and we generally expect that to
3407 * happen when the page IO is completed and the page is now uptodate. This will
3408 * queue a task_work based retry of the operation, attempting to copy the data
3409 * again. If the latter fails because the page was NOT uptodate, then we will
3410 * do a thread based blocking retry of the operation. That's the unexpected
3413 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3414 int sync, void *arg)
3416 struct wait_page_queue *wpq;
3417 struct io_kiocb *req = wait->private;
3418 struct wait_page_key *key = arg;
3421 wpq = container_of(wait, struct wait_page_queue, wait);
3423 if (!wake_page_match(wpq, key))
3426 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3427 list_del_init(&wait->entry);
3429 init_task_work(&req->task_work, io_req_task_submit);
3430 percpu_ref_get(&req->ctx->refs);
3432 /* submit ref gets dropped, acquire a new one */
3433 refcount_inc(&req->refs);
3434 ret = io_req_task_work_add(req);
3435 if (unlikely(ret)) {
3436 struct task_struct *tsk;
3438 /* queue just for cancelation */
3439 init_task_work(&req->task_work, io_req_task_cancel);
3440 tsk = io_wq_get_task(req->ctx->io_wq);
3441 task_work_add(tsk, &req->task_work, TWA_NONE);
3442 wake_up_process(tsk);
3448 * This controls whether a given IO request should be armed for async page
3449 * based retry. If we return false here, the request is handed to the async
3450 * worker threads for retry. If we're doing buffered reads on a regular file,
3451 * we prepare a private wait_page_queue entry and retry the operation. This
3452 * will either succeed because the page is now uptodate and unlocked, or it
3453 * will register a callback when the page is unlocked at IO completion. Through
3454 * that callback, io_uring uses task_work to setup a retry of the operation.
3455 * That retry will attempt the buffered read again. The retry will generally
3456 * succeed, or in rare cases where it fails, we then fall back to using the
3457 * async worker threads for a blocking retry.
3459 static bool io_rw_should_retry(struct io_kiocb *req)
3461 struct io_async_rw *rw = req->async_data;
3462 struct wait_page_queue *wait = &rw->wpq;
3463 struct kiocb *kiocb = &req->rw.kiocb;
3465 /* never retry for NOWAIT, we just complete with -EAGAIN */
3466 if (req->flags & REQ_F_NOWAIT)
3469 /* Only for buffered IO */
3470 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3474 * just use poll if we can, and don't attempt if the fs doesn't
3475 * support callback based unlocks
3477 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3480 wait->wait.func = io_async_buf_func;
3481 wait->wait.private = req;
3482 wait->wait.flags = 0;
3483 INIT_LIST_HEAD(&wait->wait.entry);
3484 kiocb->ki_flags |= IOCB_WAITQ;
3485 kiocb->ki_flags &= ~IOCB_NOWAIT;
3486 kiocb->ki_waitq = wait;
3490 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3492 if (req->file->f_op->read_iter)
3493 return call_read_iter(req->file, &req->rw.kiocb, iter);
3494 else if (req->file->f_op->read)
3495 return loop_rw_iter(READ, req, iter);
3500 static int io_read(struct io_kiocb *req, bool force_nonblock,
3501 struct io_comp_state *cs)
3503 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3504 struct kiocb *kiocb = &req->rw.kiocb;
3505 struct iov_iter __iter, *iter = &__iter;
3506 struct io_async_rw *rw = req->async_data;
3507 ssize_t io_size, ret, ret2;
3514 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3518 io_size = iov_iter_count(iter);
3519 req->result = io_size;
3522 /* Ensure we clear previously set non-block flag */
3523 if (!force_nonblock)
3524 kiocb->ki_flags &= ~IOCB_NOWAIT;
3526 kiocb->ki_flags |= IOCB_NOWAIT;
3529 /* If the file doesn't support async, just async punt */
3530 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3534 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3538 ret = io_iter_do_read(req, iter);
3542 } else if (ret == -EIOCBQUEUED) {
3545 } else if (ret == -EAGAIN) {
3546 /* IOPOLL retry should happen for io-wq threads */
3547 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3549 /* no retry on NONBLOCK marked file */
3550 if (req->file->f_flags & O_NONBLOCK)
3552 /* some cases will consume bytes even on error returns */
3553 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3556 } else if (ret < 0) {
3557 /* make sure -ERESTARTSYS -> -EINTR is done */
3561 /* read it all, or we did blocking attempt. no retry. */
3562 if (!iov_iter_count(iter) || !force_nonblock ||
3563 (req->file->f_flags & O_NONBLOCK) || !(req->flags & REQ_F_ISREG))
3568 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3575 rw = req->async_data;
3576 /* it's copied and will be cleaned with ->io */
3578 /* now use our persistent iterator, if we aren't already */
3581 rw->bytes_done += ret;
3582 /* if we can retry, do so with the callbacks armed */
3583 if (!io_rw_should_retry(req)) {
3584 kiocb->ki_flags &= ~IOCB_WAITQ;
3589 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3590 * get -EIOCBQUEUED, then we'll get a notification when the desired
3591 * page gets unlocked. We can also get a partial read here, and if we
3592 * do, then just retry at the new offset.
3594 ret = io_iter_do_read(req, iter);
3595 if (ret == -EIOCBQUEUED) {
3598 } else if (ret > 0 && ret < io_size) {
3599 /* we got some bytes, but not all. retry. */
3603 kiocb_done(kiocb, ret, cs);
3606 /* it's reportedly faster than delegating the null check to kfree() */
3612 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3616 ret = io_prep_rw(req, sqe);
3620 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3623 /* either don't need iovec imported or already have it */
3624 if (!req->async_data)
3626 return io_rw_prep_async(req, WRITE);
3629 static int io_write(struct io_kiocb *req, bool force_nonblock,
3630 struct io_comp_state *cs)
3632 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3633 struct kiocb *kiocb = &req->rw.kiocb;
3634 struct iov_iter __iter, *iter = &__iter;
3635 struct io_async_rw *rw = req->async_data;
3636 ssize_t ret, ret2, io_size;
3642 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3646 io_size = iov_iter_count(iter);
3647 req->result = io_size;
3649 /* Ensure we clear previously set non-block flag */
3650 if (!force_nonblock)
3651 kiocb->ki_flags &= ~IOCB_NOWAIT;
3653 kiocb->ki_flags |= IOCB_NOWAIT;
3655 /* If the file doesn't support async, just async punt */
3656 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3659 /* file path doesn't support NOWAIT for non-direct_IO */
3660 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3661 (req->flags & REQ_F_ISREG))
3664 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3669 * Open-code file_start_write here to grab freeze protection,
3670 * which will be released by another thread in
3671 * io_complete_rw(). Fool lockdep by telling it the lock got
3672 * released so that it doesn't complain about the held lock when
3673 * we return to userspace.
3675 if (req->flags & REQ_F_ISREG) {
3676 sb_start_write(file_inode(req->file)->i_sb);
3677 __sb_writers_release(file_inode(req->file)->i_sb,
3680 kiocb->ki_flags |= IOCB_WRITE;
3682 if (req->file->f_op->write_iter)
3683 ret2 = call_write_iter(req->file, kiocb, iter);
3684 else if (req->file->f_op->write)
3685 ret2 = loop_rw_iter(WRITE, req, iter);
3690 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3691 * retry them without IOCB_NOWAIT.
3693 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3695 /* no retry on NONBLOCK marked file */
3696 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3698 if (!force_nonblock || ret2 != -EAGAIN) {
3699 /* IOPOLL retry should happen for io-wq threads */
3700 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3703 kiocb_done(kiocb, ret2, cs);
3706 /* some cases will consume bytes even on error returns */
3707 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3708 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3713 /* it's reportedly faster than delegating the null check to kfree() */
3719 static int io_renameat_prep(struct io_kiocb *req,
3720 const struct io_uring_sqe *sqe)
3722 struct io_rename *ren = &req->rename;
3723 const char __user *oldf, *newf;
3725 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3728 ren->old_dfd = READ_ONCE(sqe->fd);
3729 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3730 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3731 ren->new_dfd = READ_ONCE(sqe->len);
3732 ren->flags = READ_ONCE(sqe->rename_flags);
3734 ren->oldpath = getname(oldf);
3735 if (IS_ERR(ren->oldpath))
3736 return PTR_ERR(ren->oldpath);
3738 ren->newpath = getname(newf);
3739 if (IS_ERR(ren->newpath)) {
3740 putname(ren->oldpath);
3741 return PTR_ERR(ren->newpath);
3744 req->flags |= REQ_F_NEED_CLEANUP;
3748 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3750 struct io_rename *ren = &req->rename;
3756 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3757 ren->newpath, ren->flags);
3759 req->flags &= ~REQ_F_NEED_CLEANUP;
3761 req_set_fail_links(req);
3762 io_req_complete(req, ret);
3766 static int io_unlinkat_prep(struct io_kiocb *req,
3767 const struct io_uring_sqe *sqe)
3769 struct io_unlink *un = &req->unlink;
3770 const char __user *fname;
3772 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3775 un->dfd = READ_ONCE(sqe->fd);
3777 un->flags = READ_ONCE(sqe->unlink_flags);
3778 if (un->flags & ~AT_REMOVEDIR)
3781 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3782 un->filename = getname(fname);
3783 if (IS_ERR(un->filename))
3784 return PTR_ERR(un->filename);
3786 req->flags |= REQ_F_NEED_CLEANUP;
3790 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3792 struct io_unlink *un = &req->unlink;
3798 if (un->flags & AT_REMOVEDIR)
3799 ret = do_rmdir(un->dfd, un->filename);
3801 ret = do_unlinkat(un->dfd, un->filename);
3803 req->flags &= ~REQ_F_NEED_CLEANUP;
3805 req_set_fail_links(req);
3806 io_req_complete(req, ret);
3810 static int io_shutdown_prep(struct io_kiocb *req,
3811 const struct io_uring_sqe *sqe)
3813 #if defined(CONFIG_NET)
3814 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3816 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3820 req->shutdown.how = READ_ONCE(sqe->len);
3827 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3829 #if defined(CONFIG_NET)
3830 struct socket *sock;
3836 sock = sock_from_file(req->file);
3837 if (unlikely(!sock))
3840 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3842 req_set_fail_links(req);
3843 io_req_complete(req, ret);
3850 static int __io_splice_prep(struct io_kiocb *req,
3851 const struct io_uring_sqe *sqe)
3853 struct io_splice* sp = &req->splice;
3854 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3856 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3860 sp->len = READ_ONCE(sqe->len);
3861 sp->flags = READ_ONCE(sqe->splice_flags);
3863 if (unlikely(sp->flags & ~valid_flags))
3866 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3867 (sp->flags & SPLICE_F_FD_IN_FIXED));
3870 req->flags |= REQ_F_NEED_CLEANUP;
3872 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3874 * Splice operation will be punted aync, and here need to
3875 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3877 io_req_init_async(req);
3878 req->work.flags |= IO_WQ_WORK_UNBOUND;
3884 static int io_tee_prep(struct io_kiocb *req,
3885 const struct io_uring_sqe *sqe)
3887 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3889 return __io_splice_prep(req, sqe);
3892 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3894 struct io_splice *sp = &req->splice;
3895 struct file *in = sp->file_in;
3896 struct file *out = sp->file_out;
3897 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3903 ret = do_tee(in, out, sp->len, flags);
3905 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3906 req->flags &= ~REQ_F_NEED_CLEANUP;
3909 req_set_fail_links(req);
3910 io_req_complete(req, ret);
3914 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3916 struct io_splice* sp = &req->splice;
3918 sp->off_in = READ_ONCE(sqe->splice_off_in);
3919 sp->off_out = READ_ONCE(sqe->off);
3920 return __io_splice_prep(req, sqe);
3923 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3925 struct io_splice *sp = &req->splice;
3926 struct file *in = sp->file_in;
3927 struct file *out = sp->file_out;
3928 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3929 loff_t *poff_in, *poff_out;
3935 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3936 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3939 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3941 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3942 req->flags &= ~REQ_F_NEED_CLEANUP;
3945 req_set_fail_links(req);
3946 io_req_complete(req, ret);
3951 * IORING_OP_NOP just posts a completion event, nothing else.
3953 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3955 struct io_ring_ctx *ctx = req->ctx;
3957 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3960 __io_req_complete(req, 0, 0, cs);
3964 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3966 struct io_ring_ctx *ctx = req->ctx;
3971 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3973 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3976 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3977 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3980 req->sync.off = READ_ONCE(sqe->off);
3981 req->sync.len = READ_ONCE(sqe->len);
3985 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3987 loff_t end = req->sync.off + req->sync.len;
3990 /* fsync always requires a blocking context */
3994 ret = vfs_fsync_range(req->file, req->sync.off,
3995 end > 0 ? end : LLONG_MAX,
3996 req->sync.flags & IORING_FSYNC_DATASYNC);
3998 req_set_fail_links(req);
3999 io_req_complete(req, ret);
4003 static int io_fallocate_prep(struct io_kiocb *req,
4004 const struct io_uring_sqe *sqe)
4006 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
4008 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4011 req->sync.off = READ_ONCE(sqe->off);
4012 req->sync.len = READ_ONCE(sqe->addr);
4013 req->sync.mode = READ_ONCE(sqe->len);
4017 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
4021 /* fallocate always requiring blocking context */
4024 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4027 req_set_fail_links(req);
4028 io_req_complete(req, ret);
4032 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4034 const char __user *fname;
4037 if (unlikely(sqe->ioprio || sqe->buf_index))
4039 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4042 /* open.how should be already initialised */
4043 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4044 req->open.how.flags |= O_LARGEFILE;
4046 req->open.dfd = READ_ONCE(sqe->fd);
4047 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4048 req->open.filename = getname(fname);
4049 if (IS_ERR(req->open.filename)) {
4050 ret = PTR_ERR(req->open.filename);
4051 req->open.filename = NULL;
4054 req->open.nofile = rlimit(RLIMIT_NOFILE);
4055 req->open.ignore_nonblock = false;
4056 req->flags |= REQ_F_NEED_CLEANUP;
4060 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4064 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4066 mode = READ_ONCE(sqe->len);
4067 flags = READ_ONCE(sqe->open_flags);
4068 req->open.how = build_open_how(flags, mode);
4069 return __io_openat_prep(req, sqe);
4072 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4074 struct open_how __user *how;
4078 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4080 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4081 len = READ_ONCE(sqe->len);
4082 if (len < OPEN_HOW_SIZE_VER0)
4085 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4090 return __io_openat_prep(req, sqe);
4093 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4095 struct open_flags op;
4099 if (force_nonblock && !req->open.ignore_nonblock)
4102 ret = build_open_flags(&req->open.how, &op);
4106 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4110 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4113 ret = PTR_ERR(file);
4115 * A work-around to ensure that /proc/self works that way
4116 * that it should - if we get -EOPNOTSUPP back, then assume
4117 * that proc_self_get_link() failed us because we're in async
4118 * context. We should be safe to retry this from the task
4119 * itself with force_nonblock == false set, as it should not
4120 * block on lookup. Would be nice to know this upfront and
4121 * avoid the async dance, but doesn't seem feasible.
4123 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4124 req->open.ignore_nonblock = true;
4125 refcount_inc(&req->refs);
4126 io_req_task_queue(req);
4130 fsnotify_open(file);
4131 fd_install(ret, file);
4134 putname(req->open.filename);
4135 req->flags &= ~REQ_F_NEED_CLEANUP;
4137 req_set_fail_links(req);
4138 io_req_complete(req, ret);
4142 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4144 return io_openat2(req, force_nonblock);
4147 static int io_remove_buffers_prep(struct io_kiocb *req,
4148 const struct io_uring_sqe *sqe)
4150 struct io_provide_buf *p = &req->pbuf;
4153 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4156 tmp = READ_ONCE(sqe->fd);
4157 if (!tmp || tmp > USHRT_MAX)
4160 memset(p, 0, sizeof(*p));
4162 p->bgid = READ_ONCE(sqe->buf_group);
4166 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4167 int bgid, unsigned nbufs)
4171 /* shouldn't happen */
4175 /* the head kbuf is the list itself */
4176 while (!list_empty(&buf->list)) {
4177 struct io_buffer *nxt;
4179 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4180 list_del(&nxt->list);
4187 idr_remove(&ctx->io_buffer_idr, bgid);
4192 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4193 struct io_comp_state *cs)
4195 struct io_provide_buf *p = &req->pbuf;
4196 struct io_ring_ctx *ctx = req->ctx;
4197 struct io_buffer *head;
4200 io_ring_submit_lock(ctx, !force_nonblock);
4202 lockdep_assert_held(&ctx->uring_lock);
4205 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4207 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4209 req_set_fail_links(req);
4211 /* need to hold the lock to complete IOPOLL requests */
4212 if (ctx->flags & IORING_SETUP_IOPOLL) {
4213 __io_req_complete(req, ret, 0, cs);
4214 io_ring_submit_unlock(ctx, !force_nonblock);
4216 io_ring_submit_unlock(ctx, !force_nonblock);
4217 __io_req_complete(req, ret, 0, cs);
4222 static int io_provide_buffers_prep(struct io_kiocb *req,
4223 const struct io_uring_sqe *sqe)
4225 struct io_provide_buf *p = &req->pbuf;
4228 if (sqe->ioprio || sqe->rw_flags)
4231 tmp = READ_ONCE(sqe->fd);
4232 if (!tmp || tmp > USHRT_MAX)
4235 p->addr = READ_ONCE(sqe->addr);
4236 p->len = READ_ONCE(sqe->len);
4238 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4241 p->bgid = READ_ONCE(sqe->buf_group);
4242 tmp = READ_ONCE(sqe->off);
4243 if (tmp > USHRT_MAX)
4249 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4251 struct io_buffer *buf;
4252 u64 addr = pbuf->addr;
4253 int i, bid = pbuf->bid;
4255 for (i = 0; i < pbuf->nbufs; i++) {
4256 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4261 buf->len = pbuf->len;
4266 INIT_LIST_HEAD(&buf->list);
4269 list_add_tail(&buf->list, &(*head)->list);
4273 return i ? i : -ENOMEM;
4276 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4277 struct io_comp_state *cs)
4279 struct io_provide_buf *p = &req->pbuf;
4280 struct io_ring_ctx *ctx = req->ctx;
4281 struct io_buffer *head, *list;
4284 io_ring_submit_lock(ctx, !force_nonblock);
4286 lockdep_assert_held(&ctx->uring_lock);
4288 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4290 ret = io_add_buffers(p, &head);
4295 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4298 __io_remove_buffers(ctx, head, p->bgid, -1U);
4304 req_set_fail_links(req);
4306 /* need to hold the lock to complete IOPOLL requests */
4307 if (ctx->flags & IORING_SETUP_IOPOLL) {
4308 __io_req_complete(req, ret, 0, cs);
4309 io_ring_submit_unlock(ctx, !force_nonblock);
4311 io_ring_submit_unlock(ctx, !force_nonblock);
4312 __io_req_complete(req, ret, 0, cs);
4317 static int io_epoll_ctl_prep(struct io_kiocb *req,
4318 const struct io_uring_sqe *sqe)
4320 #if defined(CONFIG_EPOLL)
4321 if (sqe->ioprio || sqe->buf_index)
4323 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4326 req->epoll.epfd = READ_ONCE(sqe->fd);
4327 req->epoll.op = READ_ONCE(sqe->len);
4328 req->epoll.fd = READ_ONCE(sqe->off);
4330 if (ep_op_has_event(req->epoll.op)) {
4331 struct epoll_event __user *ev;
4333 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4334 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4344 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4345 struct io_comp_state *cs)
4347 #if defined(CONFIG_EPOLL)
4348 struct io_epoll *ie = &req->epoll;
4351 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4352 if (force_nonblock && ret == -EAGAIN)
4356 req_set_fail_links(req);
4357 __io_req_complete(req, ret, 0, cs);
4364 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4366 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4367 if (sqe->ioprio || sqe->buf_index || sqe->off)
4369 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4372 req->madvise.addr = READ_ONCE(sqe->addr);
4373 req->madvise.len = READ_ONCE(sqe->len);
4374 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4381 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4383 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4384 struct io_madvise *ma = &req->madvise;
4390 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4392 req_set_fail_links(req);
4393 io_req_complete(req, ret);
4400 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4402 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4404 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4407 req->fadvise.offset = READ_ONCE(sqe->off);
4408 req->fadvise.len = READ_ONCE(sqe->len);
4409 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4413 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4415 struct io_fadvise *fa = &req->fadvise;
4418 if (force_nonblock) {
4419 switch (fa->advice) {
4420 case POSIX_FADV_NORMAL:
4421 case POSIX_FADV_RANDOM:
4422 case POSIX_FADV_SEQUENTIAL:
4429 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4431 req_set_fail_links(req);
4432 io_req_complete(req, ret);
4436 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4438 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4440 if (sqe->ioprio || sqe->buf_index)
4442 if (req->flags & REQ_F_FIXED_FILE)
4445 req->statx.dfd = READ_ONCE(sqe->fd);
4446 req->statx.mask = READ_ONCE(sqe->len);
4447 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4448 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4449 req->statx.flags = READ_ONCE(sqe->statx_flags);
4454 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4456 struct io_statx *ctx = &req->statx;
4459 if (force_nonblock) {
4460 /* only need file table for an actual valid fd */
4461 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4462 req->flags |= REQ_F_NO_FILE_TABLE;
4466 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4470 req_set_fail_links(req);
4471 io_req_complete(req, ret);
4475 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4478 * If we queue this for async, it must not be cancellable. That would
4479 * leave the 'file' in an undeterminate state, and here need to modify
4480 * io_wq_work.flags, so initialize io_wq_work firstly.
4482 io_req_init_async(req);
4484 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4486 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4487 sqe->rw_flags || sqe->buf_index)
4489 if (req->flags & REQ_F_FIXED_FILE)
4492 req->close.fd = READ_ONCE(sqe->fd);
4493 if ((req->file && req->file->f_op == &io_uring_fops))
4496 req->close.put_file = NULL;
4500 static int io_close(struct io_kiocb *req, bool force_nonblock,
4501 struct io_comp_state *cs)
4503 struct io_close *close = &req->close;
4506 /* might be already done during nonblock submission */
4507 if (!close->put_file) {
4508 ret = close_fd_get_file(close->fd, &close->put_file);
4510 return (ret == -ENOENT) ? -EBADF : ret;
4513 /* if the file has a flush method, be safe and punt to async */
4514 if (close->put_file->f_op->flush && force_nonblock) {
4515 /* not safe to cancel at this point */
4516 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4517 /* was never set, but play safe */
4518 req->flags &= ~REQ_F_NOWAIT;
4519 /* avoid grabbing files - we don't need the files */
4520 req->flags |= REQ_F_NO_FILE_TABLE;
4524 /* No ->flush() or already async, safely close from here */
4525 ret = filp_close(close->put_file, req->work.identity->files);
4527 req_set_fail_links(req);
4528 fput(close->put_file);
4529 close->put_file = NULL;
4530 __io_req_complete(req, ret, 0, cs);
4534 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4536 struct io_ring_ctx *ctx = req->ctx;
4541 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4543 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4546 req->sync.off = READ_ONCE(sqe->off);
4547 req->sync.len = READ_ONCE(sqe->len);
4548 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4552 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4556 /* sync_file_range always requires a blocking context */
4560 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4563 req_set_fail_links(req);
4564 io_req_complete(req, ret);
4568 #if defined(CONFIG_NET)
4569 static int io_setup_async_msg(struct io_kiocb *req,
4570 struct io_async_msghdr *kmsg)
4572 struct io_async_msghdr *async_msg = req->async_data;
4576 if (io_alloc_async_data(req)) {
4577 if (kmsg->iov != kmsg->fast_iov)
4581 async_msg = req->async_data;
4582 req->flags |= REQ_F_NEED_CLEANUP;
4583 memcpy(async_msg, kmsg, sizeof(*kmsg));
4587 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4588 struct io_async_msghdr *iomsg)
4590 iomsg->iov = iomsg->fast_iov;
4591 iomsg->msg.msg_name = &iomsg->addr;
4592 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4593 req->sr_msg.msg_flags, &iomsg->iov);
4596 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4598 struct io_async_msghdr *async_msg = req->async_data;
4599 struct io_sr_msg *sr = &req->sr_msg;
4602 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4605 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4606 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4607 sr->len = READ_ONCE(sqe->len);
4609 #ifdef CONFIG_COMPAT
4610 if (req->ctx->compat)
4611 sr->msg_flags |= MSG_CMSG_COMPAT;
4614 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4616 ret = io_sendmsg_copy_hdr(req, async_msg);
4618 req->flags |= REQ_F_NEED_CLEANUP;
4622 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4623 struct io_comp_state *cs)
4625 struct io_async_msghdr iomsg, *kmsg;
4626 struct socket *sock;
4630 sock = sock_from_file(req->file);
4631 if (unlikely(!sock))
4634 if (req->async_data) {
4635 kmsg = req->async_data;
4636 kmsg->msg.msg_name = &kmsg->addr;
4637 /* if iov is set, it's allocated already */
4639 kmsg->iov = kmsg->fast_iov;
4640 kmsg->msg.msg_iter.iov = kmsg->iov;
4642 ret = io_sendmsg_copy_hdr(req, &iomsg);
4648 flags = req->sr_msg.msg_flags;
4649 if (flags & MSG_DONTWAIT)
4650 req->flags |= REQ_F_NOWAIT;
4651 else if (force_nonblock)
4652 flags |= MSG_DONTWAIT;
4654 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4655 if (force_nonblock && ret == -EAGAIN)
4656 return io_setup_async_msg(req, kmsg);
4657 if (ret == -ERESTARTSYS)
4660 if (kmsg->iov != kmsg->fast_iov)
4662 req->flags &= ~REQ_F_NEED_CLEANUP;
4664 req_set_fail_links(req);
4665 __io_req_complete(req, ret, 0, cs);
4669 static int io_send(struct io_kiocb *req, bool force_nonblock,
4670 struct io_comp_state *cs)
4672 struct io_sr_msg *sr = &req->sr_msg;
4675 struct socket *sock;
4679 sock = sock_from_file(req->file);
4680 if (unlikely(!sock))
4683 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4687 msg.msg_name = NULL;
4688 msg.msg_control = NULL;
4689 msg.msg_controllen = 0;
4690 msg.msg_namelen = 0;
4692 flags = req->sr_msg.msg_flags;
4693 if (flags & MSG_DONTWAIT)
4694 req->flags |= REQ_F_NOWAIT;
4695 else if (force_nonblock)
4696 flags |= MSG_DONTWAIT;
4698 msg.msg_flags = flags;
4699 ret = sock_sendmsg(sock, &msg);
4700 if (force_nonblock && ret == -EAGAIN)
4702 if (ret == -ERESTARTSYS)
4706 req_set_fail_links(req);
4707 __io_req_complete(req, ret, 0, cs);
4711 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4712 struct io_async_msghdr *iomsg)
4714 struct io_sr_msg *sr = &req->sr_msg;
4715 struct iovec __user *uiov;
4719 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4720 &iomsg->uaddr, &uiov, &iov_len);
4724 if (req->flags & REQ_F_BUFFER_SELECT) {
4727 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4729 sr->len = iomsg->iov[0].iov_len;
4730 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4734 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4735 &iomsg->iov, &iomsg->msg.msg_iter,
4744 #ifdef CONFIG_COMPAT
4745 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4746 struct io_async_msghdr *iomsg)
4748 struct compat_msghdr __user *msg_compat;
4749 struct io_sr_msg *sr = &req->sr_msg;
4750 struct compat_iovec __user *uiov;
4755 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4756 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4761 uiov = compat_ptr(ptr);
4762 if (req->flags & REQ_F_BUFFER_SELECT) {
4763 compat_ssize_t clen;
4767 if (!access_ok(uiov, sizeof(*uiov)))
4769 if (__get_user(clen, &uiov->iov_len))
4774 iomsg->iov[0].iov_len = clen;
4777 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4778 UIO_FASTIOV, &iomsg->iov,
4779 &iomsg->msg.msg_iter, true);
4788 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4789 struct io_async_msghdr *iomsg)
4791 iomsg->msg.msg_name = &iomsg->addr;
4792 iomsg->iov = iomsg->fast_iov;
4794 #ifdef CONFIG_COMPAT
4795 if (req->ctx->compat)
4796 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4799 return __io_recvmsg_copy_hdr(req, iomsg);
4802 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4805 struct io_sr_msg *sr = &req->sr_msg;
4806 struct io_buffer *kbuf;
4808 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4813 req->flags |= REQ_F_BUFFER_SELECTED;
4817 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4819 return io_put_kbuf(req, req->sr_msg.kbuf);
4822 static int io_recvmsg_prep(struct io_kiocb *req,
4823 const struct io_uring_sqe *sqe)
4825 struct io_async_msghdr *async_msg = req->async_data;
4826 struct io_sr_msg *sr = &req->sr_msg;
4829 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4832 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4833 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4834 sr->len = READ_ONCE(sqe->len);
4835 sr->bgid = READ_ONCE(sqe->buf_group);
4837 #ifdef CONFIG_COMPAT
4838 if (req->ctx->compat)
4839 sr->msg_flags |= MSG_CMSG_COMPAT;
4842 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4844 ret = io_recvmsg_copy_hdr(req, async_msg);
4846 req->flags |= REQ_F_NEED_CLEANUP;
4850 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4851 struct io_comp_state *cs)
4853 struct io_async_msghdr iomsg, *kmsg;
4854 struct socket *sock;
4855 struct io_buffer *kbuf;
4857 int ret, cflags = 0;
4859 sock = sock_from_file(req->file);
4860 if (unlikely(!sock))
4863 if (req->async_data) {
4864 kmsg = req->async_data;
4865 kmsg->msg.msg_name = &kmsg->addr;
4866 /* if iov is set, it's allocated already */
4868 kmsg->iov = kmsg->fast_iov;
4869 kmsg->msg.msg_iter.iov = kmsg->iov;
4871 ret = io_recvmsg_copy_hdr(req, &iomsg);
4877 if (req->flags & REQ_F_BUFFER_SELECT) {
4878 kbuf = io_recv_buffer_select(req, !force_nonblock);
4880 return PTR_ERR(kbuf);
4881 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4882 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4883 1, req->sr_msg.len);
4886 flags = req->sr_msg.msg_flags;
4887 if (flags & MSG_DONTWAIT)
4888 req->flags |= REQ_F_NOWAIT;
4889 else if (force_nonblock)
4890 flags |= MSG_DONTWAIT;
4892 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4893 kmsg->uaddr, flags);
4894 if (force_nonblock && ret == -EAGAIN)
4895 return io_setup_async_msg(req, kmsg);
4896 if (ret == -ERESTARTSYS)
4899 if (req->flags & REQ_F_BUFFER_SELECTED)
4900 cflags = io_put_recv_kbuf(req);
4901 if (kmsg->iov != kmsg->fast_iov)
4903 req->flags &= ~REQ_F_NEED_CLEANUP;
4905 req_set_fail_links(req);
4906 __io_req_complete(req, ret, cflags, cs);
4910 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4911 struct io_comp_state *cs)
4913 struct io_buffer *kbuf;
4914 struct io_sr_msg *sr = &req->sr_msg;
4916 void __user *buf = sr->buf;
4917 struct socket *sock;
4920 int ret, cflags = 0;
4922 sock = sock_from_file(req->file);
4923 if (unlikely(!sock))
4926 if (req->flags & REQ_F_BUFFER_SELECT) {
4927 kbuf = io_recv_buffer_select(req, !force_nonblock);
4929 return PTR_ERR(kbuf);
4930 buf = u64_to_user_ptr(kbuf->addr);
4933 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4937 msg.msg_name = NULL;
4938 msg.msg_control = NULL;
4939 msg.msg_controllen = 0;
4940 msg.msg_namelen = 0;
4941 msg.msg_iocb = NULL;
4944 flags = req->sr_msg.msg_flags;
4945 if (flags & MSG_DONTWAIT)
4946 req->flags |= REQ_F_NOWAIT;
4947 else if (force_nonblock)
4948 flags |= MSG_DONTWAIT;
4950 ret = sock_recvmsg(sock, &msg, flags);
4951 if (force_nonblock && ret == -EAGAIN)
4953 if (ret == -ERESTARTSYS)
4956 if (req->flags & REQ_F_BUFFER_SELECTED)
4957 cflags = io_put_recv_kbuf(req);
4959 req_set_fail_links(req);
4960 __io_req_complete(req, ret, cflags, cs);
4964 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4966 struct io_accept *accept = &req->accept;
4968 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4970 if (sqe->ioprio || sqe->len || sqe->buf_index)
4973 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4974 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4975 accept->flags = READ_ONCE(sqe->accept_flags);
4976 accept->nofile = rlimit(RLIMIT_NOFILE);
4980 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4981 struct io_comp_state *cs)
4983 struct io_accept *accept = &req->accept;
4984 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4987 if (req->file->f_flags & O_NONBLOCK)
4988 req->flags |= REQ_F_NOWAIT;
4990 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4991 accept->addr_len, accept->flags,
4993 if (ret == -EAGAIN && force_nonblock)
4996 if (ret == -ERESTARTSYS)
4998 req_set_fail_links(req);
5000 __io_req_complete(req, ret, 0, cs);
5004 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5006 struct io_connect *conn = &req->connect;
5007 struct io_async_connect *io = req->async_data;
5009 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5011 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
5014 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5015 conn->addr_len = READ_ONCE(sqe->addr2);
5020 return move_addr_to_kernel(conn->addr, conn->addr_len,
5024 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5025 struct io_comp_state *cs)
5027 struct io_async_connect __io, *io;
5028 unsigned file_flags;
5031 if (req->async_data) {
5032 io = req->async_data;
5034 ret = move_addr_to_kernel(req->connect.addr,
5035 req->connect.addr_len,
5042 file_flags = force_nonblock ? O_NONBLOCK : 0;
5044 ret = __sys_connect_file(req->file, &io->address,
5045 req->connect.addr_len, file_flags);
5046 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5047 if (req->async_data)
5049 if (io_alloc_async_data(req)) {
5053 io = req->async_data;
5054 memcpy(req->async_data, &__io, sizeof(__io));
5057 if (ret == -ERESTARTSYS)
5061 req_set_fail_links(req);
5062 __io_req_complete(req, ret, 0, cs);
5065 #else /* !CONFIG_NET */
5066 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5071 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5072 struct io_comp_state *cs)
5077 static int io_send(struct io_kiocb *req, bool force_nonblock,
5078 struct io_comp_state *cs)
5083 static int io_recvmsg_prep(struct io_kiocb *req,
5084 const struct io_uring_sqe *sqe)
5089 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5090 struct io_comp_state *cs)
5095 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5096 struct io_comp_state *cs)
5101 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5106 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5107 struct io_comp_state *cs)
5112 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5117 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5118 struct io_comp_state *cs)
5122 #endif /* CONFIG_NET */
5124 struct io_poll_table {
5125 struct poll_table_struct pt;
5126 struct io_kiocb *req;
5130 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5131 __poll_t mask, task_work_func_t func)
5135 /* for instances that support it check for an event match first: */
5136 if (mask && !(mask & poll->events))
5139 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5141 list_del_init(&poll->wait.entry);
5144 init_task_work(&req->task_work, func);
5145 percpu_ref_get(&req->ctx->refs);
5148 * If this fails, then the task is exiting. When a task exits, the
5149 * work gets canceled, so just cancel this request as well instead
5150 * of executing it. We can't safely execute it anyway, as we may not
5151 * have the needed state needed for it anyway.
5153 ret = io_req_task_work_add(req);
5154 if (unlikely(ret)) {
5155 struct task_struct *tsk;
5157 WRITE_ONCE(poll->canceled, true);
5158 tsk = io_wq_get_task(req->ctx->io_wq);
5159 task_work_add(tsk, &req->task_work, TWA_NONE);
5160 wake_up_process(tsk);
5165 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5166 __acquires(&req->ctx->completion_lock)
5168 struct io_ring_ctx *ctx = req->ctx;
5170 if (!req->result && !READ_ONCE(poll->canceled)) {
5171 struct poll_table_struct pt = { ._key = poll->events };
5173 req->result = vfs_poll(req->file, &pt) & poll->events;
5176 spin_lock_irq(&ctx->completion_lock);
5177 if (!req->result && !READ_ONCE(poll->canceled)) {
5178 add_wait_queue(poll->head, &poll->wait);
5185 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5187 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5188 if (req->opcode == IORING_OP_POLL_ADD)
5189 return req->async_data;
5190 return req->apoll->double_poll;
5193 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5195 if (req->opcode == IORING_OP_POLL_ADD)
5197 return &req->apoll->poll;
5200 static void io_poll_remove_double(struct io_kiocb *req)
5202 struct io_poll_iocb *poll = io_poll_get_double(req);
5204 lockdep_assert_held(&req->ctx->completion_lock);
5206 if (poll && poll->head) {
5207 struct wait_queue_head *head = poll->head;
5209 spin_lock(&head->lock);
5210 list_del_init(&poll->wait.entry);
5211 if (poll->wait.private)
5212 refcount_dec(&req->refs);
5214 spin_unlock(&head->lock);
5218 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5220 struct io_ring_ctx *ctx = req->ctx;
5222 io_poll_remove_double(req);
5223 req->poll.done = true;
5224 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5225 io_commit_cqring(ctx);
5228 static void io_poll_task_func(struct callback_head *cb)
5230 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5231 struct io_ring_ctx *ctx = req->ctx;
5232 struct io_kiocb *nxt;
5234 if (io_poll_rewait(req, &req->poll)) {
5235 spin_unlock_irq(&ctx->completion_lock);
5237 hash_del(&req->hash_node);
5238 io_poll_complete(req, req->result, 0);
5239 spin_unlock_irq(&ctx->completion_lock);
5241 nxt = io_put_req_find_next(req);
5242 io_cqring_ev_posted(ctx);
5244 __io_req_task_submit(nxt);
5247 percpu_ref_put(&ctx->refs);
5250 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5251 int sync, void *key)
5253 struct io_kiocb *req = wait->private;
5254 struct io_poll_iocb *poll = io_poll_get_single(req);
5255 __poll_t mask = key_to_poll(key);
5257 /* for instances that support it check for an event match first: */
5258 if (mask && !(mask & poll->events))
5261 list_del_init(&wait->entry);
5263 if (poll && poll->head) {
5266 spin_lock(&poll->head->lock);
5267 done = list_empty(&poll->wait.entry);
5269 list_del_init(&poll->wait.entry);
5270 /* make sure double remove sees this as being gone */
5271 wait->private = NULL;
5272 spin_unlock(&poll->head->lock);
5274 /* use wait func handler, so it matches the rq type */
5275 poll->wait.func(&poll->wait, mode, sync, key);
5278 refcount_dec(&req->refs);
5282 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5283 wait_queue_func_t wake_func)
5287 poll->canceled = false;
5288 poll->events = events;
5289 INIT_LIST_HEAD(&poll->wait.entry);
5290 init_waitqueue_func_entry(&poll->wait, wake_func);
5293 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5294 struct wait_queue_head *head,
5295 struct io_poll_iocb **poll_ptr)
5297 struct io_kiocb *req = pt->req;
5300 * If poll->head is already set, it's because the file being polled
5301 * uses multiple waitqueues for poll handling (eg one for read, one
5302 * for write). Setup a separate io_poll_iocb if this happens.
5304 if (unlikely(poll->head)) {
5305 struct io_poll_iocb *poll_one = poll;
5307 /* already have a 2nd entry, fail a third attempt */
5309 pt->error = -EINVAL;
5312 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5314 pt->error = -ENOMEM;
5317 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5318 refcount_inc(&req->refs);
5319 poll->wait.private = req;
5326 if (poll->events & EPOLLEXCLUSIVE)
5327 add_wait_queue_exclusive(head, &poll->wait);
5329 add_wait_queue(head, &poll->wait);
5332 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5333 struct poll_table_struct *p)
5335 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5336 struct async_poll *apoll = pt->req->apoll;
5338 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5341 static void io_async_task_func(struct callback_head *cb)
5343 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5344 struct async_poll *apoll = req->apoll;
5345 struct io_ring_ctx *ctx = req->ctx;
5347 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5349 if (io_poll_rewait(req, &apoll->poll)) {
5350 spin_unlock_irq(&ctx->completion_lock);
5351 percpu_ref_put(&ctx->refs);
5355 /* If req is still hashed, it cannot have been canceled. Don't check. */
5356 if (hash_hashed(&req->hash_node))
5357 hash_del(&req->hash_node);
5359 io_poll_remove_double(req);
5360 spin_unlock_irq(&ctx->completion_lock);
5362 if (!READ_ONCE(apoll->poll.canceled))
5363 __io_req_task_submit(req);
5365 __io_req_task_cancel(req, -ECANCELED);
5367 percpu_ref_put(&ctx->refs);
5368 kfree(apoll->double_poll);
5372 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5375 struct io_kiocb *req = wait->private;
5376 struct io_poll_iocb *poll = &req->apoll->poll;
5378 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5381 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5384 static void io_poll_req_insert(struct io_kiocb *req)
5386 struct io_ring_ctx *ctx = req->ctx;
5387 struct hlist_head *list;
5389 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5390 hlist_add_head(&req->hash_node, list);
5393 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5394 struct io_poll_iocb *poll,
5395 struct io_poll_table *ipt, __poll_t mask,
5396 wait_queue_func_t wake_func)
5397 __acquires(&ctx->completion_lock)
5399 struct io_ring_ctx *ctx = req->ctx;
5400 bool cancel = false;
5402 INIT_HLIST_NODE(&req->hash_node);
5403 io_init_poll_iocb(poll, mask, wake_func);
5404 poll->file = req->file;
5405 poll->wait.private = req;
5407 ipt->pt._key = mask;
5409 ipt->error = -EINVAL;
5411 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5413 spin_lock_irq(&ctx->completion_lock);
5414 if (likely(poll->head)) {
5415 spin_lock(&poll->head->lock);
5416 if (unlikely(list_empty(&poll->wait.entry))) {
5422 if (mask || ipt->error)
5423 list_del_init(&poll->wait.entry);
5425 WRITE_ONCE(poll->canceled, true);
5426 else if (!poll->done) /* actually waiting for an event */
5427 io_poll_req_insert(req);
5428 spin_unlock(&poll->head->lock);
5434 static bool io_arm_poll_handler(struct io_kiocb *req)
5436 const struct io_op_def *def = &io_op_defs[req->opcode];
5437 struct io_ring_ctx *ctx = req->ctx;
5438 struct async_poll *apoll;
5439 struct io_poll_table ipt;
5443 if (!req->file || !file_can_poll(req->file))
5445 if (req->flags & REQ_F_POLLED)
5449 else if (def->pollout)
5453 /* if we can't nonblock try, then no point in arming a poll handler */
5454 if (!io_file_supports_async(req->file, rw))
5457 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5458 if (unlikely(!apoll))
5460 apoll->double_poll = NULL;
5462 req->flags |= REQ_F_POLLED;
5467 mask |= POLLIN | POLLRDNORM;
5469 mask |= POLLOUT | POLLWRNORM;
5471 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5472 if ((req->opcode == IORING_OP_RECVMSG) &&
5473 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5476 mask |= POLLERR | POLLPRI;
5478 ipt.pt._qproc = io_async_queue_proc;
5480 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5482 if (ret || ipt.error) {
5483 io_poll_remove_double(req);
5484 spin_unlock_irq(&ctx->completion_lock);
5485 kfree(apoll->double_poll);
5489 spin_unlock_irq(&ctx->completion_lock);
5490 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5491 apoll->poll.events);
5495 static bool __io_poll_remove_one(struct io_kiocb *req,
5496 struct io_poll_iocb *poll)
5498 bool do_complete = false;
5500 spin_lock(&poll->head->lock);
5501 WRITE_ONCE(poll->canceled, true);
5502 if (!list_empty(&poll->wait.entry)) {
5503 list_del_init(&poll->wait.entry);
5506 spin_unlock(&poll->head->lock);
5507 hash_del(&req->hash_node);
5511 static bool io_poll_remove_one(struct io_kiocb *req)
5515 io_poll_remove_double(req);
5517 if (req->opcode == IORING_OP_POLL_ADD) {
5518 do_complete = __io_poll_remove_one(req, &req->poll);
5520 struct async_poll *apoll = req->apoll;
5522 /* non-poll requests have submit ref still */
5523 do_complete = __io_poll_remove_one(req, &apoll->poll);
5526 kfree(apoll->double_poll);
5532 io_cqring_fill_event(req, -ECANCELED);
5533 io_commit_cqring(req->ctx);
5534 req_set_fail_links(req);
5535 io_put_req_deferred(req, 1);
5542 * Returns true if we found and killed one or more poll requests
5544 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5545 struct files_struct *files)
5547 struct hlist_node *tmp;
5548 struct io_kiocb *req;
5551 spin_lock_irq(&ctx->completion_lock);
5552 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5553 struct hlist_head *list;
5555 list = &ctx->cancel_hash[i];
5556 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5557 if (io_match_task(req, tsk, files))
5558 posted += io_poll_remove_one(req);
5561 spin_unlock_irq(&ctx->completion_lock);
5564 io_cqring_ev_posted(ctx);
5569 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5571 struct hlist_head *list;
5572 struct io_kiocb *req;
5574 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5575 hlist_for_each_entry(req, list, hash_node) {
5576 if (sqe_addr != req->user_data)
5578 if (io_poll_remove_one(req))
5586 static int io_poll_remove_prep(struct io_kiocb *req,
5587 const struct io_uring_sqe *sqe)
5589 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5591 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5595 req->poll_remove.addr = READ_ONCE(sqe->addr);
5600 * Find a running poll command that matches one specified in sqe->addr,
5601 * and remove it if found.
5603 static int io_poll_remove(struct io_kiocb *req)
5605 struct io_ring_ctx *ctx = req->ctx;
5608 spin_lock_irq(&ctx->completion_lock);
5609 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5610 spin_unlock_irq(&ctx->completion_lock);
5613 req_set_fail_links(req);
5614 io_req_complete(req, ret);
5618 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5621 struct io_kiocb *req = wait->private;
5622 struct io_poll_iocb *poll = &req->poll;
5624 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5627 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5628 struct poll_table_struct *p)
5630 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5632 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5635 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5637 struct io_poll_iocb *poll = &req->poll;
5640 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5642 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5645 events = READ_ONCE(sqe->poll32_events);
5647 events = swahw32(events);
5649 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5650 (events & EPOLLEXCLUSIVE);
5654 static int io_poll_add(struct io_kiocb *req)
5656 struct io_poll_iocb *poll = &req->poll;
5657 struct io_ring_ctx *ctx = req->ctx;
5658 struct io_poll_table ipt;
5661 ipt.pt._qproc = io_poll_queue_proc;
5663 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5666 if (mask) { /* no async, we'd stolen it */
5668 io_poll_complete(req, mask, 0);
5670 spin_unlock_irq(&ctx->completion_lock);
5673 io_cqring_ev_posted(ctx);
5679 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5681 struct io_timeout_data *data = container_of(timer,
5682 struct io_timeout_data, timer);
5683 struct io_kiocb *req = data->req;
5684 struct io_ring_ctx *ctx = req->ctx;
5685 unsigned long flags;
5687 spin_lock_irqsave(&ctx->completion_lock, flags);
5688 list_del_init(&req->timeout.list);
5689 atomic_set(&req->ctx->cq_timeouts,
5690 atomic_read(&req->ctx->cq_timeouts) + 1);
5692 io_cqring_fill_event(req, -ETIME);
5693 io_commit_cqring(ctx);
5694 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5696 io_cqring_ev_posted(ctx);
5697 req_set_fail_links(req);
5699 return HRTIMER_NORESTART;
5702 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5705 struct io_timeout_data *io;
5706 struct io_kiocb *req;
5709 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5710 if (user_data == req->user_data) {
5717 return ERR_PTR(ret);
5719 io = req->async_data;
5720 ret = hrtimer_try_to_cancel(&io->timer);
5722 return ERR_PTR(-EALREADY);
5723 list_del_init(&req->timeout.list);
5727 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5729 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5732 return PTR_ERR(req);
5734 req_set_fail_links(req);
5735 io_cqring_fill_event(req, -ECANCELED);
5736 io_put_req_deferred(req, 1);
5740 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5741 struct timespec64 *ts, enum hrtimer_mode mode)
5743 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5744 struct io_timeout_data *data;
5747 return PTR_ERR(req);
5749 req->timeout.off = 0; /* noseq */
5750 data = req->async_data;
5751 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5752 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5753 data->timer.function = io_timeout_fn;
5754 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5758 static int io_timeout_remove_prep(struct io_kiocb *req,
5759 const struct io_uring_sqe *sqe)
5761 struct io_timeout_rem *tr = &req->timeout_rem;
5763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5765 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5767 if (sqe->ioprio || sqe->buf_index || sqe->len)
5770 tr->addr = READ_ONCE(sqe->addr);
5771 tr->flags = READ_ONCE(sqe->timeout_flags);
5772 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5773 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5775 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5777 } else if (tr->flags) {
5778 /* timeout removal doesn't support flags */
5786 * Remove or update an existing timeout command
5788 static int io_timeout_remove(struct io_kiocb *req)
5790 struct io_timeout_rem *tr = &req->timeout_rem;
5791 struct io_ring_ctx *ctx = req->ctx;
5794 spin_lock_irq(&ctx->completion_lock);
5795 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5796 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5797 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5799 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5801 ret = io_timeout_cancel(ctx, tr->addr);
5804 io_cqring_fill_event(req, ret);
5805 io_commit_cqring(ctx);
5806 spin_unlock_irq(&ctx->completion_lock);
5807 io_cqring_ev_posted(ctx);
5809 req_set_fail_links(req);
5814 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5815 bool is_timeout_link)
5817 struct io_timeout_data *data;
5819 u32 off = READ_ONCE(sqe->off);
5821 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5823 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5825 if (off && is_timeout_link)
5827 flags = READ_ONCE(sqe->timeout_flags);
5828 if (flags & ~IORING_TIMEOUT_ABS)
5831 req->timeout.off = off;
5833 if (!req->async_data && io_alloc_async_data(req))
5836 data = req->async_data;
5839 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5842 if (flags & IORING_TIMEOUT_ABS)
5843 data->mode = HRTIMER_MODE_ABS;
5845 data->mode = HRTIMER_MODE_REL;
5847 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5851 static int io_timeout(struct io_kiocb *req)
5853 struct io_ring_ctx *ctx = req->ctx;
5854 struct io_timeout_data *data = req->async_data;
5855 struct list_head *entry;
5856 u32 tail, off = req->timeout.off;
5858 spin_lock_irq(&ctx->completion_lock);
5861 * sqe->off holds how many events that need to occur for this
5862 * timeout event to be satisfied. If it isn't set, then this is
5863 * a pure timeout request, sequence isn't used.
5865 if (io_is_timeout_noseq(req)) {
5866 entry = ctx->timeout_list.prev;
5870 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5871 req->timeout.target_seq = tail + off;
5873 /* Update the last seq here in case io_flush_timeouts() hasn't.
5874 * This is safe because ->completion_lock is held, and submissions
5875 * and completions are never mixed in the same ->completion_lock section.
5877 ctx->cq_last_tm_flush = tail;
5880 * Insertion sort, ensuring the first entry in the list is always
5881 * the one we need first.
5883 list_for_each_prev(entry, &ctx->timeout_list) {
5884 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5887 if (io_is_timeout_noseq(nxt))
5889 /* nxt.seq is behind @tail, otherwise would've been completed */
5890 if (off >= nxt->timeout.target_seq - tail)
5894 list_add(&req->timeout.list, entry);
5895 data->timer.function = io_timeout_fn;
5896 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5897 spin_unlock_irq(&ctx->completion_lock);
5901 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5903 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5905 return req->user_data == (unsigned long) data;
5908 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5910 enum io_wq_cancel cancel_ret;
5913 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5914 switch (cancel_ret) {
5915 case IO_WQ_CANCEL_OK:
5918 case IO_WQ_CANCEL_RUNNING:
5921 case IO_WQ_CANCEL_NOTFOUND:
5929 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5930 struct io_kiocb *req, __u64 sqe_addr,
5933 unsigned long flags;
5936 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5937 if (ret != -ENOENT) {
5938 spin_lock_irqsave(&ctx->completion_lock, flags);
5942 spin_lock_irqsave(&ctx->completion_lock, flags);
5943 ret = io_timeout_cancel(ctx, sqe_addr);
5946 ret = io_poll_cancel(ctx, sqe_addr);
5950 io_cqring_fill_event(req, ret);
5951 io_commit_cqring(ctx);
5952 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5953 io_cqring_ev_posted(ctx);
5956 req_set_fail_links(req);
5960 static int io_async_cancel_prep(struct io_kiocb *req,
5961 const struct io_uring_sqe *sqe)
5963 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5965 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5967 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5970 req->cancel.addr = READ_ONCE(sqe->addr);
5974 static int io_async_cancel(struct io_kiocb *req)
5976 struct io_ring_ctx *ctx = req->ctx;
5978 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5982 static int io_files_update_prep(struct io_kiocb *req,
5983 const struct io_uring_sqe *sqe)
5985 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5987 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5989 if (sqe->ioprio || sqe->rw_flags)
5992 req->files_update.offset = READ_ONCE(sqe->off);
5993 req->files_update.nr_args = READ_ONCE(sqe->len);
5994 if (!req->files_update.nr_args)
5996 req->files_update.arg = READ_ONCE(sqe->addr);
6000 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
6001 struct io_comp_state *cs)
6003 struct io_ring_ctx *ctx = req->ctx;
6004 struct io_uring_files_update up;
6010 up.offset = req->files_update.offset;
6011 up.fds = req->files_update.arg;
6013 mutex_lock(&ctx->uring_lock);
6014 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
6015 mutex_unlock(&ctx->uring_lock);
6018 req_set_fail_links(req);
6019 __io_req_complete(req, ret, 0, cs);
6023 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6025 switch (req->opcode) {
6028 case IORING_OP_READV:
6029 case IORING_OP_READ_FIXED:
6030 case IORING_OP_READ:
6031 return io_read_prep(req, sqe);
6032 case IORING_OP_WRITEV:
6033 case IORING_OP_WRITE_FIXED:
6034 case IORING_OP_WRITE:
6035 return io_write_prep(req, sqe);
6036 case IORING_OP_POLL_ADD:
6037 return io_poll_add_prep(req, sqe);
6038 case IORING_OP_POLL_REMOVE:
6039 return io_poll_remove_prep(req, sqe);
6040 case IORING_OP_FSYNC:
6041 return io_prep_fsync(req, sqe);
6042 case IORING_OP_SYNC_FILE_RANGE:
6043 return io_prep_sfr(req, sqe);
6044 case IORING_OP_SENDMSG:
6045 case IORING_OP_SEND:
6046 return io_sendmsg_prep(req, sqe);
6047 case IORING_OP_RECVMSG:
6048 case IORING_OP_RECV:
6049 return io_recvmsg_prep(req, sqe);
6050 case IORING_OP_CONNECT:
6051 return io_connect_prep(req, sqe);
6052 case IORING_OP_TIMEOUT:
6053 return io_timeout_prep(req, sqe, false);
6054 case IORING_OP_TIMEOUT_REMOVE:
6055 return io_timeout_remove_prep(req, sqe);
6056 case IORING_OP_ASYNC_CANCEL:
6057 return io_async_cancel_prep(req, sqe);
6058 case IORING_OP_LINK_TIMEOUT:
6059 return io_timeout_prep(req, sqe, true);
6060 case IORING_OP_ACCEPT:
6061 return io_accept_prep(req, sqe);
6062 case IORING_OP_FALLOCATE:
6063 return io_fallocate_prep(req, sqe);
6064 case IORING_OP_OPENAT:
6065 return io_openat_prep(req, sqe);
6066 case IORING_OP_CLOSE:
6067 return io_close_prep(req, sqe);
6068 case IORING_OP_FILES_UPDATE:
6069 return io_files_update_prep(req, sqe);
6070 case IORING_OP_STATX:
6071 return io_statx_prep(req, sqe);
6072 case IORING_OP_FADVISE:
6073 return io_fadvise_prep(req, sqe);
6074 case IORING_OP_MADVISE:
6075 return io_madvise_prep(req, sqe);
6076 case IORING_OP_OPENAT2:
6077 return io_openat2_prep(req, sqe);
6078 case IORING_OP_EPOLL_CTL:
6079 return io_epoll_ctl_prep(req, sqe);
6080 case IORING_OP_SPLICE:
6081 return io_splice_prep(req, sqe);
6082 case IORING_OP_PROVIDE_BUFFERS:
6083 return io_provide_buffers_prep(req, sqe);
6084 case IORING_OP_REMOVE_BUFFERS:
6085 return io_remove_buffers_prep(req, sqe);
6087 return io_tee_prep(req, sqe);
6088 case IORING_OP_SHUTDOWN:
6089 return io_shutdown_prep(req, sqe);
6090 case IORING_OP_RENAMEAT:
6091 return io_renameat_prep(req, sqe);
6092 case IORING_OP_UNLINKAT:
6093 return io_unlinkat_prep(req, sqe);
6096 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6101 static int io_req_defer_prep(struct io_kiocb *req,
6102 const struct io_uring_sqe *sqe)
6106 if (io_alloc_async_data(req))
6108 return io_req_prep(req, sqe);
6111 static u32 io_get_sequence(struct io_kiocb *req)
6113 struct io_kiocb *pos;
6114 struct io_ring_ctx *ctx = req->ctx;
6115 u32 total_submitted, nr_reqs = 0;
6117 io_for_each_link(pos, req)
6120 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6121 return total_submitted - nr_reqs;
6124 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6126 struct io_ring_ctx *ctx = req->ctx;
6127 struct io_defer_entry *de;
6131 /* Still need defer if there is pending req in defer list. */
6132 if (likely(list_empty_careful(&ctx->defer_list) &&
6133 !(req->flags & REQ_F_IO_DRAIN)))
6136 seq = io_get_sequence(req);
6137 /* Still a chance to pass the sequence check */
6138 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6141 if (!req->async_data) {
6142 ret = io_req_defer_prep(req, sqe);
6146 io_prep_async_link(req);
6147 de = kmalloc(sizeof(*de), GFP_KERNEL);
6151 spin_lock_irq(&ctx->completion_lock);
6152 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6153 spin_unlock_irq(&ctx->completion_lock);
6155 io_queue_async_work(req);
6156 return -EIOCBQUEUED;
6159 trace_io_uring_defer(ctx, req, req->user_data);
6162 list_add_tail(&de->list, &ctx->defer_list);
6163 spin_unlock_irq(&ctx->completion_lock);
6164 return -EIOCBQUEUED;
6167 static void io_req_drop_files(struct io_kiocb *req)
6169 struct io_ring_ctx *ctx = req->ctx;
6170 struct io_uring_task *tctx = req->task->io_uring;
6171 unsigned long flags;
6173 if (req->work.flags & IO_WQ_WORK_FILES) {
6174 put_files_struct(req->work.identity->files);
6175 put_nsproxy(req->work.identity->nsproxy);
6177 spin_lock_irqsave(&ctx->inflight_lock, flags);
6178 list_del(&req->inflight_entry);
6179 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6180 req->flags &= ~REQ_F_INFLIGHT;
6181 req->work.flags &= ~IO_WQ_WORK_FILES;
6182 if (atomic_read(&tctx->in_idle))
6183 wake_up(&tctx->wait);
6186 static void __io_clean_op(struct io_kiocb *req)
6188 if (req->flags & REQ_F_BUFFER_SELECTED) {
6189 switch (req->opcode) {
6190 case IORING_OP_READV:
6191 case IORING_OP_READ_FIXED:
6192 case IORING_OP_READ:
6193 kfree((void *)(unsigned long)req->rw.addr);
6195 case IORING_OP_RECVMSG:
6196 case IORING_OP_RECV:
6197 kfree(req->sr_msg.kbuf);
6200 req->flags &= ~REQ_F_BUFFER_SELECTED;
6203 if (req->flags & REQ_F_NEED_CLEANUP) {
6204 switch (req->opcode) {
6205 case IORING_OP_READV:
6206 case IORING_OP_READ_FIXED:
6207 case IORING_OP_READ:
6208 case IORING_OP_WRITEV:
6209 case IORING_OP_WRITE_FIXED:
6210 case IORING_OP_WRITE: {
6211 struct io_async_rw *io = req->async_data;
6213 kfree(io->free_iovec);
6216 case IORING_OP_RECVMSG:
6217 case IORING_OP_SENDMSG: {
6218 struct io_async_msghdr *io = req->async_data;
6219 if (io->iov != io->fast_iov)
6223 case IORING_OP_SPLICE:
6225 io_put_file(req, req->splice.file_in,
6226 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6228 case IORING_OP_OPENAT:
6229 case IORING_OP_OPENAT2:
6230 if (req->open.filename)
6231 putname(req->open.filename);
6233 case IORING_OP_RENAMEAT:
6234 putname(req->rename.oldpath);
6235 putname(req->rename.newpath);
6237 case IORING_OP_UNLINKAT:
6238 putname(req->unlink.filename);
6241 req->flags &= ~REQ_F_NEED_CLEANUP;
6245 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6246 struct io_comp_state *cs)
6248 struct io_ring_ctx *ctx = req->ctx;
6251 switch (req->opcode) {
6253 ret = io_nop(req, cs);
6255 case IORING_OP_READV:
6256 case IORING_OP_READ_FIXED:
6257 case IORING_OP_READ:
6258 ret = io_read(req, force_nonblock, cs);
6260 case IORING_OP_WRITEV:
6261 case IORING_OP_WRITE_FIXED:
6262 case IORING_OP_WRITE:
6263 ret = io_write(req, force_nonblock, cs);
6265 case IORING_OP_FSYNC:
6266 ret = io_fsync(req, force_nonblock);
6268 case IORING_OP_POLL_ADD:
6269 ret = io_poll_add(req);
6271 case IORING_OP_POLL_REMOVE:
6272 ret = io_poll_remove(req);
6274 case IORING_OP_SYNC_FILE_RANGE:
6275 ret = io_sync_file_range(req, force_nonblock);
6277 case IORING_OP_SENDMSG:
6278 ret = io_sendmsg(req, force_nonblock, cs);
6280 case IORING_OP_SEND:
6281 ret = io_send(req, force_nonblock, cs);
6283 case IORING_OP_RECVMSG:
6284 ret = io_recvmsg(req, force_nonblock, cs);
6286 case IORING_OP_RECV:
6287 ret = io_recv(req, force_nonblock, cs);
6289 case IORING_OP_TIMEOUT:
6290 ret = io_timeout(req);
6292 case IORING_OP_TIMEOUT_REMOVE:
6293 ret = io_timeout_remove(req);
6295 case IORING_OP_ACCEPT:
6296 ret = io_accept(req, force_nonblock, cs);
6298 case IORING_OP_CONNECT:
6299 ret = io_connect(req, force_nonblock, cs);
6301 case IORING_OP_ASYNC_CANCEL:
6302 ret = io_async_cancel(req);
6304 case IORING_OP_FALLOCATE:
6305 ret = io_fallocate(req, force_nonblock);
6307 case IORING_OP_OPENAT:
6308 ret = io_openat(req, force_nonblock);
6310 case IORING_OP_CLOSE:
6311 ret = io_close(req, force_nonblock, cs);
6313 case IORING_OP_FILES_UPDATE:
6314 ret = io_files_update(req, force_nonblock, cs);
6316 case IORING_OP_STATX:
6317 ret = io_statx(req, force_nonblock);
6319 case IORING_OP_FADVISE:
6320 ret = io_fadvise(req, force_nonblock);
6322 case IORING_OP_MADVISE:
6323 ret = io_madvise(req, force_nonblock);
6325 case IORING_OP_OPENAT2:
6326 ret = io_openat2(req, force_nonblock);
6328 case IORING_OP_EPOLL_CTL:
6329 ret = io_epoll_ctl(req, force_nonblock, cs);
6331 case IORING_OP_SPLICE:
6332 ret = io_splice(req, force_nonblock);
6334 case IORING_OP_PROVIDE_BUFFERS:
6335 ret = io_provide_buffers(req, force_nonblock, cs);
6337 case IORING_OP_REMOVE_BUFFERS:
6338 ret = io_remove_buffers(req, force_nonblock, cs);
6341 ret = io_tee(req, force_nonblock);
6343 case IORING_OP_SHUTDOWN:
6344 ret = io_shutdown(req, force_nonblock);
6346 case IORING_OP_RENAMEAT:
6347 ret = io_renameat(req, force_nonblock);
6349 case IORING_OP_UNLINKAT:
6350 ret = io_unlinkat(req, force_nonblock);
6360 /* If the op doesn't have a file, we're not polling for it */
6361 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6362 const bool in_async = io_wq_current_is_worker();
6364 /* workqueue context doesn't hold uring_lock, grab it now */
6366 mutex_lock(&ctx->uring_lock);
6368 io_iopoll_req_issued(req, in_async);
6371 mutex_unlock(&ctx->uring_lock);
6377 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6379 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6380 struct io_kiocb *timeout;
6383 timeout = io_prep_linked_timeout(req);
6385 io_queue_linked_timeout(timeout);
6387 /* if NO_CANCEL is set, we must still run the work */
6388 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6389 IO_WQ_WORK_CANCEL) {
6395 ret = io_issue_sqe(req, false, NULL);
6397 * We can get EAGAIN for polled IO even though we're
6398 * forcing a sync submission from here, since we can't
6399 * wait for request slots on the block side.
6408 struct io_ring_ctx *lock_ctx = NULL;
6410 if (req->ctx->flags & IORING_SETUP_IOPOLL)
6411 lock_ctx = req->ctx;
6414 * io_iopoll_complete() does not hold completion_lock to
6415 * complete polled io, so here for polled io, we can not call
6416 * io_req_complete() directly, otherwise there maybe concurrent
6417 * access to cqring, defer_list, etc, which is not safe. Given
6418 * that io_iopoll_complete() is always called under uring_lock,
6419 * so here for polled io, we also get uring_lock to complete
6423 mutex_lock(&lock_ctx->uring_lock);
6425 req_set_fail_links(req);
6426 io_req_complete(req, ret);
6429 mutex_unlock(&lock_ctx->uring_lock);
6432 return io_steal_work(req);
6435 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6438 struct fixed_file_table *table;
6440 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6441 return table->files[index & IORING_FILE_TABLE_MASK];
6444 static struct file *io_file_get(struct io_submit_state *state,
6445 struct io_kiocb *req, int fd, bool fixed)
6447 struct io_ring_ctx *ctx = req->ctx;
6451 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6453 fd = array_index_nospec(fd, ctx->nr_user_files);
6454 file = io_file_from_index(ctx, fd);
6455 io_set_resource_node(req);
6457 trace_io_uring_file_get(ctx, fd);
6458 file = __io_file_get(state, fd);
6461 if (file && file->f_op == &io_uring_fops) {
6462 io_req_init_async(req);
6463 req->flags |= REQ_F_INFLIGHT;
6465 spin_lock_irq(&ctx->inflight_lock);
6466 list_add(&req->inflight_entry, &ctx->inflight_list);
6467 spin_unlock_irq(&ctx->inflight_lock);
6473 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6475 struct io_timeout_data *data = container_of(timer,
6476 struct io_timeout_data, timer);
6477 struct io_kiocb *prev, *req = data->req;
6478 struct io_ring_ctx *ctx = req->ctx;
6479 unsigned long flags;
6481 spin_lock_irqsave(&ctx->completion_lock, flags);
6482 prev = req->timeout.head;
6483 req->timeout.head = NULL;
6486 * We don't expect the list to be empty, that will only happen if we
6487 * race with the completion of the linked work.
6489 if (prev && refcount_inc_not_zero(&prev->refs))
6490 io_remove_next_linked(prev);
6493 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6496 req_set_fail_links(prev);
6497 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6500 io_req_complete(req, -ETIME);
6502 return HRTIMER_NORESTART;
6505 static void __io_queue_linked_timeout(struct io_kiocb *req)
6508 * If the back reference is NULL, then our linked request finished
6509 * before we got a chance to setup the timer
6511 if (req->timeout.head) {
6512 struct io_timeout_data *data = req->async_data;
6514 data->timer.function = io_link_timeout_fn;
6515 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6520 static void io_queue_linked_timeout(struct io_kiocb *req)
6522 struct io_ring_ctx *ctx = req->ctx;
6524 spin_lock_irq(&ctx->completion_lock);
6525 __io_queue_linked_timeout(req);
6526 spin_unlock_irq(&ctx->completion_lock);
6528 /* drop submission reference */
6532 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6534 struct io_kiocb *nxt = req->link;
6536 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6537 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6540 nxt->timeout.head = req;
6541 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6542 req->flags |= REQ_F_LINK_TIMEOUT;
6546 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6548 struct io_kiocb *linked_timeout;
6549 const struct cred *old_creds = NULL;
6553 linked_timeout = io_prep_linked_timeout(req);
6555 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6556 (req->work.flags & IO_WQ_WORK_CREDS) &&
6557 req->work.identity->creds != current_cred()) {
6559 revert_creds(old_creds);
6560 if (old_creds == req->work.identity->creds)
6561 old_creds = NULL; /* restored original creds */
6563 old_creds = override_creds(req->work.identity->creds);
6566 ret = io_issue_sqe(req, true, cs);
6569 * We async punt it if the file wasn't marked NOWAIT, or if the file
6570 * doesn't support non-blocking read/write attempts
6572 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6573 if (!io_arm_poll_handler(req)) {
6575 * Queued up for async execution, worker will release
6576 * submit reference when the iocb is actually submitted.
6578 io_queue_async_work(req);
6582 io_queue_linked_timeout(linked_timeout);
6583 } else if (likely(!ret)) {
6584 /* drop submission reference */
6585 req = io_put_req_find_next(req);
6587 io_queue_linked_timeout(linked_timeout);
6590 if (!(req->flags & REQ_F_FORCE_ASYNC))
6592 io_queue_async_work(req);
6595 /* un-prep timeout, so it'll be killed as any other linked */
6596 req->flags &= ~REQ_F_LINK_TIMEOUT;
6597 req_set_fail_links(req);
6599 io_req_complete(req, ret);
6603 revert_creds(old_creds);
6606 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6607 struct io_comp_state *cs)
6611 ret = io_req_defer(req, sqe);
6613 if (ret != -EIOCBQUEUED) {
6615 req_set_fail_links(req);
6617 io_req_complete(req, ret);
6619 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6620 if (!req->async_data) {
6621 ret = io_req_defer_prep(req, sqe);
6625 io_queue_async_work(req);
6628 ret = io_req_prep(req, sqe);
6632 __io_queue_sqe(req, cs);
6636 static inline void io_queue_link_head(struct io_kiocb *req,
6637 struct io_comp_state *cs)
6639 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6641 io_req_complete(req, -ECANCELED);
6643 io_queue_sqe(req, NULL, cs);
6646 struct io_submit_link {
6647 struct io_kiocb *head;
6648 struct io_kiocb *last;
6651 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6652 struct io_submit_link *link, struct io_comp_state *cs)
6654 struct io_ring_ctx *ctx = req->ctx;
6658 * If we already have a head request, queue this one for async
6659 * submittal once the head completes. If we don't have a head but
6660 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6661 * submitted sync once the chain is complete. If none of those
6662 * conditions are true (normal request), then just queue it.
6665 struct io_kiocb *head = link->head;
6668 * Taking sequential execution of a link, draining both sides
6669 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6670 * requests in the link. So, it drains the head and the
6671 * next after the link request. The last one is done via
6672 * drain_next flag to persist the effect across calls.
6674 if (req->flags & REQ_F_IO_DRAIN) {
6675 head->flags |= REQ_F_IO_DRAIN;
6676 ctx->drain_next = 1;
6678 ret = io_req_defer_prep(req, sqe);
6679 if (unlikely(ret)) {
6680 /* fail even hard links since we don't submit */
6681 head->flags |= REQ_F_FAIL_LINK;
6684 trace_io_uring_link(ctx, req, head);
6685 link->last->link = req;
6688 /* last request of a link, enqueue the link */
6689 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6690 io_queue_link_head(head, cs);
6694 if (unlikely(ctx->drain_next)) {
6695 req->flags |= REQ_F_IO_DRAIN;
6696 ctx->drain_next = 0;
6698 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6699 ret = io_req_defer_prep(req, sqe);
6701 req->flags |= REQ_F_FAIL_LINK;
6705 io_queue_sqe(req, sqe, cs);
6713 * Batched submission is done, ensure local IO is flushed out.
6715 static void io_submit_state_end(struct io_submit_state *state)
6717 if (!list_empty(&state->comp.list))
6718 io_submit_flush_completions(&state->comp);
6719 if (state->plug_started)
6720 blk_finish_plug(&state->plug);
6721 io_state_file_put(state);
6722 if (state->free_reqs)
6723 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6727 * Start submission side cache.
6729 static void io_submit_state_start(struct io_submit_state *state,
6730 struct io_ring_ctx *ctx, unsigned int max_ios)
6732 state->plug_started = false;
6734 INIT_LIST_HEAD(&state->comp.list);
6735 state->comp.ctx = ctx;
6736 state->free_reqs = 0;
6737 state->file_refs = 0;
6738 state->ios_left = max_ios;
6741 static void io_commit_sqring(struct io_ring_ctx *ctx)
6743 struct io_rings *rings = ctx->rings;
6746 * Ensure any loads from the SQEs are done at this point,
6747 * since once we write the new head, the application could
6748 * write new data to them.
6750 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6754 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6755 * that is mapped by userspace. This means that care needs to be taken to
6756 * ensure that reads are stable, as we cannot rely on userspace always
6757 * being a good citizen. If members of the sqe are validated and then later
6758 * used, it's important that those reads are done through READ_ONCE() to
6759 * prevent a re-load down the line.
6761 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6763 u32 *sq_array = ctx->sq_array;
6767 * The cached sq head (or cq tail) serves two purposes:
6769 * 1) allows us to batch the cost of updating the user visible
6771 * 2) allows the kernel side to track the head on its own, even
6772 * though the application is the one updating it.
6774 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6775 if (likely(head < ctx->sq_entries))
6776 return &ctx->sq_sqes[head];
6778 /* drop invalid entries */
6779 ctx->cached_sq_dropped++;
6780 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6784 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6786 ctx->cached_sq_head++;
6790 * Check SQE restrictions (opcode and flags).
6792 * Returns 'true' if SQE is allowed, 'false' otherwise.
6794 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6795 struct io_kiocb *req,
6796 unsigned int sqe_flags)
6798 if (!ctx->restricted)
6801 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6804 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6805 ctx->restrictions.sqe_flags_required)
6808 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6809 ctx->restrictions.sqe_flags_required))
6815 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6816 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6817 IOSQE_BUFFER_SELECT)
6819 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6820 const struct io_uring_sqe *sqe,
6821 struct io_submit_state *state)
6823 unsigned int sqe_flags;
6826 req->opcode = READ_ONCE(sqe->opcode);
6827 req->user_data = READ_ONCE(sqe->user_data);
6828 req->async_data = NULL;
6833 req->fixed_file_refs = NULL;
6834 /* one is dropped after submission, the other at completion */
6835 refcount_set(&req->refs, 2);
6836 req->task = current;
6839 if (unlikely(req->opcode >= IORING_OP_LAST))
6842 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6845 sqe_flags = READ_ONCE(sqe->flags);
6846 /* enforce forwards compatibility on users */
6847 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6850 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6853 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6854 !io_op_defs[req->opcode].buffer_select)
6857 id = READ_ONCE(sqe->personality);
6859 struct io_identity *iod;
6861 iod = idr_find(&ctx->personality_idr, id);
6864 refcount_inc(&iod->count);
6866 __io_req_init_async(req);
6867 get_cred(iod->creds);
6868 req->work.identity = iod;
6869 req->work.flags |= IO_WQ_WORK_CREDS;
6872 /* same numerical values with corresponding REQ_F_*, safe to copy */
6873 req->flags |= sqe_flags;
6876 * Plug now if we have more than 1 IO left after this, and the target
6877 * is potentially a read/write to block based storage.
6879 if (!state->plug_started && state->ios_left > 1 &&
6880 io_op_defs[req->opcode].plug) {
6881 blk_start_plug(&state->plug);
6882 state->plug_started = true;
6886 if (io_op_defs[req->opcode].needs_file) {
6887 bool fixed = req->flags & REQ_F_FIXED_FILE;
6889 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6890 if (unlikely(!req->file &&
6891 !io_op_defs[req->opcode].needs_file_no_error))
6899 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6901 struct io_submit_state state;
6902 struct io_submit_link link;
6903 int i, submitted = 0;
6905 /* if we have a backlog and couldn't flush it all, return BUSY */
6906 if (test_bit(0, &ctx->sq_check_overflow)) {
6907 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6911 /* make sure SQ entry isn't read before tail */
6912 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6914 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6917 percpu_counter_add(¤t->io_uring->inflight, nr);
6918 refcount_add(nr, ¤t->usage);
6920 io_submit_state_start(&state, ctx, nr);
6923 for (i = 0; i < nr; i++) {
6924 const struct io_uring_sqe *sqe;
6925 struct io_kiocb *req;
6928 sqe = io_get_sqe(ctx);
6929 if (unlikely(!sqe)) {
6930 io_consume_sqe(ctx);
6933 req = io_alloc_req(ctx, &state);
6934 if (unlikely(!req)) {
6936 submitted = -EAGAIN;
6939 io_consume_sqe(ctx);
6940 /* will complete beyond this point, count as submitted */
6943 err = io_init_req(ctx, req, sqe, &state);
6944 if (unlikely(err)) {
6947 io_req_complete(req, err);
6951 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6952 true, io_async_submit(ctx));
6953 err = io_submit_sqe(req, sqe, &link, &state.comp);
6958 if (unlikely(submitted != nr)) {
6959 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6960 struct io_uring_task *tctx = current->io_uring;
6961 int unused = nr - ref_used;
6963 percpu_ref_put_many(&ctx->refs, unused);
6964 percpu_counter_sub(&tctx->inflight, unused);
6965 put_task_struct_many(current, unused);
6968 io_queue_link_head(link.head, &state.comp);
6969 io_submit_state_end(&state);
6971 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6972 io_commit_sqring(ctx);
6977 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6979 /* Tell userspace we may need a wakeup call */
6980 spin_lock_irq(&ctx->completion_lock);
6981 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6982 spin_unlock_irq(&ctx->completion_lock);
6985 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6987 spin_lock_irq(&ctx->completion_lock);
6988 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6989 spin_unlock_irq(&ctx->completion_lock);
6992 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6994 unsigned int to_submit;
6997 to_submit = io_sqring_entries(ctx);
6998 /* if we're handling multiple rings, cap submit size for fairness */
6999 if (cap_entries && to_submit > 8)
7002 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7003 unsigned nr_events = 0;
7005 mutex_lock(&ctx->uring_lock);
7006 if (!list_empty(&ctx->iopoll_list))
7007 io_do_iopoll(ctx, &nr_events, 0);
7009 if (to_submit && !ctx->sqo_dead &&
7010 likely(!percpu_ref_is_dying(&ctx->refs)))
7011 ret = io_submit_sqes(ctx, to_submit);
7012 mutex_unlock(&ctx->uring_lock);
7015 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
7016 wake_up(&ctx->sqo_sq_wait);
7021 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7023 struct io_ring_ctx *ctx;
7024 unsigned sq_thread_idle = 0;
7026 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7027 if (sq_thread_idle < ctx->sq_thread_idle)
7028 sq_thread_idle = ctx->sq_thread_idle;
7031 sqd->sq_thread_idle = sq_thread_idle;
7034 static void io_sqd_init_new(struct io_sq_data *sqd)
7036 struct io_ring_ctx *ctx;
7038 while (!list_empty(&sqd->ctx_new_list)) {
7039 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
7040 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
7041 complete(&ctx->sq_thread_comp);
7044 io_sqd_update_thread_idle(sqd);
7047 static int io_sq_thread(void *data)
7049 struct cgroup_subsys_state *cur_css = NULL;
7050 struct files_struct *old_files = current->files;
7051 struct nsproxy *old_nsproxy = current->nsproxy;
7052 const struct cred *old_cred = NULL;
7053 struct io_sq_data *sqd = data;
7054 struct io_ring_ctx *ctx;
7055 unsigned long timeout = 0;
7059 current->files = NULL;
7060 current->nsproxy = NULL;
7061 task_unlock(current);
7063 while (!kthread_should_stop()) {
7065 bool cap_entries, sqt_spin, needs_sched;
7068 * Any changes to the sqd lists are synchronized through the
7069 * kthread parking. This synchronizes the thread vs users,
7070 * the users are synchronized on the sqd->ctx_lock.
7072 if (kthread_should_park()) {
7075 * When sq thread is unparked, in case the previous park operation
7076 * comes from io_put_sq_data(), which means that sq thread is going
7077 * to be stopped, so here needs to have a check.
7079 if (kthread_should_stop())
7083 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7084 io_sqd_init_new(sqd);
7085 timeout = jiffies + sqd->sq_thread_idle;
7089 cap_entries = !list_is_singular(&sqd->ctx_list);
7090 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7091 if (current->cred != ctx->creds) {
7093 revert_creds(old_cred);
7094 old_cred = override_creds(ctx->creds);
7096 io_sq_thread_associate_blkcg(ctx, &cur_css);
7098 current->loginuid = ctx->loginuid;
7099 current->sessionid = ctx->sessionid;
7102 ret = __io_sq_thread(ctx, cap_entries);
7103 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7106 io_sq_thread_drop_mm_files();
7109 if (sqt_spin || !time_after(jiffies, timeout)) {
7111 io_sq_thread_drop_mm_files();
7114 timeout = jiffies + sqd->sq_thread_idle;
7118 if (kthread_should_park())
7122 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7123 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7124 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7125 !list_empty_careful(&ctx->iopoll_list)) {
7126 needs_sched = false;
7129 if (io_sqring_entries(ctx)) {
7130 needs_sched = false;
7136 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7137 io_ring_set_wakeup_flag(ctx);
7140 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7141 io_ring_clear_wakeup_flag(ctx);
7144 finish_wait(&sqd->wait, &wait);
7145 timeout = jiffies + sqd->sq_thread_idle;
7149 io_sq_thread_drop_mm_files();
7152 io_sq_thread_unassociate_blkcg();
7154 revert_creds(old_cred);
7157 current->files = old_files;
7158 current->nsproxy = old_nsproxy;
7159 task_unlock(current);
7166 struct io_wait_queue {
7167 struct wait_queue_entry wq;
7168 struct io_ring_ctx *ctx;
7170 unsigned nr_timeouts;
7173 static inline bool io_should_wake(struct io_wait_queue *iowq)
7175 struct io_ring_ctx *ctx = iowq->ctx;
7178 * Wake up if we have enough events, or if a timeout occurred since we
7179 * started waiting. For timeouts, we always want to return to userspace,
7180 * regardless of event count.
7182 return io_cqring_events(ctx) >= iowq->to_wait ||
7183 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7186 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7187 int wake_flags, void *key)
7189 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7193 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7194 * the task, and the next invocation will do it.
7196 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
7197 return autoremove_wake_function(curr, mode, wake_flags, key);
7201 static int io_run_task_work_sig(void)
7203 if (io_run_task_work())
7205 if (!signal_pending(current))
7207 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7208 return -ERESTARTSYS;
7213 * Wait until events become available, if we don't already have some. The
7214 * application must reap them itself, as they reside on the shared cq ring.
7216 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7217 const sigset_t __user *sig, size_t sigsz,
7218 struct __kernel_timespec __user *uts)
7220 struct io_wait_queue iowq = {
7223 .func = io_wake_function,
7224 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7227 .to_wait = min_events,
7229 struct io_rings *rings = ctx->rings;
7230 struct timespec64 ts;
7231 signed long timeout = 0;
7235 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7236 if (io_cqring_events(ctx) >= min_events)
7238 if (!io_run_task_work())
7243 #ifdef CONFIG_COMPAT
7244 if (in_compat_syscall())
7245 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7249 ret = set_user_sigmask(sig, sigsz);
7256 if (get_timespec64(&ts, uts))
7258 timeout = timespec64_to_jiffies(&ts);
7261 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7262 trace_io_uring_cqring_wait(ctx, min_events);
7264 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7265 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7266 TASK_INTERRUPTIBLE);
7267 /* make sure we run task_work before checking for signals */
7268 ret = io_run_task_work_sig();
7273 if (io_should_wake(&iowq))
7275 if (test_bit(0, &ctx->cq_check_overflow))
7278 timeout = schedule_timeout(timeout);
7287 finish_wait(&ctx->wait, &iowq.wq);
7289 restore_saved_sigmask_unless(ret == -EINTR);
7291 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7294 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7296 #if defined(CONFIG_UNIX)
7297 if (ctx->ring_sock) {
7298 struct sock *sock = ctx->ring_sock->sk;
7299 struct sk_buff *skb;
7301 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7307 for (i = 0; i < ctx->nr_user_files; i++) {
7310 file = io_file_from_index(ctx, i);
7317 static void io_file_ref_kill(struct percpu_ref *ref)
7319 struct fixed_file_data *data;
7321 data = container_of(ref, struct fixed_file_data, refs);
7322 complete(&data->done);
7325 static void io_sqe_files_set_node(struct fixed_file_data *file_data,
7326 struct fixed_file_ref_node *ref_node)
7328 spin_lock_bh(&file_data->lock);
7329 file_data->node = ref_node;
7330 list_add_tail(&ref_node->node, &file_data->ref_list);
7331 spin_unlock_bh(&file_data->lock);
7332 percpu_ref_get(&file_data->refs);
7335 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7337 struct fixed_file_data *data = ctx->file_data;
7338 struct fixed_file_ref_node *backup_node, *ref_node = NULL;
7339 unsigned nr_tables, i;
7344 backup_node = alloc_fixed_file_ref_node(ctx);
7348 spin_lock_bh(&data->lock);
7349 ref_node = data->node;
7350 spin_unlock_bh(&data->lock);
7352 percpu_ref_kill(&ref_node->refs);
7354 percpu_ref_kill(&data->refs);
7356 /* wait for all refs nodes to complete */
7357 flush_delayed_work(&ctx->file_put_work);
7359 ret = wait_for_completion_interruptible(&data->done);
7362 ret = io_run_task_work_sig();
7364 percpu_ref_resurrect(&data->refs);
7365 reinit_completion(&data->done);
7366 io_sqe_files_set_node(data, backup_node);
7371 __io_sqe_files_unregister(ctx);
7372 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7373 for (i = 0; i < nr_tables; i++)
7374 kfree(data->table[i].files);
7376 percpu_ref_exit(&data->refs);
7378 ctx->file_data = NULL;
7379 ctx->nr_user_files = 0;
7380 destroy_fixed_file_ref_node(backup_node);
7384 static void io_put_sq_data(struct io_sq_data *sqd)
7386 if (refcount_dec_and_test(&sqd->refs)) {
7388 * The park is a bit of a work-around, without it we get
7389 * warning spews on shutdown with SQPOLL set and affinity
7390 * set to a single CPU.
7393 kthread_park(sqd->thread);
7394 kthread_stop(sqd->thread);
7401 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7403 struct io_ring_ctx *ctx_attach;
7404 struct io_sq_data *sqd;
7407 f = fdget(p->wq_fd);
7409 return ERR_PTR(-ENXIO);
7410 if (f.file->f_op != &io_uring_fops) {
7412 return ERR_PTR(-EINVAL);
7415 ctx_attach = f.file->private_data;
7416 sqd = ctx_attach->sq_data;
7419 return ERR_PTR(-EINVAL);
7422 refcount_inc(&sqd->refs);
7427 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7429 struct io_sq_data *sqd;
7431 if (p->flags & IORING_SETUP_ATTACH_WQ)
7432 return io_attach_sq_data(p);
7434 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7436 return ERR_PTR(-ENOMEM);
7438 refcount_set(&sqd->refs, 1);
7439 INIT_LIST_HEAD(&sqd->ctx_list);
7440 INIT_LIST_HEAD(&sqd->ctx_new_list);
7441 mutex_init(&sqd->ctx_lock);
7442 mutex_init(&sqd->lock);
7443 init_waitqueue_head(&sqd->wait);
7447 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7448 __releases(&sqd->lock)
7452 kthread_unpark(sqd->thread);
7453 mutex_unlock(&sqd->lock);
7456 static void io_sq_thread_park(struct io_sq_data *sqd)
7457 __acquires(&sqd->lock)
7461 mutex_lock(&sqd->lock);
7462 kthread_park(sqd->thread);
7465 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7467 struct io_sq_data *sqd = ctx->sq_data;
7472 * We may arrive here from the error branch in
7473 * io_sq_offload_create() where the kthread is created
7474 * without being waked up, thus wake it up now to make
7475 * sure the wait will complete.
7477 wake_up_process(sqd->thread);
7478 wait_for_completion(&ctx->sq_thread_comp);
7480 io_sq_thread_park(sqd);
7483 mutex_lock(&sqd->ctx_lock);
7484 list_del(&ctx->sqd_list);
7485 io_sqd_update_thread_idle(sqd);
7486 mutex_unlock(&sqd->ctx_lock);
7489 io_sq_thread_unpark(sqd);
7491 io_put_sq_data(sqd);
7492 ctx->sq_data = NULL;
7496 static void io_finish_async(struct io_ring_ctx *ctx)
7498 io_sq_thread_stop(ctx);
7501 io_wq_destroy(ctx->io_wq);
7506 #if defined(CONFIG_UNIX)
7508 * Ensure the UNIX gc is aware of our file set, so we are certain that
7509 * the io_uring can be safely unregistered on process exit, even if we have
7510 * loops in the file referencing.
7512 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7514 struct sock *sk = ctx->ring_sock->sk;
7515 struct scm_fp_list *fpl;
7516 struct sk_buff *skb;
7519 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7523 skb = alloc_skb(0, GFP_KERNEL);
7532 fpl->user = get_uid(ctx->user);
7533 for (i = 0; i < nr; i++) {
7534 struct file *file = io_file_from_index(ctx, i + offset);
7538 fpl->fp[nr_files] = get_file(file);
7539 unix_inflight(fpl->user, fpl->fp[nr_files]);
7544 fpl->max = SCM_MAX_FD;
7545 fpl->count = nr_files;
7546 UNIXCB(skb).fp = fpl;
7547 skb->destructor = unix_destruct_scm;
7548 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7549 skb_queue_head(&sk->sk_receive_queue, skb);
7551 for (i = 0; i < nr_files; i++)
7562 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7563 * causes regular reference counting to break down. We rely on the UNIX
7564 * garbage collection to take care of this problem for us.
7566 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7568 unsigned left, total;
7572 left = ctx->nr_user_files;
7574 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7576 ret = __io_sqe_files_scm(ctx, this_files, total);
7580 total += this_files;
7586 while (total < ctx->nr_user_files) {
7587 struct file *file = io_file_from_index(ctx, total);
7597 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7603 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7604 unsigned nr_tables, unsigned nr_files)
7608 for (i = 0; i < nr_tables; i++) {
7609 struct fixed_file_table *table = &file_data->table[i];
7610 unsigned this_files;
7612 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7613 table->files = kcalloc(this_files, sizeof(struct file *),
7617 nr_files -= this_files;
7623 for (i = 0; i < nr_tables; i++) {
7624 struct fixed_file_table *table = &file_data->table[i];
7625 kfree(table->files);
7630 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7632 #if defined(CONFIG_UNIX)
7633 struct sock *sock = ctx->ring_sock->sk;
7634 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7635 struct sk_buff *skb;
7638 __skb_queue_head_init(&list);
7641 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7642 * remove this entry and rearrange the file array.
7644 skb = skb_dequeue(head);
7646 struct scm_fp_list *fp;
7648 fp = UNIXCB(skb).fp;
7649 for (i = 0; i < fp->count; i++) {
7652 if (fp->fp[i] != file)
7655 unix_notinflight(fp->user, fp->fp[i]);
7656 left = fp->count - 1 - i;
7658 memmove(&fp->fp[i], &fp->fp[i + 1],
7659 left * sizeof(struct file *));
7666 __skb_queue_tail(&list, skb);
7676 __skb_queue_tail(&list, skb);
7678 skb = skb_dequeue(head);
7681 if (skb_peek(&list)) {
7682 spin_lock_irq(&head->lock);
7683 while ((skb = __skb_dequeue(&list)) != NULL)
7684 __skb_queue_tail(head, skb);
7685 spin_unlock_irq(&head->lock);
7692 struct io_file_put {
7693 struct list_head list;
7697 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7699 struct fixed_file_data *file_data = ref_node->file_data;
7700 struct io_ring_ctx *ctx = file_data->ctx;
7701 struct io_file_put *pfile, *tmp;
7703 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7704 list_del(&pfile->list);
7705 io_ring_file_put(ctx, pfile->file);
7709 percpu_ref_exit(&ref_node->refs);
7711 percpu_ref_put(&file_data->refs);
7714 static void io_file_put_work(struct work_struct *work)
7716 struct io_ring_ctx *ctx;
7717 struct llist_node *node;
7719 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7720 node = llist_del_all(&ctx->file_put_llist);
7723 struct fixed_file_ref_node *ref_node;
7724 struct llist_node *next = node->next;
7726 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7727 __io_file_put_work(ref_node);
7732 static void io_file_data_ref_zero(struct percpu_ref *ref)
7734 struct fixed_file_ref_node *ref_node;
7735 struct fixed_file_data *data;
7736 struct io_ring_ctx *ctx;
7737 bool first_add = false;
7740 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7741 data = ref_node->file_data;
7744 spin_lock_bh(&data->lock);
7745 ref_node->done = true;
7747 while (!list_empty(&data->ref_list)) {
7748 ref_node = list_first_entry(&data->ref_list,
7749 struct fixed_file_ref_node, node);
7750 /* recycle ref nodes in order */
7751 if (!ref_node->done)
7753 list_del(&ref_node->node);
7754 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7756 spin_unlock_bh(&data->lock);
7758 if (percpu_ref_is_dying(&data->refs))
7762 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7764 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7767 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7768 struct io_ring_ctx *ctx)
7770 struct fixed_file_ref_node *ref_node;
7772 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7776 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7781 INIT_LIST_HEAD(&ref_node->node);
7782 INIT_LIST_HEAD(&ref_node->file_list);
7783 ref_node->file_data = ctx->file_data;
7784 ref_node->done = false;
7788 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7790 percpu_ref_exit(&ref_node->refs);
7794 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7797 __s32 __user *fds = (__s32 __user *) arg;
7798 unsigned nr_tables, i;
7800 int fd, ret = -ENOMEM;
7801 struct fixed_file_ref_node *ref_node;
7802 struct fixed_file_data *file_data;
7808 if (nr_args > IORING_MAX_FIXED_FILES)
7811 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7814 file_data->ctx = ctx;
7815 init_completion(&file_data->done);
7816 INIT_LIST_HEAD(&file_data->ref_list);
7817 spin_lock_init(&file_data->lock);
7819 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7820 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7822 if (!file_data->table)
7825 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7826 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7829 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7831 ctx->file_data = file_data;
7833 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7834 struct fixed_file_table *table;
7837 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7841 /* allow sparse sets */
7851 * Don't allow io_uring instances to be registered. If UNIX
7852 * isn't enabled, then this causes a reference cycle and this
7853 * instance can never get freed. If UNIX is enabled we'll
7854 * handle it just fine, but there's still no point in allowing
7855 * a ring fd as it doesn't support regular read/write anyway.
7857 if (file->f_op == &io_uring_fops) {
7861 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7862 index = i & IORING_FILE_TABLE_MASK;
7863 table->files[index] = file;
7866 ret = io_sqe_files_scm(ctx);
7868 io_sqe_files_unregister(ctx);
7872 ref_node = alloc_fixed_file_ref_node(ctx);
7874 io_sqe_files_unregister(ctx);
7878 io_sqe_files_set_node(file_data, ref_node);
7881 for (i = 0; i < ctx->nr_user_files; i++) {
7882 file = io_file_from_index(ctx, i);
7886 for (i = 0; i < nr_tables; i++)
7887 kfree(file_data->table[i].files);
7888 ctx->nr_user_files = 0;
7890 percpu_ref_exit(&file_data->refs);
7892 kfree(file_data->table);
7894 ctx->file_data = NULL;
7898 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7901 #if defined(CONFIG_UNIX)
7902 struct sock *sock = ctx->ring_sock->sk;
7903 struct sk_buff_head *head = &sock->sk_receive_queue;
7904 struct sk_buff *skb;
7907 * See if we can merge this file into an existing skb SCM_RIGHTS
7908 * file set. If there's no room, fall back to allocating a new skb
7909 * and filling it in.
7911 spin_lock_irq(&head->lock);
7912 skb = skb_peek(head);
7914 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7916 if (fpl->count < SCM_MAX_FD) {
7917 __skb_unlink(skb, head);
7918 spin_unlock_irq(&head->lock);
7919 fpl->fp[fpl->count] = get_file(file);
7920 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7922 spin_lock_irq(&head->lock);
7923 __skb_queue_head(head, skb);
7928 spin_unlock_irq(&head->lock);
7935 return __io_sqe_files_scm(ctx, 1, index);
7941 static int io_queue_file_removal(struct fixed_file_data *data,
7944 struct io_file_put *pfile;
7945 struct fixed_file_ref_node *ref_node = data->node;
7947 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7952 list_add(&pfile->list, &ref_node->file_list);
7957 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7958 struct io_uring_files_update *up,
7961 struct fixed_file_data *data = ctx->file_data;
7962 struct fixed_file_ref_node *ref_node;
7967 bool needs_switch = false;
7969 if (check_add_overflow(up->offset, nr_args, &done))
7971 if (done > ctx->nr_user_files)
7974 ref_node = alloc_fixed_file_ref_node(ctx);
7979 fds = u64_to_user_ptr(up->fds);
7981 struct fixed_file_table *table;
7985 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7989 i = array_index_nospec(up->offset, ctx->nr_user_files);
7990 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7991 index = i & IORING_FILE_TABLE_MASK;
7992 if (table->files[index]) {
7993 file = table->files[index];
7994 err = io_queue_file_removal(data, file);
7997 table->files[index] = NULL;
7998 needs_switch = true;
8007 * Don't allow io_uring instances to be registered. If
8008 * UNIX isn't enabled, then this causes a reference
8009 * cycle and this instance can never get freed. If UNIX
8010 * is enabled we'll handle it just fine, but there's
8011 * still no point in allowing a ring fd as it doesn't
8012 * support regular read/write anyway.
8014 if (file->f_op == &io_uring_fops) {
8019 table->files[index] = file;
8020 err = io_sqe_file_register(ctx, file, i);
8022 table->files[index] = NULL;
8033 percpu_ref_kill(&data->node->refs);
8034 io_sqe_files_set_node(data, ref_node);
8036 destroy_fixed_file_ref_node(ref_node);
8038 return done ? done : err;
8041 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
8044 struct io_uring_files_update up;
8046 if (!ctx->file_data)
8050 if (copy_from_user(&up, arg, sizeof(up)))
8055 return __io_sqe_files_update(ctx, &up, nr_args);
8058 static void io_free_work(struct io_wq_work *work)
8060 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8062 /* Consider that io_steal_work() relies on this ref */
8066 static int io_init_wq_offload(struct io_ring_ctx *ctx,
8067 struct io_uring_params *p)
8069 struct io_wq_data data;
8071 struct io_ring_ctx *ctx_attach;
8072 unsigned int concurrency;
8075 data.user = ctx->user;
8076 data.free_work = io_free_work;
8077 data.do_work = io_wq_submit_work;
8079 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
8080 /* Do QD, or 4 * CPUS, whatever is smallest */
8081 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8083 ctx->io_wq = io_wq_create(concurrency, &data);
8084 if (IS_ERR(ctx->io_wq)) {
8085 ret = PTR_ERR(ctx->io_wq);
8091 f = fdget(p->wq_fd);
8095 if (f.file->f_op != &io_uring_fops) {
8100 ctx_attach = f.file->private_data;
8101 /* @io_wq is protected by holding the fd */
8102 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8107 ctx->io_wq = ctx_attach->io_wq;
8113 static int io_uring_alloc_task_context(struct task_struct *task)
8115 struct io_uring_task *tctx;
8118 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8119 if (unlikely(!tctx))
8122 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8123 if (unlikely(ret)) {
8129 init_waitqueue_head(&tctx->wait);
8131 atomic_set(&tctx->in_idle, 0);
8132 tctx->sqpoll = false;
8133 io_init_identity(&tctx->__identity);
8134 tctx->identity = &tctx->__identity;
8135 task->io_uring = tctx;
8139 void __io_uring_free(struct task_struct *tsk)
8141 struct io_uring_task *tctx = tsk->io_uring;
8143 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8144 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8145 if (tctx->identity != &tctx->__identity)
8146 kfree(tctx->identity);
8147 percpu_counter_destroy(&tctx->inflight);
8149 tsk->io_uring = NULL;
8152 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8153 struct io_uring_params *p)
8157 if (ctx->flags & IORING_SETUP_SQPOLL) {
8158 struct io_sq_data *sqd;
8161 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8164 sqd = io_get_sq_data(p);
8171 io_sq_thread_park(sqd);
8172 mutex_lock(&sqd->ctx_lock);
8173 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8174 mutex_unlock(&sqd->ctx_lock);
8175 io_sq_thread_unpark(sqd);
8177 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8178 if (!ctx->sq_thread_idle)
8179 ctx->sq_thread_idle = HZ;
8184 if (p->flags & IORING_SETUP_SQ_AFF) {
8185 int cpu = p->sq_thread_cpu;
8188 if (cpu >= nr_cpu_ids)
8190 if (!cpu_online(cpu))
8193 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8194 cpu, "io_uring-sq");
8196 sqd->thread = kthread_create(io_sq_thread, sqd,
8199 if (IS_ERR(sqd->thread)) {
8200 ret = PTR_ERR(sqd->thread);
8204 ret = io_uring_alloc_task_context(sqd->thread);
8207 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8208 /* Can't have SQ_AFF without SQPOLL */
8214 ret = io_init_wq_offload(ctx, p);
8220 io_finish_async(ctx);
8224 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8226 struct io_sq_data *sqd = ctx->sq_data;
8228 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8229 wake_up_process(sqd->thread);
8232 static inline void __io_unaccount_mem(struct user_struct *user,
8233 unsigned long nr_pages)
8235 atomic_long_sub(nr_pages, &user->locked_vm);
8238 static inline int __io_account_mem(struct user_struct *user,
8239 unsigned long nr_pages)
8241 unsigned long page_limit, cur_pages, new_pages;
8243 /* Don't allow more pages than we can safely lock */
8244 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8247 cur_pages = atomic_long_read(&user->locked_vm);
8248 new_pages = cur_pages + nr_pages;
8249 if (new_pages > page_limit)
8251 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8252 new_pages) != cur_pages);
8257 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8258 enum io_mem_account acct)
8261 __io_unaccount_mem(ctx->user, nr_pages);
8263 if (ctx->mm_account) {
8264 if (acct == ACCT_LOCKED) {
8265 mmap_write_lock(ctx->mm_account);
8266 ctx->mm_account->locked_vm -= nr_pages;
8267 mmap_write_unlock(ctx->mm_account);
8268 }else if (acct == ACCT_PINNED) {
8269 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8274 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8275 enum io_mem_account acct)
8279 if (ctx->limit_mem) {
8280 ret = __io_account_mem(ctx->user, nr_pages);
8285 if (ctx->mm_account) {
8286 if (acct == ACCT_LOCKED) {
8287 mmap_write_lock(ctx->mm_account);
8288 ctx->mm_account->locked_vm += nr_pages;
8289 mmap_write_unlock(ctx->mm_account);
8290 } else if (acct == ACCT_PINNED) {
8291 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8298 static void io_mem_free(void *ptr)
8305 page = virt_to_head_page(ptr);
8306 if (put_page_testzero(page))
8307 free_compound_page(page);
8310 static void *io_mem_alloc(size_t size)
8312 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8315 return (void *) __get_free_pages(gfp_flags, get_order(size));
8318 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8321 struct io_rings *rings;
8322 size_t off, sq_array_size;
8324 off = struct_size(rings, cqes, cq_entries);
8325 if (off == SIZE_MAX)
8329 off = ALIGN(off, SMP_CACHE_BYTES);
8337 sq_array_size = array_size(sizeof(u32), sq_entries);
8338 if (sq_array_size == SIZE_MAX)
8341 if (check_add_overflow(off, sq_array_size, &off))
8347 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8351 pages = (size_t)1 << get_order(
8352 rings_size(sq_entries, cq_entries, NULL));
8353 pages += (size_t)1 << get_order(
8354 array_size(sizeof(struct io_uring_sqe), sq_entries));
8359 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8363 if (!ctx->user_bufs)
8366 for (i = 0; i < ctx->nr_user_bufs; i++) {
8367 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8369 for (j = 0; j < imu->nr_bvecs; j++)
8370 unpin_user_page(imu->bvec[j].bv_page);
8372 if (imu->acct_pages)
8373 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8378 kfree(ctx->user_bufs);
8379 ctx->user_bufs = NULL;
8380 ctx->nr_user_bufs = 0;
8384 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8385 void __user *arg, unsigned index)
8387 struct iovec __user *src;
8389 #ifdef CONFIG_COMPAT
8391 struct compat_iovec __user *ciovs;
8392 struct compat_iovec ciov;
8394 ciovs = (struct compat_iovec __user *) arg;
8395 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8398 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8399 dst->iov_len = ciov.iov_len;
8403 src = (struct iovec __user *) arg;
8404 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8410 * Not super efficient, but this is just a registration time. And we do cache
8411 * the last compound head, so generally we'll only do a full search if we don't
8414 * We check if the given compound head page has already been accounted, to
8415 * avoid double accounting it. This allows us to account the full size of the
8416 * page, not just the constituent pages of a huge page.
8418 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8419 int nr_pages, struct page *hpage)
8423 /* check current page array */
8424 for (i = 0; i < nr_pages; i++) {
8425 if (!PageCompound(pages[i]))
8427 if (compound_head(pages[i]) == hpage)
8431 /* check previously registered pages */
8432 for (i = 0; i < ctx->nr_user_bufs; i++) {
8433 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8435 for (j = 0; j < imu->nr_bvecs; j++) {
8436 if (!PageCompound(imu->bvec[j].bv_page))
8438 if (compound_head(imu->bvec[j].bv_page) == hpage)
8446 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8447 int nr_pages, struct io_mapped_ubuf *imu,
8448 struct page **last_hpage)
8452 for (i = 0; i < nr_pages; i++) {
8453 if (!PageCompound(pages[i])) {
8458 hpage = compound_head(pages[i]);
8459 if (hpage == *last_hpage)
8461 *last_hpage = hpage;
8462 if (headpage_already_acct(ctx, pages, i, hpage))
8464 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8468 if (!imu->acct_pages)
8471 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8473 imu->acct_pages = 0;
8477 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8480 struct vm_area_struct **vmas = NULL;
8481 struct page **pages = NULL;
8482 struct page *last_hpage = NULL;
8483 int i, j, got_pages = 0;
8488 if (!nr_args || nr_args > UIO_MAXIOV)
8491 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8493 if (!ctx->user_bufs)
8496 for (i = 0; i < nr_args; i++) {
8497 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8498 unsigned long off, start, end, ubuf;
8503 ret = io_copy_iov(ctx, &iov, arg, i);
8508 * Don't impose further limits on the size and buffer
8509 * constraints here, we'll -EINVAL later when IO is
8510 * submitted if they are wrong.
8513 if (!iov.iov_base || !iov.iov_len)
8516 /* arbitrary limit, but we need something */
8517 if (iov.iov_len > SZ_1G)
8520 ubuf = (unsigned long) iov.iov_base;
8521 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8522 start = ubuf >> PAGE_SHIFT;
8523 nr_pages = end - start;
8526 if (!pages || nr_pages > got_pages) {
8529 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8531 vmas = kvmalloc_array(nr_pages,
8532 sizeof(struct vm_area_struct *),
8534 if (!pages || !vmas) {
8538 got_pages = nr_pages;
8541 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8548 mmap_read_lock(current->mm);
8549 pret = pin_user_pages(ubuf, nr_pages,
8550 FOLL_WRITE | FOLL_LONGTERM,
8552 if (pret == nr_pages) {
8553 /* don't support file backed memory */
8554 for (j = 0; j < nr_pages; j++) {
8555 struct vm_area_struct *vma = vmas[j];
8558 !is_file_hugepages(vma->vm_file)) {
8564 ret = pret < 0 ? pret : -EFAULT;
8566 mmap_read_unlock(current->mm);
8569 * if we did partial map, or found file backed vmas,
8570 * release any pages we did get
8573 unpin_user_pages(pages, pret);
8578 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8580 unpin_user_pages(pages, pret);
8585 off = ubuf & ~PAGE_MASK;
8587 for (j = 0; j < nr_pages; j++) {
8590 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8591 imu->bvec[j].bv_page = pages[j];
8592 imu->bvec[j].bv_len = vec_len;
8593 imu->bvec[j].bv_offset = off;
8597 /* store original address for later verification */
8599 imu->len = iov.iov_len;
8600 imu->nr_bvecs = nr_pages;
8602 ctx->nr_user_bufs++;
8610 io_sqe_buffer_unregister(ctx);
8614 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8616 __s32 __user *fds = arg;
8622 if (copy_from_user(&fd, fds, sizeof(*fds)))
8625 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8626 if (IS_ERR(ctx->cq_ev_fd)) {
8627 int ret = PTR_ERR(ctx->cq_ev_fd);
8628 ctx->cq_ev_fd = NULL;
8635 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8637 if (ctx->cq_ev_fd) {
8638 eventfd_ctx_put(ctx->cq_ev_fd);
8639 ctx->cq_ev_fd = NULL;
8646 static int __io_destroy_buffers(int id, void *p, void *data)
8648 struct io_ring_ctx *ctx = data;
8649 struct io_buffer *buf = p;
8651 __io_remove_buffers(ctx, buf, id, -1U);
8655 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8657 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8658 idr_destroy(&ctx->io_buffer_idr);
8661 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8663 io_finish_async(ctx);
8664 io_sqe_buffer_unregister(ctx);
8666 if (ctx->sqo_task) {
8667 put_task_struct(ctx->sqo_task);
8668 ctx->sqo_task = NULL;
8669 mmdrop(ctx->mm_account);
8670 ctx->mm_account = NULL;
8673 #ifdef CONFIG_BLK_CGROUP
8674 if (ctx->sqo_blkcg_css)
8675 css_put(ctx->sqo_blkcg_css);
8678 io_sqe_files_unregister(ctx);
8679 io_eventfd_unregister(ctx);
8680 io_destroy_buffers(ctx);
8681 idr_destroy(&ctx->personality_idr);
8683 #if defined(CONFIG_UNIX)
8684 if (ctx->ring_sock) {
8685 ctx->ring_sock->file = NULL; /* so that iput() is called */
8686 sock_release(ctx->ring_sock);
8690 io_mem_free(ctx->rings);
8691 io_mem_free(ctx->sq_sqes);
8693 percpu_ref_exit(&ctx->refs);
8694 free_uid(ctx->user);
8695 put_cred(ctx->creds);
8696 kfree(ctx->cancel_hash);
8697 kmem_cache_free(req_cachep, ctx->fallback_req);
8701 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8703 struct io_ring_ctx *ctx = file->private_data;
8706 poll_wait(file, &ctx->cq_wait, wait);
8708 * synchronizes with barrier from wq_has_sleeper call in
8712 if (!io_sqring_full(ctx))
8713 mask |= EPOLLOUT | EPOLLWRNORM;
8714 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8715 if (io_cqring_events(ctx))
8716 mask |= EPOLLIN | EPOLLRDNORM;
8721 static int io_uring_fasync(int fd, struct file *file, int on)
8723 struct io_ring_ctx *ctx = file->private_data;
8725 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8728 static int io_remove_personalities(int id, void *p, void *data)
8730 struct io_ring_ctx *ctx = data;
8731 struct io_identity *iod;
8733 iod = idr_remove(&ctx->personality_idr, id);
8735 put_cred(iod->creds);
8736 if (refcount_dec_and_test(&iod->count))
8742 static void io_ring_exit_work(struct work_struct *work)
8744 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8748 * If we're doing polled IO and end up having requests being
8749 * submitted async (out-of-line), then completions can come in while
8750 * we're waiting for refs to drop. We need to reap these manually,
8751 * as nobody else will be looking for them.
8754 __io_uring_cancel_task_requests(ctx, NULL);
8755 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8756 io_ring_ctx_free(ctx);
8759 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8761 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8763 return req->ctx == data;
8766 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8768 mutex_lock(&ctx->uring_lock);
8769 percpu_ref_kill(&ctx->refs);
8771 if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead))
8774 /* if force is set, the ring is going away. always drop after that */
8775 ctx->cq_overflow_flushed = 1;
8777 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8778 mutex_unlock(&ctx->uring_lock);
8780 io_kill_timeouts(ctx, NULL, NULL);
8781 io_poll_remove_all(ctx, NULL, NULL);
8784 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8786 /* if we failed setting up the ctx, we might not have any rings */
8787 io_iopoll_try_reap_events(ctx);
8788 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8791 * Do this upfront, so we won't have a grace period where the ring
8792 * is closed but resources aren't reaped yet. This can cause
8793 * spurious failure in setting up a new ring.
8795 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8798 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8800 * Use system_unbound_wq to avoid spawning tons of event kworkers
8801 * if we're exiting a ton of rings at the same time. It just adds
8802 * noise and overhead, there's no discernable change in runtime
8803 * over using system_wq.
8805 queue_work(system_unbound_wq, &ctx->exit_work);
8808 static int io_uring_release(struct inode *inode, struct file *file)
8810 struct io_ring_ctx *ctx = file->private_data;
8812 file->private_data = NULL;
8813 io_ring_ctx_wait_and_kill(ctx);
8817 struct io_task_cancel {
8818 struct task_struct *task;
8819 struct files_struct *files;
8822 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8824 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8825 struct io_task_cancel *cancel = data;
8828 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8829 unsigned long flags;
8830 struct io_ring_ctx *ctx = req->ctx;
8832 /* protect against races with linked timeouts */
8833 spin_lock_irqsave(&ctx->completion_lock, flags);
8834 ret = io_match_task(req, cancel->task, cancel->files);
8835 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8837 ret = io_match_task(req, cancel->task, cancel->files);
8842 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8843 struct task_struct *task,
8844 struct files_struct *files)
8846 struct io_defer_entry *de = NULL;
8849 spin_lock_irq(&ctx->completion_lock);
8850 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8851 if (io_match_task(de->req, task, files)) {
8852 list_cut_position(&list, &ctx->defer_list, &de->list);
8856 spin_unlock_irq(&ctx->completion_lock);
8858 while (!list_empty(&list)) {
8859 de = list_first_entry(&list, struct io_defer_entry, list);
8860 list_del_init(&de->list);
8861 req_set_fail_links(de->req);
8862 io_put_req(de->req);
8863 io_req_complete(de->req, -ECANCELED);
8868 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8869 struct task_struct *task,
8870 struct files_struct *files)
8872 while (!list_empty_careful(&ctx->inflight_list)) {
8873 struct io_task_cancel cancel = { .task = task, .files = files };
8874 struct io_kiocb *req;
8878 spin_lock_irq(&ctx->inflight_lock);
8879 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8880 if (!io_match_task(req, task, files))
8886 prepare_to_wait(&task->io_uring->wait, &wait,
8887 TASK_UNINTERRUPTIBLE);
8888 spin_unlock_irq(&ctx->inflight_lock);
8890 /* We need to keep going until we don't find a matching req */
8894 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8895 io_poll_remove_all(ctx, task, files);
8896 io_kill_timeouts(ctx, task, files);
8897 io_cqring_overflow_flush(ctx, true, task, files);
8898 /* cancellations _may_ trigger task work */
8901 finish_wait(&task->io_uring->wait, &wait);
8905 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8906 struct task_struct *task)
8909 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8910 enum io_wq_cancel cret;
8914 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb,
8916 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8919 /* SQPOLL thread does its own polling */
8920 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8921 while (!list_empty_careful(&ctx->iopoll_list)) {
8922 io_iopoll_try_reap_events(ctx);
8927 ret |= io_poll_remove_all(ctx, task, NULL);
8928 ret |= io_kill_timeouts(ctx, task, NULL);
8929 ret |= io_run_task_work();
8936 static void io_disable_sqo_submit(struct io_ring_ctx *ctx)
8938 mutex_lock(&ctx->uring_lock);
8940 mutex_unlock(&ctx->uring_lock);
8942 /* make sure callers enter the ring to get error */
8944 io_ring_set_wakeup_flag(ctx);
8948 * We need to iteratively cancel requests, in case a request has dependent
8949 * hard links. These persist even for failure of cancelations, hence keep
8950 * looping until none are found.
8952 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8953 struct files_struct *files)
8955 struct task_struct *task = current;
8957 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8958 /* for SQPOLL only sqo_task has task notes */
8959 WARN_ON_ONCE(ctx->sqo_task != current);
8960 io_disable_sqo_submit(ctx);
8961 task = ctx->sq_data->thread;
8962 atomic_inc(&task->io_uring->in_idle);
8963 io_sq_thread_park(ctx->sq_data);
8966 io_cancel_defer_files(ctx, task, files);
8967 io_cqring_overflow_flush(ctx, true, task, files);
8970 __io_uring_cancel_task_requests(ctx, task);
8972 io_uring_cancel_files(ctx, task, files);
8974 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8975 atomic_dec(&task->io_uring->in_idle);
8977 * If the files that are going away are the ones in the thread
8978 * identity, clear them out.
8980 if (task->io_uring->identity->files == files)
8981 task->io_uring->identity->files = NULL;
8982 io_sq_thread_unpark(ctx->sq_data);
8987 * Note that this task has used io_uring. We use it for cancelation purposes.
8989 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8991 struct io_uring_task *tctx = current->io_uring;
8994 if (unlikely(!tctx)) {
8995 ret = io_uring_alloc_task_context(current);
8998 tctx = current->io_uring;
9000 if (tctx->last != file) {
9001 void *old = xa_load(&tctx->xa, (unsigned long)file);
9005 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
9016 * This is race safe in that the task itself is doing this, hence it
9017 * cannot be going through the exit/cancel paths at the same time.
9018 * This cannot be modified while exit/cancel is running.
9020 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
9021 tctx->sqpoll = true;
9027 * Remove this io_uring_file -> task mapping.
9029 static void io_uring_del_task_file(struct file *file)
9031 struct io_uring_task *tctx = current->io_uring;
9033 if (tctx->last == file)
9035 file = xa_erase(&tctx->xa, (unsigned long)file);
9040 static void io_uring_remove_task_files(struct io_uring_task *tctx)
9043 unsigned long index;
9045 xa_for_each(&tctx->xa, index, file)
9046 io_uring_del_task_file(file);
9049 void __io_uring_files_cancel(struct files_struct *files)
9051 struct io_uring_task *tctx = current->io_uring;
9053 unsigned long index;
9055 /* make sure overflow events are dropped */
9056 atomic_inc(&tctx->in_idle);
9057 xa_for_each(&tctx->xa, index, file)
9058 io_uring_cancel_task_requests(file->private_data, files);
9059 atomic_dec(&tctx->in_idle);
9062 io_uring_remove_task_files(tctx);
9065 static s64 tctx_inflight(struct io_uring_task *tctx)
9067 unsigned long index;
9071 inflight = percpu_counter_sum(&tctx->inflight);
9076 * If we have SQPOLL rings, then we need to iterate and find them, and
9077 * add the pending count for those.
9079 xa_for_each(&tctx->xa, index, file) {
9080 struct io_ring_ctx *ctx = file->private_data;
9082 if (ctx->flags & IORING_SETUP_SQPOLL) {
9083 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
9085 inflight += percpu_counter_sum(&__tctx->inflight);
9093 * Find any io_uring fd that this task has registered or done IO on, and cancel
9096 void __io_uring_task_cancel(void)
9098 struct io_uring_task *tctx = current->io_uring;
9102 /* make sure overflow events are dropped */
9103 atomic_inc(&tctx->in_idle);
9105 /* trigger io_disable_sqo_submit() */
9107 __io_uring_files_cancel(NULL);
9110 /* read completions before cancelations */
9111 inflight = tctx_inflight(tctx);
9114 __io_uring_files_cancel(NULL);
9116 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9119 * If we've seen completions, retry. This avoids a race where
9120 * a completion comes in before we did prepare_to_wait().
9122 if (inflight != tctx_inflight(tctx))
9125 finish_wait(&tctx->wait, &wait);
9128 finish_wait(&tctx->wait, &wait);
9129 atomic_dec(&tctx->in_idle);
9131 io_uring_remove_task_files(tctx);
9134 static int io_uring_flush(struct file *file, void *data)
9136 struct io_uring_task *tctx = current->io_uring;
9137 struct io_ring_ctx *ctx = file->private_data;
9142 /* we should have cancelled and erased it before PF_EXITING */
9143 WARN_ON_ONCE((current->flags & PF_EXITING) &&
9144 xa_load(&tctx->xa, (unsigned long)file));
9147 * fput() is pending, will be 2 if the only other ref is our potential
9148 * task file note. If the task is exiting, drop regardless of count.
9150 if (atomic_long_read(&file->f_count) != 2)
9153 if (ctx->flags & IORING_SETUP_SQPOLL) {
9154 /* there is only one file note, which is owned by sqo_task */
9155 WARN_ON_ONCE(ctx->sqo_task != current &&
9156 xa_load(&tctx->xa, (unsigned long)file));
9157 /* sqo_dead check is for when this happens after cancellation */
9158 WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead &&
9159 !xa_load(&tctx->xa, (unsigned long)file));
9161 io_disable_sqo_submit(ctx);
9164 if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current)
9165 io_uring_del_task_file(file);
9169 static void *io_uring_validate_mmap_request(struct file *file,
9170 loff_t pgoff, size_t sz)
9172 struct io_ring_ctx *ctx = file->private_data;
9173 loff_t offset = pgoff << PAGE_SHIFT;
9178 case IORING_OFF_SQ_RING:
9179 case IORING_OFF_CQ_RING:
9182 case IORING_OFF_SQES:
9186 return ERR_PTR(-EINVAL);
9189 page = virt_to_head_page(ptr);
9190 if (sz > page_size(page))
9191 return ERR_PTR(-EINVAL);
9198 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9200 size_t sz = vma->vm_end - vma->vm_start;
9204 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9206 return PTR_ERR(ptr);
9208 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9209 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9212 #else /* !CONFIG_MMU */
9214 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9216 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9219 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9221 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9224 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9225 unsigned long addr, unsigned long len,
9226 unsigned long pgoff, unsigned long flags)
9230 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9232 return PTR_ERR(ptr);
9234 return (unsigned long) ptr;
9237 #endif /* !CONFIG_MMU */
9239 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9245 if (!io_sqring_full(ctx))
9248 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9250 if (unlikely(ctx->sqo_dead)) {
9255 if (!io_sqring_full(ctx))
9259 } while (!signal_pending(current));
9261 finish_wait(&ctx->sqo_sq_wait, &wait);
9266 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9267 struct __kernel_timespec __user **ts,
9268 const sigset_t __user **sig)
9270 struct io_uring_getevents_arg arg;
9273 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9274 * is just a pointer to the sigset_t.
9276 if (!(flags & IORING_ENTER_EXT_ARG)) {
9277 *sig = (const sigset_t __user *) argp;
9283 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9284 * timespec and sigset_t pointers if good.
9286 if (*argsz != sizeof(arg))
9288 if (copy_from_user(&arg, argp, sizeof(arg)))
9290 *sig = u64_to_user_ptr(arg.sigmask);
9291 *argsz = arg.sigmask_sz;
9292 *ts = u64_to_user_ptr(arg.ts);
9296 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9297 u32, min_complete, u32, flags, const void __user *, argp,
9300 struct io_ring_ctx *ctx;
9307 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9308 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9316 if (f.file->f_op != &io_uring_fops)
9320 ctx = f.file->private_data;
9321 if (!percpu_ref_tryget(&ctx->refs))
9325 if (ctx->flags & IORING_SETUP_R_DISABLED)
9329 * For SQ polling, the thread will do all submissions and completions.
9330 * Just return the requested submit count, and wake the thread if
9334 if (ctx->flags & IORING_SETUP_SQPOLL) {
9335 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9338 if (unlikely(ctx->sqo_dead))
9340 if (flags & IORING_ENTER_SQ_WAKEUP)
9341 wake_up(&ctx->sq_data->wait);
9342 if (flags & IORING_ENTER_SQ_WAIT) {
9343 ret = io_sqpoll_wait_sq(ctx);
9347 submitted = to_submit;
9348 } else if (to_submit) {
9349 ret = io_uring_add_task_file(ctx, f.file);
9352 mutex_lock(&ctx->uring_lock);
9353 submitted = io_submit_sqes(ctx, to_submit);
9354 mutex_unlock(&ctx->uring_lock);
9356 if (submitted != to_submit)
9359 if (flags & IORING_ENTER_GETEVENTS) {
9360 const sigset_t __user *sig;
9361 struct __kernel_timespec __user *ts;
9363 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9367 min_complete = min(min_complete, ctx->cq_entries);
9370 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9371 * space applications don't need to do io completion events
9372 * polling again, they can rely on io_sq_thread to do polling
9373 * work, which can reduce cpu usage and uring_lock contention.
9375 if (ctx->flags & IORING_SETUP_IOPOLL &&
9376 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9377 ret = io_iopoll_check(ctx, min_complete);
9379 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9384 percpu_ref_put(&ctx->refs);
9387 return submitted ? submitted : ret;
9390 #ifdef CONFIG_PROC_FS
9391 static int io_uring_show_cred(int id, void *p, void *data)
9393 struct io_identity *iod = p;
9394 const struct cred *cred = iod->creds;
9395 struct seq_file *m = data;
9396 struct user_namespace *uns = seq_user_ns(m);
9397 struct group_info *gi;
9402 seq_printf(m, "%5d\n", id);
9403 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9404 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9405 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9406 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9407 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9408 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9409 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9410 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9411 seq_puts(m, "\n\tGroups:\t");
9412 gi = cred->group_info;
9413 for (g = 0; g < gi->ngroups; g++) {
9414 seq_put_decimal_ull(m, g ? " " : "",
9415 from_kgid_munged(uns, gi->gid[g]));
9417 seq_puts(m, "\n\tCapEff:\t");
9418 cap = cred->cap_effective;
9419 CAP_FOR_EACH_U32(__capi)
9420 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9425 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9427 struct io_sq_data *sq = NULL;
9432 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9433 * since fdinfo case grabs it in the opposite direction of normal use
9434 * cases. If we fail to get the lock, we just don't iterate any
9435 * structures that could be going away outside the io_uring mutex.
9437 has_lock = mutex_trylock(&ctx->uring_lock);
9439 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9442 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9443 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9444 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9445 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9446 struct fixed_file_table *table;
9449 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9450 f = table->files[i & IORING_FILE_TABLE_MASK];
9452 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9454 seq_printf(m, "%5u: <none>\n", i);
9456 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9457 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9458 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9460 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9461 (unsigned int) buf->len);
9463 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9464 seq_printf(m, "Personalities:\n");
9465 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9467 seq_printf(m, "PollList:\n");
9468 spin_lock_irq(&ctx->completion_lock);
9469 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9470 struct hlist_head *list = &ctx->cancel_hash[i];
9471 struct io_kiocb *req;
9473 hlist_for_each_entry(req, list, hash_node)
9474 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9475 req->task->task_works != NULL);
9477 spin_unlock_irq(&ctx->completion_lock);
9479 mutex_unlock(&ctx->uring_lock);
9482 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9484 struct io_ring_ctx *ctx = f->private_data;
9486 if (percpu_ref_tryget(&ctx->refs)) {
9487 __io_uring_show_fdinfo(ctx, m);
9488 percpu_ref_put(&ctx->refs);
9493 static const struct file_operations io_uring_fops = {
9494 .release = io_uring_release,
9495 .flush = io_uring_flush,
9496 .mmap = io_uring_mmap,
9498 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9499 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9501 .poll = io_uring_poll,
9502 .fasync = io_uring_fasync,
9503 #ifdef CONFIG_PROC_FS
9504 .show_fdinfo = io_uring_show_fdinfo,
9508 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9509 struct io_uring_params *p)
9511 struct io_rings *rings;
9512 size_t size, sq_array_offset;
9514 /* make sure these are sane, as we already accounted them */
9515 ctx->sq_entries = p->sq_entries;
9516 ctx->cq_entries = p->cq_entries;
9518 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9519 if (size == SIZE_MAX)
9522 rings = io_mem_alloc(size);
9527 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9528 rings->sq_ring_mask = p->sq_entries - 1;
9529 rings->cq_ring_mask = p->cq_entries - 1;
9530 rings->sq_ring_entries = p->sq_entries;
9531 rings->cq_ring_entries = p->cq_entries;
9532 ctx->sq_mask = rings->sq_ring_mask;
9533 ctx->cq_mask = rings->cq_ring_mask;
9535 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9536 if (size == SIZE_MAX) {
9537 io_mem_free(ctx->rings);
9542 ctx->sq_sqes = io_mem_alloc(size);
9543 if (!ctx->sq_sqes) {
9544 io_mem_free(ctx->rings);
9552 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9556 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9560 ret = io_uring_add_task_file(ctx, file);
9565 fd_install(fd, file);
9570 * Allocate an anonymous fd, this is what constitutes the application
9571 * visible backing of an io_uring instance. The application mmaps this
9572 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9573 * we have to tie this fd to a socket for file garbage collection purposes.
9575 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9578 #if defined(CONFIG_UNIX)
9581 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9584 return ERR_PTR(ret);
9587 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9588 O_RDWR | O_CLOEXEC);
9589 #if defined(CONFIG_UNIX)
9591 sock_release(ctx->ring_sock);
9592 ctx->ring_sock = NULL;
9594 ctx->ring_sock->file = file;
9600 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9601 struct io_uring_params __user *params)
9603 struct user_struct *user = NULL;
9604 struct io_ring_ctx *ctx;
9611 if (entries > IORING_MAX_ENTRIES) {
9612 if (!(p->flags & IORING_SETUP_CLAMP))
9614 entries = IORING_MAX_ENTRIES;
9618 * Use twice as many entries for the CQ ring. It's possible for the
9619 * application to drive a higher depth than the size of the SQ ring,
9620 * since the sqes are only used at submission time. This allows for
9621 * some flexibility in overcommitting a bit. If the application has
9622 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9623 * of CQ ring entries manually.
9625 p->sq_entries = roundup_pow_of_two(entries);
9626 if (p->flags & IORING_SETUP_CQSIZE) {
9628 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9629 * to a power-of-two, if it isn't already. We do NOT impose
9630 * any cq vs sq ring sizing.
9634 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9635 if (!(p->flags & IORING_SETUP_CLAMP))
9637 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9639 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9640 if (p->cq_entries < p->sq_entries)
9643 p->cq_entries = 2 * p->sq_entries;
9646 user = get_uid(current_user());
9647 limit_mem = !capable(CAP_IPC_LOCK);
9650 ret = __io_account_mem(user,
9651 ring_pages(p->sq_entries, p->cq_entries));
9658 ctx = io_ring_ctx_alloc(p);
9661 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9666 ctx->compat = in_compat_syscall();
9668 ctx->creds = get_current_cred();
9670 ctx->loginuid = current->loginuid;
9671 ctx->sessionid = current->sessionid;
9673 ctx->sqo_task = get_task_struct(current);
9676 * This is just grabbed for accounting purposes. When a process exits,
9677 * the mm is exited and dropped before the files, hence we need to hang
9678 * on to this mm purely for the purposes of being able to unaccount
9679 * memory (locked/pinned vm). It's not used for anything else.
9681 mmgrab(current->mm);
9682 ctx->mm_account = current->mm;
9684 #ifdef CONFIG_BLK_CGROUP
9686 * The sq thread will belong to the original cgroup it was inited in.
9687 * If the cgroup goes offline (e.g. disabling the io controller), then
9688 * issued bios will be associated with the closest cgroup later in the
9692 ctx->sqo_blkcg_css = blkcg_css();
9693 ret = css_tryget_online(ctx->sqo_blkcg_css);
9696 /* don't init against a dying cgroup, have the user try again */
9697 ctx->sqo_blkcg_css = NULL;
9704 * Account memory _before_ installing the file descriptor. Once
9705 * the descriptor is installed, it can get closed at any time. Also
9706 * do this before hitting the general error path, as ring freeing
9707 * will un-account as well.
9709 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9711 ctx->limit_mem = limit_mem;
9713 ret = io_allocate_scq_urings(ctx, p);
9717 ret = io_sq_offload_create(ctx, p);
9721 if (!(p->flags & IORING_SETUP_R_DISABLED))
9722 io_sq_offload_start(ctx);
9724 memset(&p->sq_off, 0, sizeof(p->sq_off));
9725 p->sq_off.head = offsetof(struct io_rings, sq.head);
9726 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9727 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9728 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9729 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9730 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9731 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9733 memset(&p->cq_off, 0, sizeof(p->cq_off));
9734 p->cq_off.head = offsetof(struct io_rings, cq.head);
9735 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9736 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9737 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9738 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9739 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9740 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9742 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9743 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9744 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9745 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9746 IORING_FEAT_EXT_ARG;
9748 if (copy_to_user(params, p, sizeof(*p))) {
9753 file = io_uring_get_file(ctx);
9755 ret = PTR_ERR(file);
9760 * Install ring fd as the very last thing, so we don't risk someone
9761 * having closed it before we finish setup
9763 ret = io_uring_install_fd(ctx, file);
9765 io_disable_sqo_submit(ctx);
9766 /* fput will clean it up */
9771 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9774 io_disable_sqo_submit(ctx);
9775 io_ring_ctx_wait_and_kill(ctx);
9780 * Sets up an aio uring context, and returns the fd. Applications asks for a
9781 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9782 * params structure passed in.
9784 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9786 struct io_uring_params p;
9789 if (copy_from_user(&p, params, sizeof(p)))
9791 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9796 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9797 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9798 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9799 IORING_SETUP_R_DISABLED))
9802 return io_uring_create(entries, &p, params);
9805 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9806 struct io_uring_params __user *, params)
9808 return io_uring_setup(entries, params);
9811 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9813 struct io_uring_probe *p;
9817 size = struct_size(p, ops, nr_args);
9818 if (size == SIZE_MAX)
9820 p = kzalloc(size, GFP_KERNEL);
9825 if (copy_from_user(p, arg, size))
9828 if (memchr_inv(p, 0, size))
9831 p->last_op = IORING_OP_LAST - 1;
9832 if (nr_args > IORING_OP_LAST)
9833 nr_args = IORING_OP_LAST;
9835 for (i = 0; i < nr_args; i++) {
9837 if (!io_op_defs[i].not_supported)
9838 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9843 if (copy_to_user(arg, p, size))
9850 static int io_register_personality(struct io_ring_ctx *ctx)
9852 struct io_identity *id;
9855 id = kmalloc(sizeof(*id), GFP_KERNEL);
9859 io_init_identity(id);
9860 id->creds = get_current_cred();
9862 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9864 put_cred(id->creds);
9870 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9872 struct io_identity *iod;
9874 iod = idr_remove(&ctx->personality_idr, id);
9876 put_cred(iod->creds);
9877 if (refcount_dec_and_test(&iod->count))
9885 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9886 unsigned int nr_args)
9888 struct io_uring_restriction *res;
9892 /* Restrictions allowed only if rings started disabled */
9893 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9896 /* We allow only a single restrictions registration */
9897 if (ctx->restrictions.registered)
9900 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9903 size = array_size(nr_args, sizeof(*res));
9904 if (size == SIZE_MAX)
9907 res = memdup_user(arg, size);
9909 return PTR_ERR(res);
9913 for (i = 0; i < nr_args; i++) {
9914 switch (res[i].opcode) {
9915 case IORING_RESTRICTION_REGISTER_OP:
9916 if (res[i].register_op >= IORING_REGISTER_LAST) {
9921 __set_bit(res[i].register_op,
9922 ctx->restrictions.register_op);
9924 case IORING_RESTRICTION_SQE_OP:
9925 if (res[i].sqe_op >= IORING_OP_LAST) {
9930 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9932 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9933 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9935 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9936 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9945 /* Reset all restrictions if an error happened */
9947 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9949 ctx->restrictions.registered = true;
9955 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9957 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9960 if (ctx->restrictions.registered)
9961 ctx->restricted = 1;
9963 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9965 io_sq_offload_start(ctx);
9970 static bool io_register_op_must_quiesce(int op)
9973 case IORING_UNREGISTER_FILES:
9974 case IORING_REGISTER_FILES_UPDATE:
9975 case IORING_REGISTER_PROBE:
9976 case IORING_REGISTER_PERSONALITY:
9977 case IORING_UNREGISTER_PERSONALITY:
9984 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9985 void __user *arg, unsigned nr_args)
9986 __releases(ctx->uring_lock)
9987 __acquires(ctx->uring_lock)
9992 * We're inside the ring mutex, if the ref is already dying, then
9993 * someone else killed the ctx or is already going through
9994 * io_uring_register().
9996 if (percpu_ref_is_dying(&ctx->refs))
9999 if (io_register_op_must_quiesce(opcode)) {
10000 percpu_ref_kill(&ctx->refs);
10003 * Drop uring mutex before waiting for references to exit. If
10004 * another thread is currently inside io_uring_enter() it might
10005 * need to grab the uring_lock to make progress. If we hold it
10006 * here across the drain wait, then we can deadlock. It's safe
10007 * to drop the mutex here, since no new references will come in
10008 * after we've killed the percpu ref.
10010 mutex_unlock(&ctx->uring_lock);
10012 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10015 ret = io_run_task_work_sig();
10020 mutex_lock(&ctx->uring_lock);
10023 percpu_ref_resurrect(&ctx->refs);
10028 if (ctx->restricted) {
10029 if (opcode >= IORING_REGISTER_LAST) {
10034 if (!test_bit(opcode, ctx->restrictions.register_op)) {
10041 case IORING_REGISTER_BUFFERS:
10042 ret = io_sqe_buffer_register(ctx, arg, nr_args);
10044 case IORING_UNREGISTER_BUFFERS:
10046 if (arg || nr_args)
10048 ret = io_sqe_buffer_unregister(ctx);
10050 case IORING_REGISTER_FILES:
10051 ret = io_sqe_files_register(ctx, arg, nr_args);
10053 case IORING_UNREGISTER_FILES:
10055 if (arg || nr_args)
10057 ret = io_sqe_files_unregister(ctx);
10059 case IORING_REGISTER_FILES_UPDATE:
10060 ret = io_sqe_files_update(ctx, arg, nr_args);
10062 case IORING_REGISTER_EVENTFD:
10063 case IORING_REGISTER_EVENTFD_ASYNC:
10067 ret = io_eventfd_register(ctx, arg);
10070 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10071 ctx->eventfd_async = 1;
10073 ctx->eventfd_async = 0;
10075 case IORING_UNREGISTER_EVENTFD:
10077 if (arg || nr_args)
10079 ret = io_eventfd_unregister(ctx);
10081 case IORING_REGISTER_PROBE:
10083 if (!arg || nr_args > 256)
10085 ret = io_probe(ctx, arg, nr_args);
10087 case IORING_REGISTER_PERSONALITY:
10089 if (arg || nr_args)
10091 ret = io_register_personality(ctx);
10093 case IORING_UNREGISTER_PERSONALITY:
10097 ret = io_unregister_personality(ctx, nr_args);
10099 case IORING_REGISTER_ENABLE_RINGS:
10101 if (arg || nr_args)
10103 ret = io_register_enable_rings(ctx);
10105 case IORING_REGISTER_RESTRICTIONS:
10106 ret = io_register_restrictions(ctx, arg, nr_args);
10114 if (io_register_op_must_quiesce(opcode)) {
10115 /* bring the ctx back to life */
10116 percpu_ref_reinit(&ctx->refs);
10118 reinit_completion(&ctx->ref_comp);
10123 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10124 void __user *, arg, unsigned int, nr_args)
10126 struct io_ring_ctx *ctx;
10135 if (f.file->f_op != &io_uring_fops)
10138 ctx = f.file->private_data;
10140 mutex_lock(&ctx->uring_lock);
10141 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10142 mutex_unlock(&ctx->uring_lock);
10143 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10144 ctx->cq_ev_fd != NULL, ret);
10150 static int __init io_uring_init(void)
10152 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10153 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10154 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10157 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10158 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10159 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10160 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10161 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10162 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10163 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10164 BUILD_BUG_SQE_ELEM(8, __u64, off);
10165 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10166 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10167 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10168 BUILD_BUG_SQE_ELEM(24, __u32, len);
10169 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10170 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10171 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10172 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10173 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10174 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10175 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10176 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10177 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10178 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10179 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10180 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10181 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10182 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10183 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10184 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10185 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10186 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10187 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10189 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10190 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10191 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10194 __initcall(io_uring_init);