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 |
863 [IORING_OP_RECVMSG] = {
865 .unbound_nonreg_file = 1,
868 .needs_async_data = 1,
869 .async_size = sizeof(struct io_async_msghdr),
870 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
873 [IORING_OP_TIMEOUT] = {
874 .needs_async_data = 1,
875 .async_size = sizeof(struct io_timeout_data),
876 .work_flags = IO_WQ_WORK_MM,
878 [IORING_OP_TIMEOUT_REMOVE] = {
879 /* used by timeout updates' prep() */
880 .work_flags = IO_WQ_WORK_MM,
882 [IORING_OP_ACCEPT] = {
884 .unbound_nonreg_file = 1,
886 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
888 [IORING_OP_ASYNC_CANCEL] = {},
889 [IORING_OP_LINK_TIMEOUT] = {
890 .needs_async_data = 1,
891 .async_size = sizeof(struct io_timeout_data),
892 .work_flags = IO_WQ_WORK_MM,
894 [IORING_OP_CONNECT] = {
896 .unbound_nonreg_file = 1,
898 .needs_async_data = 1,
899 .async_size = sizeof(struct io_async_connect),
900 .work_flags = IO_WQ_WORK_MM,
902 [IORING_OP_FALLOCATE] = {
904 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
906 [IORING_OP_OPENAT] = {
907 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
908 IO_WQ_WORK_FS | IO_WQ_WORK_MM,
910 [IORING_OP_CLOSE] = {
912 .needs_file_no_error = 1,
913 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
915 [IORING_OP_FILES_UPDATE] = {
916 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
918 [IORING_OP_STATX] = {
919 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
920 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
924 .unbound_nonreg_file = 1,
928 .async_size = sizeof(struct io_async_rw),
929 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
931 [IORING_OP_WRITE] = {
933 .unbound_nonreg_file = 1,
936 .async_size = sizeof(struct io_async_rw),
937 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
940 [IORING_OP_FADVISE] = {
942 .work_flags = IO_WQ_WORK_BLKCG,
944 [IORING_OP_MADVISE] = {
945 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
949 .unbound_nonreg_file = 1,
951 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
955 .unbound_nonreg_file = 1,
958 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
960 [IORING_OP_OPENAT2] = {
961 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
962 IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
964 [IORING_OP_EPOLL_CTL] = {
965 .unbound_nonreg_file = 1,
966 .work_flags = IO_WQ_WORK_FILES,
968 [IORING_OP_SPLICE] = {
971 .unbound_nonreg_file = 1,
972 .work_flags = IO_WQ_WORK_BLKCG,
974 [IORING_OP_PROVIDE_BUFFERS] = {},
975 [IORING_OP_REMOVE_BUFFERS] = {},
979 .unbound_nonreg_file = 1,
981 [IORING_OP_SHUTDOWN] = {
984 [IORING_OP_RENAMEAT] = {
985 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
986 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
988 [IORING_OP_UNLINKAT] = {
989 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
990 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
994 enum io_mem_account {
999 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
1000 struct task_struct *task);
1002 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node);
1003 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
1004 struct io_ring_ctx *ctx);
1006 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
1007 struct io_comp_state *cs);
1008 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1009 static void io_put_req(struct io_kiocb *req);
1010 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1011 static void io_double_put_req(struct io_kiocb *req);
1012 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1013 static void __io_queue_linked_timeout(struct io_kiocb *req);
1014 static void io_queue_linked_timeout(struct io_kiocb *req);
1015 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1016 struct io_uring_files_update *ip,
1018 static void __io_clean_op(struct io_kiocb *req);
1019 static struct file *io_file_get(struct io_submit_state *state,
1020 struct io_kiocb *req, int fd, bool fixed);
1021 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
1022 static void io_file_put_work(struct work_struct *work);
1024 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1025 struct iovec **iovec, struct iov_iter *iter,
1027 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1028 const struct iovec *fast_iov,
1029 struct iov_iter *iter, bool force);
1030 static void io_req_drop_files(struct io_kiocb *req);
1031 static void io_req_task_queue(struct io_kiocb *req);
1033 static struct kmem_cache *req_cachep;
1035 static const struct file_operations io_uring_fops;
1037 struct sock *io_uring_get_socket(struct file *file)
1039 #if defined(CONFIG_UNIX)
1040 if (file->f_op == &io_uring_fops) {
1041 struct io_ring_ctx *ctx = file->private_data;
1043 return ctx->ring_sock->sk;
1048 EXPORT_SYMBOL(io_uring_get_socket);
1050 #define io_for_each_link(pos, head) \
1051 for (pos = (head); pos; pos = pos->link)
1053 static inline void io_clean_op(struct io_kiocb *req)
1055 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1059 static inline void io_set_resource_node(struct io_kiocb *req)
1061 struct io_ring_ctx *ctx = req->ctx;
1063 if (!req->fixed_file_refs) {
1064 req->fixed_file_refs = &ctx->file_data->node->refs;
1065 percpu_ref_get(req->fixed_file_refs);
1069 static bool io_match_task(struct io_kiocb *head,
1070 struct task_struct *task,
1071 struct files_struct *files)
1073 struct io_kiocb *req;
1075 if (task && head->task != task) {
1076 /* in terms of cancelation, always match if req task is dead */
1077 if (head->task->flags & PF_EXITING)
1084 io_for_each_link(req, head) {
1085 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1087 if (req->file && req->file->f_op == &io_uring_fops)
1089 if ((req->work.flags & IO_WQ_WORK_FILES) &&
1090 req->work.identity->files == files)
1096 static void io_sq_thread_drop_mm_files(void)
1098 struct files_struct *files = current->files;
1099 struct mm_struct *mm = current->mm;
1102 kthread_unuse_mm(mm);
1107 struct nsproxy *nsproxy = current->nsproxy;
1110 current->files = NULL;
1111 current->nsproxy = NULL;
1112 task_unlock(current);
1113 put_files_struct(files);
1114 put_nsproxy(nsproxy);
1118 static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1120 if (current->flags & PF_EXITING)
1123 if (!current->files) {
1124 struct files_struct *files;
1125 struct nsproxy *nsproxy;
1127 task_lock(ctx->sqo_task);
1128 files = ctx->sqo_task->files;
1130 task_unlock(ctx->sqo_task);
1133 atomic_inc(&files->count);
1134 get_nsproxy(ctx->sqo_task->nsproxy);
1135 nsproxy = ctx->sqo_task->nsproxy;
1136 task_unlock(ctx->sqo_task);
1139 current->files = files;
1140 current->nsproxy = nsproxy;
1141 task_unlock(current);
1146 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1148 struct mm_struct *mm;
1150 if (current->flags & PF_EXITING)
1155 /* Should never happen */
1156 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1159 task_lock(ctx->sqo_task);
1160 mm = ctx->sqo_task->mm;
1161 if (unlikely(!mm || !mmget_not_zero(mm)))
1163 task_unlock(ctx->sqo_task);
1173 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1174 struct io_kiocb *req)
1176 const struct io_op_def *def = &io_op_defs[req->opcode];
1179 if (def->work_flags & IO_WQ_WORK_MM) {
1180 ret = __io_sq_thread_acquire_mm(ctx);
1185 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) {
1186 ret = __io_sq_thread_acquire_files(ctx);
1194 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1195 struct cgroup_subsys_state **cur_css)
1198 #ifdef CONFIG_BLK_CGROUP
1199 /* puts the old one when swapping */
1200 if (*cur_css != ctx->sqo_blkcg_css) {
1201 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1202 *cur_css = ctx->sqo_blkcg_css;
1207 static void io_sq_thread_unassociate_blkcg(void)
1209 #ifdef CONFIG_BLK_CGROUP
1210 kthread_associate_blkcg(NULL);
1214 static inline void req_set_fail_links(struct io_kiocb *req)
1216 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1217 req->flags |= REQ_F_FAIL_LINK;
1221 * None of these are dereferenced, they are simply used to check if any of
1222 * them have changed. If we're under current and check they are still the
1223 * same, we're fine to grab references to them for actual out-of-line use.
1225 static void io_init_identity(struct io_identity *id)
1227 id->files = current->files;
1228 id->mm = current->mm;
1229 #ifdef CONFIG_BLK_CGROUP
1231 id->blkcg_css = blkcg_css();
1234 id->creds = current_cred();
1235 id->nsproxy = current->nsproxy;
1236 id->fs = current->fs;
1237 id->fsize = rlimit(RLIMIT_FSIZE);
1239 id->loginuid = current->loginuid;
1240 id->sessionid = current->sessionid;
1242 refcount_set(&id->count, 1);
1245 static inline void __io_req_init_async(struct io_kiocb *req)
1247 memset(&req->work, 0, sizeof(req->work));
1248 req->flags |= REQ_F_WORK_INITIALIZED;
1252 * Note: must call io_req_init_async() for the first time you
1253 * touch any members of io_wq_work.
1255 static inline void io_req_init_async(struct io_kiocb *req)
1257 struct io_uring_task *tctx = current->io_uring;
1259 if (req->flags & REQ_F_WORK_INITIALIZED)
1262 __io_req_init_async(req);
1264 /* Grab a ref if this isn't our static identity */
1265 req->work.identity = tctx->identity;
1266 if (tctx->identity != &tctx->__identity)
1267 refcount_inc(&req->work.identity->count);
1270 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1272 return ctx->flags & IORING_SETUP_SQPOLL;
1275 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1277 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1279 complete(&ctx->ref_comp);
1282 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1284 return !req->timeout.off;
1287 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1289 struct io_ring_ctx *ctx;
1292 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1296 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1297 if (!ctx->fallback_req)
1301 * Use 5 bits less than the max cq entries, that should give us around
1302 * 32 entries per hash list if totally full and uniformly spread.
1304 hash_bits = ilog2(p->cq_entries);
1308 ctx->cancel_hash_bits = hash_bits;
1309 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1311 if (!ctx->cancel_hash)
1313 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1315 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1316 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1319 ctx->flags = p->flags;
1320 init_waitqueue_head(&ctx->sqo_sq_wait);
1321 INIT_LIST_HEAD(&ctx->sqd_list);
1322 init_waitqueue_head(&ctx->cq_wait);
1323 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1324 init_completion(&ctx->ref_comp);
1325 init_completion(&ctx->sq_thread_comp);
1326 idr_init(&ctx->io_buffer_idr);
1327 idr_init(&ctx->personality_idr);
1328 mutex_init(&ctx->uring_lock);
1329 init_waitqueue_head(&ctx->wait);
1330 spin_lock_init(&ctx->completion_lock);
1331 INIT_LIST_HEAD(&ctx->iopoll_list);
1332 INIT_LIST_HEAD(&ctx->defer_list);
1333 INIT_LIST_HEAD(&ctx->timeout_list);
1334 spin_lock_init(&ctx->inflight_lock);
1335 INIT_LIST_HEAD(&ctx->inflight_list);
1336 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1337 init_llist_head(&ctx->file_put_llist);
1340 if (ctx->fallback_req)
1341 kmem_cache_free(req_cachep, ctx->fallback_req);
1342 kfree(ctx->cancel_hash);
1347 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1349 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1350 struct io_ring_ctx *ctx = req->ctx;
1352 return seq != ctx->cached_cq_tail
1353 + READ_ONCE(ctx->cached_cq_overflow);
1359 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1361 struct io_rings *rings = ctx->rings;
1363 /* order cqe stores with ring update */
1364 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1367 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1369 if (req->work.identity == &tctx->__identity)
1371 if (refcount_dec_and_test(&req->work.identity->count))
1372 kfree(req->work.identity);
1375 static void io_req_clean_work(struct io_kiocb *req)
1377 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1380 req->flags &= ~REQ_F_WORK_INITIALIZED;
1382 if (req->work.flags & IO_WQ_WORK_MM) {
1383 mmdrop(req->work.identity->mm);
1384 req->work.flags &= ~IO_WQ_WORK_MM;
1386 #ifdef CONFIG_BLK_CGROUP
1387 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1388 css_put(req->work.identity->blkcg_css);
1389 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1392 if (req->work.flags & IO_WQ_WORK_CREDS) {
1393 put_cred(req->work.identity->creds);
1394 req->work.flags &= ~IO_WQ_WORK_CREDS;
1396 if (req->work.flags & IO_WQ_WORK_FS) {
1397 struct fs_struct *fs = req->work.identity->fs;
1399 spin_lock(&req->work.identity->fs->lock);
1402 spin_unlock(&req->work.identity->fs->lock);
1405 req->work.flags &= ~IO_WQ_WORK_FS;
1407 if (req->flags & REQ_F_INFLIGHT)
1408 io_req_drop_files(req);
1410 io_put_identity(req->task->io_uring, req);
1414 * Create a private copy of io_identity, since some fields don't match
1415 * the current context.
1417 static bool io_identity_cow(struct io_kiocb *req)
1419 struct io_uring_task *tctx = current->io_uring;
1420 const struct cred *creds = NULL;
1421 struct io_identity *id;
1423 if (req->work.flags & IO_WQ_WORK_CREDS)
1424 creds = req->work.identity->creds;
1426 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1427 if (unlikely(!id)) {
1428 req->work.flags |= IO_WQ_WORK_CANCEL;
1433 * We can safely just re-init the creds we copied Either the field
1434 * matches the current one, or we haven't grabbed it yet. The only
1435 * exception is ->creds, through registered personalities, so handle
1436 * that one separately.
1438 io_init_identity(id);
1442 /* add one for this request */
1443 refcount_inc(&id->count);
1445 /* drop tctx and req identity references, if needed */
1446 if (tctx->identity != &tctx->__identity &&
1447 refcount_dec_and_test(&tctx->identity->count))
1448 kfree(tctx->identity);
1449 if (req->work.identity != &tctx->__identity &&
1450 refcount_dec_and_test(&req->work.identity->count))
1451 kfree(req->work.identity);
1453 req->work.identity = id;
1454 tctx->identity = id;
1458 static bool io_grab_identity(struct io_kiocb *req)
1460 const struct io_op_def *def = &io_op_defs[req->opcode];
1461 struct io_identity *id = req->work.identity;
1462 struct io_ring_ctx *ctx = req->ctx;
1464 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1465 if (id->fsize != rlimit(RLIMIT_FSIZE))
1467 req->work.flags |= IO_WQ_WORK_FSIZE;
1469 #ifdef CONFIG_BLK_CGROUP
1470 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1471 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1473 if (id->blkcg_css != blkcg_css()) {
1478 * This should be rare, either the cgroup is dying or the task
1479 * is moving cgroups. Just punt to root for the handful of ios.
1481 if (css_tryget_online(id->blkcg_css))
1482 req->work.flags |= IO_WQ_WORK_BLKCG;
1486 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1487 if (id->creds != current_cred())
1489 get_cred(id->creds);
1490 req->work.flags |= IO_WQ_WORK_CREDS;
1493 if (!uid_eq(current->loginuid, id->loginuid) ||
1494 current->sessionid != id->sessionid)
1497 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1498 (def->work_flags & IO_WQ_WORK_FS)) {
1499 if (current->fs != id->fs)
1501 spin_lock(&id->fs->lock);
1502 if (!id->fs->in_exec) {
1504 req->work.flags |= IO_WQ_WORK_FS;
1506 req->work.flags |= IO_WQ_WORK_CANCEL;
1508 spin_unlock(¤t->fs->lock);
1510 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1511 (def->work_flags & IO_WQ_WORK_FILES) &&
1512 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1513 if (id->files != current->files ||
1514 id->nsproxy != current->nsproxy)
1516 atomic_inc(&id->files->count);
1517 get_nsproxy(id->nsproxy);
1519 if (!(req->flags & REQ_F_INFLIGHT)) {
1520 req->flags |= REQ_F_INFLIGHT;
1522 spin_lock_irq(&ctx->inflight_lock);
1523 list_add(&req->inflight_entry, &ctx->inflight_list);
1524 spin_unlock_irq(&ctx->inflight_lock);
1526 req->work.flags |= IO_WQ_WORK_FILES;
1528 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1529 (def->work_flags & IO_WQ_WORK_MM)) {
1530 if (id->mm != current->mm)
1533 req->work.flags |= IO_WQ_WORK_MM;
1539 static void io_prep_async_work(struct io_kiocb *req)
1541 const struct io_op_def *def = &io_op_defs[req->opcode];
1542 struct io_ring_ctx *ctx = req->ctx;
1544 io_req_init_async(req);
1546 if (req->flags & REQ_F_FORCE_ASYNC)
1547 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1549 if (req->flags & REQ_F_ISREG) {
1550 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1551 io_wq_hash_work(&req->work, file_inode(req->file));
1553 if (def->unbound_nonreg_file)
1554 req->work.flags |= IO_WQ_WORK_UNBOUND;
1557 /* if we fail grabbing identity, we must COW, regrab, and retry */
1558 if (io_grab_identity(req))
1561 if (!io_identity_cow(req))
1564 /* can't fail at this point */
1565 if (!io_grab_identity(req))
1569 static void io_prep_async_link(struct io_kiocb *req)
1571 struct io_kiocb *cur;
1573 io_for_each_link(cur, req)
1574 io_prep_async_work(cur);
1577 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1579 struct io_ring_ctx *ctx = req->ctx;
1580 struct io_kiocb *link = io_prep_linked_timeout(req);
1582 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1583 &req->work, req->flags);
1584 io_wq_enqueue(ctx->io_wq, &req->work);
1588 static void io_queue_async_work(struct io_kiocb *req)
1590 struct io_kiocb *link;
1592 /* init ->work of the whole link before punting */
1593 io_prep_async_link(req);
1594 link = __io_queue_async_work(req);
1597 io_queue_linked_timeout(link);
1600 static void io_kill_timeout(struct io_kiocb *req)
1602 struct io_timeout_data *io = req->async_data;
1605 ret = hrtimer_try_to_cancel(&io->timer);
1607 atomic_set(&req->ctx->cq_timeouts,
1608 atomic_read(&req->ctx->cq_timeouts) + 1);
1609 list_del_init(&req->timeout.list);
1610 io_cqring_fill_event(req, 0);
1611 io_put_req_deferred(req, 1);
1616 * Returns true if we found and killed one or more timeouts
1618 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1619 struct files_struct *files)
1621 struct io_kiocb *req, *tmp;
1624 spin_lock_irq(&ctx->completion_lock);
1625 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1626 if (io_match_task(req, tsk, files)) {
1627 io_kill_timeout(req);
1631 spin_unlock_irq(&ctx->completion_lock);
1632 return canceled != 0;
1635 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1638 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1639 struct io_defer_entry, list);
1641 if (req_need_defer(de->req, de->seq))
1643 list_del_init(&de->list);
1644 io_req_task_queue(de->req);
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;
1778 bool all_flushed, posted;
1781 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1785 spin_lock_irqsave(&ctx->completion_lock, flags);
1786 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1787 if (!io_match_task(req, tsk, files))
1790 cqe = io_get_cqring(ctx);
1794 list_move(&req->compl.list, &list);
1796 WRITE_ONCE(cqe->user_data, req->user_data);
1797 WRITE_ONCE(cqe->res, req->result);
1798 WRITE_ONCE(cqe->flags, req->compl.cflags);
1800 ctx->cached_cq_overflow++;
1801 WRITE_ONCE(ctx->rings->cq_overflow,
1802 ctx->cached_cq_overflow);
1807 all_flushed = list_empty(&ctx->cq_overflow_list);
1809 clear_bit(0, &ctx->sq_check_overflow);
1810 clear_bit(0, &ctx->cq_check_overflow);
1811 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1815 io_commit_cqring(ctx);
1816 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1818 io_cqring_ev_posted(ctx);
1820 while (!list_empty(&list)) {
1821 req = list_first_entry(&list, struct io_kiocb, compl.list);
1822 list_del(&req->compl.list);
1829 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1830 struct task_struct *tsk,
1831 struct files_struct *files)
1833 if (test_bit(0, &ctx->cq_check_overflow)) {
1834 /* iopoll syncs against uring_lock, not completion_lock */
1835 if (ctx->flags & IORING_SETUP_IOPOLL)
1836 mutex_lock(&ctx->uring_lock);
1837 __io_cqring_overflow_flush(ctx, force, tsk, files);
1838 if (ctx->flags & IORING_SETUP_IOPOLL)
1839 mutex_unlock(&ctx->uring_lock);
1843 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1845 struct io_ring_ctx *ctx = req->ctx;
1846 struct io_uring_cqe *cqe;
1848 trace_io_uring_complete(ctx, req->user_data, res);
1851 * If we can't get a cq entry, userspace overflowed the
1852 * submission (by quite a lot). Increment the overflow count in
1855 cqe = io_get_cqring(ctx);
1857 WRITE_ONCE(cqe->user_data, req->user_data);
1858 WRITE_ONCE(cqe->res, res);
1859 WRITE_ONCE(cqe->flags, cflags);
1860 } else if (ctx->cq_overflow_flushed ||
1861 atomic_read(&req->task->io_uring->in_idle)) {
1863 * If we're in ring overflow flush mode, or in task cancel mode,
1864 * then we cannot store the request for later flushing, we need
1865 * to drop it on the floor.
1867 ctx->cached_cq_overflow++;
1868 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1870 if (list_empty(&ctx->cq_overflow_list)) {
1871 set_bit(0, &ctx->sq_check_overflow);
1872 set_bit(0, &ctx->cq_check_overflow);
1873 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1877 req->compl.cflags = cflags;
1878 refcount_inc(&req->refs);
1879 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1883 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1885 __io_cqring_fill_event(req, res, 0);
1888 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1890 struct io_ring_ctx *ctx = req->ctx;
1891 unsigned long flags;
1893 spin_lock_irqsave(&ctx->completion_lock, flags);
1894 __io_cqring_fill_event(req, res, cflags);
1895 io_commit_cqring(ctx);
1896 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1898 io_cqring_ev_posted(ctx);
1901 static void io_submit_flush_completions(struct io_comp_state *cs)
1903 struct io_ring_ctx *ctx = cs->ctx;
1905 spin_lock_irq(&ctx->completion_lock);
1906 while (!list_empty(&cs->list)) {
1907 struct io_kiocb *req;
1909 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1910 list_del(&req->compl.list);
1911 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1914 * io_free_req() doesn't care about completion_lock unless one
1915 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1916 * because of a potential deadlock with req->work.fs->lock
1918 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1919 |REQ_F_WORK_INITIALIZED)) {
1920 spin_unlock_irq(&ctx->completion_lock);
1922 spin_lock_irq(&ctx->completion_lock);
1927 io_commit_cqring(ctx);
1928 spin_unlock_irq(&ctx->completion_lock);
1930 io_cqring_ev_posted(ctx);
1934 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1935 struct io_comp_state *cs)
1938 io_cqring_add_event(req, res, cflags);
1943 req->compl.cflags = cflags;
1944 list_add_tail(&req->compl.list, &cs->list);
1946 io_submit_flush_completions(cs);
1950 static void io_req_complete(struct io_kiocb *req, long res)
1952 __io_req_complete(req, res, 0, NULL);
1955 static inline bool io_is_fallback_req(struct io_kiocb *req)
1957 return req == (struct io_kiocb *)
1958 ((unsigned long) req->ctx->fallback_req & ~1UL);
1961 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1963 struct io_kiocb *req;
1965 req = ctx->fallback_req;
1966 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1972 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1973 struct io_submit_state *state)
1975 if (!state->free_reqs) {
1976 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1980 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1981 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1984 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1985 * retry single alloc to be on the safe side.
1987 if (unlikely(ret <= 0)) {
1988 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1989 if (!state->reqs[0])
1993 state->free_reqs = ret;
1997 return state->reqs[state->free_reqs];
1999 return io_get_fallback_req(ctx);
2002 static inline void io_put_file(struct io_kiocb *req, struct file *file,
2009 static void io_dismantle_req(struct io_kiocb *req)
2013 if (req->async_data)
2014 kfree(req->async_data);
2016 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
2017 if (req->fixed_file_refs)
2018 percpu_ref_put(req->fixed_file_refs);
2019 io_req_clean_work(req);
2022 static void __io_free_req(struct io_kiocb *req)
2024 struct io_uring_task *tctx = req->task->io_uring;
2025 struct io_ring_ctx *ctx = req->ctx;
2027 io_dismantle_req(req);
2029 percpu_counter_dec(&tctx->inflight);
2030 if (atomic_read(&tctx->in_idle))
2031 wake_up(&tctx->wait);
2032 put_task_struct(req->task);
2034 if (likely(!io_is_fallback_req(req)))
2035 kmem_cache_free(req_cachep, req);
2037 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
2038 percpu_ref_put(&ctx->refs);
2041 static inline void io_remove_next_linked(struct io_kiocb *req)
2043 struct io_kiocb *nxt = req->link;
2045 req->link = nxt->link;
2049 static void io_kill_linked_timeout(struct io_kiocb *req)
2051 struct io_ring_ctx *ctx = req->ctx;
2052 struct io_kiocb *link;
2053 bool cancelled = false;
2054 unsigned long flags;
2056 spin_lock_irqsave(&ctx->completion_lock, flags);
2060 * Can happen if a linked timeout fired and link had been like
2061 * req -> link t-out -> link t-out [-> ...]
2063 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
2064 struct io_timeout_data *io = link->async_data;
2067 io_remove_next_linked(req);
2068 link->timeout.head = NULL;
2069 ret = hrtimer_try_to_cancel(&io->timer);
2071 io_cqring_fill_event(link, -ECANCELED);
2072 io_commit_cqring(ctx);
2076 req->flags &= ~REQ_F_LINK_TIMEOUT;
2077 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2080 io_cqring_ev_posted(ctx);
2086 static void io_fail_links(struct io_kiocb *req)
2088 struct io_kiocb *link, *nxt;
2089 struct io_ring_ctx *ctx = req->ctx;
2090 unsigned long flags;
2092 spin_lock_irqsave(&ctx->completion_lock, flags);
2100 trace_io_uring_fail_link(req, link);
2101 io_cqring_fill_event(link, -ECANCELED);
2104 * It's ok to free under spinlock as they're not linked anymore,
2105 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2108 if (link->flags & REQ_F_WORK_INITIALIZED)
2109 io_put_req_deferred(link, 2);
2111 io_double_put_req(link);
2114 io_commit_cqring(ctx);
2115 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2117 io_cqring_ev_posted(ctx);
2120 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2122 if (req->flags & REQ_F_LINK_TIMEOUT)
2123 io_kill_linked_timeout(req);
2126 * If LINK is set, we have dependent requests in this chain. If we
2127 * didn't fail this request, queue the first one up, moving any other
2128 * dependencies to the next request. In case of failure, fail the rest
2131 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2132 struct io_kiocb *nxt = req->link;
2141 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2143 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2145 return __io_req_find_next(req);
2148 static int io_req_task_work_add(struct io_kiocb *req)
2150 struct task_struct *tsk = req->task;
2151 struct io_ring_ctx *ctx = req->ctx;
2152 enum task_work_notify_mode notify;
2155 if (tsk->flags & PF_EXITING)
2159 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2160 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2161 * processing task_work. There's no reliable way to tell if TWA_RESUME
2165 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2166 notify = TWA_SIGNAL;
2168 ret = task_work_add(tsk, &req->task_work, notify);
2170 wake_up_process(tsk);
2175 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2177 struct io_ring_ctx *ctx = req->ctx;
2179 spin_lock_irq(&ctx->completion_lock);
2180 io_cqring_fill_event(req, error);
2181 io_commit_cqring(ctx);
2182 spin_unlock_irq(&ctx->completion_lock);
2184 io_cqring_ev_posted(ctx);
2185 req_set_fail_links(req);
2186 io_double_put_req(req);
2189 static void io_req_task_cancel(struct callback_head *cb)
2191 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2192 struct io_ring_ctx *ctx = req->ctx;
2194 __io_req_task_cancel(req, -ECANCELED);
2195 percpu_ref_put(&ctx->refs);
2198 static void __io_req_task_submit(struct io_kiocb *req)
2200 struct io_ring_ctx *ctx = req->ctx;
2202 mutex_lock(&ctx->uring_lock);
2203 if (!ctx->sqo_dead &&
2204 !__io_sq_thread_acquire_mm(ctx) &&
2205 !__io_sq_thread_acquire_files(ctx))
2206 __io_queue_sqe(req, NULL);
2208 __io_req_task_cancel(req, -EFAULT);
2209 mutex_unlock(&ctx->uring_lock);
2211 if (ctx->flags & IORING_SETUP_SQPOLL)
2212 io_sq_thread_drop_mm_files();
2215 static void io_req_task_submit(struct callback_head *cb)
2217 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2218 struct io_ring_ctx *ctx = req->ctx;
2220 __io_req_task_submit(req);
2221 percpu_ref_put(&ctx->refs);
2224 static void io_req_task_queue(struct io_kiocb *req)
2228 init_task_work(&req->task_work, io_req_task_submit);
2229 percpu_ref_get(&req->ctx->refs);
2231 ret = io_req_task_work_add(req);
2232 if (unlikely(ret)) {
2233 struct task_struct *tsk;
2235 init_task_work(&req->task_work, io_req_task_cancel);
2236 tsk = io_wq_get_task(req->ctx->io_wq);
2237 task_work_add(tsk, &req->task_work, TWA_NONE);
2238 wake_up_process(tsk);
2242 static inline void io_queue_next(struct io_kiocb *req)
2244 struct io_kiocb *nxt = io_req_find_next(req);
2247 io_req_task_queue(nxt);
2250 static void io_free_req(struct io_kiocb *req)
2257 void *reqs[IO_IOPOLL_BATCH];
2260 struct task_struct *task;
2264 static inline void io_init_req_batch(struct req_batch *rb)
2271 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2272 struct req_batch *rb)
2274 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2275 percpu_ref_put_many(&ctx->refs, rb->to_free);
2279 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2280 struct req_batch *rb)
2283 __io_req_free_batch_flush(ctx, rb);
2285 struct io_uring_task *tctx = rb->task->io_uring;
2287 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2288 if (atomic_read(&tctx->in_idle))
2289 wake_up(&tctx->wait);
2290 put_task_struct_many(rb->task, rb->task_refs);
2295 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2297 if (unlikely(io_is_fallback_req(req))) {
2303 if (req->task != rb->task) {
2305 struct io_uring_task *tctx = rb->task->io_uring;
2307 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2308 if (atomic_read(&tctx->in_idle))
2309 wake_up(&tctx->wait);
2310 put_task_struct_many(rb->task, rb->task_refs);
2312 rb->task = req->task;
2317 io_dismantle_req(req);
2318 rb->reqs[rb->to_free++] = req;
2319 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2320 __io_req_free_batch_flush(req->ctx, rb);
2324 * Drop reference to request, return next in chain (if there is one) if this
2325 * was the last reference to this request.
2327 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2329 struct io_kiocb *nxt = NULL;
2331 if (refcount_dec_and_test(&req->refs)) {
2332 nxt = io_req_find_next(req);
2338 static void io_put_req(struct io_kiocb *req)
2340 if (refcount_dec_and_test(&req->refs))
2344 static void io_put_req_deferred_cb(struct callback_head *cb)
2346 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2351 static void io_free_req_deferred(struct io_kiocb *req)
2355 init_task_work(&req->task_work, io_put_req_deferred_cb);
2356 ret = io_req_task_work_add(req);
2357 if (unlikely(ret)) {
2358 struct task_struct *tsk;
2360 tsk = io_wq_get_task(req->ctx->io_wq);
2361 task_work_add(tsk, &req->task_work, TWA_NONE);
2362 wake_up_process(tsk);
2366 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2368 if (refcount_sub_and_test(refs, &req->refs))
2369 io_free_req_deferred(req);
2372 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2374 struct io_kiocb *nxt;
2377 * A ref is owned by io-wq in which context we're. So, if that's the
2378 * last one, it's safe to steal next work. False negatives are Ok,
2379 * it just will be re-punted async in io_put_work()
2381 if (refcount_read(&req->refs) != 1)
2384 nxt = io_req_find_next(req);
2385 return nxt ? &nxt->work : NULL;
2388 static void io_double_put_req(struct io_kiocb *req)
2390 /* drop both submit and complete references */
2391 if (refcount_sub_and_test(2, &req->refs))
2395 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2397 /* See comment at the top of this file */
2399 return __io_cqring_events(ctx);
2402 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2404 struct io_rings *rings = ctx->rings;
2406 /* make sure SQ entry isn't read before tail */
2407 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2410 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2412 unsigned int cflags;
2414 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2415 cflags |= IORING_CQE_F_BUFFER;
2416 req->flags &= ~REQ_F_BUFFER_SELECTED;
2421 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2423 struct io_buffer *kbuf;
2425 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2426 return io_put_kbuf(req, kbuf);
2429 static inline bool io_run_task_work(void)
2432 * Not safe to run on exiting task, and the task_work handling will
2433 * not add work to such a task.
2435 if (unlikely(current->flags & PF_EXITING))
2437 if (current->task_works) {
2438 __set_current_state(TASK_RUNNING);
2446 static void io_iopoll_queue(struct list_head *again)
2448 struct io_kiocb *req;
2451 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2452 list_del(&req->inflight_entry);
2453 __io_complete_rw(req, -EAGAIN, 0, NULL);
2454 } while (!list_empty(again));
2458 * Find and free completed poll iocbs
2460 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2461 struct list_head *done)
2463 struct req_batch rb;
2464 struct io_kiocb *req;
2467 /* order with ->result store in io_complete_rw_iopoll() */
2470 io_init_req_batch(&rb);
2471 while (!list_empty(done)) {
2474 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2475 if (READ_ONCE(req->result) == -EAGAIN) {
2477 req->iopoll_completed = 0;
2478 list_move_tail(&req->inflight_entry, &again);
2481 list_del(&req->inflight_entry);
2483 if (req->flags & REQ_F_BUFFER_SELECTED)
2484 cflags = io_put_rw_kbuf(req);
2486 __io_cqring_fill_event(req, req->result, cflags);
2489 if (refcount_dec_and_test(&req->refs))
2490 io_req_free_batch(&rb, req);
2493 io_commit_cqring(ctx);
2494 io_cqring_ev_posted_iopoll(ctx);
2495 io_req_free_batch_finish(ctx, &rb);
2497 if (!list_empty(&again))
2498 io_iopoll_queue(&again);
2501 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2504 struct io_kiocb *req, *tmp;
2510 * Only spin for completions if we don't have multiple devices hanging
2511 * off our complete list, and we're under the requested amount.
2513 spin = !ctx->poll_multi_file && *nr_events < min;
2516 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2517 struct kiocb *kiocb = &req->rw.kiocb;
2520 * Move completed and retryable entries to our local lists.
2521 * If we find a request that requires polling, break out
2522 * and complete those lists first, if we have entries there.
2524 if (READ_ONCE(req->iopoll_completed)) {
2525 list_move_tail(&req->inflight_entry, &done);
2528 if (!list_empty(&done))
2531 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2535 /* iopoll may have completed current req */
2536 if (READ_ONCE(req->iopoll_completed))
2537 list_move_tail(&req->inflight_entry, &done);
2544 if (!list_empty(&done))
2545 io_iopoll_complete(ctx, nr_events, &done);
2551 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2552 * non-spinning poll check - we'll still enter the driver poll loop, but only
2553 * as a non-spinning completion check.
2555 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2558 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2561 ret = io_do_iopoll(ctx, nr_events, min);
2564 if (*nr_events >= min)
2572 * We can't just wait for polled events to come to us, we have to actively
2573 * find and complete them.
2575 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2577 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2580 mutex_lock(&ctx->uring_lock);
2581 while (!list_empty(&ctx->iopoll_list)) {
2582 unsigned int nr_events = 0;
2584 io_do_iopoll(ctx, &nr_events, 0);
2586 /* let it sleep and repeat later if can't complete a request */
2590 * Ensure we allow local-to-the-cpu processing to take place,
2591 * in this case we need to ensure that we reap all events.
2592 * Also let task_work, etc. to progress by releasing the mutex
2594 if (need_resched()) {
2595 mutex_unlock(&ctx->uring_lock);
2597 mutex_lock(&ctx->uring_lock);
2600 mutex_unlock(&ctx->uring_lock);
2603 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2605 unsigned int nr_events = 0;
2606 int iters = 0, ret = 0;
2609 * We disallow the app entering submit/complete with polling, but we
2610 * still need to lock the ring to prevent racing with polled issue
2611 * that got punted to a workqueue.
2613 mutex_lock(&ctx->uring_lock);
2616 * Don't enter poll loop if we already have events pending.
2617 * If we do, we can potentially be spinning for commands that
2618 * already triggered a CQE (eg in error).
2620 if (test_bit(0, &ctx->cq_check_overflow))
2621 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2622 if (io_cqring_events(ctx))
2626 * If a submit got punted to a workqueue, we can have the
2627 * application entering polling for a command before it gets
2628 * issued. That app will hold the uring_lock for the duration
2629 * of the poll right here, so we need to take a breather every
2630 * now and then to ensure that the issue has a chance to add
2631 * the poll to the issued list. Otherwise we can spin here
2632 * forever, while the workqueue is stuck trying to acquire the
2635 if (!(++iters & 7)) {
2636 mutex_unlock(&ctx->uring_lock);
2638 mutex_lock(&ctx->uring_lock);
2641 ret = io_iopoll_getevents(ctx, &nr_events, min);
2645 } while (min && !nr_events && !need_resched());
2647 mutex_unlock(&ctx->uring_lock);
2651 static void kiocb_end_write(struct io_kiocb *req)
2654 * Tell lockdep we inherited freeze protection from submission
2657 if (req->flags & REQ_F_ISREG) {
2658 struct inode *inode = file_inode(req->file);
2660 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2662 file_end_write(req->file);
2665 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2666 struct io_comp_state *cs)
2668 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2671 if (kiocb->ki_flags & IOCB_WRITE)
2672 kiocb_end_write(req);
2674 if (res != req->result)
2675 req_set_fail_links(req);
2676 if (req->flags & REQ_F_BUFFER_SELECTED)
2677 cflags = io_put_rw_kbuf(req);
2678 __io_req_complete(req, res, cflags, cs);
2682 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2684 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2685 ssize_t ret = -ECANCELED;
2686 struct iov_iter iter;
2694 switch (req->opcode) {
2695 case IORING_OP_READV:
2696 case IORING_OP_READ_FIXED:
2697 case IORING_OP_READ:
2700 case IORING_OP_WRITEV:
2701 case IORING_OP_WRITE_FIXED:
2702 case IORING_OP_WRITE:
2706 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2711 if (!req->async_data) {
2712 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2715 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2723 req_set_fail_links(req);
2728 static bool io_rw_reissue(struct io_kiocb *req, long res)
2731 umode_t mode = file_inode(req->file)->i_mode;
2734 if (!S_ISBLK(mode) && !S_ISREG(mode))
2736 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2739 lockdep_assert_held(&req->ctx->uring_lock);
2741 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2743 if (io_resubmit_prep(req, ret)) {
2744 refcount_inc(&req->refs);
2745 io_queue_async_work(req);
2753 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2754 struct io_comp_state *cs)
2756 if (!io_rw_reissue(req, res))
2757 io_complete_rw_common(&req->rw.kiocb, res, cs);
2760 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2762 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2764 __io_complete_rw(req, res, res2, NULL);
2767 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2769 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2771 if (kiocb->ki_flags & IOCB_WRITE)
2772 kiocb_end_write(req);
2774 if (res != -EAGAIN && res != req->result)
2775 req_set_fail_links(req);
2777 WRITE_ONCE(req->result, res);
2778 /* order with io_poll_complete() checking ->result */
2780 WRITE_ONCE(req->iopoll_completed, 1);
2784 * After the iocb has been issued, it's safe to be found on the poll list.
2785 * Adding the kiocb to the list AFTER submission ensures that we don't
2786 * find it from a io_iopoll_getevents() thread before the issuer is done
2787 * accessing the kiocb cookie.
2789 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2791 struct io_ring_ctx *ctx = req->ctx;
2794 * Track whether we have multiple files in our lists. This will impact
2795 * how we do polling eventually, not spinning if we're on potentially
2796 * different devices.
2798 if (list_empty(&ctx->iopoll_list)) {
2799 ctx->poll_multi_file = false;
2800 } else if (!ctx->poll_multi_file) {
2801 struct io_kiocb *list_req;
2803 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2805 if (list_req->file != req->file)
2806 ctx->poll_multi_file = true;
2810 * For fast devices, IO may have already completed. If it has, add
2811 * it to the front so we find it first.
2813 if (READ_ONCE(req->iopoll_completed))
2814 list_add(&req->inflight_entry, &ctx->iopoll_list);
2816 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2819 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2820 * task context or in io worker task context. If current task context is
2821 * sq thread, we don't need to check whether should wake up sq thread.
2823 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2824 wq_has_sleeper(&ctx->sq_data->wait))
2825 wake_up(&ctx->sq_data->wait);
2828 static inline void __io_state_file_put(struct io_submit_state *state)
2830 fput_many(state->file, state->file_refs);
2831 state->file_refs = 0;
2834 static inline void io_state_file_put(struct io_submit_state *state)
2836 if (state->file_refs)
2837 __io_state_file_put(state);
2841 * Get as many references to a file as we have IOs left in this submission,
2842 * assuming most submissions are for one file, or at least that each file
2843 * has more than one submission.
2845 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2850 if (state->file_refs) {
2851 if (state->fd == fd) {
2855 __io_state_file_put(state);
2857 state->file = fget_many(fd, state->ios_left);
2858 if (unlikely(!state->file))
2862 state->file_refs = state->ios_left - 1;
2866 static bool io_bdev_nowait(struct block_device *bdev)
2868 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2872 * If we tracked the file through the SCM inflight mechanism, we could support
2873 * any file. For now, just ensure that anything potentially problematic is done
2876 static bool io_file_supports_async(struct file *file, int rw)
2878 umode_t mode = file_inode(file)->i_mode;
2880 if (S_ISBLK(mode)) {
2881 if (IS_ENABLED(CONFIG_BLOCK) &&
2882 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2886 if (S_ISCHR(mode) || S_ISSOCK(mode))
2888 if (S_ISREG(mode)) {
2889 if (IS_ENABLED(CONFIG_BLOCK) &&
2890 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2891 file->f_op != &io_uring_fops)
2896 /* any ->read/write should understand O_NONBLOCK */
2897 if (file->f_flags & O_NONBLOCK)
2900 if (!(file->f_mode & FMODE_NOWAIT))
2904 return file->f_op->read_iter != NULL;
2906 return file->f_op->write_iter != NULL;
2909 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2911 struct io_ring_ctx *ctx = req->ctx;
2912 struct kiocb *kiocb = &req->rw.kiocb;
2916 if (S_ISREG(file_inode(req->file)->i_mode))
2917 req->flags |= REQ_F_ISREG;
2919 kiocb->ki_pos = READ_ONCE(sqe->off);
2920 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2921 req->flags |= REQ_F_CUR_POS;
2922 kiocb->ki_pos = req->file->f_pos;
2924 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2925 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2926 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2930 ioprio = READ_ONCE(sqe->ioprio);
2932 ret = ioprio_check_cap(ioprio);
2936 kiocb->ki_ioprio = ioprio;
2938 kiocb->ki_ioprio = get_current_ioprio();
2940 /* don't allow async punt if RWF_NOWAIT was requested */
2941 if (kiocb->ki_flags & IOCB_NOWAIT)
2942 req->flags |= REQ_F_NOWAIT;
2944 if (ctx->flags & IORING_SETUP_IOPOLL) {
2945 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2946 !kiocb->ki_filp->f_op->iopoll)
2949 kiocb->ki_flags |= IOCB_HIPRI;
2950 kiocb->ki_complete = io_complete_rw_iopoll;
2951 req->iopoll_completed = 0;
2953 if (kiocb->ki_flags & IOCB_HIPRI)
2955 kiocb->ki_complete = io_complete_rw;
2958 req->rw.addr = READ_ONCE(sqe->addr);
2959 req->rw.len = READ_ONCE(sqe->len);
2960 req->buf_index = READ_ONCE(sqe->buf_index);
2964 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2970 case -ERESTARTNOINTR:
2971 case -ERESTARTNOHAND:
2972 case -ERESTART_RESTARTBLOCK:
2974 * We can't just restart the syscall, since previously
2975 * submitted sqes may already be in progress. Just fail this
2981 kiocb->ki_complete(kiocb, ret, 0);
2985 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2986 struct io_comp_state *cs)
2988 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2989 struct io_async_rw *io = req->async_data;
2991 /* add previously done IO, if any */
2992 if (io && io->bytes_done > 0) {
2994 ret = io->bytes_done;
2996 ret += io->bytes_done;
2999 if (req->flags & REQ_F_CUR_POS)
3000 req->file->f_pos = kiocb->ki_pos;
3001 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
3002 __io_complete_rw(req, ret, 0, cs);
3004 io_rw_done(kiocb, ret);
3007 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
3008 struct iov_iter *iter)
3010 struct io_ring_ctx *ctx = req->ctx;
3011 size_t len = req->rw.len;
3012 struct io_mapped_ubuf *imu;
3013 u16 index, buf_index = req->buf_index;
3017 if (unlikely(buf_index >= ctx->nr_user_bufs))
3019 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3020 imu = &ctx->user_bufs[index];
3021 buf_addr = req->rw.addr;
3024 if (buf_addr + len < buf_addr)
3026 /* not inside the mapped region */
3027 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
3031 * May not be a start of buffer, set size appropriately
3032 * and advance us to the beginning.
3034 offset = buf_addr - imu->ubuf;
3035 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3039 * Don't use iov_iter_advance() here, as it's really slow for
3040 * using the latter parts of a big fixed buffer - it iterates
3041 * over each segment manually. We can cheat a bit here, because
3044 * 1) it's a BVEC iter, we set it up
3045 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3046 * first and last bvec
3048 * So just find our index, and adjust the iterator afterwards.
3049 * If the offset is within the first bvec (or the whole first
3050 * bvec, just use iov_iter_advance(). This makes it easier
3051 * since we can just skip the first segment, which may not
3052 * be PAGE_SIZE aligned.
3054 const struct bio_vec *bvec = imu->bvec;
3056 if (offset <= bvec->bv_len) {
3057 iov_iter_advance(iter, offset);
3059 unsigned long seg_skip;
3061 /* skip first vec */
3062 offset -= bvec->bv_len;
3063 seg_skip = 1 + (offset >> PAGE_SHIFT);
3065 iter->bvec = bvec + seg_skip;
3066 iter->nr_segs -= seg_skip;
3067 iter->count -= bvec->bv_len + offset;
3068 iter->iov_offset = offset & ~PAGE_MASK;
3075 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3078 mutex_unlock(&ctx->uring_lock);
3081 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3084 * "Normal" inline submissions always hold the uring_lock, since we
3085 * grab it from the system call. Same is true for the SQPOLL offload.
3086 * The only exception is when we've detached the request and issue it
3087 * from an async worker thread, grab the lock for that case.
3090 mutex_lock(&ctx->uring_lock);
3093 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3094 int bgid, struct io_buffer *kbuf,
3097 struct io_buffer *head;
3099 if (req->flags & REQ_F_BUFFER_SELECTED)
3102 io_ring_submit_lock(req->ctx, needs_lock);
3104 lockdep_assert_held(&req->ctx->uring_lock);
3106 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3108 if (!list_empty(&head->list)) {
3109 kbuf = list_last_entry(&head->list, struct io_buffer,
3111 list_del(&kbuf->list);
3114 idr_remove(&req->ctx->io_buffer_idr, bgid);
3116 if (*len > kbuf->len)
3119 kbuf = ERR_PTR(-ENOBUFS);
3122 io_ring_submit_unlock(req->ctx, needs_lock);
3127 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3130 struct io_buffer *kbuf;
3133 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3134 bgid = req->buf_index;
3135 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3138 req->rw.addr = (u64) (unsigned long) kbuf;
3139 req->flags |= REQ_F_BUFFER_SELECTED;
3140 return u64_to_user_ptr(kbuf->addr);
3143 #ifdef CONFIG_COMPAT
3144 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3147 struct compat_iovec __user *uiov;
3148 compat_ssize_t clen;
3152 uiov = u64_to_user_ptr(req->rw.addr);
3153 if (!access_ok(uiov, sizeof(*uiov)))
3155 if (__get_user(clen, &uiov->iov_len))
3161 buf = io_rw_buffer_select(req, &len, needs_lock);
3163 return PTR_ERR(buf);
3164 iov[0].iov_base = buf;
3165 iov[0].iov_len = (compat_size_t) len;
3170 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3173 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3177 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3180 len = iov[0].iov_len;
3183 buf = io_rw_buffer_select(req, &len, needs_lock);
3185 return PTR_ERR(buf);
3186 iov[0].iov_base = buf;
3187 iov[0].iov_len = len;
3191 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3194 if (req->flags & REQ_F_BUFFER_SELECTED) {
3195 struct io_buffer *kbuf;
3197 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3198 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3199 iov[0].iov_len = kbuf->len;
3202 if (req->rw.len != 1)
3205 #ifdef CONFIG_COMPAT
3206 if (req->ctx->compat)
3207 return io_compat_import(req, iov, needs_lock);
3210 return __io_iov_buffer_select(req, iov, needs_lock);
3213 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3214 struct iovec **iovec, struct iov_iter *iter,
3217 void __user *buf = u64_to_user_ptr(req->rw.addr);
3218 size_t sqe_len = req->rw.len;
3222 opcode = req->opcode;
3223 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3225 return io_import_fixed(req, rw, iter);
3228 /* buffer index only valid with fixed read/write, or buffer select */
3229 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3232 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3233 if (req->flags & REQ_F_BUFFER_SELECT) {
3234 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3236 return PTR_ERR(buf);
3237 req->rw.len = sqe_len;
3240 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3245 if (req->flags & REQ_F_BUFFER_SELECT) {
3246 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3248 ret = (*iovec)->iov_len;
3249 iov_iter_init(iter, rw, *iovec, 1, ret);
3255 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3259 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3261 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3265 * For files that don't have ->read_iter() and ->write_iter(), handle them
3266 * by looping over ->read() or ->write() manually.
3268 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3270 struct kiocb *kiocb = &req->rw.kiocb;
3271 struct file *file = req->file;
3275 * Don't support polled IO through this interface, and we can't
3276 * support non-blocking either. For the latter, this just causes
3277 * the kiocb to be handled from an async context.
3279 if (kiocb->ki_flags & IOCB_HIPRI)
3281 if (kiocb->ki_flags & IOCB_NOWAIT)
3284 while (iov_iter_count(iter)) {
3288 if (!iov_iter_is_bvec(iter)) {
3289 iovec = iov_iter_iovec(iter);
3291 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3292 iovec.iov_len = req->rw.len;
3296 nr = file->f_op->read(file, iovec.iov_base,
3297 iovec.iov_len, io_kiocb_ppos(kiocb));
3299 nr = file->f_op->write(file, iovec.iov_base,
3300 iovec.iov_len, io_kiocb_ppos(kiocb));
3309 if (nr != iovec.iov_len)
3313 iov_iter_advance(iter, nr);
3319 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3320 const struct iovec *fast_iov, struct iov_iter *iter)
3322 struct io_async_rw *rw = req->async_data;
3324 memcpy(&rw->iter, iter, sizeof(*iter));
3325 rw->free_iovec = iovec;
3327 /* can only be fixed buffers, no need to do anything */
3328 if (iov_iter_is_bvec(iter))
3331 unsigned iov_off = 0;
3333 rw->iter.iov = rw->fast_iov;
3334 if (iter->iov != fast_iov) {
3335 iov_off = iter->iov - fast_iov;
3336 rw->iter.iov += iov_off;
3338 if (rw->fast_iov != fast_iov)
3339 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3340 sizeof(struct iovec) * iter->nr_segs);
3342 req->flags |= REQ_F_NEED_CLEANUP;
3346 static inline int __io_alloc_async_data(struct io_kiocb *req)
3348 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3349 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3350 return req->async_data == NULL;
3353 static int io_alloc_async_data(struct io_kiocb *req)
3355 if (!io_op_defs[req->opcode].needs_async_data)
3358 return __io_alloc_async_data(req);
3361 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3362 const struct iovec *fast_iov,
3363 struct iov_iter *iter, bool force)
3365 if (!force && !io_op_defs[req->opcode].needs_async_data)
3367 if (!req->async_data) {
3368 if (__io_alloc_async_data(req))
3371 io_req_map_rw(req, iovec, fast_iov, iter);
3376 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3378 struct io_async_rw *iorw = req->async_data;
3379 struct iovec *iov = iorw->fast_iov;
3382 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3383 if (unlikely(ret < 0))
3386 iorw->bytes_done = 0;
3387 iorw->free_iovec = iov;
3389 req->flags |= REQ_F_NEED_CLEANUP;
3393 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3397 ret = io_prep_rw(req, sqe);
3401 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3404 /* either don't need iovec imported or already have it */
3405 if (!req->async_data)
3407 return io_rw_prep_async(req, READ);
3411 * This is our waitqueue callback handler, registered through lock_page_async()
3412 * when we initially tried to do the IO with the iocb armed our waitqueue.
3413 * This gets called when the page is unlocked, and we generally expect that to
3414 * happen when the page IO is completed and the page is now uptodate. This will
3415 * queue a task_work based retry of the operation, attempting to copy the data
3416 * again. If the latter fails because the page was NOT uptodate, then we will
3417 * do a thread based blocking retry of the operation. That's the unexpected
3420 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3421 int sync, void *arg)
3423 struct wait_page_queue *wpq;
3424 struct io_kiocb *req = wait->private;
3425 struct wait_page_key *key = arg;
3428 wpq = container_of(wait, struct wait_page_queue, wait);
3430 if (!wake_page_match(wpq, key))
3433 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3434 list_del_init(&wait->entry);
3436 init_task_work(&req->task_work, io_req_task_submit);
3437 percpu_ref_get(&req->ctx->refs);
3439 /* submit ref gets dropped, acquire a new one */
3440 refcount_inc(&req->refs);
3441 ret = io_req_task_work_add(req);
3442 if (unlikely(ret)) {
3443 struct task_struct *tsk;
3445 /* queue just for cancelation */
3446 init_task_work(&req->task_work, io_req_task_cancel);
3447 tsk = io_wq_get_task(req->ctx->io_wq);
3448 task_work_add(tsk, &req->task_work, TWA_NONE);
3449 wake_up_process(tsk);
3455 * This controls whether a given IO request should be armed for async page
3456 * based retry. If we return false here, the request is handed to the async
3457 * worker threads for retry. If we're doing buffered reads on a regular file,
3458 * we prepare a private wait_page_queue entry and retry the operation. This
3459 * will either succeed because the page is now uptodate and unlocked, or it
3460 * will register a callback when the page is unlocked at IO completion. Through
3461 * that callback, io_uring uses task_work to setup a retry of the operation.
3462 * That retry will attempt the buffered read again. The retry will generally
3463 * succeed, or in rare cases where it fails, we then fall back to using the
3464 * async worker threads for a blocking retry.
3466 static bool io_rw_should_retry(struct io_kiocb *req)
3468 struct io_async_rw *rw = req->async_data;
3469 struct wait_page_queue *wait = &rw->wpq;
3470 struct kiocb *kiocb = &req->rw.kiocb;
3472 /* never retry for NOWAIT, we just complete with -EAGAIN */
3473 if (req->flags & REQ_F_NOWAIT)
3476 /* Only for buffered IO */
3477 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3481 * just use poll if we can, and don't attempt if the fs doesn't
3482 * support callback based unlocks
3484 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3487 wait->wait.func = io_async_buf_func;
3488 wait->wait.private = req;
3489 wait->wait.flags = 0;
3490 INIT_LIST_HEAD(&wait->wait.entry);
3491 kiocb->ki_flags |= IOCB_WAITQ;
3492 kiocb->ki_flags &= ~IOCB_NOWAIT;
3493 kiocb->ki_waitq = wait;
3497 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3499 if (req->file->f_op->read_iter)
3500 return call_read_iter(req->file, &req->rw.kiocb, iter);
3501 else if (req->file->f_op->read)
3502 return loop_rw_iter(READ, req, iter);
3507 static int io_read(struct io_kiocb *req, bool force_nonblock,
3508 struct io_comp_state *cs)
3510 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3511 struct kiocb *kiocb = &req->rw.kiocb;
3512 struct iov_iter __iter, *iter = &__iter;
3513 struct io_async_rw *rw = req->async_data;
3514 ssize_t io_size, ret, ret2;
3521 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3525 io_size = iov_iter_count(iter);
3526 req->result = io_size;
3529 /* Ensure we clear previously set non-block flag */
3530 if (!force_nonblock)
3531 kiocb->ki_flags &= ~IOCB_NOWAIT;
3533 kiocb->ki_flags |= IOCB_NOWAIT;
3536 /* If the file doesn't support async, just async punt */
3537 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3541 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3545 ret = io_iter_do_read(req, iter);
3549 } else if (ret == -EIOCBQUEUED) {
3552 } else if (ret == -EAGAIN) {
3553 /* IOPOLL retry should happen for io-wq threads */
3554 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3556 /* no retry on NONBLOCK marked file */
3557 if (req->file->f_flags & O_NONBLOCK)
3559 /* some cases will consume bytes even on error returns */
3560 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3563 } else if (ret < 0) {
3564 /* make sure -ERESTARTSYS -> -EINTR is done */
3568 /* read it all, or we did blocking attempt. no retry. */
3569 if (!iov_iter_count(iter) || !force_nonblock ||
3570 (req->file->f_flags & O_NONBLOCK) || !(req->flags & REQ_F_ISREG))
3575 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3582 rw = req->async_data;
3583 /* it's copied and will be cleaned with ->io */
3585 /* now use our persistent iterator, if we aren't already */
3588 rw->bytes_done += ret;
3589 /* if we can retry, do so with the callbacks armed */
3590 if (!io_rw_should_retry(req)) {
3591 kiocb->ki_flags &= ~IOCB_WAITQ;
3596 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3597 * get -EIOCBQUEUED, then we'll get a notification when the desired
3598 * page gets unlocked. We can also get a partial read here, and if we
3599 * do, then just retry at the new offset.
3601 ret = io_iter_do_read(req, iter);
3602 if (ret == -EIOCBQUEUED) {
3605 } else if (ret > 0 && ret < io_size) {
3606 /* we got some bytes, but not all. retry. */
3610 kiocb_done(kiocb, ret, cs);
3613 /* it's reportedly faster than delegating the null check to kfree() */
3619 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3623 ret = io_prep_rw(req, sqe);
3627 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3630 /* either don't need iovec imported or already have it */
3631 if (!req->async_data)
3633 return io_rw_prep_async(req, WRITE);
3636 static int io_write(struct io_kiocb *req, bool force_nonblock,
3637 struct io_comp_state *cs)
3639 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3640 struct kiocb *kiocb = &req->rw.kiocb;
3641 struct iov_iter __iter, *iter = &__iter;
3642 struct io_async_rw *rw = req->async_data;
3643 ssize_t ret, ret2, io_size;
3649 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3653 io_size = iov_iter_count(iter);
3654 req->result = io_size;
3656 /* Ensure we clear previously set non-block flag */
3657 if (!force_nonblock)
3658 kiocb->ki_flags &= ~IOCB_NOWAIT;
3660 kiocb->ki_flags |= IOCB_NOWAIT;
3662 /* If the file doesn't support async, just async punt */
3663 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3666 /* file path doesn't support NOWAIT for non-direct_IO */
3667 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3668 (req->flags & REQ_F_ISREG))
3671 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3676 * Open-code file_start_write here to grab freeze protection,
3677 * which will be released by another thread in
3678 * io_complete_rw(). Fool lockdep by telling it the lock got
3679 * released so that it doesn't complain about the held lock when
3680 * we return to userspace.
3682 if (req->flags & REQ_F_ISREG) {
3683 sb_start_write(file_inode(req->file)->i_sb);
3684 __sb_writers_release(file_inode(req->file)->i_sb,
3687 kiocb->ki_flags |= IOCB_WRITE;
3689 if (req->file->f_op->write_iter)
3690 ret2 = call_write_iter(req->file, kiocb, iter);
3691 else if (req->file->f_op->write)
3692 ret2 = loop_rw_iter(WRITE, req, iter);
3697 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3698 * retry them without IOCB_NOWAIT.
3700 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3702 /* no retry on NONBLOCK marked file */
3703 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3705 if (!force_nonblock || ret2 != -EAGAIN) {
3706 /* IOPOLL retry should happen for io-wq threads */
3707 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3710 kiocb_done(kiocb, ret2, cs);
3713 /* some cases will consume bytes even on error returns */
3714 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3715 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3720 /* it's reportedly faster than delegating the null check to kfree() */
3726 static int io_renameat_prep(struct io_kiocb *req,
3727 const struct io_uring_sqe *sqe)
3729 struct io_rename *ren = &req->rename;
3730 const char __user *oldf, *newf;
3732 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3735 ren->old_dfd = READ_ONCE(sqe->fd);
3736 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3737 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3738 ren->new_dfd = READ_ONCE(sqe->len);
3739 ren->flags = READ_ONCE(sqe->rename_flags);
3741 ren->oldpath = getname(oldf);
3742 if (IS_ERR(ren->oldpath))
3743 return PTR_ERR(ren->oldpath);
3745 ren->newpath = getname(newf);
3746 if (IS_ERR(ren->newpath)) {
3747 putname(ren->oldpath);
3748 return PTR_ERR(ren->newpath);
3751 req->flags |= REQ_F_NEED_CLEANUP;
3755 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3757 struct io_rename *ren = &req->rename;
3763 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3764 ren->newpath, ren->flags);
3766 req->flags &= ~REQ_F_NEED_CLEANUP;
3768 req_set_fail_links(req);
3769 io_req_complete(req, ret);
3773 static int io_unlinkat_prep(struct io_kiocb *req,
3774 const struct io_uring_sqe *sqe)
3776 struct io_unlink *un = &req->unlink;
3777 const char __user *fname;
3779 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3782 un->dfd = READ_ONCE(sqe->fd);
3784 un->flags = READ_ONCE(sqe->unlink_flags);
3785 if (un->flags & ~AT_REMOVEDIR)
3788 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3789 un->filename = getname(fname);
3790 if (IS_ERR(un->filename))
3791 return PTR_ERR(un->filename);
3793 req->flags |= REQ_F_NEED_CLEANUP;
3797 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3799 struct io_unlink *un = &req->unlink;
3805 if (un->flags & AT_REMOVEDIR)
3806 ret = do_rmdir(un->dfd, un->filename);
3808 ret = do_unlinkat(un->dfd, un->filename);
3810 req->flags &= ~REQ_F_NEED_CLEANUP;
3812 req_set_fail_links(req);
3813 io_req_complete(req, ret);
3817 static int io_shutdown_prep(struct io_kiocb *req,
3818 const struct io_uring_sqe *sqe)
3820 #if defined(CONFIG_NET)
3821 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3823 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3827 req->shutdown.how = READ_ONCE(sqe->len);
3834 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3836 #if defined(CONFIG_NET)
3837 struct socket *sock;
3843 sock = sock_from_file(req->file);
3844 if (unlikely(!sock))
3847 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3849 req_set_fail_links(req);
3850 io_req_complete(req, ret);
3857 static int __io_splice_prep(struct io_kiocb *req,
3858 const struct io_uring_sqe *sqe)
3860 struct io_splice* sp = &req->splice;
3861 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3863 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3867 sp->len = READ_ONCE(sqe->len);
3868 sp->flags = READ_ONCE(sqe->splice_flags);
3870 if (unlikely(sp->flags & ~valid_flags))
3873 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3874 (sp->flags & SPLICE_F_FD_IN_FIXED));
3877 req->flags |= REQ_F_NEED_CLEANUP;
3879 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3881 * Splice operation will be punted aync, and here need to
3882 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3884 io_req_init_async(req);
3885 req->work.flags |= IO_WQ_WORK_UNBOUND;
3891 static int io_tee_prep(struct io_kiocb *req,
3892 const struct io_uring_sqe *sqe)
3894 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3896 return __io_splice_prep(req, sqe);
3899 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3901 struct io_splice *sp = &req->splice;
3902 struct file *in = sp->file_in;
3903 struct file *out = sp->file_out;
3904 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3910 ret = do_tee(in, out, sp->len, flags);
3912 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3913 req->flags &= ~REQ_F_NEED_CLEANUP;
3916 req_set_fail_links(req);
3917 io_req_complete(req, ret);
3921 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3923 struct io_splice* sp = &req->splice;
3925 sp->off_in = READ_ONCE(sqe->splice_off_in);
3926 sp->off_out = READ_ONCE(sqe->off);
3927 return __io_splice_prep(req, sqe);
3930 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3932 struct io_splice *sp = &req->splice;
3933 struct file *in = sp->file_in;
3934 struct file *out = sp->file_out;
3935 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3936 loff_t *poff_in, *poff_out;
3942 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3943 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3946 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3948 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3949 req->flags &= ~REQ_F_NEED_CLEANUP;
3952 req_set_fail_links(req);
3953 io_req_complete(req, ret);
3958 * IORING_OP_NOP just posts a completion event, nothing else.
3960 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3962 struct io_ring_ctx *ctx = req->ctx;
3964 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3967 __io_req_complete(req, 0, 0, cs);
3971 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3973 struct io_ring_ctx *ctx = req->ctx;
3978 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3980 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3983 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3984 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3987 req->sync.off = READ_ONCE(sqe->off);
3988 req->sync.len = READ_ONCE(sqe->len);
3992 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3994 loff_t end = req->sync.off + req->sync.len;
3997 /* fsync always requires a blocking context */
4001 ret = vfs_fsync_range(req->file, req->sync.off,
4002 end > 0 ? end : LLONG_MAX,
4003 req->sync.flags & IORING_FSYNC_DATASYNC);
4005 req_set_fail_links(req);
4006 io_req_complete(req, ret);
4010 static int io_fallocate_prep(struct io_kiocb *req,
4011 const struct io_uring_sqe *sqe)
4013 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
4015 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4018 req->sync.off = READ_ONCE(sqe->off);
4019 req->sync.len = READ_ONCE(sqe->addr);
4020 req->sync.mode = READ_ONCE(sqe->len);
4024 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
4028 /* fallocate always requiring blocking context */
4031 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4034 req_set_fail_links(req);
4035 io_req_complete(req, ret);
4039 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4041 const char __user *fname;
4044 if (unlikely(sqe->ioprio || sqe->buf_index))
4046 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4049 /* open.how should be already initialised */
4050 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4051 req->open.how.flags |= O_LARGEFILE;
4053 req->open.dfd = READ_ONCE(sqe->fd);
4054 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4055 req->open.filename = getname(fname);
4056 if (IS_ERR(req->open.filename)) {
4057 ret = PTR_ERR(req->open.filename);
4058 req->open.filename = NULL;
4061 req->open.nofile = rlimit(RLIMIT_NOFILE);
4062 req->open.ignore_nonblock = false;
4063 req->flags |= REQ_F_NEED_CLEANUP;
4067 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4071 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4073 mode = READ_ONCE(sqe->len);
4074 flags = READ_ONCE(sqe->open_flags);
4075 req->open.how = build_open_how(flags, mode);
4076 return __io_openat_prep(req, sqe);
4079 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4081 struct open_how __user *how;
4085 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4087 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4088 len = READ_ONCE(sqe->len);
4089 if (len < OPEN_HOW_SIZE_VER0)
4092 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4097 return __io_openat_prep(req, sqe);
4100 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4102 struct open_flags op;
4106 if (force_nonblock && !req->open.ignore_nonblock)
4109 ret = build_open_flags(&req->open.how, &op);
4113 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4117 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4120 ret = PTR_ERR(file);
4122 * A work-around to ensure that /proc/self works that way
4123 * that it should - if we get -EOPNOTSUPP back, then assume
4124 * that proc_self_get_link() failed us because we're in async
4125 * context. We should be safe to retry this from the task
4126 * itself with force_nonblock == false set, as it should not
4127 * block on lookup. Would be nice to know this upfront and
4128 * avoid the async dance, but doesn't seem feasible.
4130 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4131 req->open.ignore_nonblock = true;
4132 refcount_inc(&req->refs);
4133 io_req_task_queue(req);
4137 fsnotify_open(file);
4138 fd_install(ret, file);
4141 putname(req->open.filename);
4142 req->flags &= ~REQ_F_NEED_CLEANUP;
4144 req_set_fail_links(req);
4145 io_req_complete(req, ret);
4149 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4151 return io_openat2(req, force_nonblock);
4154 static int io_remove_buffers_prep(struct io_kiocb *req,
4155 const struct io_uring_sqe *sqe)
4157 struct io_provide_buf *p = &req->pbuf;
4160 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4163 tmp = READ_ONCE(sqe->fd);
4164 if (!tmp || tmp > USHRT_MAX)
4167 memset(p, 0, sizeof(*p));
4169 p->bgid = READ_ONCE(sqe->buf_group);
4173 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4174 int bgid, unsigned nbufs)
4178 /* shouldn't happen */
4182 /* the head kbuf is the list itself */
4183 while (!list_empty(&buf->list)) {
4184 struct io_buffer *nxt;
4186 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4187 list_del(&nxt->list);
4194 idr_remove(&ctx->io_buffer_idr, bgid);
4199 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4200 struct io_comp_state *cs)
4202 struct io_provide_buf *p = &req->pbuf;
4203 struct io_ring_ctx *ctx = req->ctx;
4204 struct io_buffer *head;
4207 io_ring_submit_lock(ctx, !force_nonblock);
4209 lockdep_assert_held(&ctx->uring_lock);
4212 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4214 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4216 req_set_fail_links(req);
4218 /* need to hold the lock to complete IOPOLL requests */
4219 if (ctx->flags & IORING_SETUP_IOPOLL) {
4220 __io_req_complete(req, ret, 0, cs);
4221 io_ring_submit_unlock(ctx, !force_nonblock);
4223 io_ring_submit_unlock(ctx, !force_nonblock);
4224 __io_req_complete(req, ret, 0, cs);
4229 static int io_provide_buffers_prep(struct io_kiocb *req,
4230 const struct io_uring_sqe *sqe)
4232 struct io_provide_buf *p = &req->pbuf;
4235 if (sqe->ioprio || sqe->rw_flags)
4238 tmp = READ_ONCE(sqe->fd);
4239 if (!tmp || tmp > USHRT_MAX)
4242 p->addr = READ_ONCE(sqe->addr);
4243 p->len = READ_ONCE(sqe->len);
4245 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4248 p->bgid = READ_ONCE(sqe->buf_group);
4249 tmp = READ_ONCE(sqe->off);
4250 if (tmp > USHRT_MAX)
4256 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4258 struct io_buffer *buf;
4259 u64 addr = pbuf->addr;
4260 int i, bid = pbuf->bid;
4262 for (i = 0; i < pbuf->nbufs; i++) {
4263 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4268 buf->len = pbuf->len;
4273 INIT_LIST_HEAD(&buf->list);
4276 list_add_tail(&buf->list, &(*head)->list);
4280 return i ? i : -ENOMEM;
4283 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4284 struct io_comp_state *cs)
4286 struct io_provide_buf *p = &req->pbuf;
4287 struct io_ring_ctx *ctx = req->ctx;
4288 struct io_buffer *head, *list;
4291 io_ring_submit_lock(ctx, !force_nonblock);
4293 lockdep_assert_held(&ctx->uring_lock);
4295 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4297 ret = io_add_buffers(p, &head);
4302 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4305 __io_remove_buffers(ctx, head, p->bgid, -1U);
4311 req_set_fail_links(req);
4313 /* need to hold the lock to complete IOPOLL requests */
4314 if (ctx->flags & IORING_SETUP_IOPOLL) {
4315 __io_req_complete(req, ret, 0, cs);
4316 io_ring_submit_unlock(ctx, !force_nonblock);
4318 io_ring_submit_unlock(ctx, !force_nonblock);
4319 __io_req_complete(req, ret, 0, cs);
4324 static int io_epoll_ctl_prep(struct io_kiocb *req,
4325 const struct io_uring_sqe *sqe)
4327 #if defined(CONFIG_EPOLL)
4328 if (sqe->ioprio || sqe->buf_index)
4330 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4333 req->epoll.epfd = READ_ONCE(sqe->fd);
4334 req->epoll.op = READ_ONCE(sqe->len);
4335 req->epoll.fd = READ_ONCE(sqe->off);
4337 if (ep_op_has_event(req->epoll.op)) {
4338 struct epoll_event __user *ev;
4340 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4341 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4351 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4352 struct io_comp_state *cs)
4354 #if defined(CONFIG_EPOLL)
4355 struct io_epoll *ie = &req->epoll;
4358 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4359 if (force_nonblock && ret == -EAGAIN)
4363 req_set_fail_links(req);
4364 __io_req_complete(req, ret, 0, cs);
4371 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4373 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4374 if (sqe->ioprio || sqe->buf_index || sqe->off)
4376 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4379 req->madvise.addr = READ_ONCE(sqe->addr);
4380 req->madvise.len = READ_ONCE(sqe->len);
4381 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4388 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4390 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4391 struct io_madvise *ma = &req->madvise;
4397 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4399 req_set_fail_links(req);
4400 io_req_complete(req, ret);
4407 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4409 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4411 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4414 req->fadvise.offset = READ_ONCE(sqe->off);
4415 req->fadvise.len = READ_ONCE(sqe->len);
4416 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4420 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4422 struct io_fadvise *fa = &req->fadvise;
4425 if (force_nonblock) {
4426 switch (fa->advice) {
4427 case POSIX_FADV_NORMAL:
4428 case POSIX_FADV_RANDOM:
4429 case POSIX_FADV_SEQUENTIAL:
4436 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4438 req_set_fail_links(req);
4439 io_req_complete(req, ret);
4443 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4445 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4447 if (sqe->ioprio || sqe->buf_index)
4449 if (req->flags & REQ_F_FIXED_FILE)
4452 req->statx.dfd = READ_ONCE(sqe->fd);
4453 req->statx.mask = READ_ONCE(sqe->len);
4454 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4455 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4456 req->statx.flags = READ_ONCE(sqe->statx_flags);
4461 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4463 struct io_statx *ctx = &req->statx;
4466 if (force_nonblock) {
4467 /* only need file table for an actual valid fd */
4468 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4469 req->flags |= REQ_F_NO_FILE_TABLE;
4473 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4477 req_set_fail_links(req);
4478 io_req_complete(req, ret);
4482 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4485 * If we queue this for async, it must not be cancellable. That would
4486 * leave the 'file' in an undeterminate state, and here need to modify
4487 * io_wq_work.flags, so initialize io_wq_work firstly.
4489 io_req_init_async(req);
4491 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4493 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4494 sqe->rw_flags || sqe->buf_index)
4496 if (req->flags & REQ_F_FIXED_FILE)
4499 req->close.fd = READ_ONCE(sqe->fd);
4500 if ((req->file && req->file->f_op == &io_uring_fops))
4503 req->close.put_file = NULL;
4507 static int io_close(struct io_kiocb *req, bool force_nonblock,
4508 struct io_comp_state *cs)
4510 struct io_close *close = &req->close;
4513 /* might be already done during nonblock submission */
4514 if (!close->put_file) {
4515 ret = close_fd_get_file(close->fd, &close->put_file);
4517 return (ret == -ENOENT) ? -EBADF : ret;
4520 /* if the file has a flush method, be safe and punt to async */
4521 if (close->put_file->f_op->flush && force_nonblock) {
4522 /* not safe to cancel at this point */
4523 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4524 /* was never set, but play safe */
4525 req->flags &= ~REQ_F_NOWAIT;
4526 /* avoid grabbing files - we don't need the files */
4527 req->flags |= REQ_F_NO_FILE_TABLE;
4531 /* No ->flush() or already async, safely close from here */
4532 ret = filp_close(close->put_file, req->work.identity->files);
4534 req_set_fail_links(req);
4535 fput(close->put_file);
4536 close->put_file = NULL;
4537 __io_req_complete(req, ret, 0, cs);
4541 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4543 struct io_ring_ctx *ctx = req->ctx;
4548 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4550 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4553 req->sync.off = READ_ONCE(sqe->off);
4554 req->sync.len = READ_ONCE(sqe->len);
4555 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4559 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4563 /* sync_file_range always requires a blocking context */
4567 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4570 req_set_fail_links(req);
4571 io_req_complete(req, ret);
4575 #if defined(CONFIG_NET)
4576 static int io_setup_async_msg(struct io_kiocb *req,
4577 struct io_async_msghdr *kmsg)
4579 struct io_async_msghdr *async_msg = req->async_data;
4583 if (io_alloc_async_data(req)) {
4584 if (kmsg->iov != kmsg->fast_iov)
4588 async_msg = req->async_data;
4589 req->flags |= REQ_F_NEED_CLEANUP;
4590 memcpy(async_msg, kmsg, sizeof(*kmsg));
4594 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4595 struct io_async_msghdr *iomsg)
4597 iomsg->iov = iomsg->fast_iov;
4598 iomsg->msg.msg_name = &iomsg->addr;
4599 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4600 req->sr_msg.msg_flags, &iomsg->iov);
4603 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4605 struct io_async_msghdr *async_msg = req->async_data;
4606 struct io_sr_msg *sr = &req->sr_msg;
4609 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4612 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4613 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4614 sr->len = READ_ONCE(sqe->len);
4616 #ifdef CONFIG_COMPAT
4617 if (req->ctx->compat)
4618 sr->msg_flags |= MSG_CMSG_COMPAT;
4621 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4623 ret = io_sendmsg_copy_hdr(req, async_msg);
4625 req->flags |= REQ_F_NEED_CLEANUP;
4629 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4630 struct io_comp_state *cs)
4632 struct io_async_msghdr iomsg, *kmsg;
4633 struct socket *sock;
4637 sock = sock_from_file(req->file);
4638 if (unlikely(!sock))
4641 if (req->async_data) {
4642 kmsg = req->async_data;
4643 kmsg->msg.msg_name = &kmsg->addr;
4644 /* if iov is set, it's allocated already */
4646 kmsg->iov = kmsg->fast_iov;
4647 kmsg->msg.msg_iter.iov = kmsg->iov;
4649 ret = io_sendmsg_copy_hdr(req, &iomsg);
4655 flags = req->sr_msg.msg_flags;
4656 if (flags & MSG_DONTWAIT)
4657 req->flags |= REQ_F_NOWAIT;
4658 else if (force_nonblock)
4659 flags |= MSG_DONTWAIT;
4661 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4662 if (force_nonblock && ret == -EAGAIN)
4663 return io_setup_async_msg(req, kmsg);
4664 if (ret == -ERESTARTSYS)
4667 if (kmsg->iov != kmsg->fast_iov)
4669 req->flags &= ~REQ_F_NEED_CLEANUP;
4671 req_set_fail_links(req);
4672 __io_req_complete(req, ret, 0, cs);
4676 static int io_send(struct io_kiocb *req, bool force_nonblock,
4677 struct io_comp_state *cs)
4679 struct io_sr_msg *sr = &req->sr_msg;
4682 struct socket *sock;
4686 sock = sock_from_file(req->file);
4687 if (unlikely(!sock))
4690 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4694 msg.msg_name = NULL;
4695 msg.msg_control = NULL;
4696 msg.msg_controllen = 0;
4697 msg.msg_namelen = 0;
4699 flags = req->sr_msg.msg_flags;
4700 if (flags & MSG_DONTWAIT)
4701 req->flags |= REQ_F_NOWAIT;
4702 else if (force_nonblock)
4703 flags |= MSG_DONTWAIT;
4705 msg.msg_flags = flags;
4706 ret = sock_sendmsg(sock, &msg);
4707 if (force_nonblock && ret == -EAGAIN)
4709 if (ret == -ERESTARTSYS)
4713 req_set_fail_links(req);
4714 __io_req_complete(req, ret, 0, cs);
4718 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4719 struct io_async_msghdr *iomsg)
4721 struct io_sr_msg *sr = &req->sr_msg;
4722 struct iovec __user *uiov;
4726 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4727 &iomsg->uaddr, &uiov, &iov_len);
4731 if (req->flags & REQ_F_BUFFER_SELECT) {
4734 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4736 sr->len = iomsg->iov[0].iov_len;
4737 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4741 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4742 &iomsg->iov, &iomsg->msg.msg_iter,
4751 #ifdef CONFIG_COMPAT
4752 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4753 struct io_async_msghdr *iomsg)
4755 struct compat_msghdr __user *msg_compat;
4756 struct io_sr_msg *sr = &req->sr_msg;
4757 struct compat_iovec __user *uiov;
4762 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4763 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4768 uiov = compat_ptr(ptr);
4769 if (req->flags & REQ_F_BUFFER_SELECT) {
4770 compat_ssize_t clen;
4774 if (!access_ok(uiov, sizeof(*uiov)))
4776 if (__get_user(clen, &uiov->iov_len))
4781 iomsg->iov[0].iov_len = clen;
4784 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4785 UIO_FASTIOV, &iomsg->iov,
4786 &iomsg->msg.msg_iter, true);
4795 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4796 struct io_async_msghdr *iomsg)
4798 iomsg->msg.msg_name = &iomsg->addr;
4799 iomsg->iov = iomsg->fast_iov;
4801 #ifdef CONFIG_COMPAT
4802 if (req->ctx->compat)
4803 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4806 return __io_recvmsg_copy_hdr(req, iomsg);
4809 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4812 struct io_sr_msg *sr = &req->sr_msg;
4813 struct io_buffer *kbuf;
4815 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4820 req->flags |= REQ_F_BUFFER_SELECTED;
4824 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4826 return io_put_kbuf(req, req->sr_msg.kbuf);
4829 static int io_recvmsg_prep(struct io_kiocb *req,
4830 const struct io_uring_sqe *sqe)
4832 struct io_async_msghdr *async_msg = req->async_data;
4833 struct io_sr_msg *sr = &req->sr_msg;
4836 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4839 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4840 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4841 sr->len = READ_ONCE(sqe->len);
4842 sr->bgid = READ_ONCE(sqe->buf_group);
4844 #ifdef CONFIG_COMPAT
4845 if (req->ctx->compat)
4846 sr->msg_flags |= MSG_CMSG_COMPAT;
4849 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4851 ret = io_recvmsg_copy_hdr(req, async_msg);
4853 req->flags |= REQ_F_NEED_CLEANUP;
4857 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4858 struct io_comp_state *cs)
4860 struct io_async_msghdr iomsg, *kmsg;
4861 struct socket *sock;
4862 struct io_buffer *kbuf;
4864 int ret, cflags = 0;
4866 sock = sock_from_file(req->file);
4867 if (unlikely(!sock))
4870 if (req->async_data) {
4871 kmsg = req->async_data;
4872 kmsg->msg.msg_name = &kmsg->addr;
4873 /* if iov is set, it's allocated already */
4875 kmsg->iov = kmsg->fast_iov;
4876 kmsg->msg.msg_iter.iov = kmsg->iov;
4878 ret = io_recvmsg_copy_hdr(req, &iomsg);
4884 if (req->flags & REQ_F_BUFFER_SELECT) {
4885 kbuf = io_recv_buffer_select(req, !force_nonblock);
4887 return PTR_ERR(kbuf);
4888 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4889 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4890 1, req->sr_msg.len);
4893 flags = req->sr_msg.msg_flags;
4894 if (flags & MSG_DONTWAIT)
4895 req->flags |= REQ_F_NOWAIT;
4896 else if (force_nonblock)
4897 flags |= MSG_DONTWAIT;
4899 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4900 kmsg->uaddr, flags);
4901 if (force_nonblock && ret == -EAGAIN)
4902 return io_setup_async_msg(req, kmsg);
4903 if (ret == -ERESTARTSYS)
4906 if (req->flags & REQ_F_BUFFER_SELECTED)
4907 cflags = io_put_recv_kbuf(req);
4908 if (kmsg->iov != kmsg->fast_iov)
4910 req->flags &= ~REQ_F_NEED_CLEANUP;
4912 req_set_fail_links(req);
4913 __io_req_complete(req, ret, cflags, cs);
4917 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4918 struct io_comp_state *cs)
4920 struct io_buffer *kbuf;
4921 struct io_sr_msg *sr = &req->sr_msg;
4923 void __user *buf = sr->buf;
4924 struct socket *sock;
4927 int ret, cflags = 0;
4929 sock = sock_from_file(req->file);
4930 if (unlikely(!sock))
4933 if (req->flags & REQ_F_BUFFER_SELECT) {
4934 kbuf = io_recv_buffer_select(req, !force_nonblock);
4936 return PTR_ERR(kbuf);
4937 buf = u64_to_user_ptr(kbuf->addr);
4940 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4944 msg.msg_name = NULL;
4945 msg.msg_control = NULL;
4946 msg.msg_controllen = 0;
4947 msg.msg_namelen = 0;
4948 msg.msg_iocb = NULL;
4951 flags = req->sr_msg.msg_flags;
4952 if (flags & MSG_DONTWAIT)
4953 req->flags |= REQ_F_NOWAIT;
4954 else if (force_nonblock)
4955 flags |= MSG_DONTWAIT;
4957 ret = sock_recvmsg(sock, &msg, flags);
4958 if (force_nonblock && ret == -EAGAIN)
4960 if (ret == -ERESTARTSYS)
4963 if (req->flags & REQ_F_BUFFER_SELECTED)
4964 cflags = io_put_recv_kbuf(req);
4966 req_set_fail_links(req);
4967 __io_req_complete(req, ret, cflags, cs);
4971 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4973 struct io_accept *accept = &req->accept;
4975 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4977 if (sqe->ioprio || sqe->len || sqe->buf_index)
4980 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4981 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4982 accept->flags = READ_ONCE(sqe->accept_flags);
4983 accept->nofile = rlimit(RLIMIT_NOFILE);
4987 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4988 struct io_comp_state *cs)
4990 struct io_accept *accept = &req->accept;
4991 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4994 if (req->file->f_flags & O_NONBLOCK)
4995 req->flags |= REQ_F_NOWAIT;
4997 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4998 accept->addr_len, accept->flags,
5000 if (ret == -EAGAIN && force_nonblock)
5003 if (ret == -ERESTARTSYS)
5005 req_set_fail_links(req);
5007 __io_req_complete(req, ret, 0, cs);
5011 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5013 struct io_connect *conn = &req->connect;
5014 struct io_async_connect *io = req->async_data;
5016 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5018 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
5021 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5022 conn->addr_len = READ_ONCE(sqe->addr2);
5027 return move_addr_to_kernel(conn->addr, conn->addr_len,
5031 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5032 struct io_comp_state *cs)
5034 struct io_async_connect __io, *io;
5035 unsigned file_flags;
5038 if (req->async_data) {
5039 io = req->async_data;
5041 ret = move_addr_to_kernel(req->connect.addr,
5042 req->connect.addr_len,
5049 file_flags = force_nonblock ? O_NONBLOCK : 0;
5051 ret = __sys_connect_file(req->file, &io->address,
5052 req->connect.addr_len, file_flags);
5053 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5054 if (req->async_data)
5056 if (io_alloc_async_data(req)) {
5060 io = req->async_data;
5061 memcpy(req->async_data, &__io, sizeof(__io));
5064 if (ret == -ERESTARTSYS)
5068 req_set_fail_links(req);
5069 __io_req_complete(req, ret, 0, cs);
5072 #else /* !CONFIG_NET */
5073 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5078 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5079 struct io_comp_state *cs)
5084 static int io_send(struct io_kiocb *req, bool force_nonblock,
5085 struct io_comp_state *cs)
5090 static int io_recvmsg_prep(struct io_kiocb *req,
5091 const struct io_uring_sqe *sqe)
5096 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5097 struct io_comp_state *cs)
5102 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5103 struct io_comp_state *cs)
5108 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5113 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5114 struct io_comp_state *cs)
5119 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5124 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5125 struct io_comp_state *cs)
5129 #endif /* CONFIG_NET */
5131 struct io_poll_table {
5132 struct poll_table_struct pt;
5133 struct io_kiocb *req;
5137 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5138 __poll_t mask, task_work_func_t func)
5142 /* for instances that support it check for an event match first: */
5143 if (mask && !(mask & poll->events))
5146 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5148 list_del_init(&poll->wait.entry);
5151 init_task_work(&req->task_work, func);
5152 percpu_ref_get(&req->ctx->refs);
5155 * If this fails, then the task is exiting. When a task exits, the
5156 * work gets canceled, so just cancel this request as well instead
5157 * of executing it. We can't safely execute it anyway, as we may not
5158 * have the needed state needed for it anyway.
5160 ret = io_req_task_work_add(req);
5161 if (unlikely(ret)) {
5162 struct task_struct *tsk;
5164 WRITE_ONCE(poll->canceled, true);
5165 tsk = io_wq_get_task(req->ctx->io_wq);
5166 task_work_add(tsk, &req->task_work, TWA_NONE);
5167 wake_up_process(tsk);
5172 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5173 __acquires(&req->ctx->completion_lock)
5175 struct io_ring_ctx *ctx = req->ctx;
5177 if (!req->result && !READ_ONCE(poll->canceled)) {
5178 struct poll_table_struct pt = { ._key = poll->events };
5180 req->result = vfs_poll(req->file, &pt) & poll->events;
5183 spin_lock_irq(&ctx->completion_lock);
5184 if (!req->result && !READ_ONCE(poll->canceled)) {
5185 add_wait_queue(poll->head, &poll->wait);
5192 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5194 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5195 if (req->opcode == IORING_OP_POLL_ADD)
5196 return req->async_data;
5197 return req->apoll->double_poll;
5200 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5202 if (req->opcode == IORING_OP_POLL_ADD)
5204 return &req->apoll->poll;
5207 static void io_poll_remove_double(struct io_kiocb *req)
5209 struct io_poll_iocb *poll = io_poll_get_double(req);
5211 lockdep_assert_held(&req->ctx->completion_lock);
5213 if (poll && poll->head) {
5214 struct wait_queue_head *head = poll->head;
5216 spin_lock(&head->lock);
5217 list_del_init(&poll->wait.entry);
5218 if (poll->wait.private)
5219 refcount_dec(&req->refs);
5221 spin_unlock(&head->lock);
5225 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5227 struct io_ring_ctx *ctx = req->ctx;
5229 io_poll_remove_double(req);
5230 req->poll.done = true;
5231 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5232 io_commit_cqring(ctx);
5235 static void io_poll_task_func(struct callback_head *cb)
5237 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5238 struct io_ring_ctx *ctx = req->ctx;
5239 struct io_kiocb *nxt;
5241 if (io_poll_rewait(req, &req->poll)) {
5242 spin_unlock_irq(&ctx->completion_lock);
5244 hash_del(&req->hash_node);
5245 io_poll_complete(req, req->result, 0);
5246 spin_unlock_irq(&ctx->completion_lock);
5248 nxt = io_put_req_find_next(req);
5249 io_cqring_ev_posted(ctx);
5251 __io_req_task_submit(nxt);
5254 percpu_ref_put(&ctx->refs);
5257 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5258 int sync, void *key)
5260 struct io_kiocb *req = wait->private;
5261 struct io_poll_iocb *poll = io_poll_get_single(req);
5262 __poll_t mask = key_to_poll(key);
5264 /* for instances that support it check for an event match first: */
5265 if (mask && !(mask & poll->events))
5268 list_del_init(&wait->entry);
5270 if (poll && poll->head) {
5273 spin_lock(&poll->head->lock);
5274 done = list_empty(&poll->wait.entry);
5276 list_del_init(&poll->wait.entry);
5277 /* make sure double remove sees this as being gone */
5278 wait->private = NULL;
5279 spin_unlock(&poll->head->lock);
5281 /* use wait func handler, so it matches the rq type */
5282 poll->wait.func(&poll->wait, mode, sync, key);
5285 refcount_dec(&req->refs);
5289 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5290 wait_queue_func_t wake_func)
5294 poll->canceled = false;
5295 poll->events = events;
5296 INIT_LIST_HEAD(&poll->wait.entry);
5297 init_waitqueue_func_entry(&poll->wait, wake_func);
5300 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5301 struct wait_queue_head *head,
5302 struct io_poll_iocb **poll_ptr)
5304 struct io_kiocb *req = pt->req;
5307 * If poll->head is already set, it's because the file being polled
5308 * uses multiple waitqueues for poll handling (eg one for read, one
5309 * for write). Setup a separate io_poll_iocb if this happens.
5311 if (unlikely(poll->head)) {
5312 struct io_poll_iocb *poll_one = poll;
5314 /* already have a 2nd entry, fail a third attempt */
5316 pt->error = -EINVAL;
5319 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5321 pt->error = -ENOMEM;
5324 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5325 refcount_inc(&req->refs);
5326 poll->wait.private = req;
5333 if (poll->events & EPOLLEXCLUSIVE)
5334 add_wait_queue_exclusive(head, &poll->wait);
5336 add_wait_queue(head, &poll->wait);
5339 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5340 struct poll_table_struct *p)
5342 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5343 struct async_poll *apoll = pt->req->apoll;
5345 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5348 static void io_async_task_func(struct callback_head *cb)
5350 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5351 struct async_poll *apoll = req->apoll;
5352 struct io_ring_ctx *ctx = req->ctx;
5354 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5356 if (io_poll_rewait(req, &apoll->poll)) {
5357 spin_unlock_irq(&ctx->completion_lock);
5358 percpu_ref_put(&ctx->refs);
5362 /* If req is still hashed, it cannot have been canceled. Don't check. */
5363 if (hash_hashed(&req->hash_node))
5364 hash_del(&req->hash_node);
5366 io_poll_remove_double(req);
5367 spin_unlock_irq(&ctx->completion_lock);
5369 if (!READ_ONCE(apoll->poll.canceled))
5370 __io_req_task_submit(req);
5372 __io_req_task_cancel(req, -ECANCELED);
5374 percpu_ref_put(&ctx->refs);
5375 kfree(apoll->double_poll);
5379 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5382 struct io_kiocb *req = wait->private;
5383 struct io_poll_iocb *poll = &req->apoll->poll;
5385 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5388 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5391 static void io_poll_req_insert(struct io_kiocb *req)
5393 struct io_ring_ctx *ctx = req->ctx;
5394 struct hlist_head *list;
5396 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5397 hlist_add_head(&req->hash_node, list);
5400 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5401 struct io_poll_iocb *poll,
5402 struct io_poll_table *ipt, __poll_t mask,
5403 wait_queue_func_t wake_func)
5404 __acquires(&ctx->completion_lock)
5406 struct io_ring_ctx *ctx = req->ctx;
5407 bool cancel = false;
5409 INIT_HLIST_NODE(&req->hash_node);
5410 io_init_poll_iocb(poll, mask, wake_func);
5411 poll->file = req->file;
5412 poll->wait.private = req;
5414 ipt->pt._key = mask;
5416 ipt->error = -EINVAL;
5418 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5420 spin_lock_irq(&ctx->completion_lock);
5421 if (likely(poll->head)) {
5422 spin_lock(&poll->head->lock);
5423 if (unlikely(list_empty(&poll->wait.entry))) {
5429 if (mask || ipt->error)
5430 list_del_init(&poll->wait.entry);
5432 WRITE_ONCE(poll->canceled, true);
5433 else if (!poll->done) /* actually waiting for an event */
5434 io_poll_req_insert(req);
5435 spin_unlock(&poll->head->lock);
5441 static bool io_arm_poll_handler(struct io_kiocb *req)
5443 const struct io_op_def *def = &io_op_defs[req->opcode];
5444 struct io_ring_ctx *ctx = req->ctx;
5445 struct async_poll *apoll;
5446 struct io_poll_table ipt;
5450 if (!req->file || !file_can_poll(req->file))
5452 if (req->flags & REQ_F_POLLED)
5456 else if (def->pollout)
5460 /* if we can't nonblock try, then no point in arming a poll handler */
5461 if (!io_file_supports_async(req->file, rw))
5464 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5465 if (unlikely(!apoll))
5467 apoll->double_poll = NULL;
5469 req->flags |= REQ_F_POLLED;
5474 mask |= POLLIN | POLLRDNORM;
5476 mask |= POLLOUT | POLLWRNORM;
5478 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5479 if ((req->opcode == IORING_OP_RECVMSG) &&
5480 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5483 mask |= POLLERR | POLLPRI;
5485 ipt.pt._qproc = io_async_queue_proc;
5487 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5489 if (ret || ipt.error) {
5490 io_poll_remove_double(req);
5491 spin_unlock_irq(&ctx->completion_lock);
5492 kfree(apoll->double_poll);
5496 spin_unlock_irq(&ctx->completion_lock);
5497 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5498 apoll->poll.events);
5502 static bool __io_poll_remove_one(struct io_kiocb *req,
5503 struct io_poll_iocb *poll)
5505 bool do_complete = false;
5507 spin_lock(&poll->head->lock);
5508 WRITE_ONCE(poll->canceled, true);
5509 if (!list_empty(&poll->wait.entry)) {
5510 list_del_init(&poll->wait.entry);
5513 spin_unlock(&poll->head->lock);
5514 hash_del(&req->hash_node);
5518 static bool io_poll_remove_one(struct io_kiocb *req)
5522 io_poll_remove_double(req);
5524 if (req->opcode == IORING_OP_POLL_ADD) {
5525 do_complete = __io_poll_remove_one(req, &req->poll);
5527 struct async_poll *apoll = req->apoll;
5529 /* non-poll requests have submit ref still */
5530 do_complete = __io_poll_remove_one(req, &apoll->poll);
5533 kfree(apoll->double_poll);
5539 io_cqring_fill_event(req, -ECANCELED);
5540 io_commit_cqring(req->ctx);
5541 req_set_fail_links(req);
5542 io_put_req_deferred(req, 1);
5549 * Returns true if we found and killed one or more poll requests
5551 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5552 struct files_struct *files)
5554 struct hlist_node *tmp;
5555 struct io_kiocb *req;
5558 spin_lock_irq(&ctx->completion_lock);
5559 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5560 struct hlist_head *list;
5562 list = &ctx->cancel_hash[i];
5563 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5564 if (io_match_task(req, tsk, files))
5565 posted += io_poll_remove_one(req);
5568 spin_unlock_irq(&ctx->completion_lock);
5571 io_cqring_ev_posted(ctx);
5576 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5578 struct hlist_head *list;
5579 struct io_kiocb *req;
5581 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5582 hlist_for_each_entry(req, list, hash_node) {
5583 if (sqe_addr != req->user_data)
5585 if (io_poll_remove_one(req))
5593 static int io_poll_remove_prep(struct io_kiocb *req,
5594 const struct io_uring_sqe *sqe)
5596 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5598 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5602 req->poll_remove.addr = READ_ONCE(sqe->addr);
5607 * Find a running poll command that matches one specified in sqe->addr,
5608 * and remove it if found.
5610 static int io_poll_remove(struct io_kiocb *req)
5612 struct io_ring_ctx *ctx = req->ctx;
5615 spin_lock_irq(&ctx->completion_lock);
5616 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5617 spin_unlock_irq(&ctx->completion_lock);
5620 req_set_fail_links(req);
5621 io_req_complete(req, ret);
5625 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5628 struct io_kiocb *req = wait->private;
5629 struct io_poll_iocb *poll = &req->poll;
5631 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5634 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5635 struct poll_table_struct *p)
5637 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5639 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5642 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5644 struct io_poll_iocb *poll = &req->poll;
5647 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5649 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5652 events = READ_ONCE(sqe->poll32_events);
5654 events = swahw32(events);
5656 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5657 (events & EPOLLEXCLUSIVE);
5661 static int io_poll_add(struct io_kiocb *req)
5663 struct io_poll_iocb *poll = &req->poll;
5664 struct io_ring_ctx *ctx = req->ctx;
5665 struct io_poll_table ipt;
5668 ipt.pt._qproc = io_poll_queue_proc;
5670 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5673 if (mask) { /* no async, we'd stolen it */
5675 io_poll_complete(req, mask, 0);
5677 spin_unlock_irq(&ctx->completion_lock);
5680 io_cqring_ev_posted(ctx);
5686 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5688 struct io_timeout_data *data = container_of(timer,
5689 struct io_timeout_data, timer);
5690 struct io_kiocb *req = data->req;
5691 struct io_ring_ctx *ctx = req->ctx;
5692 unsigned long flags;
5694 spin_lock_irqsave(&ctx->completion_lock, flags);
5695 list_del_init(&req->timeout.list);
5696 atomic_set(&req->ctx->cq_timeouts,
5697 atomic_read(&req->ctx->cq_timeouts) + 1);
5699 io_cqring_fill_event(req, -ETIME);
5700 io_commit_cqring(ctx);
5701 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5703 io_cqring_ev_posted(ctx);
5704 req_set_fail_links(req);
5706 return HRTIMER_NORESTART;
5709 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5712 struct io_timeout_data *io;
5713 struct io_kiocb *req;
5716 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5717 if (user_data == req->user_data) {
5724 return ERR_PTR(ret);
5726 io = req->async_data;
5727 ret = hrtimer_try_to_cancel(&io->timer);
5729 return ERR_PTR(-EALREADY);
5730 list_del_init(&req->timeout.list);
5734 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5736 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5739 return PTR_ERR(req);
5741 req_set_fail_links(req);
5742 io_cqring_fill_event(req, -ECANCELED);
5743 io_put_req_deferred(req, 1);
5747 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5748 struct timespec64 *ts, enum hrtimer_mode mode)
5750 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5751 struct io_timeout_data *data;
5754 return PTR_ERR(req);
5756 req->timeout.off = 0; /* noseq */
5757 data = req->async_data;
5758 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5759 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5760 data->timer.function = io_timeout_fn;
5761 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5765 static int io_timeout_remove_prep(struct io_kiocb *req,
5766 const struct io_uring_sqe *sqe)
5768 struct io_timeout_rem *tr = &req->timeout_rem;
5770 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5772 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5774 if (sqe->ioprio || sqe->buf_index || sqe->len)
5777 tr->addr = READ_ONCE(sqe->addr);
5778 tr->flags = READ_ONCE(sqe->timeout_flags);
5779 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5780 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5782 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5784 } else if (tr->flags) {
5785 /* timeout removal doesn't support flags */
5793 * Remove or update an existing timeout command
5795 static int io_timeout_remove(struct io_kiocb *req)
5797 struct io_timeout_rem *tr = &req->timeout_rem;
5798 struct io_ring_ctx *ctx = req->ctx;
5801 spin_lock_irq(&ctx->completion_lock);
5802 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5803 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5804 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5806 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5808 ret = io_timeout_cancel(ctx, tr->addr);
5811 io_cqring_fill_event(req, ret);
5812 io_commit_cqring(ctx);
5813 spin_unlock_irq(&ctx->completion_lock);
5814 io_cqring_ev_posted(ctx);
5816 req_set_fail_links(req);
5821 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5822 bool is_timeout_link)
5824 struct io_timeout_data *data;
5826 u32 off = READ_ONCE(sqe->off);
5828 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5830 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5832 if (off && is_timeout_link)
5834 flags = READ_ONCE(sqe->timeout_flags);
5835 if (flags & ~IORING_TIMEOUT_ABS)
5838 req->timeout.off = off;
5840 if (!req->async_data && io_alloc_async_data(req))
5843 data = req->async_data;
5846 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5849 if (flags & IORING_TIMEOUT_ABS)
5850 data->mode = HRTIMER_MODE_ABS;
5852 data->mode = HRTIMER_MODE_REL;
5854 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5858 static int io_timeout(struct io_kiocb *req)
5860 struct io_ring_ctx *ctx = req->ctx;
5861 struct io_timeout_data *data = req->async_data;
5862 struct list_head *entry;
5863 u32 tail, off = req->timeout.off;
5865 spin_lock_irq(&ctx->completion_lock);
5868 * sqe->off holds how many events that need to occur for this
5869 * timeout event to be satisfied. If it isn't set, then this is
5870 * a pure timeout request, sequence isn't used.
5872 if (io_is_timeout_noseq(req)) {
5873 entry = ctx->timeout_list.prev;
5877 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5878 req->timeout.target_seq = tail + off;
5880 /* Update the last seq here in case io_flush_timeouts() hasn't.
5881 * This is safe because ->completion_lock is held, and submissions
5882 * and completions are never mixed in the same ->completion_lock section.
5884 ctx->cq_last_tm_flush = tail;
5887 * Insertion sort, ensuring the first entry in the list is always
5888 * the one we need first.
5890 list_for_each_prev(entry, &ctx->timeout_list) {
5891 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5894 if (io_is_timeout_noseq(nxt))
5896 /* nxt.seq is behind @tail, otherwise would've been completed */
5897 if (off >= nxt->timeout.target_seq - tail)
5901 list_add(&req->timeout.list, entry);
5902 data->timer.function = io_timeout_fn;
5903 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5904 spin_unlock_irq(&ctx->completion_lock);
5908 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5910 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5912 return req->user_data == (unsigned long) data;
5915 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5917 enum io_wq_cancel cancel_ret;
5920 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5921 switch (cancel_ret) {
5922 case IO_WQ_CANCEL_OK:
5925 case IO_WQ_CANCEL_RUNNING:
5928 case IO_WQ_CANCEL_NOTFOUND:
5936 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5937 struct io_kiocb *req, __u64 sqe_addr,
5940 unsigned long flags;
5943 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5944 if (ret != -ENOENT) {
5945 spin_lock_irqsave(&ctx->completion_lock, flags);
5949 spin_lock_irqsave(&ctx->completion_lock, flags);
5950 ret = io_timeout_cancel(ctx, sqe_addr);
5953 ret = io_poll_cancel(ctx, sqe_addr);
5957 io_cqring_fill_event(req, ret);
5958 io_commit_cqring(ctx);
5959 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5960 io_cqring_ev_posted(ctx);
5963 req_set_fail_links(req);
5967 static int io_async_cancel_prep(struct io_kiocb *req,
5968 const struct io_uring_sqe *sqe)
5970 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5972 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5974 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5977 req->cancel.addr = READ_ONCE(sqe->addr);
5981 static int io_async_cancel(struct io_kiocb *req)
5983 struct io_ring_ctx *ctx = req->ctx;
5985 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5989 static int io_files_update_prep(struct io_kiocb *req,
5990 const struct io_uring_sqe *sqe)
5992 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5994 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5996 if (sqe->ioprio || sqe->rw_flags)
5999 req->files_update.offset = READ_ONCE(sqe->off);
6000 req->files_update.nr_args = READ_ONCE(sqe->len);
6001 if (!req->files_update.nr_args)
6003 req->files_update.arg = READ_ONCE(sqe->addr);
6007 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
6008 struct io_comp_state *cs)
6010 struct io_ring_ctx *ctx = req->ctx;
6011 struct io_uring_files_update up;
6017 up.offset = req->files_update.offset;
6018 up.fds = req->files_update.arg;
6020 mutex_lock(&ctx->uring_lock);
6021 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
6022 mutex_unlock(&ctx->uring_lock);
6025 req_set_fail_links(req);
6026 __io_req_complete(req, ret, 0, cs);
6030 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6032 switch (req->opcode) {
6035 case IORING_OP_READV:
6036 case IORING_OP_READ_FIXED:
6037 case IORING_OP_READ:
6038 return io_read_prep(req, sqe);
6039 case IORING_OP_WRITEV:
6040 case IORING_OP_WRITE_FIXED:
6041 case IORING_OP_WRITE:
6042 return io_write_prep(req, sqe);
6043 case IORING_OP_POLL_ADD:
6044 return io_poll_add_prep(req, sqe);
6045 case IORING_OP_POLL_REMOVE:
6046 return io_poll_remove_prep(req, sqe);
6047 case IORING_OP_FSYNC:
6048 return io_prep_fsync(req, sqe);
6049 case IORING_OP_SYNC_FILE_RANGE:
6050 return io_prep_sfr(req, sqe);
6051 case IORING_OP_SENDMSG:
6052 case IORING_OP_SEND:
6053 return io_sendmsg_prep(req, sqe);
6054 case IORING_OP_RECVMSG:
6055 case IORING_OP_RECV:
6056 return io_recvmsg_prep(req, sqe);
6057 case IORING_OP_CONNECT:
6058 return io_connect_prep(req, sqe);
6059 case IORING_OP_TIMEOUT:
6060 return io_timeout_prep(req, sqe, false);
6061 case IORING_OP_TIMEOUT_REMOVE:
6062 return io_timeout_remove_prep(req, sqe);
6063 case IORING_OP_ASYNC_CANCEL:
6064 return io_async_cancel_prep(req, sqe);
6065 case IORING_OP_LINK_TIMEOUT:
6066 return io_timeout_prep(req, sqe, true);
6067 case IORING_OP_ACCEPT:
6068 return io_accept_prep(req, sqe);
6069 case IORING_OP_FALLOCATE:
6070 return io_fallocate_prep(req, sqe);
6071 case IORING_OP_OPENAT:
6072 return io_openat_prep(req, sqe);
6073 case IORING_OP_CLOSE:
6074 return io_close_prep(req, sqe);
6075 case IORING_OP_FILES_UPDATE:
6076 return io_files_update_prep(req, sqe);
6077 case IORING_OP_STATX:
6078 return io_statx_prep(req, sqe);
6079 case IORING_OP_FADVISE:
6080 return io_fadvise_prep(req, sqe);
6081 case IORING_OP_MADVISE:
6082 return io_madvise_prep(req, sqe);
6083 case IORING_OP_OPENAT2:
6084 return io_openat2_prep(req, sqe);
6085 case IORING_OP_EPOLL_CTL:
6086 return io_epoll_ctl_prep(req, sqe);
6087 case IORING_OP_SPLICE:
6088 return io_splice_prep(req, sqe);
6089 case IORING_OP_PROVIDE_BUFFERS:
6090 return io_provide_buffers_prep(req, sqe);
6091 case IORING_OP_REMOVE_BUFFERS:
6092 return io_remove_buffers_prep(req, sqe);
6094 return io_tee_prep(req, sqe);
6095 case IORING_OP_SHUTDOWN:
6096 return io_shutdown_prep(req, sqe);
6097 case IORING_OP_RENAMEAT:
6098 return io_renameat_prep(req, sqe);
6099 case IORING_OP_UNLINKAT:
6100 return io_unlinkat_prep(req, sqe);
6103 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6108 static int io_req_defer_prep(struct io_kiocb *req,
6109 const struct io_uring_sqe *sqe)
6113 if (io_alloc_async_data(req))
6115 return io_req_prep(req, sqe);
6118 static u32 io_get_sequence(struct io_kiocb *req)
6120 struct io_kiocb *pos;
6121 struct io_ring_ctx *ctx = req->ctx;
6122 u32 total_submitted, nr_reqs = 0;
6124 io_for_each_link(pos, req)
6127 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6128 return total_submitted - nr_reqs;
6131 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6133 struct io_ring_ctx *ctx = req->ctx;
6134 struct io_defer_entry *de;
6138 /* Still need defer if there is pending req in defer list. */
6139 if (likely(list_empty_careful(&ctx->defer_list) &&
6140 !(req->flags & REQ_F_IO_DRAIN)))
6143 seq = io_get_sequence(req);
6144 /* Still a chance to pass the sequence check */
6145 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6148 if (!req->async_data) {
6149 ret = io_req_defer_prep(req, sqe);
6153 io_prep_async_link(req);
6154 de = kmalloc(sizeof(*de), GFP_KERNEL);
6158 spin_lock_irq(&ctx->completion_lock);
6159 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6160 spin_unlock_irq(&ctx->completion_lock);
6162 io_queue_async_work(req);
6163 return -EIOCBQUEUED;
6166 trace_io_uring_defer(ctx, req, req->user_data);
6169 list_add_tail(&de->list, &ctx->defer_list);
6170 spin_unlock_irq(&ctx->completion_lock);
6171 return -EIOCBQUEUED;
6174 static void io_req_drop_files(struct io_kiocb *req)
6176 struct io_ring_ctx *ctx = req->ctx;
6177 struct io_uring_task *tctx = req->task->io_uring;
6178 unsigned long flags;
6180 if (req->work.flags & IO_WQ_WORK_FILES) {
6181 put_files_struct(req->work.identity->files);
6182 put_nsproxy(req->work.identity->nsproxy);
6184 spin_lock_irqsave(&ctx->inflight_lock, flags);
6185 list_del(&req->inflight_entry);
6186 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6187 req->flags &= ~REQ_F_INFLIGHT;
6188 req->work.flags &= ~IO_WQ_WORK_FILES;
6189 if (atomic_read(&tctx->in_idle))
6190 wake_up(&tctx->wait);
6193 static void __io_clean_op(struct io_kiocb *req)
6195 if (req->flags & REQ_F_BUFFER_SELECTED) {
6196 switch (req->opcode) {
6197 case IORING_OP_READV:
6198 case IORING_OP_READ_FIXED:
6199 case IORING_OP_READ:
6200 kfree((void *)(unsigned long)req->rw.addr);
6202 case IORING_OP_RECVMSG:
6203 case IORING_OP_RECV:
6204 kfree(req->sr_msg.kbuf);
6207 req->flags &= ~REQ_F_BUFFER_SELECTED;
6210 if (req->flags & REQ_F_NEED_CLEANUP) {
6211 switch (req->opcode) {
6212 case IORING_OP_READV:
6213 case IORING_OP_READ_FIXED:
6214 case IORING_OP_READ:
6215 case IORING_OP_WRITEV:
6216 case IORING_OP_WRITE_FIXED:
6217 case IORING_OP_WRITE: {
6218 struct io_async_rw *io = req->async_data;
6220 kfree(io->free_iovec);
6223 case IORING_OP_RECVMSG:
6224 case IORING_OP_SENDMSG: {
6225 struct io_async_msghdr *io = req->async_data;
6226 if (io->iov != io->fast_iov)
6230 case IORING_OP_SPLICE:
6232 io_put_file(req, req->splice.file_in,
6233 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6235 case IORING_OP_OPENAT:
6236 case IORING_OP_OPENAT2:
6237 if (req->open.filename)
6238 putname(req->open.filename);
6240 case IORING_OP_RENAMEAT:
6241 putname(req->rename.oldpath);
6242 putname(req->rename.newpath);
6244 case IORING_OP_UNLINKAT:
6245 putname(req->unlink.filename);
6248 req->flags &= ~REQ_F_NEED_CLEANUP;
6252 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6253 struct io_comp_state *cs)
6255 struct io_ring_ctx *ctx = req->ctx;
6258 switch (req->opcode) {
6260 ret = io_nop(req, cs);
6262 case IORING_OP_READV:
6263 case IORING_OP_READ_FIXED:
6264 case IORING_OP_READ:
6265 ret = io_read(req, force_nonblock, cs);
6267 case IORING_OP_WRITEV:
6268 case IORING_OP_WRITE_FIXED:
6269 case IORING_OP_WRITE:
6270 ret = io_write(req, force_nonblock, cs);
6272 case IORING_OP_FSYNC:
6273 ret = io_fsync(req, force_nonblock);
6275 case IORING_OP_POLL_ADD:
6276 ret = io_poll_add(req);
6278 case IORING_OP_POLL_REMOVE:
6279 ret = io_poll_remove(req);
6281 case IORING_OP_SYNC_FILE_RANGE:
6282 ret = io_sync_file_range(req, force_nonblock);
6284 case IORING_OP_SENDMSG:
6285 ret = io_sendmsg(req, force_nonblock, cs);
6287 case IORING_OP_SEND:
6288 ret = io_send(req, force_nonblock, cs);
6290 case IORING_OP_RECVMSG:
6291 ret = io_recvmsg(req, force_nonblock, cs);
6293 case IORING_OP_RECV:
6294 ret = io_recv(req, force_nonblock, cs);
6296 case IORING_OP_TIMEOUT:
6297 ret = io_timeout(req);
6299 case IORING_OP_TIMEOUT_REMOVE:
6300 ret = io_timeout_remove(req);
6302 case IORING_OP_ACCEPT:
6303 ret = io_accept(req, force_nonblock, cs);
6305 case IORING_OP_CONNECT:
6306 ret = io_connect(req, force_nonblock, cs);
6308 case IORING_OP_ASYNC_CANCEL:
6309 ret = io_async_cancel(req);
6311 case IORING_OP_FALLOCATE:
6312 ret = io_fallocate(req, force_nonblock);
6314 case IORING_OP_OPENAT:
6315 ret = io_openat(req, force_nonblock);
6317 case IORING_OP_CLOSE:
6318 ret = io_close(req, force_nonblock, cs);
6320 case IORING_OP_FILES_UPDATE:
6321 ret = io_files_update(req, force_nonblock, cs);
6323 case IORING_OP_STATX:
6324 ret = io_statx(req, force_nonblock);
6326 case IORING_OP_FADVISE:
6327 ret = io_fadvise(req, force_nonblock);
6329 case IORING_OP_MADVISE:
6330 ret = io_madvise(req, force_nonblock);
6332 case IORING_OP_OPENAT2:
6333 ret = io_openat2(req, force_nonblock);
6335 case IORING_OP_EPOLL_CTL:
6336 ret = io_epoll_ctl(req, force_nonblock, cs);
6338 case IORING_OP_SPLICE:
6339 ret = io_splice(req, force_nonblock);
6341 case IORING_OP_PROVIDE_BUFFERS:
6342 ret = io_provide_buffers(req, force_nonblock, cs);
6344 case IORING_OP_REMOVE_BUFFERS:
6345 ret = io_remove_buffers(req, force_nonblock, cs);
6348 ret = io_tee(req, force_nonblock);
6350 case IORING_OP_SHUTDOWN:
6351 ret = io_shutdown(req, force_nonblock);
6353 case IORING_OP_RENAMEAT:
6354 ret = io_renameat(req, force_nonblock);
6356 case IORING_OP_UNLINKAT:
6357 ret = io_unlinkat(req, force_nonblock);
6367 /* If the op doesn't have a file, we're not polling for it */
6368 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6369 const bool in_async = io_wq_current_is_worker();
6371 /* workqueue context doesn't hold uring_lock, grab it now */
6373 mutex_lock(&ctx->uring_lock);
6375 io_iopoll_req_issued(req, in_async);
6378 mutex_unlock(&ctx->uring_lock);
6384 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6386 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6387 struct io_kiocb *timeout;
6390 timeout = io_prep_linked_timeout(req);
6392 io_queue_linked_timeout(timeout);
6394 /* if NO_CANCEL is set, we must still run the work */
6395 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6396 IO_WQ_WORK_CANCEL) {
6402 ret = io_issue_sqe(req, false, NULL);
6404 * We can get EAGAIN for polled IO even though we're
6405 * forcing a sync submission from here, since we can't
6406 * wait for request slots on the block side.
6415 struct io_ring_ctx *lock_ctx = NULL;
6417 if (req->ctx->flags & IORING_SETUP_IOPOLL)
6418 lock_ctx = req->ctx;
6421 * io_iopoll_complete() does not hold completion_lock to
6422 * complete polled io, so here for polled io, we can not call
6423 * io_req_complete() directly, otherwise there maybe concurrent
6424 * access to cqring, defer_list, etc, which is not safe. Given
6425 * that io_iopoll_complete() is always called under uring_lock,
6426 * so here for polled io, we also get uring_lock to complete
6430 mutex_lock(&lock_ctx->uring_lock);
6432 req_set_fail_links(req);
6433 io_req_complete(req, ret);
6436 mutex_unlock(&lock_ctx->uring_lock);
6439 return io_steal_work(req);
6442 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6445 struct fixed_file_table *table;
6447 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6448 return table->files[index & IORING_FILE_TABLE_MASK];
6451 static struct file *io_file_get(struct io_submit_state *state,
6452 struct io_kiocb *req, int fd, bool fixed)
6454 struct io_ring_ctx *ctx = req->ctx;
6458 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6460 fd = array_index_nospec(fd, ctx->nr_user_files);
6461 file = io_file_from_index(ctx, fd);
6462 io_set_resource_node(req);
6464 trace_io_uring_file_get(ctx, fd);
6465 file = __io_file_get(state, fd);
6468 if (file && file->f_op == &io_uring_fops &&
6469 !(req->flags & REQ_F_INFLIGHT)) {
6470 io_req_init_async(req);
6471 req->flags |= REQ_F_INFLIGHT;
6473 spin_lock_irq(&ctx->inflight_lock);
6474 list_add(&req->inflight_entry, &ctx->inflight_list);
6475 spin_unlock_irq(&ctx->inflight_lock);
6481 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6483 struct io_timeout_data *data = container_of(timer,
6484 struct io_timeout_data, timer);
6485 struct io_kiocb *prev, *req = data->req;
6486 struct io_ring_ctx *ctx = req->ctx;
6487 unsigned long flags;
6489 spin_lock_irqsave(&ctx->completion_lock, flags);
6490 prev = req->timeout.head;
6491 req->timeout.head = NULL;
6494 * We don't expect the list to be empty, that will only happen if we
6495 * race with the completion of the linked work.
6497 if (prev && refcount_inc_not_zero(&prev->refs))
6498 io_remove_next_linked(prev);
6501 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6504 req_set_fail_links(prev);
6505 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6508 io_req_complete(req, -ETIME);
6510 return HRTIMER_NORESTART;
6513 static void __io_queue_linked_timeout(struct io_kiocb *req)
6516 * If the back reference is NULL, then our linked request finished
6517 * before we got a chance to setup the timer
6519 if (req->timeout.head) {
6520 struct io_timeout_data *data = req->async_data;
6522 data->timer.function = io_link_timeout_fn;
6523 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6528 static void io_queue_linked_timeout(struct io_kiocb *req)
6530 struct io_ring_ctx *ctx = req->ctx;
6532 spin_lock_irq(&ctx->completion_lock);
6533 __io_queue_linked_timeout(req);
6534 spin_unlock_irq(&ctx->completion_lock);
6536 /* drop submission reference */
6540 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6542 struct io_kiocb *nxt = req->link;
6544 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6545 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6548 nxt->timeout.head = req;
6549 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6550 req->flags |= REQ_F_LINK_TIMEOUT;
6554 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6556 struct io_kiocb *linked_timeout;
6557 const struct cred *old_creds = NULL;
6561 linked_timeout = io_prep_linked_timeout(req);
6563 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6564 (req->work.flags & IO_WQ_WORK_CREDS) &&
6565 req->work.identity->creds != current_cred()) {
6567 revert_creds(old_creds);
6568 if (old_creds == req->work.identity->creds)
6569 old_creds = NULL; /* restored original creds */
6571 old_creds = override_creds(req->work.identity->creds);
6574 ret = io_issue_sqe(req, true, cs);
6577 * We async punt it if the file wasn't marked NOWAIT, or if the file
6578 * doesn't support non-blocking read/write attempts
6580 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6581 if (!io_arm_poll_handler(req)) {
6583 * Queued up for async execution, worker will release
6584 * submit reference when the iocb is actually submitted.
6586 io_queue_async_work(req);
6590 io_queue_linked_timeout(linked_timeout);
6591 } else if (likely(!ret)) {
6592 /* drop submission reference */
6593 req = io_put_req_find_next(req);
6595 io_queue_linked_timeout(linked_timeout);
6598 if (!(req->flags & REQ_F_FORCE_ASYNC))
6600 io_queue_async_work(req);
6603 /* un-prep timeout, so it'll be killed as any other linked */
6604 req->flags &= ~REQ_F_LINK_TIMEOUT;
6605 req_set_fail_links(req);
6607 io_req_complete(req, ret);
6611 revert_creds(old_creds);
6614 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6615 struct io_comp_state *cs)
6619 ret = io_req_defer(req, sqe);
6621 if (ret != -EIOCBQUEUED) {
6623 req_set_fail_links(req);
6625 io_req_complete(req, ret);
6627 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6628 if (!req->async_data) {
6629 ret = io_req_defer_prep(req, sqe);
6633 io_queue_async_work(req);
6636 ret = io_req_prep(req, sqe);
6640 __io_queue_sqe(req, cs);
6644 static inline void io_queue_link_head(struct io_kiocb *req,
6645 struct io_comp_state *cs)
6647 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6649 io_req_complete(req, -ECANCELED);
6651 io_queue_sqe(req, NULL, cs);
6654 struct io_submit_link {
6655 struct io_kiocb *head;
6656 struct io_kiocb *last;
6659 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6660 struct io_submit_link *link, struct io_comp_state *cs)
6662 struct io_ring_ctx *ctx = req->ctx;
6666 * If we already have a head request, queue this one for async
6667 * submittal once the head completes. If we don't have a head but
6668 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6669 * submitted sync once the chain is complete. If none of those
6670 * conditions are true (normal request), then just queue it.
6673 struct io_kiocb *head = link->head;
6676 * Taking sequential execution of a link, draining both sides
6677 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6678 * requests in the link. So, it drains the head and the
6679 * next after the link request. The last one is done via
6680 * drain_next flag to persist the effect across calls.
6682 if (req->flags & REQ_F_IO_DRAIN) {
6683 head->flags |= REQ_F_IO_DRAIN;
6684 ctx->drain_next = 1;
6686 ret = io_req_defer_prep(req, sqe);
6687 if (unlikely(ret)) {
6688 /* fail even hard links since we don't submit */
6689 head->flags |= REQ_F_FAIL_LINK;
6692 trace_io_uring_link(ctx, req, head);
6693 link->last->link = req;
6696 /* last request of a link, enqueue the link */
6697 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6698 io_queue_link_head(head, cs);
6702 if (unlikely(ctx->drain_next)) {
6703 req->flags |= REQ_F_IO_DRAIN;
6704 ctx->drain_next = 0;
6706 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6707 ret = io_req_defer_prep(req, sqe);
6709 req->flags |= REQ_F_FAIL_LINK;
6713 io_queue_sqe(req, sqe, cs);
6721 * Batched submission is done, ensure local IO is flushed out.
6723 static void io_submit_state_end(struct io_submit_state *state)
6725 if (!list_empty(&state->comp.list))
6726 io_submit_flush_completions(&state->comp);
6727 if (state->plug_started)
6728 blk_finish_plug(&state->plug);
6729 io_state_file_put(state);
6730 if (state->free_reqs)
6731 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6735 * Start submission side cache.
6737 static void io_submit_state_start(struct io_submit_state *state,
6738 struct io_ring_ctx *ctx, unsigned int max_ios)
6740 state->plug_started = false;
6742 INIT_LIST_HEAD(&state->comp.list);
6743 state->comp.ctx = ctx;
6744 state->free_reqs = 0;
6745 state->file_refs = 0;
6746 state->ios_left = max_ios;
6749 static void io_commit_sqring(struct io_ring_ctx *ctx)
6751 struct io_rings *rings = ctx->rings;
6754 * Ensure any loads from the SQEs are done at this point,
6755 * since once we write the new head, the application could
6756 * write new data to them.
6758 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6762 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6763 * that is mapped by userspace. This means that care needs to be taken to
6764 * ensure that reads are stable, as we cannot rely on userspace always
6765 * being a good citizen. If members of the sqe are validated and then later
6766 * used, it's important that those reads are done through READ_ONCE() to
6767 * prevent a re-load down the line.
6769 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6771 u32 *sq_array = ctx->sq_array;
6775 * The cached sq head (or cq tail) serves two purposes:
6777 * 1) allows us to batch the cost of updating the user visible
6779 * 2) allows the kernel side to track the head on its own, even
6780 * though the application is the one updating it.
6782 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6783 if (likely(head < ctx->sq_entries))
6784 return &ctx->sq_sqes[head];
6786 /* drop invalid entries */
6787 ctx->cached_sq_dropped++;
6788 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6792 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6794 ctx->cached_sq_head++;
6798 * Check SQE restrictions (opcode and flags).
6800 * Returns 'true' if SQE is allowed, 'false' otherwise.
6802 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6803 struct io_kiocb *req,
6804 unsigned int sqe_flags)
6806 if (!ctx->restricted)
6809 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6812 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6813 ctx->restrictions.sqe_flags_required)
6816 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6817 ctx->restrictions.sqe_flags_required))
6823 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6824 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6825 IOSQE_BUFFER_SELECT)
6827 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6828 const struct io_uring_sqe *sqe,
6829 struct io_submit_state *state)
6831 unsigned int sqe_flags;
6834 req->opcode = READ_ONCE(sqe->opcode);
6835 req->user_data = READ_ONCE(sqe->user_data);
6836 req->async_data = NULL;
6841 req->fixed_file_refs = NULL;
6842 /* one is dropped after submission, the other at completion */
6843 refcount_set(&req->refs, 2);
6844 req->task = current;
6847 if (unlikely(req->opcode >= IORING_OP_LAST))
6850 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6853 sqe_flags = READ_ONCE(sqe->flags);
6854 /* enforce forwards compatibility on users */
6855 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6858 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6861 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6862 !io_op_defs[req->opcode].buffer_select)
6865 id = READ_ONCE(sqe->personality);
6867 struct io_identity *iod;
6869 iod = idr_find(&ctx->personality_idr, id);
6872 refcount_inc(&iod->count);
6874 __io_req_init_async(req);
6875 get_cred(iod->creds);
6876 req->work.identity = iod;
6877 req->work.flags |= IO_WQ_WORK_CREDS;
6880 /* same numerical values with corresponding REQ_F_*, safe to copy */
6881 req->flags |= sqe_flags;
6884 * Plug now if we have more than 1 IO left after this, and the target
6885 * is potentially a read/write to block based storage.
6887 if (!state->plug_started && state->ios_left > 1 &&
6888 io_op_defs[req->opcode].plug) {
6889 blk_start_plug(&state->plug);
6890 state->plug_started = true;
6894 if (io_op_defs[req->opcode].needs_file) {
6895 bool fixed = req->flags & REQ_F_FIXED_FILE;
6897 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6898 if (unlikely(!req->file &&
6899 !io_op_defs[req->opcode].needs_file_no_error))
6907 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6909 struct io_submit_state state;
6910 struct io_submit_link link;
6911 int i, submitted = 0;
6913 /* if we have a backlog and couldn't flush it all, return BUSY */
6914 if (test_bit(0, &ctx->sq_check_overflow)) {
6915 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6919 /* make sure SQ entry isn't read before tail */
6920 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6922 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6925 percpu_counter_add(¤t->io_uring->inflight, nr);
6926 refcount_add(nr, ¤t->usage);
6928 io_submit_state_start(&state, ctx, nr);
6931 for (i = 0; i < nr; i++) {
6932 const struct io_uring_sqe *sqe;
6933 struct io_kiocb *req;
6936 sqe = io_get_sqe(ctx);
6937 if (unlikely(!sqe)) {
6938 io_consume_sqe(ctx);
6941 req = io_alloc_req(ctx, &state);
6942 if (unlikely(!req)) {
6944 submitted = -EAGAIN;
6947 io_consume_sqe(ctx);
6948 /* will complete beyond this point, count as submitted */
6951 err = io_init_req(ctx, req, sqe, &state);
6952 if (unlikely(err)) {
6955 io_req_complete(req, err);
6959 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6960 true, io_async_submit(ctx));
6961 err = io_submit_sqe(req, sqe, &link, &state.comp);
6966 if (unlikely(submitted != nr)) {
6967 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6968 struct io_uring_task *tctx = current->io_uring;
6969 int unused = nr - ref_used;
6971 percpu_ref_put_many(&ctx->refs, unused);
6972 percpu_counter_sub(&tctx->inflight, unused);
6973 put_task_struct_many(current, unused);
6976 io_queue_link_head(link.head, &state.comp);
6977 io_submit_state_end(&state);
6979 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6980 io_commit_sqring(ctx);
6985 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6987 /* Tell userspace we may need a wakeup call */
6988 spin_lock_irq(&ctx->completion_lock);
6989 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6990 spin_unlock_irq(&ctx->completion_lock);
6993 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6995 spin_lock_irq(&ctx->completion_lock);
6996 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6997 spin_unlock_irq(&ctx->completion_lock);
7000 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7002 unsigned int to_submit;
7005 to_submit = io_sqring_entries(ctx);
7006 /* if we're handling multiple rings, cap submit size for fairness */
7007 if (cap_entries && to_submit > 8)
7010 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7011 unsigned nr_events = 0;
7013 mutex_lock(&ctx->uring_lock);
7014 if (!list_empty(&ctx->iopoll_list))
7015 io_do_iopoll(ctx, &nr_events, 0);
7017 if (to_submit && !ctx->sqo_dead &&
7018 likely(!percpu_ref_is_dying(&ctx->refs)))
7019 ret = io_submit_sqes(ctx, to_submit);
7020 mutex_unlock(&ctx->uring_lock);
7023 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
7024 wake_up(&ctx->sqo_sq_wait);
7029 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7031 struct io_ring_ctx *ctx;
7032 unsigned sq_thread_idle = 0;
7034 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7035 if (sq_thread_idle < ctx->sq_thread_idle)
7036 sq_thread_idle = ctx->sq_thread_idle;
7039 sqd->sq_thread_idle = sq_thread_idle;
7042 static void io_sqd_init_new(struct io_sq_data *sqd)
7044 struct io_ring_ctx *ctx;
7046 while (!list_empty(&sqd->ctx_new_list)) {
7047 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
7048 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
7049 complete(&ctx->sq_thread_comp);
7052 io_sqd_update_thread_idle(sqd);
7055 static int io_sq_thread(void *data)
7057 struct cgroup_subsys_state *cur_css = NULL;
7058 struct files_struct *old_files = current->files;
7059 struct nsproxy *old_nsproxy = current->nsproxy;
7060 const struct cred *old_cred = NULL;
7061 struct io_sq_data *sqd = data;
7062 struct io_ring_ctx *ctx;
7063 unsigned long timeout = 0;
7067 current->files = NULL;
7068 current->nsproxy = NULL;
7069 task_unlock(current);
7071 while (!kthread_should_stop()) {
7073 bool cap_entries, sqt_spin, needs_sched;
7076 * Any changes to the sqd lists are synchronized through the
7077 * kthread parking. This synchronizes the thread vs users,
7078 * the users are synchronized on the sqd->ctx_lock.
7080 if (kthread_should_park()) {
7083 * When sq thread is unparked, in case the previous park operation
7084 * comes from io_put_sq_data(), which means that sq thread is going
7085 * to be stopped, so here needs to have a check.
7087 if (kthread_should_stop())
7091 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7092 io_sqd_init_new(sqd);
7093 timeout = jiffies + sqd->sq_thread_idle;
7097 cap_entries = !list_is_singular(&sqd->ctx_list);
7098 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7099 if (current->cred != ctx->creds) {
7101 revert_creds(old_cred);
7102 old_cred = override_creds(ctx->creds);
7104 io_sq_thread_associate_blkcg(ctx, &cur_css);
7106 current->loginuid = ctx->loginuid;
7107 current->sessionid = ctx->sessionid;
7110 ret = __io_sq_thread(ctx, cap_entries);
7111 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7114 io_sq_thread_drop_mm_files();
7117 if (sqt_spin || !time_after(jiffies, timeout)) {
7119 io_sq_thread_drop_mm_files();
7122 timeout = jiffies + sqd->sq_thread_idle;
7126 if (kthread_should_park())
7130 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7131 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7132 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7133 !list_empty_careful(&ctx->iopoll_list)) {
7134 needs_sched = false;
7137 if (io_sqring_entries(ctx)) {
7138 needs_sched = false;
7144 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7145 io_ring_set_wakeup_flag(ctx);
7148 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7149 io_ring_clear_wakeup_flag(ctx);
7152 finish_wait(&sqd->wait, &wait);
7153 timeout = jiffies + sqd->sq_thread_idle;
7157 io_sq_thread_drop_mm_files();
7160 io_sq_thread_unassociate_blkcg();
7162 revert_creds(old_cred);
7165 current->files = old_files;
7166 current->nsproxy = old_nsproxy;
7167 task_unlock(current);
7174 struct io_wait_queue {
7175 struct wait_queue_entry wq;
7176 struct io_ring_ctx *ctx;
7178 unsigned nr_timeouts;
7181 static inline bool io_should_wake(struct io_wait_queue *iowq)
7183 struct io_ring_ctx *ctx = iowq->ctx;
7186 * Wake up if we have enough events, or if a timeout occurred since we
7187 * started waiting. For timeouts, we always want to return to userspace,
7188 * regardless of event count.
7190 return io_cqring_events(ctx) >= iowq->to_wait ||
7191 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7194 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7195 int wake_flags, void *key)
7197 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7201 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7202 * the task, and the next invocation will do it.
7204 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
7205 return autoremove_wake_function(curr, mode, wake_flags, key);
7209 static int io_run_task_work_sig(void)
7211 if (io_run_task_work())
7213 if (!signal_pending(current))
7215 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7216 return -ERESTARTSYS;
7221 * Wait until events become available, if we don't already have some. The
7222 * application must reap them itself, as they reside on the shared cq ring.
7224 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7225 const sigset_t __user *sig, size_t sigsz,
7226 struct __kernel_timespec __user *uts)
7228 struct io_wait_queue iowq = {
7231 .func = io_wake_function,
7232 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7235 .to_wait = min_events,
7237 struct io_rings *rings = ctx->rings;
7238 struct timespec64 ts;
7239 signed long timeout = 0;
7243 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7244 if (io_cqring_events(ctx) >= min_events)
7246 if (!io_run_task_work())
7251 #ifdef CONFIG_COMPAT
7252 if (in_compat_syscall())
7253 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7257 ret = set_user_sigmask(sig, sigsz);
7264 if (get_timespec64(&ts, uts))
7266 timeout = timespec64_to_jiffies(&ts);
7269 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7270 trace_io_uring_cqring_wait(ctx, min_events);
7272 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7273 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7274 TASK_INTERRUPTIBLE);
7275 /* make sure we run task_work before checking for signals */
7276 ret = io_run_task_work_sig();
7278 finish_wait(&ctx->wait, &iowq.wq);
7283 if (io_should_wake(&iowq))
7285 if (test_bit(0, &ctx->cq_check_overflow)) {
7286 finish_wait(&ctx->wait, &iowq.wq);
7290 timeout = schedule_timeout(timeout);
7299 finish_wait(&ctx->wait, &iowq.wq);
7301 restore_saved_sigmask_unless(ret == -EINTR);
7303 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7306 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7308 #if defined(CONFIG_UNIX)
7309 if (ctx->ring_sock) {
7310 struct sock *sock = ctx->ring_sock->sk;
7311 struct sk_buff *skb;
7313 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7319 for (i = 0; i < ctx->nr_user_files; i++) {
7322 file = io_file_from_index(ctx, i);
7329 static void io_file_ref_kill(struct percpu_ref *ref)
7331 struct fixed_file_data *data;
7333 data = container_of(ref, struct fixed_file_data, refs);
7334 complete(&data->done);
7337 static void io_sqe_files_set_node(struct fixed_file_data *file_data,
7338 struct fixed_file_ref_node *ref_node)
7340 spin_lock_bh(&file_data->lock);
7341 file_data->node = ref_node;
7342 list_add_tail(&ref_node->node, &file_data->ref_list);
7343 spin_unlock_bh(&file_data->lock);
7344 percpu_ref_get(&file_data->refs);
7347 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7349 struct fixed_file_data *data = ctx->file_data;
7350 struct fixed_file_ref_node *backup_node, *ref_node = NULL;
7351 unsigned nr_tables, i;
7356 backup_node = alloc_fixed_file_ref_node(ctx);
7360 spin_lock_bh(&data->lock);
7361 ref_node = data->node;
7362 spin_unlock_bh(&data->lock);
7364 percpu_ref_kill(&ref_node->refs);
7366 percpu_ref_kill(&data->refs);
7368 /* wait for all refs nodes to complete */
7369 flush_delayed_work(&ctx->file_put_work);
7371 ret = wait_for_completion_interruptible(&data->done);
7374 ret = io_run_task_work_sig();
7376 percpu_ref_resurrect(&data->refs);
7377 reinit_completion(&data->done);
7378 io_sqe_files_set_node(data, backup_node);
7383 __io_sqe_files_unregister(ctx);
7384 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7385 for (i = 0; i < nr_tables; i++)
7386 kfree(data->table[i].files);
7388 percpu_ref_exit(&data->refs);
7390 ctx->file_data = NULL;
7391 ctx->nr_user_files = 0;
7392 destroy_fixed_file_ref_node(backup_node);
7396 static void io_put_sq_data(struct io_sq_data *sqd)
7398 if (refcount_dec_and_test(&sqd->refs)) {
7400 * The park is a bit of a work-around, without it we get
7401 * warning spews on shutdown with SQPOLL set and affinity
7402 * set to a single CPU.
7405 kthread_park(sqd->thread);
7406 kthread_stop(sqd->thread);
7413 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7415 struct io_ring_ctx *ctx_attach;
7416 struct io_sq_data *sqd;
7419 f = fdget(p->wq_fd);
7421 return ERR_PTR(-ENXIO);
7422 if (f.file->f_op != &io_uring_fops) {
7424 return ERR_PTR(-EINVAL);
7427 ctx_attach = f.file->private_data;
7428 sqd = ctx_attach->sq_data;
7431 return ERR_PTR(-EINVAL);
7434 refcount_inc(&sqd->refs);
7439 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7441 struct io_sq_data *sqd;
7443 if (p->flags & IORING_SETUP_ATTACH_WQ)
7444 return io_attach_sq_data(p);
7446 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7448 return ERR_PTR(-ENOMEM);
7450 refcount_set(&sqd->refs, 1);
7451 INIT_LIST_HEAD(&sqd->ctx_list);
7452 INIT_LIST_HEAD(&sqd->ctx_new_list);
7453 mutex_init(&sqd->ctx_lock);
7454 mutex_init(&sqd->lock);
7455 init_waitqueue_head(&sqd->wait);
7459 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7460 __releases(&sqd->lock)
7464 kthread_unpark(sqd->thread);
7465 mutex_unlock(&sqd->lock);
7468 static void io_sq_thread_park(struct io_sq_data *sqd)
7469 __acquires(&sqd->lock)
7473 mutex_lock(&sqd->lock);
7474 kthread_park(sqd->thread);
7477 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7479 struct io_sq_data *sqd = ctx->sq_data;
7484 * We may arrive here from the error branch in
7485 * io_sq_offload_create() where the kthread is created
7486 * without being waked up, thus wake it up now to make
7487 * sure the wait will complete.
7489 wake_up_process(sqd->thread);
7490 wait_for_completion(&ctx->sq_thread_comp);
7492 io_sq_thread_park(sqd);
7495 mutex_lock(&sqd->ctx_lock);
7496 list_del(&ctx->sqd_list);
7497 io_sqd_update_thread_idle(sqd);
7498 mutex_unlock(&sqd->ctx_lock);
7501 io_sq_thread_unpark(sqd);
7503 io_put_sq_data(sqd);
7504 ctx->sq_data = NULL;
7508 static void io_finish_async(struct io_ring_ctx *ctx)
7510 io_sq_thread_stop(ctx);
7513 io_wq_destroy(ctx->io_wq);
7518 #if defined(CONFIG_UNIX)
7520 * Ensure the UNIX gc is aware of our file set, so we are certain that
7521 * the io_uring can be safely unregistered on process exit, even if we have
7522 * loops in the file referencing.
7524 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7526 struct sock *sk = ctx->ring_sock->sk;
7527 struct scm_fp_list *fpl;
7528 struct sk_buff *skb;
7531 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7535 skb = alloc_skb(0, GFP_KERNEL);
7544 fpl->user = get_uid(ctx->user);
7545 for (i = 0; i < nr; i++) {
7546 struct file *file = io_file_from_index(ctx, i + offset);
7550 fpl->fp[nr_files] = get_file(file);
7551 unix_inflight(fpl->user, fpl->fp[nr_files]);
7556 fpl->max = SCM_MAX_FD;
7557 fpl->count = nr_files;
7558 UNIXCB(skb).fp = fpl;
7559 skb->destructor = unix_destruct_scm;
7560 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7561 skb_queue_head(&sk->sk_receive_queue, skb);
7563 for (i = 0; i < nr_files; i++)
7574 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7575 * causes regular reference counting to break down. We rely on the UNIX
7576 * garbage collection to take care of this problem for us.
7578 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7580 unsigned left, total;
7584 left = ctx->nr_user_files;
7586 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7588 ret = __io_sqe_files_scm(ctx, this_files, total);
7592 total += this_files;
7598 while (total < ctx->nr_user_files) {
7599 struct file *file = io_file_from_index(ctx, total);
7609 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7615 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7616 unsigned nr_tables, unsigned nr_files)
7620 for (i = 0; i < nr_tables; i++) {
7621 struct fixed_file_table *table = &file_data->table[i];
7622 unsigned this_files;
7624 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7625 table->files = kcalloc(this_files, sizeof(struct file *),
7629 nr_files -= this_files;
7635 for (i = 0; i < nr_tables; i++) {
7636 struct fixed_file_table *table = &file_data->table[i];
7637 kfree(table->files);
7642 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7644 #if defined(CONFIG_UNIX)
7645 struct sock *sock = ctx->ring_sock->sk;
7646 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7647 struct sk_buff *skb;
7650 __skb_queue_head_init(&list);
7653 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7654 * remove this entry and rearrange the file array.
7656 skb = skb_dequeue(head);
7658 struct scm_fp_list *fp;
7660 fp = UNIXCB(skb).fp;
7661 for (i = 0; i < fp->count; i++) {
7664 if (fp->fp[i] != file)
7667 unix_notinflight(fp->user, fp->fp[i]);
7668 left = fp->count - 1 - i;
7670 memmove(&fp->fp[i], &fp->fp[i + 1],
7671 left * sizeof(struct file *));
7678 __skb_queue_tail(&list, skb);
7688 __skb_queue_tail(&list, skb);
7690 skb = skb_dequeue(head);
7693 if (skb_peek(&list)) {
7694 spin_lock_irq(&head->lock);
7695 while ((skb = __skb_dequeue(&list)) != NULL)
7696 __skb_queue_tail(head, skb);
7697 spin_unlock_irq(&head->lock);
7704 struct io_file_put {
7705 struct list_head list;
7709 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7711 struct fixed_file_data *file_data = ref_node->file_data;
7712 struct io_ring_ctx *ctx = file_data->ctx;
7713 struct io_file_put *pfile, *tmp;
7715 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7716 list_del(&pfile->list);
7717 io_ring_file_put(ctx, pfile->file);
7721 percpu_ref_exit(&ref_node->refs);
7723 percpu_ref_put(&file_data->refs);
7726 static void io_file_put_work(struct work_struct *work)
7728 struct io_ring_ctx *ctx;
7729 struct llist_node *node;
7731 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7732 node = llist_del_all(&ctx->file_put_llist);
7735 struct fixed_file_ref_node *ref_node;
7736 struct llist_node *next = node->next;
7738 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7739 __io_file_put_work(ref_node);
7744 static void io_file_data_ref_zero(struct percpu_ref *ref)
7746 struct fixed_file_ref_node *ref_node;
7747 struct fixed_file_data *data;
7748 struct io_ring_ctx *ctx;
7749 bool first_add = false;
7752 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7753 data = ref_node->file_data;
7756 spin_lock_bh(&data->lock);
7757 ref_node->done = true;
7759 while (!list_empty(&data->ref_list)) {
7760 ref_node = list_first_entry(&data->ref_list,
7761 struct fixed_file_ref_node, node);
7762 /* recycle ref nodes in order */
7763 if (!ref_node->done)
7765 list_del(&ref_node->node);
7766 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7768 spin_unlock_bh(&data->lock);
7770 if (percpu_ref_is_dying(&data->refs))
7774 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7776 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7779 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7780 struct io_ring_ctx *ctx)
7782 struct fixed_file_ref_node *ref_node;
7784 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7788 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7793 INIT_LIST_HEAD(&ref_node->node);
7794 INIT_LIST_HEAD(&ref_node->file_list);
7795 ref_node->file_data = ctx->file_data;
7796 ref_node->done = false;
7800 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7802 percpu_ref_exit(&ref_node->refs);
7806 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7809 __s32 __user *fds = (__s32 __user *) arg;
7810 unsigned nr_tables, i;
7812 int fd, ret = -ENOMEM;
7813 struct fixed_file_ref_node *ref_node;
7814 struct fixed_file_data *file_data;
7820 if (nr_args > IORING_MAX_FIXED_FILES)
7823 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7826 file_data->ctx = ctx;
7827 init_completion(&file_data->done);
7828 INIT_LIST_HEAD(&file_data->ref_list);
7829 spin_lock_init(&file_data->lock);
7831 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7832 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7834 if (!file_data->table)
7837 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7838 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7841 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7843 ctx->file_data = file_data;
7845 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7846 struct fixed_file_table *table;
7849 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7853 /* allow sparse sets */
7863 * Don't allow io_uring instances to be registered. If UNIX
7864 * isn't enabled, then this causes a reference cycle and this
7865 * instance can never get freed. If UNIX is enabled we'll
7866 * handle it just fine, but there's still no point in allowing
7867 * a ring fd as it doesn't support regular read/write anyway.
7869 if (file->f_op == &io_uring_fops) {
7873 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7874 index = i & IORING_FILE_TABLE_MASK;
7875 table->files[index] = file;
7878 ret = io_sqe_files_scm(ctx);
7880 io_sqe_files_unregister(ctx);
7884 ref_node = alloc_fixed_file_ref_node(ctx);
7886 io_sqe_files_unregister(ctx);
7890 io_sqe_files_set_node(file_data, ref_node);
7893 for (i = 0; i < ctx->nr_user_files; i++) {
7894 file = io_file_from_index(ctx, i);
7898 for (i = 0; i < nr_tables; i++)
7899 kfree(file_data->table[i].files);
7900 ctx->nr_user_files = 0;
7902 percpu_ref_exit(&file_data->refs);
7904 kfree(file_data->table);
7906 ctx->file_data = NULL;
7910 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7913 #if defined(CONFIG_UNIX)
7914 struct sock *sock = ctx->ring_sock->sk;
7915 struct sk_buff_head *head = &sock->sk_receive_queue;
7916 struct sk_buff *skb;
7919 * See if we can merge this file into an existing skb SCM_RIGHTS
7920 * file set. If there's no room, fall back to allocating a new skb
7921 * and filling it in.
7923 spin_lock_irq(&head->lock);
7924 skb = skb_peek(head);
7926 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7928 if (fpl->count < SCM_MAX_FD) {
7929 __skb_unlink(skb, head);
7930 spin_unlock_irq(&head->lock);
7931 fpl->fp[fpl->count] = get_file(file);
7932 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7934 spin_lock_irq(&head->lock);
7935 __skb_queue_head(head, skb);
7940 spin_unlock_irq(&head->lock);
7947 return __io_sqe_files_scm(ctx, 1, index);
7953 static int io_queue_file_removal(struct fixed_file_data *data,
7956 struct io_file_put *pfile;
7957 struct fixed_file_ref_node *ref_node = data->node;
7959 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7964 list_add(&pfile->list, &ref_node->file_list);
7969 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7970 struct io_uring_files_update *up,
7973 struct fixed_file_data *data = ctx->file_data;
7974 struct fixed_file_ref_node *ref_node;
7979 bool needs_switch = false;
7981 if (check_add_overflow(up->offset, nr_args, &done))
7983 if (done > ctx->nr_user_files)
7986 ref_node = alloc_fixed_file_ref_node(ctx);
7991 fds = u64_to_user_ptr(up->fds);
7993 struct fixed_file_table *table;
7997 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
8001 i = array_index_nospec(up->offset, ctx->nr_user_files);
8002 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8003 index = i & IORING_FILE_TABLE_MASK;
8004 if (table->files[index]) {
8005 file = table->files[index];
8006 err = io_queue_file_removal(data, file);
8009 table->files[index] = NULL;
8010 needs_switch = true;
8019 * Don't allow io_uring instances to be registered. If
8020 * UNIX isn't enabled, then this causes a reference
8021 * cycle and this instance can never get freed. If UNIX
8022 * is enabled we'll handle it just fine, but there's
8023 * still no point in allowing a ring fd as it doesn't
8024 * support regular read/write anyway.
8026 if (file->f_op == &io_uring_fops) {
8031 table->files[index] = file;
8032 err = io_sqe_file_register(ctx, file, i);
8034 table->files[index] = NULL;
8045 percpu_ref_kill(&data->node->refs);
8046 io_sqe_files_set_node(data, ref_node);
8048 destroy_fixed_file_ref_node(ref_node);
8050 return done ? done : err;
8053 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
8056 struct io_uring_files_update up;
8058 if (!ctx->file_data)
8062 if (copy_from_user(&up, arg, sizeof(up)))
8067 return __io_sqe_files_update(ctx, &up, nr_args);
8070 static void io_free_work(struct io_wq_work *work)
8072 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8074 /* Consider that io_steal_work() relies on this ref */
8078 static int io_init_wq_offload(struct io_ring_ctx *ctx,
8079 struct io_uring_params *p)
8081 struct io_wq_data data;
8083 struct io_ring_ctx *ctx_attach;
8084 unsigned int concurrency;
8087 data.user = ctx->user;
8088 data.free_work = io_free_work;
8089 data.do_work = io_wq_submit_work;
8091 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
8092 /* Do QD, or 4 * CPUS, whatever is smallest */
8093 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8095 ctx->io_wq = io_wq_create(concurrency, &data);
8096 if (IS_ERR(ctx->io_wq)) {
8097 ret = PTR_ERR(ctx->io_wq);
8103 f = fdget(p->wq_fd);
8107 if (f.file->f_op != &io_uring_fops) {
8112 ctx_attach = f.file->private_data;
8113 /* @io_wq is protected by holding the fd */
8114 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8119 ctx->io_wq = ctx_attach->io_wq;
8125 static int io_uring_alloc_task_context(struct task_struct *task)
8127 struct io_uring_task *tctx;
8130 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8131 if (unlikely(!tctx))
8134 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8135 if (unlikely(ret)) {
8141 init_waitqueue_head(&tctx->wait);
8143 atomic_set(&tctx->in_idle, 0);
8144 tctx->sqpoll = false;
8145 io_init_identity(&tctx->__identity);
8146 tctx->identity = &tctx->__identity;
8147 task->io_uring = tctx;
8151 void __io_uring_free(struct task_struct *tsk)
8153 struct io_uring_task *tctx = tsk->io_uring;
8155 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8156 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8157 if (tctx->identity != &tctx->__identity)
8158 kfree(tctx->identity);
8159 percpu_counter_destroy(&tctx->inflight);
8161 tsk->io_uring = NULL;
8164 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8165 struct io_uring_params *p)
8169 if (ctx->flags & IORING_SETUP_SQPOLL) {
8170 struct io_sq_data *sqd;
8173 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8176 sqd = io_get_sq_data(p);
8183 io_sq_thread_park(sqd);
8184 mutex_lock(&sqd->ctx_lock);
8185 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8186 mutex_unlock(&sqd->ctx_lock);
8187 io_sq_thread_unpark(sqd);
8189 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8190 if (!ctx->sq_thread_idle)
8191 ctx->sq_thread_idle = HZ;
8196 if (p->flags & IORING_SETUP_SQ_AFF) {
8197 int cpu = p->sq_thread_cpu;
8200 if (cpu >= nr_cpu_ids)
8202 if (!cpu_online(cpu))
8205 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8206 cpu, "io_uring-sq");
8208 sqd->thread = kthread_create(io_sq_thread, sqd,
8211 if (IS_ERR(sqd->thread)) {
8212 ret = PTR_ERR(sqd->thread);
8216 ret = io_uring_alloc_task_context(sqd->thread);
8219 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8220 /* Can't have SQ_AFF without SQPOLL */
8226 ret = io_init_wq_offload(ctx, p);
8232 io_finish_async(ctx);
8236 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8238 struct io_sq_data *sqd = ctx->sq_data;
8240 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8241 wake_up_process(sqd->thread);
8244 static inline void __io_unaccount_mem(struct user_struct *user,
8245 unsigned long nr_pages)
8247 atomic_long_sub(nr_pages, &user->locked_vm);
8250 static inline int __io_account_mem(struct user_struct *user,
8251 unsigned long nr_pages)
8253 unsigned long page_limit, cur_pages, new_pages;
8255 /* Don't allow more pages than we can safely lock */
8256 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8259 cur_pages = atomic_long_read(&user->locked_vm);
8260 new_pages = cur_pages + nr_pages;
8261 if (new_pages > page_limit)
8263 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8264 new_pages) != cur_pages);
8269 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8270 enum io_mem_account acct)
8273 __io_unaccount_mem(ctx->user, nr_pages);
8275 if (ctx->mm_account) {
8276 if (acct == ACCT_LOCKED) {
8277 mmap_write_lock(ctx->mm_account);
8278 ctx->mm_account->locked_vm -= nr_pages;
8279 mmap_write_unlock(ctx->mm_account);
8280 }else if (acct == ACCT_PINNED) {
8281 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8286 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8287 enum io_mem_account acct)
8291 if (ctx->limit_mem) {
8292 ret = __io_account_mem(ctx->user, nr_pages);
8297 if (ctx->mm_account) {
8298 if (acct == ACCT_LOCKED) {
8299 mmap_write_lock(ctx->mm_account);
8300 ctx->mm_account->locked_vm += nr_pages;
8301 mmap_write_unlock(ctx->mm_account);
8302 } else if (acct == ACCT_PINNED) {
8303 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8310 static void io_mem_free(void *ptr)
8317 page = virt_to_head_page(ptr);
8318 if (put_page_testzero(page))
8319 free_compound_page(page);
8322 static void *io_mem_alloc(size_t size)
8324 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8327 return (void *) __get_free_pages(gfp_flags, get_order(size));
8330 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8333 struct io_rings *rings;
8334 size_t off, sq_array_size;
8336 off = struct_size(rings, cqes, cq_entries);
8337 if (off == SIZE_MAX)
8341 off = ALIGN(off, SMP_CACHE_BYTES);
8349 sq_array_size = array_size(sizeof(u32), sq_entries);
8350 if (sq_array_size == SIZE_MAX)
8353 if (check_add_overflow(off, sq_array_size, &off))
8359 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8363 pages = (size_t)1 << get_order(
8364 rings_size(sq_entries, cq_entries, NULL));
8365 pages += (size_t)1 << get_order(
8366 array_size(sizeof(struct io_uring_sqe), sq_entries));
8371 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8375 if (!ctx->user_bufs)
8378 for (i = 0; i < ctx->nr_user_bufs; i++) {
8379 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8381 for (j = 0; j < imu->nr_bvecs; j++)
8382 unpin_user_page(imu->bvec[j].bv_page);
8384 if (imu->acct_pages)
8385 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8390 kfree(ctx->user_bufs);
8391 ctx->user_bufs = NULL;
8392 ctx->nr_user_bufs = 0;
8396 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8397 void __user *arg, unsigned index)
8399 struct iovec __user *src;
8401 #ifdef CONFIG_COMPAT
8403 struct compat_iovec __user *ciovs;
8404 struct compat_iovec ciov;
8406 ciovs = (struct compat_iovec __user *) arg;
8407 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8410 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8411 dst->iov_len = ciov.iov_len;
8415 src = (struct iovec __user *) arg;
8416 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8422 * Not super efficient, but this is just a registration time. And we do cache
8423 * the last compound head, so generally we'll only do a full search if we don't
8426 * We check if the given compound head page has already been accounted, to
8427 * avoid double accounting it. This allows us to account the full size of the
8428 * page, not just the constituent pages of a huge page.
8430 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8431 int nr_pages, struct page *hpage)
8435 /* check current page array */
8436 for (i = 0; i < nr_pages; i++) {
8437 if (!PageCompound(pages[i]))
8439 if (compound_head(pages[i]) == hpage)
8443 /* check previously registered pages */
8444 for (i = 0; i < ctx->nr_user_bufs; i++) {
8445 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8447 for (j = 0; j < imu->nr_bvecs; j++) {
8448 if (!PageCompound(imu->bvec[j].bv_page))
8450 if (compound_head(imu->bvec[j].bv_page) == hpage)
8458 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8459 int nr_pages, struct io_mapped_ubuf *imu,
8460 struct page **last_hpage)
8464 for (i = 0; i < nr_pages; i++) {
8465 if (!PageCompound(pages[i])) {
8470 hpage = compound_head(pages[i]);
8471 if (hpage == *last_hpage)
8473 *last_hpage = hpage;
8474 if (headpage_already_acct(ctx, pages, i, hpage))
8476 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8480 if (!imu->acct_pages)
8483 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8485 imu->acct_pages = 0;
8489 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8492 struct vm_area_struct **vmas = NULL;
8493 struct page **pages = NULL;
8494 struct page *last_hpage = NULL;
8495 int i, j, got_pages = 0;
8500 if (!nr_args || nr_args > UIO_MAXIOV)
8503 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8505 if (!ctx->user_bufs)
8508 for (i = 0; i < nr_args; i++) {
8509 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8510 unsigned long off, start, end, ubuf;
8515 ret = io_copy_iov(ctx, &iov, arg, i);
8520 * Don't impose further limits on the size and buffer
8521 * constraints here, we'll -EINVAL later when IO is
8522 * submitted if they are wrong.
8525 if (!iov.iov_base || !iov.iov_len)
8528 /* arbitrary limit, but we need something */
8529 if (iov.iov_len > SZ_1G)
8532 ubuf = (unsigned long) iov.iov_base;
8533 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8534 start = ubuf >> PAGE_SHIFT;
8535 nr_pages = end - start;
8538 if (!pages || nr_pages > got_pages) {
8541 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8543 vmas = kvmalloc_array(nr_pages,
8544 sizeof(struct vm_area_struct *),
8546 if (!pages || !vmas) {
8550 got_pages = nr_pages;
8553 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8560 mmap_read_lock(current->mm);
8561 pret = pin_user_pages(ubuf, nr_pages,
8562 FOLL_WRITE | FOLL_LONGTERM,
8564 if (pret == nr_pages) {
8565 /* don't support file backed memory */
8566 for (j = 0; j < nr_pages; j++) {
8567 struct vm_area_struct *vma = vmas[j];
8570 !is_file_hugepages(vma->vm_file)) {
8576 ret = pret < 0 ? pret : -EFAULT;
8578 mmap_read_unlock(current->mm);
8581 * if we did partial map, or found file backed vmas,
8582 * release any pages we did get
8585 unpin_user_pages(pages, pret);
8590 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8592 unpin_user_pages(pages, pret);
8597 off = ubuf & ~PAGE_MASK;
8599 for (j = 0; j < nr_pages; j++) {
8602 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8603 imu->bvec[j].bv_page = pages[j];
8604 imu->bvec[j].bv_len = vec_len;
8605 imu->bvec[j].bv_offset = off;
8609 /* store original address for later verification */
8611 imu->len = iov.iov_len;
8612 imu->nr_bvecs = nr_pages;
8614 ctx->nr_user_bufs++;
8622 io_sqe_buffer_unregister(ctx);
8626 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8628 __s32 __user *fds = arg;
8634 if (copy_from_user(&fd, fds, sizeof(*fds)))
8637 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8638 if (IS_ERR(ctx->cq_ev_fd)) {
8639 int ret = PTR_ERR(ctx->cq_ev_fd);
8640 ctx->cq_ev_fd = NULL;
8647 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8649 if (ctx->cq_ev_fd) {
8650 eventfd_ctx_put(ctx->cq_ev_fd);
8651 ctx->cq_ev_fd = NULL;
8658 static int __io_destroy_buffers(int id, void *p, void *data)
8660 struct io_ring_ctx *ctx = data;
8661 struct io_buffer *buf = p;
8663 __io_remove_buffers(ctx, buf, id, -1U);
8667 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8669 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8670 idr_destroy(&ctx->io_buffer_idr);
8673 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8675 io_finish_async(ctx);
8676 io_sqe_buffer_unregister(ctx);
8678 if (ctx->sqo_task) {
8679 put_task_struct(ctx->sqo_task);
8680 ctx->sqo_task = NULL;
8681 mmdrop(ctx->mm_account);
8682 ctx->mm_account = NULL;
8685 #ifdef CONFIG_BLK_CGROUP
8686 if (ctx->sqo_blkcg_css)
8687 css_put(ctx->sqo_blkcg_css);
8690 io_sqe_files_unregister(ctx);
8691 io_eventfd_unregister(ctx);
8692 io_destroy_buffers(ctx);
8693 idr_destroy(&ctx->personality_idr);
8695 #if defined(CONFIG_UNIX)
8696 if (ctx->ring_sock) {
8697 ctx->ring_sock->file = NULL; /* so that iput() is called */
8698 sock_release(ctx->ring_sock);
8702 io_mem_free(ctx->rings);
8703 io_mem_free(ctx->sq_sqes);
8705 percpu_ref_exit(&ctx->refs);
8706 free_uid(ctx->user);
8707 put_cred(ctx->creds);
8708 kfree(ctx->cancel_hash);
8709 kmem_cache_free(req_cachep, ctx->fallback_req);
8713 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8715 struct io_ring_ctx *ctx = file->private_data;
8718 poll_wait(file, &ctx->cq_wait, wait);
8720 * synchronizes with barrier from wq_has_sleeper call in
8724 if (!io_sqring_full(ctx))
8725 mask |= EPOLLOUT | EPOLLWRNORM;
8726 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8727 if (io_cqring_events(ctx))
8728 mask |= EPOLLIN | EPOLLRDNORM;
8733 static int io_uring_fasync(int fd, struct file *file, int on)
8735 struct io_ring_ctx *ctx = file->private_data;
8737 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8740 static int io_remove_personalities(int id, void *p, void *data)
8742 struct io_ring_ctx *ctx = data;
8743 struct io_identity *iod;
8745 iod = idr_remove(&ctx->personality_idr, id);
8747 put_cred(iod->creds);
8748 if (refcount_dec_and_test(&iod->count))
8754 static void io_ring_exit_work(struct work_struct *work)
8756 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8760 * If we're doing polled IO and end up having requests being
8761 * submitted async (out-of-line), then completions can come in while
8762 * we're waiting for refs to drop. We need to reap these manually,
8763 * as nobody else will be looking for them.
8766 __io_uring_cancel_task_requests(ctx, NULL);
8767 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8768 io_ring_ctx_free(ctx);
8771 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8773 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8775 return req->ctx == data;
8778 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8780 mutex_lock(&ctx->uring_lock);
8781 percpu_ref_kill(&ctx->refs);
8783 if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead))
8786 /* if force is set, the ring is going away. always drop after that */
8787 ctx->cq_overflow_flushed = 1;
8789 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8790 mutex_unlock(&ctx->uring_lock);
8792 io_kill_timeouts(ctx, NULL, NULL);
8793 io_poll_remove_all(ctx, NULL, NULL);
8796 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8798 /* if we failed setting up the ctx, we might not have any rings */
8799 io_iopoll_try_reap_events(ctx);
8800 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8803 * Do this upfront, so we won't have a grace period where the ring
8804 * is closed but resources aren't reaped yet. This can cause
8805 * spurious failure in setting up a new ring.
8807 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8810 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8812 * Use system_unbound_wq to avoid spawning tons of event kworkers
8813 * if we're exiting a ton of rings at the same time. It just adds
8814 * noise and overhead, there's no discernable change in runtime
8815 * over using system_wq.
8817 queue_work(system_unbound_wq, &ctx->exit_work);
8820 static int io_uring_release(struct inode *inode, struct file *file)
8822 struct io_ring_ctx *ctx = file->private_data;
8824 file->private_data = NULL;
8825 io_ring_ctx_wait_and_kill(ctx);
8829 struct io_task_cancel {
8830 struct task_struct *task;
8831 struct files_struct *files;
8834 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8836 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8837 struct io_task_cancel *cancel = data;
8840 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8841 unsigned long flags;
8842 struct io_ring_ctx *ctx = req->ctx;
8844 /* protect against races with linked timeouts */
8845 spin_lock_irqsave(&ctx->completion_lock, flags);
8846 ret = io_match_task(req, cancel->task, cancel->files);
8847 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8849 ret = io_match_task(req, cancel->task, cancel->files);
8854 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8855 struct task_struct *task,
8856 struct files_struct *files)
8858 struct io_defer_entry *de = NULL;
8861 spin_lock_irq(&ctx->completion_lock);
8862 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8863 if (io_match_task(de->req, task, files)) {
8864 list_cut_position(&list, &ctx->defer_list, &de->list);
8868 spin_unlock_irq(&ctx->completion_lock);
8870 while (!list_empty(&list)) {
8871 de = list_first_entry(&list, struct io_defer_entry, list);
8872 list_del_init(&de->list);
8873 req_set_fail_links(de->req);
8874 io_put_req(de->req);
8875 io_req_complete(de->req, -ECANCELED);
8880 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8881 struct task_struct *task,
8882 struct files_struct *files)
8884 struct io_kiocb *req;
8887 spin_lock_irq(&ctx->inflight_lock);
8888 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8889 cnt += io_match_task(req, task, files);
8890 spin_unlock_irq(&ctx->inflight_lock);
8894 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8895 struct task_struct *task,
8896 struct files_struct *files)
8898 while (!list_empty_careful(&ctx->inflight_list)) {
8899 struct io_task_cancel cancel = { .task = task, .files = files };
8903 inflight = io_uring_count_inflight(ctx, task, files);
8907 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8908 io_poll_remove_all(ctx, task, files);
8909 io_kill_timeouts(ctx, task, files);
8910 io_cqring_overflow_flush(ctx, true, task, files);
8911 /* cancellations _may_ trigger task work */
8914 prepare_to_wait(&task->io_uring->wait, &wait,
8915 TASK_UNINTERRUPTIBLE);
8916 if (inflight == io_uring_count_inflight(ctx, task, files))
8918 finish_wait(&task->io_uring->wait, &wait);
8922 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8923 struct task_struct *task)
8926 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8927 enum io_wq_cancel cret;
8931 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb,
8933 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8936 /* SQPOLL thread does its own polling */
8937 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8938 while (!list_empty_careful(&ctx->iopoll_list)) {
8939 io_iopoll_try_reap_events(ctx);
8944 ret |= io_poll_remove_all(ctx, task, NULL);
8945 ret |= io_kill_timeouts(ctx, task, NULL);
8946 ret |= io_run_task_work();
8953 static void io_disable_sqo_submit(struct io_ring_ctx *ctx)
8955 mutex_lock(&ctx->uring_lock);
8957 mutex_unlock(&ctx->uring_lock);
8959 /* make sure callers enter the ring to get error */
8961 io_ring_set_wakeup_flag(ctx);
8965 * We need to iteratively cancel requests, in case a request has dependent
8966 * hard links. These persist even for failure of cancelations, hence keep
8967 * looping until none are found.
8969 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8970 struct files_struct *files)
8972 struct task_struct *task = current;
8974 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8975 io_disable_sqo_submit(ctx);
8976 task = ctx->sq_data->thread;
8977 atomic_inc(&task->io_uring->in_idle);
8978 io_sq_thread_park(ctx->sq_data);
8981 io_cancel_defer_files(ctx, task, files);
8982 io_cqring_overflow_flush(ctx, true, task, files);
8984 io_uring_cancel_files(ctx, task, files);
8986 __io_uring_cancel_task_requests(ctx, task);
8988 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8989 atomic_dec(&task->io_uring->in_idle);
8990 io_sq_thread_unpark(ctx->sq_data);
8995 * Note that this task has used io_uring. We use it for cancelation purposes.
8997 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8999 struct io_uring_task *tctx = current->io_uring;
9002 if (unlikely(!tctx)) {
9003 ret = io_uring_alloc_task_context(current);
9006 tctx = current->io_uring;
9008 if (tctx->last != file) {
9009 void *old = xa_load(&tctx->xa, (unsigned long)file);
9013 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
9024 * This is race safe in that the task itself is doing this, hence it
9025 * cannot be going through the exit/cancel paths at the same time.
9026 * This cannot be modified while exit/cancel is running.
9028 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
9029 tctx->sqpoll = true;
9035 * Remove this io_uring_file -> task mapping.
9037 static void io_uring_del_task_file(struct file *file)
9039 struct io_uring_task *tctx = current->io_uring;
9041 if (tctx->last == file)
9043 file = xa_erase(&tctx->xa, (unsigned long)file);
9048 static void io_uring_remove_task_files(struct io_uring_task *tctx)
9051 unsigned long index;
9053 xa_for_each(&tctx->xa, index, file)
9054 io_uring_del_task_file(file);
9057 void __io_uring_files_cancel(struct files_struct *files)
9059 struct io_uring_task *tctx = current->io_uring;
9061 unsigned long index;
9063 /* make sure overflow events are dropped */
9064 atomic_inc(&tctx->in_idle);
9065 xa_for_each(&tctx->xa, index, file)
9066 io_uring_cancel_task_requests(file->private_data, files);
9067 atomic_dec(&tctx->in_idle);
9070 io_uring_remove_task_files(tctx);
9073 static s64 tctx_inflight(struct io_uring_task *tctx)
9075 unsigned long index;
9079 inflight = percpu_counter_sum(&tctx->inflight);
9084 * If we have SQPOLL rings, then we need to iterate and find them, and
9085 * add the pending count for those.
9087 xa_for_each(&tctx->xa, index, file) {
9088 struct io_ring_ctx *ctx = file->private_data;
9090 if (ctx->flags & IORING_SETUP_SQPOLL) {
9091 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
9093 inflight += percpu_counter_sum(&__tctx->inflight);
9101 * Find any io_uring fd that this task has registered or done IO on, and cancel
9104 void __io_uring_task_cancel(void)
9106 struct io_uring_task *tctx = current->io_uring;
9110 /* make sure overflow events are dropped */
9111 atomic_inc(&tctx->in_idle);
9113 /* trigger io_disable_sqo_submit() */
9115 __io_uring_files_cancel(NULL);
9118 /* read completions before cancelations */
9119 inflight = tctx_inflight(tctx);
9122 __io_uring_files_cancel(NULL);
9124 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9127 * If we've seen completions, retry without waiting. This
9128 * avoids a race where a completion comes in before we did
9129 * prepare_to_wait().
9131 if (inflight == tctx_inflight(tctx))
9133 finish_wait(&tctx->wait, &wait);
9136 atomic_dec(&tctx->in_idle);
9138 io_uring_remove_task_files(tctx);
9141 static int io_uring_flush(struct file *file, void *data)
9143 struct io_uring_task *tctx = current->io_uring;
9144 struct io_ring_ctx *ctx = file->private_data;
9146 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
9147 io_uring_cancel_task_requests(ctx, NULL);
9152 /* we should have cancelled and erased it before PF_EXITING */
9153 WARN_ON_ONCE((current->flags & PF_EXITING) &&
9154 xa_load(&tctx->xa, (unsigned long)file));
9157 * fput() is pending, will be 2 if the only other ref is our potential
9158 * task file note. If the task is exiting, drop regardless of count.
9160 if (atomic_long_read(&file->f_count) != 2)
9163 if (ctx->flags & IORING_SETUP_SQPOLL) {
9164 /* there is only one file note, which is owned by sqo_task */
9165 WARN_ON_ONCE(ctx->sqo_task != current &&
9166 xa_load(&tctx->xa, (unsigned long)file));
9167 /* sqo_dead check is for when this happens after cancellation */
9168 WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead &&
9169 !xa_load(&tctx->xa, (unsigned long)file));
9171 io_disable_sqo_submit(ctx);
9174 if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current)
9175 io_uring_del_task_file(file);
9179 static void *io_uring_validate_mmap_request(struct file *file,
9180 loff_t pgoff, size_t sz)
9182 struct io_ring_ctx *ctx = file->private_data;
9183 loff_t offset = pgoff << PAGE_SHIFT;
9188 case IORING_OFF_SQ_RING:
9189 case IORING_OFF_CQ_RING:
9192 case IORING_OFF_SQES:
9196 return ERR_PTR(-EINVAL);
9199 page = virt_to_head_page(ptr);
9200 if (sz > page_size(page))
9201 return ERR_PTR(-EINVAL);
9208 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9210 size_t sz = vma->vm_end - vma->vm_start;
9214 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9216 return PTR_ERR(ptr);
9218 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9219 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9222 #else /* !CONFIG_MMU */
9224 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9226 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9229 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9231 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9234 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9235 unsigned long addr, unsigned long len,
9236 unsigned long pgoff, unsigned long flags)
9240 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9242 return PTR_ERR(ptr);
9244 return (unsigned long) ptr;
9247 #endif /* !CONFIG_MMU */
9249 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9255 if (!io_sqring_full(ctx))
9258 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9260 if (unlikely(ctx->sqo_dead)) {
9265 if (!io_sqring_full(ctx))
9269 } while (!signal_pending(current));
9271 finish_wait(&ctx->sqo_sq_wait, &wait);
9276 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9277 struct __kernel_timespec __user **ts,
9278 const sigset_t __user **sig)
9280 struct io_uring_getevents_arg arg;
9283 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9284 * is just a pointer to the sigset_t.
9286 if (!(flags & IORING_ENTER_EXT_ARG)) {
9287 *sig = (const sigset_t __user *) argp;
9293 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9294 * timespec and sigset_t pointers if good.
9296 if (*argsz != sizeof(arg))
9298 if (copy_from_user(&arg, argp, sizeof(arg)))
9300 *sig = u64_to_user_ptr(arg.sigmask);
9301 *argsz = arg.sigmask_sz;
9302 *ts = u64_to_user_ptr(arg.ts);
9306 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9307 u32, min_complete, u32, flags, const void __user *, argp,
9310 struct io_ring_ctx *ctx;
9317 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9318 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9326 if (f.file->f_op != &io_uring_fops)
9330 ctx = f.file->private_data;
9331 if (!percpu_ref_tryget(&ctx->refs))
9335 if (ctx->flags & IORING_SETUP_R_DISABLED)
9339 * For SQ polling, the thread will do all submissions and completions.
9340 * Just return the requested submit count, and wake the thread if
9344 if (ctx->flags & IORING_SETUP_SQPOLL) {
9345 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9348 if (unlikely(ctx->sqo_dead))
9350 if (flags & IORING_ENTER_SQ_WAKEUP)
9351 wake_up(&ctx->sq_data->wait);
9352 if (flags & IORING_ENTER_SQ_WAIT) {
9353 ret = io_sqpoll_wait_sq(ctx);
9357 submitted = to_submit;
9358 } else if (to_submit) {
9359 ret = io_uring_add_task_file(ctx, f.file);
9362 mutex_lock(&ctx->uring_lock);
9363 submitted = io_submit_sqes(ctx, to_submit);
9364 mutex_unlock(&ctx->uring_lock);
9366 if (submitted != to_submit)
9369 if (flags & IORING_ENTER_GETEVENTS) {
9370 const sigset_t __user *sig;
9371 struct __kernel_timespec __user *ts;
9373 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9377 min_complete = min(min_complete, ctx->cq_entries);
9380 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9381 * space applications don't need to do io completion events
9382 * polling again, they can rely on io_sq_thread to do polling
9383 * work, which can reduce cpu usage and uring_lock contention.
9385 if (ctx->flags & IORING_SETUP_IOPOLL &&
9386 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9387 ret = io_iopoll_check(ctx, min_complete);
9389 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9394 percpu_ref_put(&ctx->refs);
9397 return submitted ? submitted : ret;
9400 #ifdef CONFIG_PROC_FS
9401 static int io_uring_show_cred(int id, void *p, void *data)
9403 struct io_identity *iod = p;
9404 const struct cred *cred = iod->creds;
9405 struct seq_file *m = data;
9406 struct user_namespace *uns = seq_user_ns(m);
9407 struct group_info *gi;
9412 seq_printf(m, "%5d\n", id);
9413 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9414 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9415 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9416 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9417 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9418 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9419 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9420 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9421 seq_puts(m, "\n\tGroups:\t");
9422 gi = cred->group_info;
9423 for (g = 0; g < gi->ngroups; g++) {
9424 seq_put_decimal_ull(m, g ? " " : "",
9425 from_kgid_munged(uns, gi->gid[g]));
9427 seq_puts(m, "\n\tCapEff:\t");
9428 cap = cred->cap_effective;
9429 CAP_FOR_EACH_U32(__capi)
9430 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9435 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9437 struct io_sq_data *sq = NULL;
9442 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9443 * since fdinfo case grabs it in the opposite direction of normal use
9444 * cases. If we fail to get the lock, we just don't iterate any
9445 * structures that could be going away outside the io_uring mutex.
9447 has_lock = mutex_trylock(&ctx->uring_lock);
9449 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9452 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9453 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9454 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9455 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9456 struct fixed_file_table *table;
9459 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9460 f = table->files[i & IORING_FILE_TABLE_MASK];
9462 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9464 seq_printf(m, "%5u: <none>\n", i);
9466 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9467 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9468 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9470 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9471 (unsigned int) buf->len);
9473 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9474 seq_printf(m, "Personalities:\n");
9475 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9477 seq_printf(m, "PollList:\n");
9478 spin_lock_irq(&ctx->completion_lock);
9479 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9480 struct hlist_head *list = &ctx->cancel_hash[i];
9481 struct io_kiocb *req;
9483 hlist_for_each_entry(req, list, hash_node)
9484 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9485 req->task->task_works != NULL);
9487 spin_unlock_irq(&ctx->completion_lock);
9489 mutex_unlock(&ctx->uring_lock);
9492 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9494 struct io_ring_ctx *ctx = f->private_data;
9496 if (percpu_ref_tryget(&ctx->refs)) {
9497 __io_uring_show_fdinfo(ctx, m);
9498 percpu_ref_put(&ctx->refs);
9503 static const struct file_operations io_uring_fops = {
9504 .release = io_uring_release,
9505 .flush = io_uring_flush,
9506 .mmap = io_uring_mmap,
9508 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9509 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9511 .poll = io_uring_poll,
9512 .fasync = io_uring_fasync,
9513 #ifdef CONFIG_PROC_FS
9514 .show_fdinfo = io_uring_show_fdinfo,
9518 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9519 struct io_uring_params *p)
9521 struct io_rings *rings;
9522 size_t size, sq_array_offset;
9524 /* make sure these are sane, as we already accounted them */
9525 ctx->sq_entries = p->sq_entries;
9526 ctx->cq_entries = p->cq_entries;
9528 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9529 if (size == SIZE_MAX)
9532 rings = io_mem_alloc(size);
9537 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9538 rings->sq_ring_mask = p->sq_entries - 1;
9539 rings->cq_ring_mask = p->cq_entries - 1;
9540 rings->sq_ring_entries = p->sq_entries;
9541 rings->cq_ring_entries = p->cq_entries;
9542 ctx->sq_mask = rings->sq_ring_mask;
9543 ctx->cq_mask = rings->cq_ring_mask;
9545 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9546 if (size == SIZE_MAX) {
9547 io_mem_free(ctx->rings);
9552 ctx->sq_sqes = io_mem_alloc(size);
9553 if (!ctx->sq_sqes) {
9554 io_mem_free(ctx->rings);
9562 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9566 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9570 ret = io_uring_add_task_file(ctx, file);
9575 fd_install(fd, file);
9580 * Allocate an anonymous fd, this is what constitutes the application
9581 * visible backing of an io_uring instance. The application mmaps this
9582 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9583 * we have to tie this fd to a socket for file garbage collection purposes.
9585 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9588 #if defined(CONFIG_UNIX)
9591 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9594 return ERR_PTR(ret);
9597 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9598 O_RDWR | O_CLOEXEC);
9599 #if defined(CONFIG_UNIX)
9601 sock_release(ctx->ring_sock);
9602 ctx->ring_sock = NULL;
9604 ctx->ring_sock->file = file;
9610 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9611 struct io_uring_params __user *params)
9613 struct user_struct *user = NULL;
9614 struct io_ring_ctx *ctx;
9621 if (entries > IORING_MAX_ENTRIES) {
9622 if (!(p->flags & IORING_SETUP_CLAMP))
9624 entries = IORING_MAX_ENTRIES;
9628 * Use twice as many entries for the CQ ring. It's possible for the
9629 * application to drive a higher depth than the size of the SQ ring,
9630 * since the sqes are only used at submission time. This allows for
9631 * some flexibility in overcommitting a bit. If the application has
9632 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9633 * of CQ ring entries manually.
9635 p->sq_entries = roundup_pow_of_two(entries);
9636 if (p->flags & IORING_SETUP_CQSIZE) {
9638 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9639 * to a power-of-two, if it isn't already. We do NOT impose
9640 * any cq vs sq ring sizing.
9644 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9645 if (!(p->flags & IORING_SETUP_CLAMP))
9647 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9649 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9650 if (p->cq_entries < p->sq_entries)
9653 p->cq_entries = 2 * p->sq_entries;
9656 user = get_uid(current_user());
9657 limit_mem = !capable(CAP_IPC_LOCK);
9660 ret = __io_account_mem(user,
9661 ring_pages(p->sq_entries, p->cq_entries));
9668 ctx = io_ring_ctx_alloc(p);
9671 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9676 ctx->compat = in_compat_syscall();
9678 ctx->creds = get_current_cred();
9680 ctx->loginuid = current->loginuid;
9681 ctx->sessionid = current->sessionid;
9683 ctx->sqo_task = get_task_struct(current);
9686 * This is just grabbed for accounting purposes. When a process exits,
9687 * the mm is exited and dropped before the files, hence we need to hang
9688 * on to this mm purely for the purposes of being able to unaccount
9689 * memory (locked/pinned vm). It's not used for anything else.
9691 mmgrab(current->mm);
9692 ctx->mm_account = current->mm;
9694 #ifdef CONFIG_BLK_CGROUP
9696 * The sq thread will belong to the original cgroup it was inited in.
9697 * If the cgroup goes offline (e.g. disabling the io controller), then
9698 * issued bios will be associated with the closest cgroup later in the
9702 ctx->sqo_blkcg_css = blkcg_css();
9703 ret = css_tryget_online(ctx->sqo_blkcg_css);
9706 /* don't init against a dying cgroup, have the user try again */
9707 ctx->sqo_blkcg_css = NULL;
9714 * Account memory _before_ installing the file descriptor. Once
9715 * the descriptor is installed, it can get closed at any time. Also
9716 * do this before hitting the general error path, as ring freeing
9717 * will un-account as well.
9719 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9721 ctx->limit_mem = limit_mem;
9723 ret = io_allocate_scq_urings(ctx, p);
9727 ret = io_sq_offload_create(ctx, p);
9731 if (!(p->flags & IORING_SETUP_R_DISABLED))
9732 io_sq_offload_start(ctx);
9734 memset(&p->sq_off, 0, sizeof(p->sq_off));
9735 p->sq_off.head = offsetof(struct io_rings, sq.head);
9736 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9737 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9738 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9739 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9740 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9741 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9743 memset(&p->cq_off, 0, sizeof(p->cq_off));
9744 p->cq_off.head = offsetof(struct io_rings, cq.head);
9745 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9746 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9747 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9748 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9749 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9750 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9752 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9753 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9754 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9755 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9756 IORING_FEAT_EXT_ARG;
9758 if (copy_to_user(params, p, sizeof(*p))) {
9763 file = io_uring_get_file(ctx);
9765 ret = PTR_ERR(file);
9770 * Install ring fd as the very last thing, so we don't risk someone
9771 * having closed it before we finish setup
9773 ret = io_uring_install_fd(ctx, file);
9775 io_disable_sqo_submit(ctx);
9776 /* fput will clean it up */
9781 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9784 io_disable_sqo_submit(ctx);
9785 io_ring_ctx_wait_and_kill(ctx);
9790 * Sets up an aio uring context, and returns the fd. Applications asks for a
9791 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9792 * params structure passed in.
9794 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9796 struct io_uring_params p;
9799 if (copy_from_user(&p, params, sizeof(p)))
9801 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9806 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9807 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9808 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9809 IORING_SETUP_R_DISABLED))
9812 return io_uring_create(entries, &p, params);
9815 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9816 struct io_uring_params __user *, params)
9818 return io_uring_setup(entries, params);
9821 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9823 struct io_uring_probe *p;
9827 size = struct_size(p, ops, nr_args);
9828 if (size == SIZE_MAX)
9830 p = kzalloc(size, GFP_KERNEL);
9835 if (copy_from_user(p, arg, size))
9838 if (memchr_inv(p, 0, size))
9841 p->last_op = IORING_OP_LAST - 1;
9842 if (nr_args > IORING_OP_LAST)
9843 nr_args = IORING_OP_LAST;
9845 for (i = 0; i < nr_args; i++) {
9847 if (!io_op_defs[i].not_supported)
9848 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9853 if (copy_to_user(arg, p, size))
9860 static int io_register_personality(struct io_ring_ctx *ctx)
9862 struct io_identity *id;
9865 id = kmalloc(sizeof(*id), GFP_KERNEL);
9869 io_init_identity(id);
9870 id->creds = get_current_cred();
9872 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9874 put_cred(id->creds);
9880 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9882 struct io_identity *iod;
9884 iod = idr_remove(&ctx->personality_idr, id);
9886 put_cred(iod->creds);
9887 if (refcount_dec_and_test(&iod->count))
9895 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9896 unsigned int nr_args)
9898 struct io_uring_restriction *res;
9902 /* Restrictions allowed only if rings started disabled */
9903 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9906 /* We allow only a single restrictions registration */
9907 if (ctx->restrictions.registered)
9910 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9913 size = array_size(nr_args, sizeof(*res));
9914 if (size == SIZE_MAX)
9917 res = memdup_user(arg, size);
9919 return PTR_ERR(res);
9923 for (i = 0; i < nr_args; i++) {
9924 switch (res[i].opcode) {
9925 case IORING_RESTRICTION_REGISTER_OP:
9926 if (res[i].register_op >= IORING_REGISTER_LAST) {
9931 __set_bit(res[i].register_op,
9932 ctx->restrictions.register_op);
9934 case IORING_RESTRICTION_SQE_OP:
9935 if (res[i].sqe_op >= IORING_OP_LAST) {
9940 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9942 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9943 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9945 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9946 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9955 /* Reset all restrictions if an error happened */
9957 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9959 ctx->restrictions.registered = true;
9965 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9967 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9970 if (ctx->restrictions.registered)
9971 ctx->restricted = 1;
9973 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9975 io_sq_offload_start(ctx);
9980 static bool io_register_op_must_quiesce(int op)
9983 case IORING_UNREGISTER_FILES:
9984 case IORING_REGISTER_FILES_UPDATE:
9985 case IORING_REGISTER_PROBE:
9986 case IORING_REGISTER_PERSONALITY:
9987 case IORING_UNREGISTER_PERSONALITY:
9994 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9995 void __user *arg, unsigned nr_args)
9996 __releases(ctx->uring_lock)
9997 __acquires(ctx->uring_lock)
10002 * We're inside the ring mutex, if the ref is already dying, then
10003 * someone else killed the ctx or is already going through
10004 * io_uring_register().
10006 if (percpu_ref_is_dying(&ctx->refs))
10009 if (io_register_op_must_quiesce(opcode)) {
10010 percpu_ref_kill(&ctx->refs);
10013 * Drop uring mutex before waiting for references to exit. If
10014 * another thread is currently inside io_uring_enter() it might
10015 * need to grab the uring_lock to make progress. If we hold it
10016 * here across the drain wait, then we can deadlock. It's safe
10017 * to drop the mutex here, since no new references will come in
10018 * after we've killed the percpu ref.
10020 mutex_unlock(&ctx->uring_lock);
10022 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10025 ret = io_run_task_work_sig();
10030 mutex_lock(&ctx->uring_lock);
10033 percpu_ref_resurrect(&ctx->refs);
10038 if (ctx->restricted) {
10039 if (opcode >= IORING_REGISTER_LAST) {
10044 if (!test_bit(opcode, ctx->restrictions.register_op)) {
10051 case IORING_REGISTER_BUFFERS:
10052 ret = io_sqe_buffer_register(ctx, arg, nr_args);
10054 case IORING_UNREGISTER_BUFFERS:
10056 if (arg || nr_args)
10058 ret = io_sqe_buffer_unregister(ctx);
10060 case IORING_REGISTER_FILES:
10061 ret = io_sqe_files_register(ctx, arg, nr_args);
10063 case IORING_UNREGISTER_FILES:
10065 if (arg || nr_args)
10067 ret = io_sqe_files_unregister(ctx);
10069 case IORING_REGISTER_FILES_UPDATE:
10070 ret = io_sqe_files_update(ctx, arg, nr_args);
10072 case IORING_REGISTER_EVENTFD:
10073 case IORING_REGISTER_EVENTFD_ASYNC:
10077 ret = io_eventfd_register(ctx, arg);
10080 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10081 ctx->eventfd_async = 1;
10083 ctx->eventfd_async = 0;
10085 case IORING_UNREGISTER_EVENTFD:
10087 if (arg || nr_args)
10089 ret = io_eventfd_unregister(ctx);
10091 case IORING_REGISTER_PROBE:
10093 if (!arg || nr_args > 256)
10095 ret = io_probe(ctx, arg, nr_args);
10097 case IORING_REGISTER_PERSONALITY:
10099 if (arg || nr_args)
10101 ret = io_register_personality(ctx);
10103 case IORING_UNREGISTER_PERSONALITY:
10107 ret = io_unregister_personality(ctx, nr_args);
10109 case IORING_REGISTER_ENABLE_RINGS:
10111 if (arg || nr_args)
10113 ret = io_register_enable_rings(ctx);
10115 case IORING_REGISTER_RESTRICTIONS:
10116 ret = io_register_restrictions(ctx, arg, nr_args);
10124 if (io_register_op_must_quiesce(opcode)) {
10125 /* bring the ctx back to life */
10126 percpu_ref_reinit(&ctx->refs);
10128 reinit_completion(&ctx->ref_comp);
10133 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10134 void __user *, arg, unsigned int, nr_args)
10136 struct io_ring_ctx *ctx;
10145 if (f.file->f_op != &io_uring_fops)
10148 ctx = f.file->private_data;
10150 mutex_lock(&ctx->uring_lock);
10151 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10152 mutex_unlock(&ctx->uring_lock);
10153 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10154 ctx->cq_ev_fd != NULL, ret);
10160 static int __init io_uring_init(void)
10162 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10163 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10164 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10167 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10168 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10169 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10170 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10171 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10172 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10173 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10174 BUILD_BUG_SQE_ELEM(8, __u64, off);
10175 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10176 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10177 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10178 BUILD_BUG_SQE_ELEM(24, __u32, len);
10179 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10180 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10181 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10182 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10183 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10184 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10185 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10186 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10187 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10188 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10189 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10190 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10191 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10192 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10193 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10194 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10195 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10196 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10197 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10199 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10200 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10201 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10204 __initcall(io_uring_init);