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_cqe (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/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head ltimeout_list;
379 struct list_head cq_overflow_list;
380 struct xarray io_buffers;
381 struct xarray personalities;
383 unsigned sq_thread_idle;
384 } ____cacheline_aligned_in_smp;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list;
388 unsigned int locked_free_nr;
390 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
396 unsigned long check_cq_overflow;
399 unsigned cached_cq_tail;
401 struct eventfd_ctx *cq_ev_fd;
402 struct wait_queue_head poll_wait;
403 struct wait_queue_head cq_wait;
405 atomic_t cq_timeouts;
406 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
460 bool iowq_limits_set;
464 struct io_uring_task {
465 /* submission side */
468 struct wait_queue_head wait;
469 const struct io_ring_ctx *last;
471 struct percpu_counter inflight;
472 atomic_t inflight_tracked;
475 spinlock_t task_lock;
476 struct io_wq_work_list task_list;
477 struct callback_head task_work;
482 * First field must be the file pointer in all the
483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
485 struct io_poll_iocb {
487 struct wait_queue_head *head;
491 struct wait_queue_entry wait;
494 struct io_poll_update {
500 bool update_user_data;
509 struct io_timeout_data {
510 struct io_kiocb *req;
511 struct hrtimer timer;
512 struct timespec64 ts;
513 enum hrtimer_mode mode;
519 struct sockaddr __user *addr;
520 int __user *addr_len;
523 unsigned long nofile;
543 struct list_head list;
544 /* head of the link, used by linked timeouts only */
545 struct io_kiocb *head;
546 /* for linked completions */
547 struct io_kiocb *prev;
550 struct io_timeout_rem {
555 struct timespec64 ts;
561 /* NOTE: kiocb has the file as the first member, so don't do it here */
569 struct sockaddr __user *addr;
576 struct compat_msghdr __user *umsg_compat;
577 struct user_msghdr __user *umsg;
583 struct io_buffer *kbuf;
590 struct filename *filename;
592 unsigned long nofile;
595 struct io_rsrc_update {
621 struct epoll_event event;
625 struct file *file_out;
626 struct file *file_in;
633 struct io_provide_buf {
647 const char __user *filename;
648 struct statx __user *buffer;
660 struct filename *oldpath;
661 struct filename *newpath;
669 struct filename *filename;
676 struct filename *filename;
682 struct filename *oldpath;
683 struct filename *newpath;
690 struct filename *oldpath;
691 struct filename *newpath;
695 struct io_completion {
700 struct io_async_connect {
701 struct sockaddr_storage address;
704 struct io_async_msghdr {
705 struct iovec fast_iov[UIO_FASTIOV];
706 /* points to an allocated iov, if NULL we use fast_iov instead */
707 struct iovec *free_iov;
708 struct sockaddr __user *uaddr;
710 struct sockaddr_storage addr;
714 struct iovec fast_iov[UIO_FASTIOV];
715 const struct iovec *free_iovec;
716 struct iov_iter iter;
717 struct iov_iter_state iter_state;
719 struct wait_page_queue wpq;
723 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
724 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
725 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
726 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
727 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
728 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
730 /* first byte is taken by user flags, shift it to not overlap */
735 REQ_F_LINK_TIMEOUT_BIT,
736 REQ_F_NEED_CLEANUP_BIT,
738 REQ_F_BUFFER_SELECTED_BIT,
739 REQ_F_COMPLETE_INLINE_BIT,
743 REQ_F_ARM_LTIMEOUT_BIT,
744 /* keep async read/write and isreg together and in order */
745 REQ_F_NOWAIT_READ_BIT,
746 REQ_F_NOWAIT_WRITE_BIT,
749 /* not a real bit, just to check we're not overflowing the space */
755 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
756 /* drain existing IO first */
757 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
759 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
760 /* doesn't sever on completion < 0 */
761 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
763 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
764 /* IOSQE_BUFFER_SELECT */
765 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
767 /* fail rest of links */
768 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
769 /* on inflight list, should be cancelled and waited on exit reliably */
770 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
771 /* read/write uses file position */
772 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
773 /* must not punt to workers */
774 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
775 /* has or had linked timeout */
776 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
778 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
779 /* already went through poll handler */
780 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
781 /* buffer already selected */
782 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
783 /* completion is deferred through io_comp_state */
784 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
785 /* caller should reissue async */
786 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
787 /* supports async reads */
788 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
789 /* supports async writes */
790 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
792 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
793 /* has creds assigned */
794 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
795 /* skip refcounting if not set */
796 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
797 /* there is a linked timeout that has to be armed */
798 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
802 struct io_poll_iocb poll;
803 struct io_poll_iocb *double_poll;
806 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
808 struct io_task_work {
810 struct io_wq_work_node node;
811 struct llist_node fallback_node;
813 io_req_tw_func_t func;
817 IORING_RSRC_FILE = 0,
818 IORING_RSRC_BUFFER = 1,
822 * NOTE! Each of the iocb union members has the file pointer
823 * as the first entry in their struct definition. So you can
824 * access the file pointer through any of the sub-structs,
825 * or directly as just 'ki_filp' in this struct.
831 struct io_poll_iocb poll;
832 struct io_poll_update poll_update;
833 struct io_accept accept;
835 struct io_cancel cancel;
836 struct io_timeout timeout;
837 struct io_timeout_rem timeout_rem;
838 struct io_connect connect;
839 struct io_sr_msg sr_msg;
841 struct io_close close;
842 struct io_rsrc_update rsrc_update;
843 struct io_fadvise fadvise;
844 struct io_madvise madvise;
845 struct io_epoll epoll;
846 struct io_splice splice;
847 struct io_provide_buf pbuf;
848 struct io_statx statx;
849 struct io_shutdown shutdown;
850 struct io_rename rename;
851 struct io_unlink unlink;
852 struct io_mkdir mkdir;
853 struct io_symlink symlink;
854 struct io_hardlink hardlink;
855 /* use only after cleaning per-op data, see io_clean_op() */
856 struct io_completion compl;
859 /* opcode allocated if it needs to store data for async defer */
862 /* polled IO has completed */
868 struct io_ring_ctx *ctx;
871 struct task_struct *task;
874 struct io_kiocb *link;
875 struct percpu_ref *fixed_rsrc_refs;
877 /* used with ctx->iopoll_list with reads/writes */
878 struct list_head inflight_entry;
879 struct io_task_work io_task_work;
880 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
881 struct hlist_node hash_node;
882 struct async_poll *apoll;
883 struct io_wq_work work;
884 const struct cred *creds;
886 /* store used ubuf, so we can prevent reloading */
887 struct io_mapped_ubuf *imu;
890 struct io_tctx_node {
891 struct list_head ctx_node;
892 struct task_struct *task;
893 struct io_ring_ctx *ctx;
896 struct io_defer_entry {
897 struct list_head list;
898 struct io_kiocb *req;
903 /* needs req->file assigned */
904 unsigned needs_file : 1;
905 /* hash wq insertion if file is a regular file */
906 unsigned hash_reg_file : 1;
907 /* unbound wq insertion if file is a non-regular file */
908 unsigned unbound_nonreg_file : 1;
909 /* opcode is not supported by this kernel */
910 unsigned not_supported : 1;
911 /* set if opcode supports polled "wait" */
913 unsigned pollout : 1;
914 /* op supports buffer selection */
915 unsigned buffer_select : 1;
916 /* do prep async if is going to be punted */
917 unsigned needs_async_setup : 1;
918 /* should block plug */
920 /* size of async data needed, if any */
921 unsigned short async_size;
924 static const struct io_op_def io_op_defs[] = {
925 [IORING_OP_NOP] = {},
926 [IORING_OP_READV] = {
928 .unbound_nonreg_file = 1,
931 .needs_async_setup = 1,
933 .async_size = sizeof(struct io_async_rw),
935 [IORING_OP_WRITEV] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_setup = 1,
942 .async_size = sizeof(struct io_async_rw),
944 [IORING_OP_FSYNC] = {
947 [IORING_OP_READ_FIXED] = {
949 .unbound_nonreg_file = 1,
952 .async_size = sizeof(struct io_async_rw),
954 [IORING_OP_WRITE_FIXED] = {
957 .unbound_nonreg_file = 1,
960 .async_size = sizeof(struct io_async_rw),
962 [IORING_OP_POLL_ADD] = {
964 .unbound_nonreg_file = 1,
966 [IORING_OP_POLL_REMOVE] = {},
967 [IORING_OP_SYNC_FILE_RANGE] = {
970 [IORING_OP_SENDMSG] = {
972 .unbound_nonreg_file = 1,
974 .needs_async_setup = 1,
975 .async_size = sizeof(struct io_async_msghdr),
977 [IORING_OP_RECVMSG] = {
979 .unbound_nonreg_file = 1,
982 .needs_async_setup = 1,
983 .async_size = sizeof(struct io_async_msghdr),
985 [IORING_OP_TIMEOUT] = {
986 .async_size = sizeof(struct io_timeout_data),
988 [IORING_OP_TIMEOUT_REMOVE] = {
989 /* used by timeout updates' prep() */
991 [IORING_OP_ACCEPT] = {
993 .unbound_nonreg_file = 1,
996 [IORING_OP_ASYNC_CANCEL] = {},
997 [IORING_OP_LINK_TIMEOUT] = {
998 .async_size = sizeof(struct io_timeout_data),
1000 [IORING_OP_CONNECT] = {
1002 .unbound_nonreg_file = 1,
1004 .needs_async_setup = 1,
1005 .async_size = sizeof(struct io_async_connect),
1007 [IORING_OP_FALLOCATE] = {
1010 [IORING_OP_OPENAT] = {},
1011 [IORING_OP_CLOSE] = {},
1012 [IORING_OP_FILES_UPDATE] = {},
1013 [IORING_OP_STATX] = {},
1014 [IORING_OP_READ] = {
1016 .unbound_nonreg_file = 1,
1020 .async_size = sizeof(struct io_async_rw),
1022 [IORING_OP_WRITE] = {
1025 .unbound_nonreg_file = 1,
1028 .async_size = sizeof(struct io_async_rw),
1030 [IORING_OP_FADVISE] = {
1033 [IORING_OP_MADVISE] = {},
1034 [IORING_OP_SEND] = {
1036 .unbound_nonreg_file = 1,
1039 [IORING_OP_RECV] = {
1041 .unbound_nonreg_file = 1,
1045 [IORING_OP_OPENAT2] = {
1047 [IORING_OP_EPOLL_CTL] = {
1048 .unbound_nonreg_file = 1,
1050 [IORING_OP_SPLICE] = {
1053 .unbound_nonreg_file = 1,
1055 [IORING_OP_PROVIDE_BUFFERS] = {},
1056 [IORING_OP_REMOVE_BUFFERS] = {},
1060 .unbound_nonreg_file = 1,
1062 [IORING_OP_SHUTDOWN] = {
1065 [IORING_OP_RENAMEAT] = {},
1066 [IORING_OP_UNLINKAT] = {},
1067 [IORING_OP_MKDIRAT] = {},
1068 [IORING_OP_SYMLINKAT] = {},
1069 [IORING_OP_LINKAT] = {},
1072 /* requests with any of those set should undergo io_disarm_next() */
1073 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1075 static bool io_disarm_next(struct io_kiocb *req);
1076 static void io_uring_del_tctx_node(unsigned long index);
1077 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1078 struct task_struct *task,
1080 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1082 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1083 long res, unsigned int cflags);
1084 static void io_put_req(struct io_kiocb *req);
1085 static void io_put_req_deferred(struct io_kiocb *req);
1086 static void io_dismantle_req(struct io_kiocb *req);
1087 static void io_queue_linked_timeout(struct io_kiocb *req);
1088 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1089 struct io_uring_rsrc_update2 *up,
1091 static void io_clean_op(struct io_kiocb *req);
1092 static struct file *io_file_get(struct io_ring_ctx *ctx,
1093 struct io_kiocb *req, int fd, bool fixed);
1094 static void __io_queue_sqe(struct io_kiocb *req);
1095 static void io_rsrc_put_work(struct work_struct *work);
1097 static void io_req_task_queue(struct io_kiocb *req);
1098 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1099 static int io_req_prep_async(struct io_kiocb *req);
1101 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1102 unsigned int issue_flags, u32 slot_index);
1103 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1105 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1107 static struct kmem_cache *req_cachep;
1109 static const struct file_operations io_uring_fops;
1111 struct sock *io_uring_get_socket(struct file *file)
1113 #if defined(CONFIG_UNIX)
1114 if (file->f_op == &io_uring_fops) {
1115 struct io_ring_ctx *ctx = file->private_data;
1117 return ctx->ring_sock->sk;
1122 EXPORT_SYMBOL(io_uring_get_socket);
1124 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1127 mutex_lock(&ctx->uring_lock);
1132 #define io_for_each_link(pos, head) \
1133 for (pos = (head); pos; pos = pos->link)
1136 * Shamelessly stolen from the mm implementation of page reference checking,
1137 * see commit f958d7b528b1 for details.
1139 #define req_ref_zero_or_close_to_overflow(req) \
1140 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1142 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1144 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1145 return atomic_inc_not_zero(&req->refs);
1148 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1150 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1153 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1154 return atomic_dec_and_test(&req->refs);
1157 static inline void req_ref_put(struct io_kiocb *req)
1159 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1160 WARN_ON_ONCE(req_ref_put_and_test(req));
1163 static inline void req_ref_get(struct io_kiocb *req)
1165 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1166 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1167 atomic_inc(&req->refs);
1170 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1172 if (!(req->flags & REQ_F_REFCOUNT)) {
1173 req->flags |= REQ_F_REFCOUNT;
1174 atomic_set(&req->refs, nr);
1178 static inline void io_req_set_refcount(struct io_kiocb *req)
1180 __io_req_set_refcount(req, 1);
1183 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1185 struct io_ring_ctx *ctx = req->ctx;
1187 if (!req->fixed_rsrc_refs) {
1188 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1189 percpu_ref_get(req->fixed_rsrc_refs);
1193 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1195 bool got = percpu_ref_tryget(ref);
1197 /* already at zero, wait for ->release() */
1199 wait_for_completion(compl);
1200 percpu_ref_resurrect(ref);
1202 percpu_ref_put(ref);
1205 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1208 struct io_kiocb *req;
1210 if (task && head->task != task)
1215 io_for_each_link(req, head) {
1216 if (req->flags & REQ_F_INFLIGHT)
1222 static inline void req_set_fail(struct io_kiocb *req)
1224 req->flags |= REQ_F_FAIL;
1227 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1233 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1235 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1237 complete(&ctx->ref_comp);
1240 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1242 return !req->timeout.off;
1245 static void io_fallback_req_func(struct work_struct *work)
1247 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1248 fallback_work.work);
1249 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1250 struct io_kiocb *req, *tmp;
1251 bool locked = false;
1253 percpu_ref_get(&ctx->refs);
1254 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1255 req->io_task_work.func(req, &locked);
1258 if (ctx->submit_state.compl_nr)
1259 io_submit_flush_completions(ctx);
1260 mutex_unlock(&ctx->uring_lock);
1262 percpu_ref_put(&ctx->refs);
1266 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1268 struct io_ring_ctx *ctx;
1271 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1276 * Use 5 bits less than the max cq entries, that should give us around
1277 * 32 entries per hash list if totally full and uniformly spread.
1279 hash_bits = ilog2(p->cq_entries);
1283 ctx->cancel_hash_bits = hash_bits;
1284 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1286 if (!ctx->cancel_hash)
1288 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1290 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1291 if (!ctx->dummy_ubuf)
1293 /* set invalid range, so io_import_fixed() fails meeting it */
1294 ctx->dummy_ubuf->ubuf = -1UL;
1296 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1297 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1300 ctx->flags = p->flags;
1301 init_waitqueue_head(&ctx->sqo_sq_wait);
1302 INIT_LIST_HEAD(&ctx->sqd_list);
1303 init_waitqueue_head(&ctx->poll_wait);
1304 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1305 init_completion(&ctx->ref_comp);
1306 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1307 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1308 mutex_init(&ctx->uring_lock);
1309 init_waitqueue_head(&ctx->cq_wait);
1310 spin_lock_init(&ctx->completion_lock);
1311 spin_lock_init(&ctx->timeout_lock);
1312 INIT_LIST_HEAD(&ctx->iopoll_list);
1313 INIT_LIST_HEAD(&ctx->defer_list);
1314 INIT_LIST_HEAD(&ctx->timeout_list);
1315 INIT_LIST_HEAD(&ctx->ltimeout_list);
1316 spin_lock_init(&ctx->rsrc_ref_lock);
1317 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1318 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1319 init_llist_head(&ctx->rsrc_put_llist);
1320 INIT_LIST_HEAD(&ctx->tctx_list);
1321 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1322 INIT_LIST_HEAD(&ctx->locked_free_list);
1323 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1326 kfree(ctx->dummy_ubuf);
1327 kfree(ctx->cancel_hash);
1332 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1334 struct io_rings *r = ctx->rings;
1336 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1340 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1342 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1343 struct io_ring_ctx *ctx = req->ctx;
1345 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1351 #define FFS_ASYNC_READ 0x1UL
1352 #define FFS_ASYNC_WRITE 0x2UL
1354 #define FFS_ISREG 0x4UL
1356 #define FFS_ISREG 0x0UL
1358 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1360 static inline bool io_req_ffs_set(struct io_kiocb *req)
1362 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1365 static void io_req_track_inflight(struct io_kiocb *req)
1367 if (!(req->flags & REQ_F_INFLIGHT)) {
1368 req->flags |= REQ_F_INFLIGHT;
1369 atomic_inc(¤t->io_uring->inflight_tracked);
1373 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1375 if (WARN_ON_ONCE(!req->link))
1378 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1379 req->flags |= REQ_F_LINK_TIMEOUT;
1381 /* linked timeouts should have two refs once prep'ed */
1382 io_req_set_refcount(req);
1383 __io_req_set_refcount(req->link, 2);
1387 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1389 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1391 return __io_prep_linked_timeout(req);
1394 static void io_prep_async_work(struct io_kiocb *req)
1396 const struct io_op_def *def = &io_op_defs[req->opcode];
1397 struct io_ring_ctx *ctx = req->ctx;
1399 if (!(req->flags & REQ_F_CREDS)) {
1400 req->flags |= REQ_F_CREDS;
1401 req->creds = get_current_cred();
1404 req->work.list.next = NULL;
1405 req->work.flags = 0;
1406 if (req->flags & REQ_F_FORCE_ASYNC)
1407 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1409 if (req->flags & REQ_F_ISREG) {
1410 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1411 io_wq_hash_work(&req->work, file_inode(req->file));
1412 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1413 if (def->unbound_nonreg_file)
1414 req->work.flags |= IO_WQ_WORK_UNBOUND;
1417 switch (req->opcode) {
1418 case IORING_OP_SPLICE:
1420 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1421 req->work.flags |= IO_WQ_WORK_UNBOUND;
1426 static void io_prep_async_link(struct io_kiocb *req)
1428 struct io_kiocb *cur;
1430 if (req->flags & REQ_F_LINK_TIMEOUT) {
1431 struct io_ring_ctx *ctx = req->ctx;
1433 spin_lock_irq(&ctx->timeout_lock);
1434 io_for_each_link(cur, req)
1435 io_prep_async_work(cur);
1436 spin_unlock_irq(&ctx->timeout_lock);
1438 io_for_each_link(cur, req)
1439 io_prep_async_work(cur);
1443 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1445 struct io_ring_ctx *ctx = req->ctx;
1446 struct io_kiocb *link = io_prep_linked_timeout(req);
1447 struct io_uring_task *tctx = req->task->io_uring;
1449 /* must not take the lock, NULL it as a precaution */
1453 BUG_ON(!tctx->io_wq);
1455 /* init ->work of the whole link before punting */
1456 io_prep_async_link(req);
1459 * Not expected to happen, but if we do have a bug where this _can_
1460 * happen, catch it here and ensure the request is marked as
1461 * canceled. That will make io-wq go through the usual work cancel
1462 * procedure rather than attempt to run this request (or create a new
1465 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1466 req->work.flags |= IO_WQ_WORK_CANCEL;
1468 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1469 &req->work, req->flags);
1470 io_wq_enqueue(tctx->io_wq, &req->work);
1472 io_queue_linked_timeout(link);
1475 static void io_kill_timeout(struct io_kiocb *req, int status)
1476 __must_hold(&req->ctx->completion_lock)
1477 __must_hold(&req->ctx->timeout_lock)
1479 struct io_timeout_data *io = req->async_data;
1481 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1484 atomic_set(&req->ctx->cq_timeouts,
1485 atomic_read(&req->ctx->cq_timeouts) + 1);
1486 list_del_init(&req->timeout.list);
1487 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1488 io_put_req_deferred(req);
1492 static void io_queue_deferred(struct io_ring_ctx *ctx)
1494 while (!list_empty(&ctx->defer_list)) {
1495 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1496 struct io_defer_entry, list);
1498 if (req_need_defer(de->req, de->seq))
1500 list_del_init(&de->list);
1501 io_req_task_queue(de->req);
1506 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1507 __must_hold(&ctx->completion_lock)
1509 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1511 spin_lock_irq(&ctx->timeout_lock);
1512 while (!list_empty(&ctx->timeout_list)) {
1513 u32 events_needed, events_got;
1514 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1515 struct io_kiocb, timeout.list);
1517 if (io_is_timeout_noseq(req))
1521 * Since seq can easily wrap around over time, subtract
1522 * the last seq at which timeouts were flushed before comparing.
1523 * Assuming not more than 2^31-1 events have happened since,
1524 * these subtractions won't have wrapped, so we can check if
1525 * target is in [last_seq, current_seq] by comparing the two.
1527 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1528 events_got = seq - ctx->cq_last_tm_flush;
1529 if (events_got < events_needed)
1532 list_del_init(&req->timeout.list);
1533 io_kill_timeout(req, 0);
1535 ctx->cq_last_tm_flush = seq;
1536 spin_unlock_irq(&ctx->timeout_lock);
1539 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1541 if (ctx->off_timeout_used)
1542 io_flush_timeouts(ctx);
1543 if (ctx->drain_active)
1544 io_queue_deferred(ctx);
1547 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1549 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1550 __io_commit_cqring_flush(ctx);
1551 /* order cqe stores with ring update */
1552 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1555 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1557 struct io_rings *r = ctx->rings;
1559 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1562 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1564 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1567 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1569 struct io_rings *rings = ctx->rings;
1570 unsigned tail, mask = ctx->cq_entries - 1;
1573 * writes to the cq entry need to come after reading head; the
1574 * control dependency is enough as we're using WRITE_ONCE to
1577 if (__io_cqring_events(ctx) == ctx->cq_entries)
1580 tail = ctx->cached_cq_tail++;
1581 return &rings->cqes[tail & mask];
1584 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1586 if (likely(!ctx->cq_ev_fd))
1588 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1590 return !ctx->eventfd_async || io_wq_current_is_worker();
1594 * This should only get called when at least one event has been posted.
1595 * Some applications rely on the eventfd notification count only changing
1596 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1597 * 1:1 relationship between how many times this function is called (and
1598 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1600 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1603 * wake_up_all() may seem excessive, but io_wake_function() and
1604 * io_should_wake() handle the termination of the loop and only
1605 * wake as many waiters as we need to.
1607 if (wq_has_sleeper(&ctx->cq_wait))
1608 wake_up_all(&ctx->cq_wait);
1609 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1610 wake_up(&ctx->sq_data->wait);
1611 if (io_should_trigger_evfd(ctx))
1612 eventfd_signal(ctx->cq_ev_fd, 1);
1613 if (waitqueue_active(&ctx->poll_wait))
1614 wake_up_interruptible(&ctx->poll_wait);
1617 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1619 /* see waitqueue_active() comment */
1622 if (ctx->flags & IORING_SETUP_SQPOLL) {
1623 if (waitqueue_active(&ctx->cq_wait))
1624 wake_up_all(&ctx->cq_wait);
1626 if (io_should_trigger_evfd(ctx))
1627 eventfd_signal(ctx->cq_ev_fd, 1);
1628 if (waitqueue_active(&ctx->poll_wait))
1629 wake_up_interruptible(&ctx->poll_wait);
1632 /* Returns true if there are no backlogged entries after the flush */
1633 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1635 bool all_flushed, posted;
1637 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1641 spin_lock(&ctx->completion_lock);
1642 while (!list_empty(&ctx->cq_overflow_list)) {
1643 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1644 struct io_overflow_cqe *ocqe;
1648 ocqe = list_first_entry(&ctx->cq_overflow_list,
1649 struct io_overflow_cqe, list);
1651 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1653 io_account_cq_overflow(ctx);
1656 list_del(&ocqe->list);
1660 all_flushed = list_empty(&ctx->cq_overflow_list);
1662 clear_bit(0, &ctx->check_cq_overflow);
1663 WRITE_ONCE(ctx->rings->sq_flags,
1664 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1668 io_commit_cqring(ctx);
1669 spin_unlock(&ctx->completion_lock);
1671 io_cqring_ev_posted(ctx);
1675 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1679 if (test_bit(0, &ctx->check_cq_overflow)) {
1680 /* iopoll syncs against uring_lock, not completion_lock */
1681 if (ctx->flags & IORING_SETUP_IOPOLL)
1682 mutex_lock(&ctx->uring_lock);
1683 ret = __io_cqring_overflow_flush(ctx, false);
1684 if (ctx->flags & IORING_SETUP_IOPOLL)
1685 mutex_unlock(&ctx->uring_lock);
1691 /* must to be called somewhat shortly after putting a request */
1692 static inline void io_put_task(struct task_struct *task, int nr)
1694 struct io_uring_task *tctx = task->io_uring;
1696 if (likely(task == current)) {
1697 tctx->cached_refs += nr;
1699 percpu_counter_sub(&tctx->inflight, nr);
1700 if (unlikely(atomic_read(&tctx->in_idle)))
1701 wake_up(&tctx->wait);
1702 put_task_struct_many(task, nr);
1706 static void io_task_refs_refill(struct io_uring_task *tctx)
1708 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1710 percpu_counter_add(&tctx->inflight, refill);
1711 refcount_add(refill, ¤t->usage);
1712 tctx->cached_refs += refill;
1715 static inline void io_get_task_refs(int nr)
1717 struct io_uring_task *tctx = current->io_uring;
1719 tctx->cached_refs -= nr;
1720 if (unlikely(tctx->cached_refs < 0))
1721 io_task_refs_refill(tctx);
1724 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1725 long res, unsigned int cflags)
1727 struct io_overflow_cqe *ocqe;
1729 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1732 * If we're in ring overflow flush mode, or in task cancel mode,
1733 * or cannot allocate an overflow entry, then we need to drop it
1736 io_account_cq_overflow(ctx);
1739 if (list_empty(&ctx->cq_overflow_list)) {
1740 set_bit(0, &ctx->check_cq_overflow);
1741 WRITE_ONCE(ctx->rings->sq_flags,
1742 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1745 ocqe->cqe.user_data = user_data;
1746 ocqe->cqe.res = res;
1747 ocqe->cqe.flags = cflags;
1748 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1752 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1753 long res, unsigned int cflags)
1755 struct io_uring_cqe *cqe;
1757 trace_io_uring_complete(ctx, user_data, res, cflags);
1760 * If we can't get a cq entry, userspace overflowed the
1761 * submission (by quite a lot). Increment the overflow count in
1764 cqe = io_get_cqe(ctx);
1766 WRITE_ONCE(cqe->user_data, user_data);
1767 WRITE_ONCE(cqe->res, res);
1768 WRITE_ONCE(cqe->flags, cflags);
1771 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1774 /* not as hot to bloat with inlining */
1775 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1776 long res, unsigned int cflags)
1778 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1781 static void io_req_complete_post(struct io_kiocb *req, long res,
1782 unsigned int cflags)
1784 struct io_ring_ctx *ctx = req->ctx;
1786 spin_lock(&ctx->completion_lock);
1787 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1789 * If we're the last reference to this request, add to our locked
1792 if (req_ref_put_and_test(req)) {
1793 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1794 if (req->flags & IO_DISARM_MASK)
1795 io_disarm_next(req);
1797 io_req_task_queue(req->link);
1801 io_dismantle_req(req);
1802 io_put_task(req->task, 1);
1803 list_add(&req->inflight_entry, &ctx->locked_free_list);
1804 ctx->locked_free_nr++;
1806 if (!percpu_ref_tryget(&ctx->refs))
1809 io_commit_cqring(ctx);
1810 spin_unlock(&ctx->completion_lock);
1813 io_cqring_ev_posted(ctx);
1814 percpu_ref_put(&ctx->refs);
1818 static inline bool io_req_needs_clean(struct io_kiocb *req)
1820 return req->flags & IO_REQ_CLEAN_FLAGS;
1823 static void io_req_complete_state(struct io_kiocb *req, long res,
1824 unsigned int cflags)
1826 if (io_req_needs_clean(req))
1829 req->compl.cflags = cflags;
1830 req->flags |= REQ_F_COMPLETE_INLINE;
1833 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1834 long res, unsigned cflags)
1836 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1837 io_req_complete_state(req, res, cflags);
1839 io_req_complete_post(req, res, cflags);
1842 static inline void io_req_complete(struct io_kiocb *req, long res)
1844 __io_req_complete(req, 0, res, 0);
1847 static void io_req_complete_failed(struct io_kiocb *req, long res)
1850 io_req_complete_post(req, res, 0);
1853 static void io_req_complete_fail_submit(struct io_kiocb *req)
1856 * We don't submit, fail them all, for that replace hardlinks with
1857 * normal links. Extra REQ_F_LINK is tolerated.
1859 req->flags &= ~REQ_F_HARDLINK;
1860 req->flags |= REQ_F_LINK;
1861 io_req_complete_failed(req, req->result);
1865 * Don't initialise the fields below on every allocation, but do that in
1866 * advance and keep them valid across allocations.
1868 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1872 req->async_data = NULL;
1873 /* not necessary, but safer to zero */
1877 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1878 struct io_submit_state *state)
1880 spin_lock(&ctx->completion_lock);
1881 list_splice_init(&ctx->locked_free_list, &state->free_list);
1882 ctx->locked_free_nr = 0;
1883 spin_unlock(&ctx->completion_lock);
1886 /* Returns true IFF there are requests in the cache */
1887 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1889 struct io_submit_state *state = &ctx->submit_state;
1893 * If we have more than a batch's worth of requests in our IRQ side
1894 * locked cache, grab the lock and move them over to our submission
1897 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1898 io_flush_cached_locked_reqs(ctx, state);
1900 nr = state->free_reqs;
1901 while (!list_empty(&state->free_list)) {
1902 struct io_kiocb *req = list_first_entry(&state->free_list,
1903 struct io_kiocb, inflight_entry);
1905 list_del(&req->inflight_entry);
1906 state->reqs[nr++] = req;
1907 if (nr == ARRAY_SIZE(state->reqs))
1911 state->free_reqs = nr;
1916 * A request might get retired back into the request caches even before opcode
1917 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1918 * Because of that, io_alloc_req() should be called only under ->uring_lock
1919 * and with extra caution to not get a request that is still worked on.
1921 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1922 __must_hold(&ctx->uring_lock)
1924 struct io_submit_state *state = &ctx->submit_state;
1925 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1928 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1930 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1933 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1937 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1938 * retry single alloc to be on the safe side.
1940 if (unlikely(ret <= 0)) {
1941 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1942 if (!state->reqs[0])
1947 for (i = 0; i < ret; i++)
1948 io_preinit_req(state->reqs[i], ctx);
1949 state->free_reqs = ret;
1952 return state->reqs[state->free_reqs];
1955 static inline void io_put_file(struct file *file)
1961 static void io_dismantle_req(struct io_kiocb *req)
1963 unsigned int flags = req->flags;
1965 if (io_req_needs_clean(req))
1967 if (!(flags & REQ_F_FIXED_FILE))
1968 io_put_file(req->file);
1969 if (req->fixed_rsrc_refs)
1970 percpu_ref_put(req->fixed_rsrc_refs);
1971 if (req->async_data) {
1972 kfree(req->async_data);
1973 req->async_data = NULL;
1977 static void __io_free_req(struct io_kiocb *req)
1979 struct io_ring_ctx *ctx = req->ctx;
1981 io_dismantle_req(req);
1982 io_put_task(req->task, 1);
1984 spin_lock(&ctx->completion_lock);
1985 list_add(&req->inflight_entry, &ctx->locked_free_list);
1986 ctx->locked_free_nr++;
1987 spin_unlock(&ctx->completion_lock);
1989 percpu_ref_put(&ctx->refs);
1992 static inline void io_remove_next_linked(struct io_kiocb *req)
1994 struct io_kiocb *nxt = req->link;
1996 req->link = nxt->link;
2000 static bool io_kill_linked_timeout(struct io_kiocb *req)
2001 __must_hold(&req->ctx->completion_lock)
2002 __must_hold(&req->ctx->timeout_lock)
2004 struct io_kiocb *link = req->link;
2006 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2007 struct io_timeout_data *io = link->async_data;
2009 io_remove_next_linked(req);
2010 link->timeout.head = NULL;
2011 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2012 list_del(&link->timeout.list);
2013 io_cqring_fill_event(link->ctx, link->user_data,
2015 io_put_req_deferred(link);
2022 static void io_fail_links(struct io_kiocb *req)
2023 __must_hold(&req->ctx->completion_lock)
2025 struct io_kiocb *nxt, *link = req->link;
2029 long res = -ECANCELED;
2031 if (link->flags & REQ_F_FAIL)
2037 trace_io_uring_fail_link(req, link);
2038 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2039 io_put_req_deferred(link);
2044 static bool io_disarm_next(struct io_kiocb *req)
2045 __must_hold(&req->ctx->completion_lock)
2047 bool posted = false;
2049 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2050 struct io_kiocb *link = req->link;
2052 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2053 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2054 io_remove_next_linked(req);
2055 io_cqring_fill_event(link->ctx, link->user_data,
2057 io_put_req_deferred(link);
2060 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2061 struct io_ring_ctx *ctx = req->ctx;
2063 spin_lock_irq(&ctx->timeout_lock);
2064 posted = io_kill_linked_timeout(req);
2065 spin_unlock_irq(&ctx->timeout_lock);
2067 if (unlikely((req->flags & REQ_F_FAIL) &&
2068 !(req->flags & REQ_F_HARDLINK))) {
2069 posted |= (req->link != NULL);
2075 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2077 struct io_kiocb *nxt;
2080 * If LINK is set, we have dependent requests in this chain. If we
2081 * didn't fail this request, queue the first one up, moving any other
2082 * dependencies to the next request. In case of failure, fail the rest
2085 if (req->flags & IO_DISARM_MASK) {
2086 struct io_ring_ctx *ctx = req->ctx;
2089 spin_lock(&ctx->completion_lock);
2090 posted = io_disarm_next(req);
2092 io_commit_cqring(req->ctx);
2093 spin_unlock(&ctx->completion_lock);
2095 io_cqring_ev_posted(ctx);
2102 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2104 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2106 return __io_req_find_next(req);
2109 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2114 if (ctx->submit_state.compl_nr)
2115 io_submit_flush_completions(ctx);
2116 mutex_unlock(&ctx->uring_lock);
2119 percpu_ref_put(&ctx->refs);
2122 static void tctx_task_work(struct callback_head *cb)
2124 bool locked = false;
2125 struct io_ring_ctx *ctx = NULL;
2126 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2130 struct io_wq_work_node *node;
2132 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2133 io_submit_flush_completions(ctx);
2135 spin_lock_irq(&tctx->task_lock);
2136 node = tctx->task_list.first;
2137 INIT_WQ_LIST(&tctx->task_list);
2139 tctx->task_running = false;
2140 spin_unlock_irq(&tctx->task_lock);
2145 struct io_wq_work_node *next = node->next;
2146 struct io_kiocb *req = container_of(node, struct io_kiocb,
2149 if (req->ctx != ctx) {
2150 ctx_flush_and_put(ctx, &locked);
2152 /* if not contended, grab and improve batching */
2153 locked = mutex_trylock(&ctx->uring_lock);
2154 percpu_ref_get(&ctx->refs);
2156 req->io_task_work.func(req, &locked);
2163 ctx_flush_and_put(ctx, &locked);
2166 static void io_req_task_work_add(struct io_kiocb *req)
2168 struct task_struct *tsk = req->task;
2169 struct io_uring_task *tctx = tsk->io_uring;
2170 enum task_work_notify_mode notify;
2171 struct io_wq_work_node *node;
2172 unsigned long flags;
2175 WARN_ON_ONCE(!tctx);
2177 spin_lock_irqsave(&tctx->task_lock, flags);
2178 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2179 running = tctx->task_running;
2181 tctx->task_running = true;
2182 spin_unlock_irqrestore(&tctx->task_lock, flags);
2184 /* task_work already pending, we're done */
2189 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2190 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2191 * processing task_work. There's no reliable way to tell if TWA_RESUME
2194 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2195 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2196 wake_up_process(tsk);
2200 spin_lock_irqsave(&tctx->task_lock, flags);
2201 tctx->task_running = false;
2202 node = tctx->task_list.first;
2203 INIT_WQ_LIST(&tctx->task_list);
2204 spin_unlock_irqrestore(&tctx->task_lock, flags);
2207 req = container_of(node, struct io_kiocb, io_task_work.node);
2209 if (llist_add(&req->io_task_work.fallback_node,
2210 &req->ctx->fallback_llist))
2211 schedule_delayed_work(&req->ctx->fallback_work, 1);
2215 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2217 struct io_ring_ctx *ctx = req->ctx;
2219 /* not needed for normal modes, but SQPOLL depends on it */
2220 io_tw_lock(ctx, locked);
2221 io_req_complete_failed(req, req->result);
2224 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2226 struct io_ring_ctx *ctx = req->ctx;
2228 io_tw_lock(ctx, locked);
2229 /* req->task == current here, checking PF_EXITING is safe */
2230 if (likely(!(req->task->flags & PF_EXITING)))
2231 __io_queue_sqe(req);
2233 io_req_complete_failed(req, -EFAULT);
2236 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2239 req->io_task_work.func = io_req_task_cancel;
2240 io_req_task_work_add(req);
2243 static void io_req_task_queue(struct io_kiocb *req)
2245 req->io_task_work.func = io_req_task_submit;
2246 io_req_task_work_add(req);
2249 static void io_req_task_queue_reissue(struct io_kiocb *req)
2251 req->io_task_work.func = io_queue_async_work;
2252 io_req_task_work_add(req);
2255 static inline void io_queue_next(struct io_kiocb *req)
2257 struct io_kiocb *nxt = io_req_find_next(req);
2260 io_req_task_queue(nxt);
2263 static void io_free_req(struct io_kiocb *req)
2269 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2275 struct task_struct *task;
2280 static inline void io_init_req_batch(struct req_batch *rb)
2287 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2288 struct req_batch *rb)
2291 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2293 io_put_task(rb->task, rb->task_refs);
2296 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2297 struct io_submit_state *state)
2300 io_dismantle_req(req);
2302 if (req->task != rb->task) {
2304 io_put_task(rb->task, rb->task_refs);
2305 rb->task = req->task;
2311 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2312 state->reqs[state->free_reqs++] = req;
2314 list_add(&req->inflight_entry, &state->free_list);
2317 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2318 __must_hold(&ctx->uring_lock)
2320 struct io_submit_state *state = &ctx->submit_state;
2321 int i, nr = state->compl_nr;
2322 struct req_batch rb;
2324 spin_lock(&ctx->completion_lock);
2325 for (i = 0; i < nr; i++) {
2326 struct io_kiocb *req = state->compl_reqs[i];
2328 __io_cqring_fill_event(ctx, req->user_data, req->result,
2331 io_commit_cqring(ctx);
2332 spin_unlock(&ctx->completion_lock);
2333 io_cqring_ev_posted(ctx);
2335 io_init_req_batch(&rb);
2336 for (i = 0; i < nr; i++) {
2337 struct io_kiocb *req = state->compl_reqs[i];
2339 if (req_ref_put_and_test(req))
2340 io_req_free_batch(&rb, req, &ctx->submit_state);
2343 io_req_free_batch_finish(ctx, &rb);
2344 state->compl_nr = 0;
2348 * Drop reference to request, return next in chain (if there is one) if this
2349 * was the last reference to this request.
2351 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2353 struct io_kiocb *nxt = NULL;
2355 if (req_ref_put_and_test(req)) {
2356 nxt = io_req_find_next(req);
2362 static inline void io_put_req(struct io_kiocb *req)
2364 if (req_ref_put_and_test(req))
2368 static inline void io_put_req_deferred(struct io_kiocb *req)
2370 if (req_ref_put_and_test(req)) {
2371 req->io_task_work.func = io_free_req_work;
2372 io_req_task_work_add(req);
2376 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2378 /* See comment at the top of this file */
2380 return __io_cqring_events(ctx);
2383 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2385 struct io_rings *rings = ctx->rings;
2387 /* make sure SQ entry isn't read before tail */
2388 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2391 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2393 unsigned int cflags;
2395 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2396 cflags |= IORING_CQE_F_BUFFER;
2397 req->flags &= ~REQ_F_BUFFER_SELECTED;
2402 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2404 struct io_buffer *kbuf;
2406 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2408 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2409 return io_put_kbuf(req, kbuf);
2412 static inline bool io_run_task_work(void)
2414 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2415 __set_current_state(TASK_RUNNING);
2416 tracehook_notify_signal();
2424 * Find and free completed poll iocbs
2426 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2427 struct list_head *done)
2429 struct req_batch rb;
2430 struct io_kiocb *req;
2432 /* order with ->result store in io_complete_rw_iopoll() */
2435 io_init_req_batch(&rb);
2436 while (!list_empty(done)) {
2437 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2438 list_del(&req->inflight_entry);
2440 __io_cqring_fill_event(ctx, req->user_data, req->result,
2441 io_put_rw_kbuf(req));
2444 if (req_ref_put_and_test(req))
2445 io_req_free_batch(&rb, req, &ctx->submit_state);
2448 io_commit_cqring(ctx);
2449 io_cqring_ev_posted_iopoll(ctx);
2450 io_req_free_batch_finish(ctx, &rb);
2453 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2456 struct io_kiocb *req, *tmp;
2461 * Only spin for completions if we don't have multiple devices hanging
2462 * off our complete list, and we're under the requested amount.
2464 spin = !ctx->poll_multi_queue && *nr_events < min;
2466 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2467 struct kiocb *kiocb = &req->rw.kiocb;
2471 * Move completed and retryable entries to our local lists.
2472 * If we find a request that requires polling, break out
2473 * and complete those lists first, if we have entries there.
2475 if (READ_ONCE(req->iopoll_completed)) {
2476 list_move_tail(&req->inflight_entry, &done);
2479 if (!list_empty(&done))
2482 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2483 if (unlikely(ret < 0))
2488 /* iopoll may have completed current req */
2489 if (READ_ONCE(req->iopoll_completed))
2490 list_move_tail(&req->inflight_entry, &done);
2493 if (!list_empty(&done))
2494 io_iopoll_complete(ctx, nr_events, &done);
2500 * We can't just wait for polled events to come to us, we have to actively
2501 * find and complete them.
2503 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2505 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2508 mutex_lock(&ctx->uring_lock);
2509 while (!list_empty(&ctx->iopoll_list)) {
2510 unsigned int nr_events = 0;
2512 io_do_iopoll(ctx, &nr_events, 0);
2514 /* let it sleep and repeat later if can't complete a request */
2518 * Ensure we allow local-to-the-cpu processing to take place,
2519 * in this case we need to ensure that we reap all events.
2520 * Also let task_work, etc. to progress by releasing the mutex
2522 if (need_resched()) {
2523 mutex_unlock(&ctx->uring_lock);
2525 mutex_lock(&ctx->uring_lock);
2528 mutex_unlock(&ctx->uring_lock);
2531 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2533 unsigned int nr_events = 0;
2537 * We disallow the app entering submit/complete with polling, but we
2538 * still need to lock the ring to prevent racing with polled issue
2539 * that got punted to a workqueue.
2541 mutex_lock(&ctx->uring_lock);
2543 * Don't enter poll loop if we already have events pending.
2544 * If we do, we can potentially be spinning for commands that
2545 * already triggered a CQE (eg in error).
2547 if (test_bit(0, &ctx->check_cq_overflow))
2548 __io_cqring_overflow_flush(ctx, false);
2549 if (io_cqring_events(ctx))
2553 * If a submit got punted to a workqueue, we can have the
2554 * application entering polling for a command before it gets
2555 * issued. That app will hold the uring_lock for the duration
2556 * of the poll right here, so we need to take a breather every
2557 * now and then to ensure that the issue has a chance to add
2558 * the poll to the issued list. Otherwise we can spin here
2559 * forever, while the workqueue is stuck trying to acquire the
2562 if (list_empty(&ctx->iopoll_list)) {
2563 u32 tail = ctx->cached_cq_tail;
2565 mutex_unlock(&ctx->uring_lock);
2567 mutex_lock(&ctx->uring_lock);
2569 /* some requests don't go through iopoll_list */
2570 if (tail != ctx->cached_cq_tail ||
2571 list_empty(&ctx->iopoll_list))
2574 ret = io_do_iopoll(ctx, &nr_events, min);
2575 } while (!ret && nr_events < min && !need_resched());
2577 mutex_unlock(&ctx->uring_lock);
2581 static void kiocb_end_write(struct io_kiocb *req)
2584 * Tell lockdep we inherited freeze protection from submission
2587 if (req->flags & REQ_F_ISREG) {
2588 struct super_block *sb = file_inode(req->file)->i_sb;
2590 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2596 static bool io_resubmit_prep(struct io_kiocb *req)
2598 struct io_async_rw *rw = req->async_data;
2601 return !io_req_prep_async(req);
2602 iov_iter_restore(&rw->iter, &rw->iter_state);
2606 static bool io_rw_should_reissue(struct io_kiocb *req)
2608 umode_t mode = file_inode(req->file)->i_mode;
2609 struct io_ring_ctx *ctx = req->ctx;
2611 if (!S_ISBLK(mode) && !S_ISREG(mode))
2613 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2614 !(ctx->flags & IORING_SETUP_IOPOLL)))
2617 * If ref is dying, we might be running poll reap from the exit work.
2618 * Don't attempt to reissue from that path, just let it fail with
2621 if (percpu_ref_is_dying(&ctx->refs))
2624 * Play it safe and assume not safe to re-import and reissue if we're
2625 * not in the original thread group (or in task context).
2627 if (!same_thread_group(req->task, current) || !in_task())
2632 static bool io_resubmit_prep(struct io_kiocb *req)
2636 static bool io_rw_should_reissue(struct io_kiocb *req)
2642 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2644 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2645 kiocb_end_write(req);
2646 if (res != req->result) {
2647 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2648 io_rw_should_reissue(req)) {
2649 req->flags |= REQ_F_REISSUE;
2658 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2660 unsigned int cflags = io_put_rw_kbuf(req);
2661 long res = req->result;
2664 struct io_ring_ctx *ctx = req->ctx;
2665 struct io_submit_state *state = &ctx->submit_state;
2667 io_req_complete_state(req, res, cflags);
2668 state->compl_reqs[state->compl_nr++] = req;
2669 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2670 io_submit_flush_completions(ctx);
2672 io_req_complete_post(req, res, cflags);
2676 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2677 unsigned int issue_flags)
2679 if (__io_complete_rw_common(req, res))
2681 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2684 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2686 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2688 if (__io_complete_rw_common(req, res))
2691 req->io_task_work.func = io_req_task_complete;
2692 io_req_task_work_add(req);
2695 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2697 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2699 if (kiocb->ki_flags & IOCB_WRITE)
2700 kiocb_end_write(req);
2701 if (unlikely(res != req->result)) {
2702 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2703 req->flags |= REQ_F_REISSUE;
2708 WRITE_ONCE(req->result, res);
2709 /* order with io_iopoll_complete() checking ->result */
2711 WRITE_ONCE(req->iopoll_completed, 1);
2715 * After the iocb has been issued, it's safe to be found on the poll list.
2716 * Adding the kiocb to the list AFTER submission ensures that we don't
2717 * find it from a io_do_iopoll() thread before the issuer is done
2718 * accessing the kiocb cookie.
2720 static void io_iopoll_req_issued(struct io_kiocb *req)
2722 struct io_ring_ctx *ctx = req->ctx;
2723 const bool in_async = io_wq_current_is_worker();
2725 /* workqueue context doesn't hold uring_lock, grab it now */
2726 if (unlikely(in_async))
2727 mutex_lock(&ctx->uring_lock);
2730 * Track whether we have multiple files in our lists. This will impact
2731 * how we do polling eventually, not spinning if we're on potentially
2732 * different devices.
2734 if (list_empty(&ctx->iopoll_list)) {
2735 ctx->poll_multi_queue = false;
2736 } else if (!ctx->poll_multi_queue) {
2737 struct io_kiocb *list_req;
2738 unsigned int queue_num0, queue_num1;
2740 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2743 if (list_req->file != req->file) {
2744 ctx->poll_multi_queue = true;
2746 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2747 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2748 if (queue_num0 != queue_num1)
2749 ctx->poll_multi_queue = true;
2754 * For fast devices, IO may have already completed. If it has, add
2755 * it to the front so we find it first.
2757 if (READ_ONCE(req->iopoll_completed))
2758 list_add(&req->inflight_entry, &ctx->iopoll_list);
2760 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2762 if (unlikely(in_async)) {
2764 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2765 * in sq thread task context or in io worker task context. If
2766 * current task context is sq thread, we don't need to check
2767 * whether should wake up sq thread.
2769 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2770 wq_has_sleeper(&ctx->sq_data->wait))
2771 wake_up(&ctx->sq_data->wait);
2773 mutex_unlock(&ctx->uring_lock);
2777 static bool io_bdev_nowait(struct block_device *bdev)
2779 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2783 * If we tracked the file through the SCM inflight mechanism, we could support
2784 * any file. For now, just ensure that anything potentially problematic is done
2787 static bool __io_file_supports_nowait(struct file *file, int rw)
2789 umode_t mode = file_inode(file)->i_mode;
2791 if (S_ISBLK(mode)) {
2792 if (IS_ENABLED(CONFIG_BLOCK) &&
2793 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2799 if (S_ISREG(mode)) {
2800 if (IS_ENABLED(CONFIG_BLOCK) &&
2801 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2802 file->f_op != &io_uring_fops)
2807 /* any ->read/write should understand O_NONBLOCK */
2808 if (file->f_flags & O_NONBLOCK)
2811 if (!(file->f_mode & FMODE_NOWAIT))
2815 return file->f_op->read_iter != NULL;
2817 return file->f_op->write_iter != NULL;
2820 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2822 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2824 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2827 return __io_file_supports_nowait(req->file, rw);
2830 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2833 struct io_ring_ctx *ctx = req->ctx;
2834 struct kiocb *kiocb = &req->rw.kiocb;
2835 struct file *file = req->file;
2839 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2840 req->flags |= REQ_F_ISREG;
2842 kiocb->ki_pos = READ_ONCE(sqe->off);
2843 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2844 req->flags |= REQ_F_CUR_POS;
2845 kiocb->ki_pos = file->f_pos;
2847 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2848 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2849 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2854 * If the file is marked O_NONBLOCK, still allow retry for it if it
2855 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2856 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2858 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2859 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2860 req->flags |= REQ_F_NOWAIT;
2862 ioprio = READ_ONCE(sqe->ioprio);
2864 ret = ioprio_check_cap(ioprio);
2868 kiocb->ki_ioprio = ioprio;
2870 kiocb->ki_ioprio = get_current_ioprio();
2872 if (ctx->flags & IORING_SETUP_IOPOLL) {
2873 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2874 !kiocb->ki_filp->f_op->iopoll)
2877 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2878 kiocb->ki_complete = io_complete_rw_iopoll;
2879 req->iopoll_completed = 0;
2881 if (kiocb->ki_flags & IOCB_HIPRI)
2883 kiocb->ki_complete = io_complete_rw;
2886 if (req->opcode == IORING_OP_READ_FIXED ||
2887 req->opcode == IORING_OP_WRITE_FIXED) {
2889 io_req_set_rsrc_node(req);
2892 req->rw.addr = READ_ONCE(sqe->addr);
2893 req->rw.len = READ_ONCE(sqe->len);
2894 req->buf_index = READ_ONCE(sqe->buf_index);
2898 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2904 case -ERESTARTNOINTR:
2905 case -ERESTARTNOHAND:
2906 case -ERESTART_RESTARTBLOCK:
2908 * We can't just restart the syscall, since previously
2909 * submitted sqes may already be in progress. Just fail this
2915 kiocb->ki_complete(kiocb, ret, 0);
2919 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2920 unsigned int issue_flags)
2922 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2923 struct io_async_rw *io = req->async_data;
2925 /* add previously done IO, if any */
2926 if (io && io->bytes_done > 0) {
2928 ret = io->bytes_done;
2930 ret += io->bytes_done;
2933 if (req->flags & REQ_F_CUR_POS)
2934 req->file->f_pos = kiocb->ki_pos;
2935 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2936 __io_complete_rw(req, ret, 0, issue_flags);
2938 io_rw_done(kiocb, ret);
2940 if (req->flags & REQ_F_REISSUE) {
2941 req->flags &= ~REQ_F_REISSUE;
2942 if (io_resubmit_prep(req)) {
2943 io_req_task_queue_reissue(req);
2945 unsigned int cflags = io_put_rw_kbuf(req);
2946 struct io_ring_ctx *ctx = req->ctx;
2949 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
2950 mutex_lock(&ctx->uring_lock);
2951 __io_req_complete(req, issue_flags, ret, cflags);
2952 mutex_unlock(&ctx->uring_lock);
2954 __io_req_complete(req, issue_flags, ret, cflags);
2960 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2961 struct io_mapped_ubuf *imu)
2963 size_t len = req->rw.len;
2964 u64 buf_end, buf_addr = req->rw.addr;
2967 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2969 /* not inside the mapped region */
2970 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2974 * May not be a start of buffer, set size appropriately
2975 * and advance us to the beginning.
2977 offset = buf_addr - imu->ubuf;
2978 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2982 * Don't use iov_iter_advance() here, as it's really slow for
2983 * using the latter parts of a big fixed buffer - it iterates
2984 * over each segment manually. We can cheat a bit here, because
2987 * 1) it's a BVEC iter, we set it up
2988 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2989 * first and last bvec
2991 * So just find our index, and adjust the iterator afterwards.
2992 * If the offset is within the first bvec (or the whole first
2993 * bvec, just use iov_iter_advance(). This makes it easier
2994 * since we can just skip the first segment, which may not
2995 * be PAGE_SIZE aligned.
2997 const struct bio_vec *bvec = imu->bvec;
2999 if (offset <= bvec->bv_len) {
3000 iov_iter_advance(iter, offset);
3002 unsigned long seg_skip;
3004 /* skip first vec */
3005 offset -= bvec->bv_len;
3006 seg_skip = 1 + (offset >> PAGE_SHIFT);
3008 iter->bvec = bvec + seg_skip;
3009 iter->nr_segs -= seg_skip;
3010 iter->count -= bvec->bv_len + offset;
3011 iter->iov_offset = offset & ~PAGE_MASK;
3018 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3020 struct io_ring_ctx *ctx = req->ctx;
3021 struct io_mapped_ubuf *imu = req->imu;
3022 u16 index, buf_index = req->buf_index;
3025 if (unlikely(buf_index >= ctx->nr_user_bufs))
3027 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3028 imu = READ_ONCE(ctx->user_bufs[index]);
3031 return __io_import_fixed(req, rw, iter, imu);
3034 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3037 mutex_unlock(&ctx->uring_lock);
3040 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3043 * "Normal" inline submissions always hold the uring_lock, since we
3044 * grab it from the system call. Same is true for the SQPOLL offload.
3045 * The only exception is when we've detached the request and issue it
3046 * from an async worker thread, grab the lock for that case.
3049 mutex_lock(&ctx->uring_lock);
3052 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3053 int bgid, struct io_buffer *kbuf,
3056 struct io_buffer *head;
3058 if (req->flags & REQ_F_BUFFER_SELECTED)
3061 io_ring_submit_lock(req->ctx, needs_lock);
3063 lockdep_assert_held(&req->ctx->uring_lock);
3065 head = xa_load(&req->ctx->io_buffers, bgid);
3067 if (!list_empty(&head->list)) {
3068 kbuf = list_last_entry(&head->list, struct io_buffer,
3070 list_del(&kbuf->list);
3073 xa_erase(&req->ctx->io_buffers, bgid);
3075 if (*len > kbuf->len)
3078 kbuf = ERR_PTR(-ENOBUFS);
3081 io_ring_submit_unlock(req->ctx, needs_lock);
3086 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3089 struct io_buffer *kbuf;
3092 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3093 bgid = req->buf_index;
3094 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3097 req->rw.addr = (u64) (unsigned long) kbuf;
3098 req->flags |= REQ_F_BUFFER_SELECTED;
3099 return u64_to_user_ptr(kbuf->addr);
3102 #ifdef CONFIG_COMPAT
3103 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3106 struct compat_iovec __user *uiov;
3107 compat_ssize_t clen;
3111 uiov = u64_to_user_ptr(req->rw.addr);
3112 if (!access_ok(uiov, sizeof(*uiov)))
3114 if (__get_user(clen, &uiov->iov_len))
3120 buf = io_rw_buffer_select(req, &len, needs_lock);
3122 return PTR_ERR(buf);
3123 iov[0].iov_base = buf;
3124 iov[0].iov_len = (compat_size_t) len;
3129 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3132 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3136 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3139 len = iov[0].iov_len;
3142 buf = io_rw_buffer_select(req, &len, needs_lock);
3144 return PTR_ERR(buf);
3145 iov[0].iov_base = buf;
3146 iov[0].iov_len = len;
3150 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3153 if (req->flags & REQ_F_BUFFER_SELECTED) {
3154 struct io_buffer *kbuf;
3156 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3157 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3158 iov[0].iov_len = kbuf->len;
3161 if (req->rw.len != 1)
3164 #ifdef CONFIG_COMPAT
3165 if (req->ctx->compat)
3166 return io_compat_import(req, iov, needs_lock);
3169 return __io_iov_buffer_select(req, iov, needs_lock);
3172 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3173 struct iov_iter *iter, bool needs_lock)
3175 void __user *buf = u64_to_user_ptr(req->rw.addr);
3176 size_t sqe_len = req->rw.len;
3177 u8 opcode = req->opcode;
3180 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3182 return io_import_fixed(req, rw, iter);
3185 /* buffer index only valid with fixed read/write, or buffer select */
3186 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3189 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3190 if (req->flags & REQ_F_BUFFER_SELECT) {
3191 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3193 return PTR_ERR(buf);
3194 req->rw.len = sqe_len;
3197 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3202 if (req->flags & REQ_F_BUFFER_SELECT) {
3203 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3205 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3210 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3214 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3216 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3220 * For files that don't have ->read_iter() and ->write_iter(), handle them
3221 * by looping over ->read() or ->write() manually.
3223 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3225 struct kiocb *kiocb = &req->rw.kiocb;
3226 struct file *file = req->file;
3230 * Don't support polled IO through this interface, and we can't
3231 * support non-blocking either. For the latter, this just causes
3232 * the kiocb to be handled from an async context.
3234 if (kiocb->ki_flags & IOCB_HIPRI)
3236 if (kiocb->ki_flags & IOCB_NOWAIT)
3239 while (iov_iter_count(iter)) {
3243 if (!iov_iter_is_bvec(iter)) {
3244 iovec = iov_iter_iovec(iter);
3246 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3247 iovec.iov_len = req->rw.len;
3251 nr = file->f_op->read(file, iovec.iov_base,
3252 iovec.iov_len, io_kiocb_ppos(kiocb));
3254 nr = file->f_op->write(file, iovec.iov_base,
3255 iovec.iov_len, io_kiocb_ppos(kiocb));
3263 if (!iov_iter_is_bvec(iter)) {
3264 iov_iter_advance(iter, nr);
3270 if (nr != iovec.iov_len)
3277 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3278 const struct iovec *fast_iov, struct iov_iter *iter)
3280 struct io_async_rw *rw = req->async_data;
3282 memcpy(&rw->iter, iter, sizeof(*iter));
3283 rw->free_iovec = iovec;
3285 /* can only be fixed buffers, no need to do anything */
3286 if (iov_iter_is_bvec(iter))
3289 unsigned iov_off = 0;
3291 rw->iter.iov = rw->fast_iov;
3292 if (iter->iov != fast_iov) {
3293 iov_off = iter->iov - fast_iov;
3294 rw->iter.iov += iov_off;
3296 if (rw->fast_iov != fast_iov)
3297 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3298 sizeof(struct iovec) * iter->nr_segs);
3300 req->flags |= REQ_F_NEED_CLEANUP;
3304 static inline int io_alloc_async_data(struct io_kiocb *req)
3306 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3307 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3308 return req->async_data == NULL;
3311 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3312 const struct iovec *fast_iov,
3313 struct iov_iter *iter, bool force)
3315 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3317 if (!req->async_data) {
3318 struct io_async_rw *iorw;
3320 if (io_alloc_async_data(req)) {
3325 io_req_map_rw(req, iovec, fast_iov, iter);
3326 iorw = req->async_data;
3327 /* we've copied and mapped the iter, ensure state is saved */
3328 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3333 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3335 struct io_async_rw *iorw = req->async_data;
3336 struct iovec *iov = iorw->fast_iov;
3339 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3340 if (unlikely(ret < 0))
3343 iorw->bytes_done = 0;
3344 iorw->free_iovec = iov;
3346 req->flags |= REQ_F_NEED_CLEANUP;
3347 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3351 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3353 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3355 return io_prep_rw(req, sqe, READ);
3359 * This is our waitqueue callback handler, registered through lock_page_async()
3360 * when we initially tried to do the IO with the iocb armed our waitqueue.
3361 * This gets called when the page is unlocked, and we generally expect that to
3362 * happen when the page IO is completed and the page is now uptodate. This will
3363 * queue a task_work based retry of the operation, attempting to copy the data
3364 * again. If the latter fails because the page was NOT uptodate, then we will
3365 * do a thread based blocking retry of the operation. That's the unexpected
3368 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3369 int sync, void *arg)
3371 struct wait_page_queue *wpq;
3372 struct io_kiocb *req = wait->private;
3373 struct wait_page_key *key = arg;
3375 wpq = container_of(wait, struct wait_page_queue, wait);
3377 if (!wake_page_match(wpq, key))
3380 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3381 list_del_init(&wait->entry);
3382 io_req_task_queue(req);
3387 * This controls whether a given IO request should be armed for async page
3388 * based retry. If we return false here, the request is handed to the async
3389 * worker threads for retry. If we're doing buffered reads on a regular file,
3390 * we prepare a private wait_page_queue entry and retry the operation. This
3391 * will either succeed because the page is now uptodate and unlocked, or it
3392 * will register a callback when the page is unlocked at IO completion. Through
3393 * that callback, io_uring uses task_work to setup a retry of the operation.
3394 * That retry will attempt the buffered read again. The retry will generally
3395 * succeed, or in rare cases where it fails, we then fall back to using the
3396 * async worker threads for a blocking retry.
3398 static bool io_rw_should_retry(struct io_kiocb *req)
3400 struct io_async_rw *rw = req->async_data;
3401 struct wait_page_queue *wait = &rw->wpq;
3402 struct kiocb *kiocb = &req->rw.kiocb;
3404 /* never retry for NOWAIT, we just complete with -EAGAIN */
3405 if (req->flags & REQ_F_NOWAIT)
3408 /* Only for buffered IO */
3409 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3413 * just use poll if we can, and don't attempt if the fs doesn't
3414 * support callback based unlocks
3416 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3419 wait->wait.func = io_async_buf_func;
3420 wait->wait.private = req;
3421 wait->wait.flags = 0;
3422 INIT_LIST_HEAD(&wait->wait.entry);
3423 kiocb->ki_flags |= IOCB_WAITQ;
3424 kiocb->ki_flags &= ~IOCB_NOWAIT;
3425 kiocb->ki_waitq = wait;
3429 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3431 if (req->file->f_op->read_iter)
3432 return call_read_iter(req->file, &req->rw.kiocb, iter);
3433 else if (req->file->f_op->read)
3434 return loop_rw_iter(READ, req, iter);
3439 static bool need_read_all(struct io_kiocb *req)
3441 return req->flags & REQ_F_ISREG ||
3442 S_ISBLK(file_inode(req->file)->i_mode);
3445 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3447 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3448 struct kiocb *kiocb = &req->rw.kiocb;
3449 struct iov_iter __iter, *iter = &__iter;
3450 struct io_async_rw *rw = req->async_data;
3451 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3452 struct iov_iter_state __state, *state;
3457 state = &rw->iter_state;
3459 * We come here from an earlier attempt, restore our state to
3460 * match in case it doesn't. It's cheap enough that we don't
3461 * need to make this conditional.
3463 iov_iter_restore(iter, state);
3466 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3470 iov_iter_save_state(iter, state);
3472 req->result = iov_iter_count(iter);
3474 /* Ensure we clear previously set non-block flag */
3475 if (!force_nonblock)
3476 kiocb->ki_flags &= ~IOCB_NOWAIT;
3478 kiocb->ki_flags |= IOCB_NOWAIT;
3480 /* If the file doesn't support async, just async punt */
3481 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3482 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3483 return ret ?: -EAGAIN;
3486 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3487 if (unlikely(ret)) {
3492 ret = io_iter_do_read(req, iter);
3494 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3495 req->flags &= ~REQ_F_REISSUE;
3496 /* IOPOLL retry should happen for io-wq threads */
3497 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3499 /* no retry on NONBLOCK nor RWF_NOWAIT */
3500 if (req->flags & REQ_F_NOWAIT)
3503 } else if (ret == -EIOCBQUEUED) {
3505 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3506 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3507 /* read all, failed, already did sync or don't want to retry */
3512 * Don't depend on the iter state matching what was consumed, or being
3513 * untouched in case of error. Restore it and we'll advance it
3514 * manually if we need to.
3516 iov_iter_restore(iter, state);
3518 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3523 rw = req->async_data;
3525 * Now use our persistent iterator and state, if we aren't already.
3526 * We've restored and mapped the iter to match.
3528 if (iter != &rw->iter) {
3530 state = &rw->iter_state;
3535 * We end up here because of a partial read, either from
3536 * above or inside this loop. Advance the iter by the bytes
3537 * that were consumed.
3539 iov_iter_advance(iter, ret);
3540 if (!iov_iter_count(iter))
3542 rw->bytes_done += ret;
3543 iov_iter_save_state(iter, state);
3545 /* if we can retry, do so with the callbacks armed */
3546 if (!io_rw_should_retry(req)) {
3547 kiocb->ki_flags &= ~IOCB_WAITQ;
3552 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3553 * we get -EIOCBQUEUED, then we'll get a notification when the
3554 * desired page gets unlocked. We can also get a partial read
3555 * here, and if we do, then just retry at the new offset.
3557 ret = io_iter_do_read(req, iter);
3558 if (ret == -EIOCBQUEUED)
3560 /* we got some bytes, but not all. retry. */
3561 kiocb->ki_flags &= ~IOCB_WAITQ;
3562 iov_iter_restore(iter, state);
3565 kiocb_done(kiocb, ret, issue_flags);
3567 /* it's faster to check here then delegate to kfree */
3573 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3575 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3577 return io_prep_rw(req, sqe, WRITE);
3580 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3582 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3583 struct kiocb *kiocb = &req->rw.kiocb;
3584 struct iov_iter __iter, *iter = &__iter;
3585 struct io_async_rw *rw = req->async_data;
3586 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3587 struct iov_iter_state __state, *state;
3592 state = &rw->iter_state;
3593 iov_iter_restore(iter, state);
3596 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3600 iov_iter_save_state(iter, state);
3602 req->result = iov_iter_count(iter);
3604 /* Ensure we clear previously set non-block flag */
3605 if (!force_nonblock)
3606 kiocb->ki_flags &= ~IOCB_NOWAIT;
3608 kiocb->ki_flags |= IOCB_NOWAIT;
3610 /* If the file doesn't support async, just async punt */
3611 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3614 /* file path doesn't support NOWAIT for non-direct_IO */
3615 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3616 (req->flags & REQ_F_ISREG))
3619 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3624 * Open-code file_start_write here to grab freeze protection,
3625 * which will be released by another thread in
3626 * io_complete_rw(). Fool lockdep by telling it the lock got
3627 * released so that it doesn't complain about the held lock when
3628 * we return to userspace.
3630 if (req->flags & REQ_F_ISREG) {
3631 sb_start_write(file_inode(req->file)->i_sb);
3632 __sb_writers_release(file_inode(req->file)->i_sb,
3635 kiocb->ki_flags |= IOCB_WRITE;
3637 if (req->file->f_op->write_iter)
3638 ret2 = call_write_iter(req->file, kiocb, iter);
3639 else if (req->file->f_op->write)
3640 ret2 = loop_rw_iter(WRITE, req, iter);
3644 if (req->flags & REQ_F_REISSUE) {
3645 req->flags &= ~REQ_F_REISSUE;
3650 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3651 * retry them without IOCB_NOWAIT.
3653 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3655 /* no retry on NONBLOCK nor RWF_NOWAIT */
3656 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3658 if (!force_nonblock || ret2 != -EAGAIN) {
3659 /* IOPOLL retry should happen for io-wq threads */
3660 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3663 kiocb_done(kiocb, ret2, issue_flags);
3666 iov_iter_restore(iter, state);
3667 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3668 return ret ?: -EAGAIN;
3671 /* it's reportedly faster than delegating the null check to kfree() */
3677 static int io_renameat_prep(struct io_kiocb *req,
3678 const struct io_uring_sqe *sqe)
3680 struct io_rename *ren = &req->rename;
3681 const char __user *oldf, *newf;
3683 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3685 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3687 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3690 ren->old_dfd = READ_ONCE(sqe->fd);
3691 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3692 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3693 ren->new_dfd = READ_ONCE(sqe->len);
3694 ren->flags = READ_ONCE(sqe->rename_flags);
3696 ren->oldpath = getname(oldf);
3697 if (IS_ERR(ren->oldpath))
3698 return PTR_ERR(ren->oldpath);
3700 ren->newpath = getname(newf);
3701 if (IS_ERR(ren->newpath)) {
3702 putname(ren->oldpath);
3703 return PTR_ERR(ren->newpath);
3706 req->flags |= REQ_F_NEED_CLEANUP;
3710 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3712 struct io_rename *ren = &req->rename;
3715 if (issue_flags & IO_URING_F_NONBLOCK)
3718 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3719 ren->newpath, ren->flags);
3721 req->flags &= ~REQ_F_NEED_CLEANUP;
3724 io_req_complete(req, ret);
3728 static int io_unlinkat_prep(struct io_kiocb *req,
3729 const struct io_uring_sqe *sqe)
3731 struct io_unlink *un = &req->unlink;
3732 const char __user *fname;
3734 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3736 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3739 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3742 un->dfd = READ_ONCE(sqe->fd);
3744 un->flags = READ_ONCE(sqe->unlink_flags);
3745 if (un->flags & ~AT_REMOVEDIR)
3748 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3749 un->filename = getname(fname);
3750 if (IS_ERR(un->filename))
3751 return PTR_ERR(un->filename);
3753 req->flags |= REQ_F_NEED_CLEANUP;
3757 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3759 struct io_unlink *un = &req->unlink;
3762 if (issue_flags & IO_URING_F_NONBLOCK)
3765 if (un->flags & AT_REMOVEDIR)
3766 ret = do_rmdir(un->dfd, un->filename);
3768 ret = do_unlinkat(un->dfd, un->filename);
3770 req->flags &= ~REQ_F_NEED_CLEANUP;
3773 io_req_complete(req, ret);
3777 static int io_mkdirat_prep(struct io_kiocb *req,
3778 const struct io_uring_sqe *sqe)
3780 struct io_mkdir *mkd = &req->mkdir;
3781 const char __user *fname;
3783 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3785 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3788 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3791 mkd->dfd = READ_ONCE(sqe->fd);
3792 mkd->mode = READ_ONCE(sqe->len);
3794 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3795 mkd->filename = getname(fname);
3796 if (IS_ERR(mkd->filename))
3797 return PTR_ERR(mkd->filename);
3799 req->flags |= REQ_F_NEED_CLEANUP;
3803 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3805 struct io_mkdir *mkd = &req->mkdir;
3808 if (issue_flags & IO_URING_F_NONBLOCK)
3811 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3813 req->flags &= ~REQ_F_NEED_CLEANUP;
3816 io_req_complete(req, ret);
3820 static int io_symlinkat_prep(struct io_kiocb *req,
3821 const struct io_uring_sqe *sqe)
3823 struct io_symlink *sl = &req->symlink;
3824 const char __user *oldpath, *newpath;
3826 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3828 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3831 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3834 sl->new_dfd = READ_ONCE(sqe->fd);
3835 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3836 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3838 sl->oldpath = getname(oldpath);
3839 if (IS_ERR(sl->oldpath))
3840 return PTR_ERR(sl->oldpath);
3842 sl->newpath = getname(newpath);
3843 if (IS_ERR(sl->newpath)) {
3844 putname(sl->oldpath);
3845 return PTR_ERR(sl->newpath);
3848 req->flags |= REQ_F_NEED_CLEANUP;
3852 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3854 struct io_symlink *sl = &req->symlink;
3857 if (issue_flags & IO_URING_F_NONBLOCK)
3860 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3862 req->flags &= ~REQ_F_NEED_CLEANUP;
3865 io_req_complete(req, ret);
3869 static int io_linkat_prep(struct io_kiocb *req,
3870 const struct io_uring_sqe *sqe)
3872 struct io_hardlink *lnk = &req->hardlink;
3873 const char __user *oldf, *newf;
3875 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3877 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3879 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3882 lnk->old_dfd = READ_ONCE(sqe->fd);
3883 lnk->new_dfd = READ_ONCE(sqe->len);
3884 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3885 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3886 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3888 lnk->oldpath = getname(oldf);
3889 if (IS_ERR(lnk->oldpath))
3890 return PTR_ERR(lnk->oldpath);
3892 lnk->newpath = getname(newf);
3893 if (IS_ERR(lnk->newpath)) {
3894 putname(lnk->oldpath);
3895 return PTR_ERR(lnk->newpath);
3898 req->flags |= REQ_F_NEED_CLEANUP;
3902 static int io_linkat(struct io_kiocb *req, int issue_flags)
3904 struct io_hardlink *lnk = &req->hardlink;
3907 if (issue_flags & IO_URING_F_NONBLOCK)
3910 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3911 lnk->newpath, lnk->flags);
3913 req->flags &= ~REQ_F_NEED_CLEANUP;
3916 io_req_complete(req, ret);
3920 static int io_shutdown_prep(struct io_kiocb *req,
3921 const struct io_uring_sqe *sqe)
3923 #if defined(CONFIG_NET)
3924 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3926 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3927 sqe->buf_index || sqe->splice_fd_in))
3930 req->shutdown.how = READ_ONCE(sqe->len);
3937 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3939 #if defined(CONFIG_NET)
3940 struct socket *sock;
3943 if (issue_flags & IO_URING_F_NONBLOCK)
3946 sock = sock_from_file(req->file);
3947 if (unlikely(!sock))
3950 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3953 io_req_complete(req, ret);
3960 static int __io_splice_prep(struct io_kiocb *req,
3961 const struct io_uring_sqe *sqe)
3963 struct io_splice *sp = &req->splice;
3964 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3966 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3970 sp->len = READ_ONCE(sqe->len);
3971 sp->flags = READ_ONCE(sqe->splice_flags);
3973 if (unlikely(sp->flags & ~valid_flags))
3976 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3977 (sp->flags & SPLICE_F_FD_IN_FIXED));
3980 req->flags |= REQ_F_NEED_CLEANUP;
3984 static int io_tee_prep(struct io_kiocb *req,
3985 const struct io_uring_sqe *sqe)
3987 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3989 return __io_splice_prep(req, sqe);
3992 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3994 struct io_splice *sp = &req->splice;
3995 struct file *in = sp->file_in;
3996 struct file *out = sp->file_out;
3997 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4000 if (issue_flags & IO_URING_F_NONBLOCK)
4003 ret = do_tee(in, out, sp->len, flags);
4005 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4007 req->flags &= ~REQ_F_NEED_CLEANUP;
4011 io_req_complete(req, ret);
4015 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4017 struct io_splice *sp = &req->splice;
4019 sp->off_in = READ_ONCE(sqe->splice_off_in);
4020 sp->off_out = READ_ONCE(sqe->off);
4021 return __io_splice_prep(req, sqe);
4024 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4026 struct io_splice *sp = &req->splice;
4027 struct file *in = sp->file_in;
4028 struct file *out = sp->file_out;
4029 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4030 loff_t *poff_in, *poff_out;
4033 if (issue_flags & IO_URING_F_NONBLOCK)
4036 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4037 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4040 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4042 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4044 req->flags &= ~REQ_F_NEED_CLEANUP;
4048 io_req_complete(req, ret);
4053 * IORING_OP_NOP just posts a completion event, nothing else.
4055 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4057 struct io_ring_ctx *ctx = req->ctx;
4059 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4062 __io_req_complete(req, issue_flags, 0, 0);
4066 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4068 struct io_ring_ctx *ctx = req->ctx;
4073 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4075 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4079 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4080 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4083 req->sync.off = READ_ONCE(sqe->off);
4084 req->sync.len = READ_ONCE(sqe->len);
4088 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4090 loff_t end = req->sync.off + req->sync.len;
4093 /* fsync always requires a blocking context */
4094 if (issue_flags & IO_URING_F_NONBLOCK)
4097 ret = vfs_fsync_range(req->file, req->sync.off,
4098 end > 0 ? end : LLONG_MAX,
4099 req->sync.flags & IORING_FSYNC_DATASYNC);
4102 io_req_complete(req, ret);
4106 static int io_fallocate_prep(struct io_kiocb *req,
4107 const struct io_uring_sqe *sqe)
4109 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4112 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4115 req->sync.off = READ_ONCE(sqe->off);
4116 req->sync.len = READ_ONCE(sqe->addr);
4117 req->sync.mode = READ_ONCE(sqe->len);
4121 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4125 /* fallocate always requiring blocking context */
4126 if (issue_flags & IO_URING_F_NONBLOCK)
4128 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4132 io_req_complete(req, ret);
4136 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4138 const char __user *fname;
4141 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4143 if (unlikely(sqe->ioprio || sqe->buf_index))
4145 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4148 /* open.how should be already initialised */
4149 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4150 req->open.how.flags |= O_LARGEFILE;
4152 req->open.dfd = READ_ONCE(sqe->fd);
4153 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4154 req->open.filename = getname(fname);
4155 if (IS_ERR(req->open.filename)) {
4156 ret = PTR_ERR(req->open.filename);
4157 req->open.filename = NULL;
4161 req->open.file_slot = READ_ONCE(sqe->file_index);
4162 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4165 req->open.nofile = rlimit(RLIMIT_NOFILE);
4166 req->flags |= REQ_F_NEED_CLEANUP;
4170 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4172 u64 mode = READ_ONCE(sqe->len);
4173 u64 flags = READ_ONCE(sqe->open_flags);
4175 req->open.how = build_open_how(flags, mode);
4176 return __io_openat_prep(req, sqe);
4179 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4181 struct open_how __user *how;
4185 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4186 len = READ_ONCE(sqe->len);
4187 if (len < OPEN_HOW_SIZE_VER0)
4190 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4195 return __io_openat_prep(req, sqe);
4198 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4200 struct open_flags op;
4202 bool resolve_nonblock, nonblock_set;
4203 bool fixed = !!req->open.file_slot;
4206 ret = build_open_flags(&req->open.how, &op);
4209 nonblock_set = op.open_flag & O_NONBLOCK;
4210 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4211 if (issue_flags & IO_URING_F_NONBLOCK) {
4213 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4214 * it'll always -EAGAIN
4216 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4218 op.lookup_flags |= LOOKUP_CACHED;
4219 op.open_flag |= O_NONBLOCK;
4223 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4228 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4231 * We could hang on to this 'fd' on retrying, but seems like
4232 * marginal gain for something that is now known to be a slower
4233 * path. So just put it, and we'll get a new one when we retry.
4238 ret = PTR_ERR(file);
4239 /* only retry if RESOLVE_CACHED wasn't already set by application */
4240 if (ret == -EAGAIN &&
4241 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4246 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4247 file->f_flags &= ~O_NONBLOCK;
4248 fsnotify_open(file);
4251 fd_install(ret, file);
4253 ret = io_install_fixed_file(req, file, issue_flags,
4254 req->open.file_slot - 1);
4256 putname(req->open.filename);
4257 req->flags &= ~REQ_F_NEED_CLEANUP;
4260 __io_req_complete(req, issue_flags, ret, 0);
4264 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4266 return io_openat2(req, issue_flags);
4269 static int io_remove_buffers_prep(struct io_kiocb *req,
4270 const struct io_uring_sqe *sqe)
4272 struct io_provide_buf *p = &req->pbuf;
4275 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4279 tmp = READ_ONCE(sqe->fd);
4280 if (!tmp || tmp > USHRT_MAX)
4283 memset(p, 0, sizeof(*p));
4285 p->bgid = READ_ONCE(sqe->buf_group);
4289 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4290 int bgid, unsigned nbufs)
4294 /* shouldn't happen */
4298 /* the head kbuf is the list itself */
4299 while (!list_empty(&buf->list)) {
4300 struct io_buffer *nxt;
4302 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4303 list_del(&nxt->list);
4310 xa_erase(&ctx->io_buffers, bgid);
4315 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4317 struct io_provide_buf *p = &req->pbuf;
4318 struct io_ring_ctx *ctx = req->ctx;
4319 struct io_buffer *head;
4321 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4323 io_ring_submit_lock(ctx, !force_nonblock);
4325 lockdep_assert_held(&ctx->uring_lock);
4328 head = xa_load(&ctx->io_buffers, p->bgid);
4330 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4334 /* complete before unlock, IOPOLL may need the lock */
4335 __io_req_complete(req, issue_flags, ret, 0);
4336 io_ring_submit_unlock(ctx, !force_nonblock);
4340 static int io_provide_buffers_prep(struct io_kiocb *req,
4341 const struct io_uring_sqe *sqe)
4343 unsigned long size, tmp_check;
4344 struct io_provide_buf *p = &req->pbuf;
4347 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4350 tmp = READ_ONCE(sqe->fd);
4351 if (!tmp || tmp > USHRT_MAX)
4354 p->addr = READ_ONCE(sqe->addr);
4355 p->len = READ_ONCE(sqe->len);
4357 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4360 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4363 size = (unsigned long)p->len * p->nbufs;
4364 if (!access_ok(u64_to_user_ptr(p->addr), size))
4367 p->bgid = READ_ONCE(sqe->buf_group);
4368 tmp = READ_ONCE(sqe->off);
4369 if (tmp > USHRT_MAX)
4375 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4377 struct io_buffer *buf;
4378 u64 addr = pbuf->addr;
4379 int i, bid = pbuf->bid;
4381 for (i = 0; i < pbuf->nbufs; i++) {
4382 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4387 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4392 INIT_LIST_HEAD(&buf->list);
4395 list_add_tail(&buf->list, &(*head)->list);
4399 return i ? i : -ENOMEM;
4402 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4404 struct io_provide_buf *p = &req->pbuf;
4405 struct io_ring_ctx *ctx = req->ctx;
4406 struct io_buffer *head, *list;
4408 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4410 io_ring_submit_lock(ctx, !force_nonblock);
4412 lockdep_assert_held(&ctx->uring_lock);
4414 list = head = xa_load(&ctx->io_buffers, p->bgid);
4416 ret = io_add_buffers(p, &head);
4417 if (ret >= 0 && !list) {
4418 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4420 __io_remove_buffers(ctx, head, p->bgid, -1U);
4424 /* complete before unlock, IOPOLL may need the lock */
4425 __io_req_complete(req, issue_flags, ret, 0);
4426 io_ring_submit_unlock(ctx, !force_nonblock);
4430 static int io_epoll_ctl_prep(struct io_kiocb *req,
4431 const struct io_uring_sqe *sqe)
4433 #if defined(CONFIG_EPOLL)
4434 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4436 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4439 req->epoll.epfd = READ_ONCE(sqe->fd);
4440 req->epoll.op = READ_ONCE(sqe->len);
4441 req->epoll.fd = READ_ONCE(sqe->off);
4443 if (ep_op_has_event(req->epoll.op)) {
4444 struct epoll_event __user *ev;
4446 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4447 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4457 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4459 #if defined(CONFIG_EPOLL)
4460 struct io_epoll *ie = &req->epoll;
4462 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4464 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4465 if (force_nonblock && ret == -EAGAIN)
4470 __io_req_complete(req, issue_flags, ret, 0);
4477 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4479 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4480 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4482 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4485 req->madvise.addr = READ_ONCE(sqe->addr);
4486 req->madvise.len = READ_ONCE(sqe->len);
4487 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4494 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4496 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4497 struct io_madvise *ma = &req->madvise;
4500 if (issue_flags & IO_URING_F_NONBLOCK)
4503 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4506 io_req_complete(req, ret);
4513 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4515 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4517 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4520 req->fadvise.offset = READ_ONCE(sqe->off);
4521 req->fadvise.len = READ_ONCE(sqe->len);
4522 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4526 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4528 struct io_fadvise *fa = &req->fadvise;
4531 if (issue_flags & IO_URING_F_NONBLOCK) {
4532 switch (fa->advice) {
4533 case POSIX_FADV_NORMAL:
4534 case POSIX_FADV_RANDOM:
4535 case POSIX_FADV_SEQUENTIAL:
4542 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4545 __io_req_complete(req, issue_flags, ret, 0);
4549 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4551 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4553 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4555 if (req->flags & REQ_F_FIXED_FILE)
4558 req->statx.dfd = READ_ONCE(sqe->fd);
4559 req->statx.mask = READ_ONCE(sqe->len);
4560 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4561 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4562 req->statx.flags = READ_ONCE(sqe->statx_flags);
4567 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4569 struct io_statx *ctx = &req->statx;
4572 if (issue_flags & IO_URING_F_NONBLOCK)
4575 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4580 io_req_complete(req, ret);
4584 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4586 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4588 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4589 sqe->rw_flags || sqe->buf_index)
4591 if (req->flags & REQ_F_FIXED_FILE)
4594 req->close.fd = READ_ONCE(sqe->fd);
4595 req->close.file_slot = READ_ONCE(sqe->file_index);
4596 if (req->close.file_slot && req->close.fd)
4602 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4604 struct files_struct *files = current->files;
4605 struct io_close *close = &req->close;
4606 struct fdtable *fdt;
4607 struct file *file = NULL;
4610 if (req->close.file_slot) {
4611 ret = io_close_fixed(req, issue_flags);
4615 spin_lock(&files->file_lock);
4616 fdt = files_fdtable(files);
4617 if (close->fd >= fdt->max_fds) {
4618 spin_unlock(&files->file_lock);
4621 file = fdt->fd[close->fd];
4622 if (!file || file->f_op == &io_uring_fops) {
4623 spin_unlock(&files->file_lock);
4628 /* if the file has a flush method, be safe and punt to async */
4629 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4630 spin_unlock(&files->file_lock);
4634 ret = __close_fd_get_file(close->fd, &file);
4635 spin_unlock(&files->file_lock);
4642 /* No ->flush() or already async, safely close from here */
4643 ret = filp_close(file, current->files);
4649 __io_req_complete(req, issue_flags, ret, 0);
4653 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4655 struct io_ring_ctx *ctx = req->ctx;
4657 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4659 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4663 req->sync.off = READ_ONCE(sqe->off);
4664 req->sync.len = READ_ONCE(sqe->len);
4665 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4669 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4673 /* sync_file_range always requires a blocking context */
4674 if (issue_flags & IO_URING_F_NONBLOCK)
4677 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4681 io_req_complete(req, ret);
4685 #if defined(CONFIG_NET)
4686 static int io_setup_async_msg(struct io_kiocb *req,
4687 struct io_async_msghdr *kmsg)
4689 struct io_async_msghdr *async_msg = req->async_data;
4693 if (io_alloc_async_data(req)) {
4694 kfree(kmsg->free_iov);
4697 async_msg = req->async_data;
4698 req->flags |= REQ_F_NEED_CLEANUP;
4699 memcpy(async_msg, kmsg, sizeof(*kmsg));
4700 async_msg->msg.msg_name = &async_msg->addr;
4701 /* if were using fast_iov, set it to the new one */
4702 if (!async_msg->free_iov)
4703 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4708 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4709 struct io_async_msghdr *iomsg)
4711 iomsg->msg.msg_name = &iomsg->addr;
4712 iomsg->free_iov = iomsg->fast_iov;
4713 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4714 req->sr_msg.msg_flags, &iomsg->free_iov);
4717 static int io_sendmsg_prep_async(struct io_kiocb *req)
4721 ret = io_sendmsg_copy_hdr(req, req->async_data);
4723 req->flags |= REQ_F_NEED_CLEANUP;
4727 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4729 struct io_sr_msg *sr = &req->sr_msg;
4731 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4734 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4735 sr->len = READ_ONCE(sqe->len);
4736 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4737 if (sr->msg_flags & MSG_DONTWAIT)
4738 req->flags |= REQ_F_NOWAIT;
4740 #ifdef CONFIG_COMPAT
4741 if (req->ctx->compat)
4742 sr->msg_flags |= MSG_CMSG_COMPAT;
4747 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4749 struct io_async_msghdr iomsg, *kmsg;
4750 struct socket *sock;
4755 sock = sock_from_file(req->file);
4756 if (unlikely(!sock))
4759 kmsg = req->async_data;
4761 ret = io_sendmsg_copy_hdr(req, &iomsg);
4767 flags = req->sr_msg.msg_flags;
4768 if (issue_flags & IO_URING_F_NONBLOCK)
4769 flags |= MSG_DONTWAIT;
4770 if (flags & MSG_WAITALL)
4771 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4773 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4774 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4775 return io_setup_async_msg(req, kmsg);
4776 if (ret == -ERESTARTSYS)
4779 /* fast path, check for non-NULL to avoid function call */
4781 kfree(kmsg->free_iov);
4782 req->flags &= ~REQ_F_NEED_CLEANUP;
4785 __io_req_complete(req, issue_flags, ret, 0);
4789 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4791 struct io_sr_msg *sr = &req->sr_msg;
4794 struct socket *sock;
4799 sock = sock_from_file(req->file);
4800 if (unlikely(!sock))
4803 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4807 msg.msg_name = NULL;
4808 msg.msg_control = NULL;
4809 msg.msg_controllen = 0;
4810 msg.msg_namelen = 0;
4812 flags = req->sr_msg.msg_flags;
4813 if (issue_flags & IO_URING_F_NONBLOCK)
4814 flags |= MSG_DONTWAIT;
4815 if (flags & MSG_WAITALL)
4816 min_ret = iov_iter_count(&msg.msg_iter);
4818 msg.msg_flags = flags;
4819 ret = sock_sendmsg(sock, &msg);
4820 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4822 if (ret == -ERESTARTSYS)
4827 __io_req_complete(req, issue_flags, ret, 0);
4831 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4832 struct io_async_msghdr *iomsg)
4834 struct io_sr_msg *sr = &req->sr_msg;
4835 struct iovec __user *uiov;
4839 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4840 &iomsg->uaddr, &uiov, &iov_len);
4844 if (req->flags & REQ_F_BUFFER_SELECT) {
4847 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4849 sr->len = iomsg->fast_iov[0].iov_len;
4850 iomsg->free_iov = NULL;
4852 iomsg->free_iov = iomsg->fast_iov;
4853 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4854 &iomsg->free_iov, &iomsg->msg.msg_iter,
4863 #ifdef CONFIG_COMPAT
4864 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4865 struct io_async_msghdr *iomsg)
4867 struct io_sr_msg *sr = &req->sr_msg;
4868 struct compat_iovec __user *uiov;
4873 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4878 uiov = compat_ptr(ptr);
4879 if (req->flags & REQ_F_BUFFER_SELECT) {
4880 compat_ssize_t clen;
4884 if (!access_ok(uiov, sizeof(*uiov)))
4886 if (__get_user(clen, &uiov->iov_len))
4891 iomsg->free_iov = NULL;
4893 iomsg->free_iov = iomsg->fast_iov;
4894 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4895 UIO_FASTIOV, &iomsg->free_iov,
4896 &iomsg->msg.msg_iter, true);
4905 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4906 struct io_async_msghdr *iomsg)
4908 iomsg->msg.msg_name = &iomsg->addr;
4910 #ifdef CONFIG_COMPAT
4911 if (req->ctx->compat)
4912 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4915 return __io_recvmsg_copy_hdr(req, iomsg);
4918 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4921 struct io_sr_msg *sr = &req->sr_msg;
4922 struct io_buffer *kbuf;
4924 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4929 req->flags |= REQ_F_BUFFER_SELECTED;
4933 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4935 return io_put_kbuf(req, req->sr_msg.kbuf);
4938 static int io_recvmsg_prep_async(struct io_kiocb *req)
4942 ret = io_recvmsg_copy_hdr(req, req->async_data);
4944 req->flags |= REQ_F_NEED_CLEANUP;
4948 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4950 struct io_sr_msg *sr = &req->sr_msg;
4952 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4955 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4956 sr->len = READ_ONCE(sqe->len);
4957 sr->bgid = READ_ONCE(sqe->buf_group);
4958 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4959 if (sr->msg_flags & MSG_DONTWAIT)
4960 req->flags |= REQ_F_NOWAIT;
4962 #ifdef CONFIG_COMPAT
4963 if (req->ctx->compat)
4964 sr->msg_flags |= MSG_CMSG_COMPAT;
4969 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4971 struct io_async_msghdr iomsg, *kmsg;
4972 struct socket *sock;
4973 struct io_buffer *kbuf;
4976 int ret, cflags = 0;
4977 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4979 sock = sock_from_file(req->file);
4980 if (unlikely(!sock))
4983 kmsg = req->async_data;
4985 ret = io_recvmsg_copy_hdr(req, &iomsg);
4991 if (req->flags & REQ_F_BUFFER_SELECT) {
4992 kbuf = io_recv_buffer_select(req, !force_nonblock);
4994 return PTR_ERR(kbuf);
4995 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4996 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4997 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4998 1, req->sr_msg.len);
5001 flags = req->sr_msg.msg_flags;
5003 flags |= MSG_DONTWAIT;
5004 if (flags & MSG_WAITALL)
5005 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5007 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5008 kmsg->uaddr, flags);
5009 if (force_nonblock && ret == -EAGAIN)
5010 return io_setup_async_msg(req, kmsg);
5011 if (ret == -ERESTARTSYS)
5014 if (req->flags & REQ_F_BUFFER_SELECTED)
5015 cflags = io_put_recv_kbuf(req);
5016 /* fast path, check for non-NULL to avoid function call */
5018 kfree(kmsg->free_iov);
5019 req->flags &= ~REQ_F_NEED_CLEANUP;
5020 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5022 __io_req_complete(req, issue_flags, ret, cflags);
5026 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5028 struct io_buffer *kbuf;
5029 struct io_sr_msg *sr = &req->sr_msg;
5031 void __user *buf = sr->buf;
5032 struct socket *sock;
5036 int ret, cflags = 0;
5037 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5039 sock = sock_from_file(req->file);
5040 if (unlikely(!sock))
5043 if (req->flags & REQ_F_BUFFER_SELECT) {
5044 kbuf = io_recv_buffer_select(req, !force_nonblock);
5046 return PTR_ERR(kbuf);
5047 buf = u64_to_user_ptr(kbuf->addr);
5050 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5054 msg.msg_name = NULL;
5055 msg.msg_control = NULL;
5056 msg.msg_controllen = 0;
5057 msg.msg_namelen = 0;
5058 msg.msg_iocb = NULL;
5061 flags = req->sr_msg.msg_flags;
5063 flags |= MSG_DONTWAIT;
5064 if (flags & MSG_WAITALL)
5065 min_ret = iov_iter_count(&msg.msg_iter);
5067 ret = sock_recvmsg(sock, &msg, flags);
5068 if (force_nonblock && ret == -EAGAIN)
5070 if (ret == -ERESTARTSYS)
5073 if (req->flags & REQ_F_BUFFER_SELECTED)
5074 cflags = io_put_recv_kbuf(req);
5075 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5077 __io_req_complete(req, issue_flags, ret, cflags);
5081 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5083 struct io_accept *accept = &req->accept;
5085 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5087 if (sqe->ioprio || sqe->len || sqe->buf_index)
5090 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5091 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5092 accept->flags = READ_ONCE(sqe->accept_flags);
5093 accept->nofile = rlimit(RLIMIT_NOFILE);
5095 accept->file_slot = READ_ONCE(sqe->file_index);
5096 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5097 (accept->flags & SOCK_CLOEXEC)))
5099 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5101 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5102 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5106 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5108 struct io_accept *accept = &req->accept;
5109 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5110 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5111 bool fixed = !!accept->file_slot;
5115 if (req->file->f_flags & O_NONBLOCK)
5116 req->flags |= REQ_F_NOWAIT;
5119 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5120 if (unlikely(fd < 0))
5123 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5128 ret = PTR_ERR(file);
5129 if (ret == -EAGAIN && force_nonblock)
5131 if (ret == -ERESTARTSYS)
5134 } else if (!fixed) {
5135 fd_install(fd, file);
5138 ret = io_install_fixed_file(req, file, issue_flags,
5139 accept->file_slot - 1);
5141 __io_req_complete(req, issue_flags, ret, 0);
5145 static int io_connect_prep_async(struct io_kiocb *req)
5147 struct io_async_connect *io = req->async_data;
5148 struct io_connect *conn = &req->connect;
5150 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5153 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5155 struct io_connect *conn = &req->connect;
5157 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5159 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5163 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5164 conn->addr_len = READ_ONCE(sqe->addr2);
5168 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5170 struct io_async_connect __io, *io;
5171 unsigned file_flags;
5173 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5175 if (req->async_data) {
5176 io = req->async_data;
5178 ret = move_addr_to_kernel(req->connect.addr,
5179 req->connect.addr_len,
5186 file_flags = force_nonblock ? O_NONBLOCK : 0;
5188 ret = __sys_connect_file(req->file, &io->address,
5189 req->connect.addr_len, file_flags);
5190 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5191 if (req->async_data)
5193 if (io_alloc_async_data(req)) {
5197 memcpy(req->async_data, &__io, sizeof(__io));
5200 if (ret == -ERESTARTSYS)
5205 __io_req_complete(req, issue_flags, ret, 0);
5208 #else /* !CONFIG_NET */
5209 #define IO_NETOP_FN(op) \
5210 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5212 return -EOPNOTSUPP; \
5215 #define IO_NETOP_PREP(op) \
5217 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5219 return -EOPNOTSUPP; \
5222 #define IO_NETOP_PREP_ASYNC(op) \
5224 static int io_##op##_prep_async(struct io_kiocb *req) \
5226 return -EOPNOTSUPP; \
5229 IO_NETOP_PREP_ASYNC(sendmsg);
5230 IO_NETOP_PREP_ASYNC(recvmsg);
5231 IO_NETOP_PREP_ASYNC(connect);
5232 IO_NETOP_PREP(accept);
5235 #endif /* CONFIG_NET */
5237 struct io_poll_table {
5238 struct poll_table_struct pt;
5239 struct io_kiocb *req;
5244 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5245 __poll_t mask, io_req_tw_func_t func)
5247 /* for instances that support it check for an event match first: */
5248 if (mask && !(mask & poll->events))
5251 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5253 list_del_init(&poll->wait.entry);
5256 req->io_task_work.func = func;
5259 * If this fails, then the task is exiting. When a task exits, the
5260 * work gets canceled, so just cancel this request as well instead
5261 * of executing it. We can't safely execute it anyway, as we may not
5262 * have the needed state needed for it anyway.
5264 io_req_task_work_add(req);
5268 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5269 __acquires(&req->ctx->completion_lock)
5271 struct io_ring_ctx *ctx = req->ctx;
5273 /* req->task == current here, checking PF_EXITING is safe */
5274 if (unlikely(req->task->flags & PF_EXITING))
5275 WRITE_ONCE(poll->canceled, true);
5277 if (!req->result && !READ_ONCE(poll->canceled)) {
5278 struct poll_table_struct pt = { ._key = poll->events };
5280 req->result = vfs_poll(req->file, &pt) & poll->events;
5283 spin_lock(&ctx->completion_lock);
5284 if (!req->result && !READ_ONCE(poll->canceled)) {
5285 add_wait_queue(poll->head, &poll->wait);
5292 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5294 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5295 if (req->opcode == IORING_OP_POLL_ADD)
5296 return req->async_data;
5297 return req->apoll->double_poll;
5300 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5302 if (req->opcode == IORING_OP_POLL_ADD)
5304 return &req->apoll->poll;
5307 static void io_poll_remove_double(struct io_kiocb *req)
5308 __must_hold(&req->ctx->completion_lock)
5310 struct io_poll_iocb *poll = io_poll_get_double(req);
5312 lockdep_assert_held(&req->ctx->completion_lock);
5314 if (poll && poll->head) {
5315 struct wait_queue_head *head = poll->head;
5317 spin_lock_irq(&head->lock);
5318 list_del_init(&poll->wait.entry);
5319 if (poll->wait.private)
5322 spin_unlock_irq(&head->lock);
5326 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5327 __must_hold(&req->ctx->completion_lock)
5329 struct io_ring_ctx *ctx = req->ctx;
5330 unsigned flags = IORING_CQE_F_MORE;
5333 if (READ_ONCE(req->poll.canceled)) {
5335 req->poll.events |= EPOLLONESHOT;
5337 error = mangle_poll(mask);
5339 if (req->poll.events & EPOLLONESHOT)
5341 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5342 req->poll.events |= EPOLLONESHOT;
5345 if (flags & IORING_CQE_F_MORE)
5348 return !(flags & IORING_CQE_F_MORE);
5351 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5352 __must_hold(&req->ctx->completion_lock)
5356 done = __io_poll_complete(req, mask);
5357 io_commit_cqring(req->ctx);
5361 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5363 struct io_ring_ctx *ctx = req->ctx;
5364 struct io_kiocb *nxt;
5366 if (io_poll_rewait(req, &req->poll)) {
5367 spin_unlock(&ctx->completion_lock);
5371 if (req->poll.done) {
5372 spin_unlock(&ctx->completion_lock);
5375 done = __io_poll_complete(req, req->result);
5377 io_poll_remove_double(req);
5378 hash_del(&req->hash_node);
5379 req->poll.done = true;
5382 add_wait_queue(req->poll.head, &req->poll.wait);
5384 io_commit_cqring(ctx);
5385 spin_unlock(&ctx->completion_lock);
5386 io_cqring_ev_posted(ctx);
5389 nxt = io_put_req_find_next(req);
5391 io_req_task_submit(nxt, locked);
5396 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5397 int sync, void *key)
5399 struct io_kiocb *req = wait->private;
5400 struct io_poll_iocb *poll = io_poll_get_single(req);
5401 __poll_t mask = key_to_poll(key);
5402 unsigned long flags;
5404 /* for instances that support it check for an event match first: */
5405 if (mask && !(mask & poll->events))
5407 if (!(poll->events & EPOLLONESHOT))
5408 return poll->wait.func(&poll->wait, mode, sync, key);
5410 list_del_init(&wait->entry);
5415 spin_lock_irqsave(&poll->head->lock, flags);
5416 done = list_empty(&poll->wait.entry);
5418 list_del_init(&poll->wait.entry);
5419 /* make sure double remove sees this as being gone */
5420 wait->private = NULL;
5421 spin_unlock_irqrestore(&poll->head->lock, flags);
5423 /* use wait func handler, so it matches the rq type */
5424 poll->wait.func(&poll->wait, mode, sync, key);
5431 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5432 wait_queue_func_t wake_func)
5436 poll->canceled = false;
5437 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5438 /* mask in events that we always want/need */
5439 poll->events = events | IO_POLL_UNMASK;
5440 INIT_LIST_HEAD(&poll->wait.entry);
5441 init_waitqueue_func_entry(&poll->wait, wake_func);
5444 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5445 struct wait_queue_head *head,
5446 struct io_poll_iocb **poll_ptr)
5448 struct io_kiocb *req = pt->req;
5451 * The file being polled uses multiple waitqueues for poll handling
5452 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5455 if (unlikely(pt->nr_entries)) {
5456 struct io_poll_iocb *poll_one = poll;
5458 /* double add on the same waitqueue head, ignore */
5459 if (poll_one->head == head)
5461 /* already have a 2nd entry, fail a third attempt */
5463 if ((*poll_ptr)->head == head)
5465 pt->error = -EINVAL;
5469 * Can't handle multishot for double wait for now, turn it
5470 * into one-shot mode.
5472 if (!(poll_one->events & EPOLLONESHOT))
5473 poll_one->events |= EPOLLONESHOT;
5474 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5476 pt->error = -ENOMEM;
5479 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5481 poll->wait.private = req;
5488 if (poll->events & EPOLLEXCLUSIVE)
5489 add_wait_queue_exclusive(head, &poll->wait);
5491 add_wait_queue(head, &poll->wait);
5494 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5495 struct poll_table_struct *p)
5497 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5498 struct async_poll *apoll = pt->req->apoll;
5500 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5503 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5505 struct async_poll *apoll = req->apoll;
5506 struct io_ring_ctx *ctx = req->ctx;
5508 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5510 if (io_poll_rewait(req, &apoll->poll)) {
5511 spin_unlock(&ctx->completion_lock);
5515 hash_del(&req->hash_node);
5516 io_poll_remove_double(req);
5517 apoll->poll.done = true;
5518 spin_unlock(&ctx->completion_lock);
5520 if (!READ_ONCE(apoll->poll.canceled))
5521 io_req_task_submit(req, locked);
5523 io_req_complete_failed(req, -ECANCELED);
5526 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5529 struct io_kiocb *req = wait->private;
5530 struct io_poll_iocb *poll = &req->apoll->poll;
5532 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5535 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5538 static void io_poll_req_insert(struct io_kiocb *req)
5540 struct io_ring_ctx *ctx = req->ctx;
5541 struct hlist_head *list;
5543 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5544 hlist_add_head(&req->hash_node, list);
5547 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5548 struct io_poll_iocb *poll,
5549 struct io_poll_table *ipt, __poll_t mask,
5550 wait_queue_func_t wake_func)
5551 __acquires(&ctx->completion_lock)
5553 struct io_ring_ctx *ctx = req->ctx;
5554 bool cancel = false;
5556 INIT_HLIST_NODE(&req->hash_node);
5557 io_init_poll_iocb(poll, mask, wake_func);
5558 poll->file = req->file;
5559 poll->wait.private = req;
5561 ipt->pt._key = mask;
5564 ipt->nr_entries = 0;
5566 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5567 if (unlikely(!ipt->nr_entries) && !ipt->error)
5568 ipt->error = -EINVAL;
5570 spin_lock(&ctx->completion_lock);
5571 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5572 io_poll_remove_double(req);
5573 if (likely(poll->head)) {
5574 spin_lock_irq(&poll->head->lock);
5575 if (unlikely(list_empty(&poll->wait.entry))) {
5581 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5582 list_del_init(&poll->wait.entry);
5584 WRITE_ONCE(poll->canceled, true);
5585 else if (!poll->done) /* actually waiting for an event */
5586 io_poll_req_insert(req);
5587 spin_unlock_irq(&poll->head->lock);
5599 static int io_arm_poll_handler(struct io_kiocb *req)
5601 const struct io_op_def *def = &io_op_defs[req->opcode];
5602 struct io_ring_ctx *ctx = req->ctx;
5603 struct async_poll *apoll;
5604 struct io_poll_table ipt;
5605 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5608 if (!req->file || !file_can_poll(req->file))
5609 return IO_APOLL_ABORTED;
5610 if (req->flags & REQ_F_POLLED)
5611 return IO_APOLL_ABORTED;
5612 if (!def->pollin && !def->pollout)
5613 return IO_APOLL_ABORTED;
5617 mask |= POLLIN | POLLRDNORM;
5619 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5620 if ((req->opcode == IORING_OP_RECVMSG) &&
5621 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5625 mask |= POLLOUT | POLLWRNORM;
5628 /* if we can't nonblock try, then no point in arming a poll handler */
5629 if (!io_file_supports_nowait(req, rw))
5630 return IO_APOLL_ABORTED;
5632 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5633 if (unlikely(!apoll))
5634 return IO_APOLL_ABORTED;
5635 apoll->double_poll = NULL;
5637 req->flags |= REQ_F_POLLED;
5638 ipt.pt._qproc = io_async_queue_proc;
5639 io_req_set_refcount(req);
5641 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5643 spin_unlock(&ctx->completion_lock);
5644 if (ret || ipt.error)
5645 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5647 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5648 mask, apoll->poll.events);
5652 static bool __io_poll_remove_one(struct io_kiocb *req,
5653 struct io_poll_iocb *poll, bool do_cancel)
5654 __must_hold(&req->ctx->completion_lock)
5656 bool do_complete = false;
5660 spin_lock_irq(&poll->head->lock);
5662 WRITE_ONCE(poll->canceled, true);
5663 if (!list_empty(&poll->wait.entry)) {
5664 list_del_init(&poll->wait.entry);
5667 spin_unlock_irq(&poll->head->lock);
5668 hash_del(&req->hash_node);
5672 static bool io_poll_remove_one(struct io_kiocb *req)
5673 __must_hold(&req->ctx->completion_lock)
5677 io_poll_remove_double(req);
5678 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5681 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5682 io_commit_cqring(req->ctx);
5684 io_put_req_deferred(req);
5690 * Returns true if we found and killed one or more poll requests
5692 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5695 struct hlist_node *tmp;
5696 struct io_kiocb *req;
5699 spin_lock(&ctx->completion_lock);
5700 spin_lock_irq(&ctx->timeout_lock);
5701 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5702 struct hlist_head *list;
5704 list = &ctx->cancel_hash[i];
5705 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5706 if (io_match_task(req, tsk, cancel_all))
5707 posted += io_poll_remove_one(req);
5710 spin_unlock_irq(&ctx->timeout_lock);
5711 spin_unlock(&ctx->completion_lock);
5714 io_cqring_ev_posted(ctx);
5719 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5721 __must_hold(&ctx->completion_lock)
5723 struct hlist_head *list;
5724 struct io_kiocb *req;
5726 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5727 hlist_for_each_entry(req, list, hash_node) {
5728 if (sqe_addr != req->user_data)
5730 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5737 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5739 __must_hold(&ctx->completion_lock)
5741 struct io_kiocb *req;
5743 req = io_poll_find(ctx, sqe_addr, poll_only);
5746 if (io_poll_remove_one(req))
5752 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5757 events = READ_ONCE(sqe->poll32_events);
5759 events = swahw32(events);
5761 if (!(flags & IORING_POLL_ADD_MULTI))
5762 events |= EPOLLONESHOT;
5763 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5766 static int io_poll_update_prep(struct io_kiocb *req,
5767 const struct io_uring_sqe *sqe)
5769 struct io_poll_update *upd = &req->poll_update;
5772 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5774 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5776 flags = READ_ONCE(sqe->len);
5777 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5778 IORING_POLL_ADD_MULTI))
5780 /* meaningless without update */
5781 if (flags == IORING_POLL_ADD_MULTI)
5784 upd->old_user_data = READ_ONCE(sqe->addr);
5785 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5786 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5788 upd->new_user_data = READ_ONCE(sqe->off);
5789 if (!upd->update_user_data && upd->new_user_data)
5791 if (upd->update_events)
5792 upd->events = io_poll_parse_events(sqe, flags);
5793 else if (sqe->poll32_events)
5799 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5802 struct io_kiocb *req = wait->private;
5803 struct io_poll_iocb *poll = &req->poll;
5805 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5808 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5809 struct poll_table_struct *p)
5811 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5813 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5816 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5818 struct io_poll_iocb *poll = &req->poll;
5821 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5823 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5825 flags = READ_ONCE(sqe->len);
5826 if (flags & ~IORING_POLL_ADD_MULTI)
5829 io_req_set_refcount(req);
5830 poll->events = io_poll_parse_events(sqe, flags);
5834 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5836 struct io_poll_iocb *poll = &req->poll;
5837 struct io_ring_ctx *ctx = req->ctx;
5838 struct io_poll_table ipt;
5842 ipt.pt._qproc = io_poll_queue_proc;
5844 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5847 if (mask) { /* no async, we'd stolen it */
5849 done = io_poll_complete(req, mask);
5851 spin_unlock(&ctx->completion_lock);
5854 io_cqring_ev_posted(ctx);
5861 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5863 struct io_ring_ctx *ctx = req->ctx;
5864 struct io_kiocb *preq;
5868 spin_lock(&ctx->completion_lock);
5869 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5875 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5877 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5882 * Don't allow racy completion with singleshot, as we cannot safely
5883 * update those. For multishot, if we're racing with completion, just
5884 * let completion re-add it.
5886 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5887 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5891 /* we now have a detached poll request. reissue. */
5895 spin_unlock(&ctx->completion_lock);
5897 io_req_complete(req, ret);
5900 /* only mask one event flags, keep behavior flags */
5901 if (req->poll_update.update_events) {
5902 preq->poll.events &= ~0xffff;
5903 preq->poll.events |= req->poll_update.events & 0xffff;
5904 preq->poll.events |= IO_POLL_UNMASK;
5906 if (req->poll_update.update_user_data)
5907 preq->user_data = req->poll_update.new_user_data;
5908 spin_unlock(&ctx->completion_lock);
5910 /* complete update request, we're done with it */
5911 io_req_complete(req, ret);
5914 ret = io_poll_add(preq, issue_flags);
5917 io_req_complete(preq, ret);
5923 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5926 io_req_complete_post(req, -ETIME, 0);
5929 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5931 struct io_timeout_data *data = container_of(timer,
5932 struct io_timeout_data, timer);
5933 struct io_kiocb *req = data->req;
5934 struct io_ring_ctx *ctx = req->ctx;
5935 unsigned long flags;
5937 spin_lock_irqsave(&ctx->timeout_lock, flags);
5938 list_del_init(&req->timeout.list);
5939 atomic_set(&req->ctx->cq_timeouts,
5940 atomic_read(&req->ctx->cq_timeouts) + 1);
5941 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5943 req->io_task_work.func = io_req_task_timeout;
5944 io_req_task_work_add(req);
5945 return HRTIMER_NORESTART;
5948 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5950 __must_hold(&ctx->timeout_lock)
5952 struct io_timeout_data *io;
5953 struct io_kiocb *req;
5956 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5957 found = user_data == req->user_data;
5962 return ERR_PTR(-ENOENT);
5964 io = req->async_data;
5965 if (hrtimer_try_to_cancel(&io->timer) == -1)
5966 return ERR_PTR(-EALREADY);
5967 list_del_init(&req->timeout.list);
5971 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5972 __must_hold(&ctx->completion_lock)
5973 __must_hold(&ctx->timeout_lock)
5975 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5978 return PTR_ERR(req);
5981 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5982 io_put_req_deferred(req);
5986 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5988 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5989 case IORING_TIMEOUT_BOOTTIME:
5990 return CLOCK_BOOTTIME;
5991 case IORING_TIMEOUT_REALTIME:
5992 return CLOCK_REALTIME;
5994 /* can't happen, vetted at prep time */
5998 return CLOCK_MONOTONIC;
6002 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6003 struct timespec64 *ts, enum hrtimer_mode mode)
6004 __must_hold(&ctx->timeout_lock)
6006 struct io_timeout_data *io;
6007 struct io_kiocb *req;
6010 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6011 found = user_data == req->user_data;
6018 io = req->async_data;
6019 if (hrtimer_try_to_cancel(&io->timer) == -1)
6021 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6022 io->timer.function = io_link_timeout_fn;
6023 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6027 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6028 struct timespec64 *ts, enum hrtimer_mode mode)
6029 __must_hold(&ctx->timeout_lock)
6031 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6032 struct io_timeout_data *data;
6035 return PTR_ERR(req);
6037 req->timeout.off = 0; /* noseq */
6038 data = req->async_data;
6039 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6040 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6041 data->timer.function = io_timeout_fn;
6042 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6046 static int io_timeout_remove_prep(struct io_kiocb *req,
6047 const struct io_uring_sqe *sqe)
6049 struct io_timeout_rem *tr = &req->timeout_rem;
6051 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6053 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6055 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6058 tr->ltimeout = false;
6059 tr->addr = READ_ONCE(sqe->addr);
6060 tr->flags = READ_ONCE(sqe->timeout_flags);
6061 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6062 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6064 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6065 tr->ltimeout = true;
6066 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6068 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6070 } else if (tr->flags) {
6071 /* timeout removal doesn't support flags */
6078 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6080 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6085 * Remove or update an existing timeout command
6087 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6089 struct io_timeout_rem *tr = &req->timeout_rem;
6090 struct io_ring_ctx *ctx = req->ctx;
6093 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6094 spin_lock(&ctx->completion_lock);
6095 spin_lock_irq(&ctx->timeout_lock);
6096 ret = io_timeout_cancel(ctx, tr->addr);
6097 spin_unlock_irq(&ctx->timeout_lock);
6098 spin_unlock(&ctx->completion_lock);
6100 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6102 spin_lock_irq(&ctx->timeout_lock);
6104 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6106 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6107 spin_unlock_irq(&ctx->timeout_lock);
6112 io_req_complete_post(req, ret, 0);
6116 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6117 bool is_timeout_link)
6119 struct io_timeout_data *data;
6121 u32 off = READ_ONCE(sqe->off);
6123 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6125 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6128 if (off && is_timeout_link)
6130 flags = READ_ONCE(sqe->timeout_flags);
6131 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6133 /* more than one clock specified is invalid, obviously */
6134 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6137 INIT_LIST_HEAD(&req->timeout.list);
6138 req->timeout.off = off;
6139 if (unlikely(off && !req->ctx->off_timeout_used))
6140 req->ctx->off_timeout_used = true;
6142 if (!req->async_data && io_alloc_async_data(req))
6145 data = req->async_data;
6147 data->flags = flags;
6149 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6152 data->mode = io_translate_timeout_mode(flags);
6153 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6155 if (is_timeout_link) {
6156 struct io_submit_link *link = &req->ctx->submit_state.link;
6160 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6162 req->timeout.head = link->last;
6163 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6168 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6170 struct io_ring_ctx *ctx = req->ctx;
6171 struct io_timeout_data *data = req->async_data;
6172 struct list_head *entry;
6173 u32 tail, off = req->timeout.off;
6175 spin_lock_irq(&ctx->timeout_lock);
6178 * sqe->off holds how many events that need to occur for this
6179 * timeout event to be satisfied. If it isn't set, then this is
6180 * a pure timeout request, sequence isn't used.
6182 if (io_is_timeout_noseq(req)) {
6183 entry = ctx->timeout_list.prev;
6187 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6188 req->timeout.target_seq = tail + off;
6190 /* Update the last seq here in case io_flush_timeouts() hasn't.
6191 * This is safe because ->completion_lock is held, and submissions
6192 * and completions are never mixed in the same ->completion_lock section.
6194 ctx->cq_last_tm_flush = tail;
6197 * Insertion sort, ensuring the first entry in the list is always
6198 * the one we need first.
6200 list_for_each_prev(entry, &ctx->timeout_list) {
6201 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6204 if (io_is_timeout_noseq(nxt))
6206 /* nxt.seq is behind @tail, otherwise would've been completed */
6207 if (off >= nxt->timeout.target_seq - tail)
6211 list_add(&req->timeout.list, entry);
6212 data->timer.function = io_timeout_fn;
6213 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6214 spin_unlock_irq(&ctx->timeout_lock);
6218 struct io_cancel_data {
6219 struct io_ring_ctx *ctx;
6223 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6225 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6226 struct io_cancel_data *cd = data;
6228 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6231 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6232 struct io_ring_ctx *ctx)
6234 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6235 enum io_wq_cancel cancel_ret;
6238 if (!tctx || !tctx->io_wq)
6241 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6242 switch (cancel_ret) {
6243 case IO_WQ_CANCEL_OK:
6246 case IO_WQ_CANCEL_RUNNING:
6249 case IO_WQ_CANCEL_NOTFOUND:
6257 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6259 struct io_ring_ctx *ctx = req->ctx;
6262 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6264 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6268 spin_lock(&ctx->completion_lock);
6269 spin_lock_irq(&ctx->timeout_lock);
6270 ret = io_timeout_cancel(ctx, sqe_addr);
6271 spin_unlock_irq(&ctx->timeout_lock);
6274 ret = io_poll_cancel(ctx, sqe_addr, false);
6276 spin_unlock(&ctx->completion_lock);
6280 static int io_async_cancel_prep(struct io_kiocb *req,
6281 const struct io_uring_sqe *sqe)
6283 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6285 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6287 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6291 req->cancel.addr = READ_ONCE(sqe->addr);
6295 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6297 struct io_ring_ctx *ctx = req->ctx;
6298 u64 sqe_addr = req->cancel.addr;
6299 struct io_tctx_node *node;
6302 ret = io_try_cancel_userdata(req, sqe_addr);
6306 /* slow path, try all io-wq's */
6307 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6309 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6310 struct io_uring_task *tctx = node->task->io_uring;
6312 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6316 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6320 io_req_complete_post(req, ret, 0);
6324 static int io_rsrc_update_prep(struct io_kiocb *req,
6325 const struct io_uring_sqe *sqe)
6327 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6329 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6332 req->rsrc_update.offset = READ_ONCE(sqe->off);
6333 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6334 if (!req->rsrc_update.nr_args)
6336 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6340 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6342 struct io_ring_ctx *ctx = req->ctx;
6343 struct io_uring_rsrc_update2 up;
6346 up.offset = req->rsrc_update.offset;
6347 up.data = req->rsrc_update.arg;
6352 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6353 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6354 &up, req->rsrc_update.nr_args);
6355 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6359 __io_req_complete(req, issue_flags, ret, 0);
6363 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6365 switch (req->opcode) {
6368 case IORING_OP_READV:
6369 case IORING_OP_READ_FIXED:
6370 case IORING_OP_READ:
6371 return io_read_prep(req, sqe);
6372 case IORING_OP_WRITEV:
6373 case IORING_OP_WRITE_FIXED:
6374 case IORING_OP_WRITE:
6375 return io_write_prep(req, sqe);
6376 case IORING_OP_POLL_ADD:
6377 return io_poll_add_prep(req, sqe);
6378 case IORING_OP_POLL_REMOVE:
6379 return io_poll_update_prep(req, sqe);
6380 case IORING_OP_FSYNC:
6381 return io_fsync_prep(req, sqe);
6382 case IORING_OP_SYNC_FILE_RANGE:
6383 return io_sfr_prep(req, sqe);
6384 case IORING_OP_SENDMSG:
6385 case IORING_OP_SEND:
6386 return io_sendmsg_prep(req, sqe);
6387 case IORING_OP_RECVMSG:
6388 case IORING_OP_RECV:
6389 return io_recvmsg_prep(req, sqe);
6390 case IORING_OP_CONNECT:
6391 return io_connect_prep(req, sqe);
6392 case IORING_OP_TIMEOUT:
6393 return io_timeout_prep(req, sqe, false);
6394 case IORING_OP_TIMEOUT_REMOVE:
6395 return io_timeout_remove_prep(req, sqe);
6396 case IORING_OP_ASYNC_CANCEL:
6397 return io_async_cancel_prep(req, sqe);
6398 case IORING_OP_LINK_TIMEOUT:
6399 return io_timeout_prep(req, sqe, true);
6400 case IORING_OP_ACCEPT:
6401 return io_accept_prep(req, sqe);
6402 case IORING_OP_FALLOCATE:
6403 return io_fallocate_prep(req, sqe);
6404 case IORING_OP_OPENAT:
6405 return io_openat_prep(req, sqe);
6406 case IORING_OP_CLOSE:
6407 return io_close_prep(req, sqe);
6408 case IORING_OP_FILES_UPDATE:
6409 return io_rsrc_update_prep(req, sqe);
6410 case IORING_OP_STATX:
6411 return io_statx_prep(req, sqe);
6412 case IORING_OP_FADVISE:
6413 return io_fadvise_prep(req, sqe);
6414 case IORING_OP_MADVISE:
6415 return io_madvise_prep(req, sqe);
6416 case IORING_OP_OPENAT2:
6417 return io_openat2_prep(req, sqe);
6418 case IORING_OP_EPOLL_CTL:
6419 return io_epoll_ctl_prep(req, sqe);
6420 case IORING_OP_SPLICE:
6421 return io_splice_prep(req, sqe);
6422 case IORING_OP_PROVIDE_BUFFERS:
6423 return io_provide_buffers_prep(req, sqe);
6424 case IORING_OP_REMOVE_BUFFERS:
6425 return io_remove_buffers_prep(req, sqe);
6427 return io_tee_prep(req, sqe);
6428 case IORING_OP_SHUTDOWN:
6429 return io_shutdown_prep(req, sqe);
6430 case IORING_OP_RENAMEAT:
6431 return io_renameat_prep(req, sqe);
6432 case IORING_OP_UNLINKAT:
6433 return io_unlinkat_prep(req, sqe);
6434 case IORING_OP_MKDIRAT:
6435 return io_mkdirat_prep(req, sqe);
6436 case IORING_OP_SYMLINKAT:
6437 return io_symlinkat_prep(req, sqe);
6438 case IORING_OP_LINKAT:
6439 return io_linkat_prep(req, sqe);
6442 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6447 static int io_req_prep_async(struct io_kiocb *req)
6449 if (!io_op_defs[req->opcode].needs_async_setup)
6451 if (WARN_ON_ONCE(req->async_data))
6453 if (io_alloc_async_data(req))
6456 switch (req->opcode) {
6457 case IORING_OP_READV:
6458 return io_rw_prep_async(req, READ);
6459 case IORING_OP_WRITEV:
6460 return io_rw_prep_async(req, WRITE);
6461 case IORING_OP_SENDMSG:
6462 return io_sendmsg_prep_async(req);
6463 case IORING_OP_RECVMSG:
6464 return io_recvmsg_prep_async(req);
6465 case IORING_OP_CONNECT:
6466 return io_connect_prep_async(req);
6468 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6473 static u32 io_get_sequence(struct io_kiocb *req)
6475 u32 seq = req->ctx->cached_sq_head;
6477 /* need original cached_sq_head, but it was increased for each req */
6478 io_for_each_link(req, req)
6483 static bool io_drain_req(struct io_kiocb *req)
6485 struct io_kiocb *pos;
6486 struct io_ring_ctx *ctx = req->ctx;
6487 struct io_defer_entry *de;
6491 if (req->flags & REQ_F_FAIL) {
6492 io_req_complete_fail_submit(req);
6497 * If we need to drain a request in the middle of a link, drain the
6498 * head request and the next request/link after the current link.
6499 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6500 * maintained for every request of our link.
6502 if (ctx->drain_next) {
6503 req->flags |= REQ_F_IO_DRAIN;
6504 ctx->drain_next = false;
6506 /* not interested in head, start from the first linked */
6507 io_for_each_link(pos, req->link) {
6508 if (pos->flags & REQ_F_IO_DRAIN) {
6509 ctx->drain_next = true;
6510 req->flags |= REQ_F_IO_DRAIN;
6515 /* Still need defer if there is pending req in defer list. */
6516 if (likely(list_empty_careful(&ctx->defer_list) &&
6517 !(req->flags & REQ_F_IO_DRAIN))) {
6518 ctx->drain_active = false;
6522 seq = io_get_sequence(req);
6523 /* Still a chance to pass the sequence check */
6524 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6527 ret = io_req_prep_async(req);
6530 io_prep_async_link(req);
6531 de = kmalloc(sizeof(*de), GFP_KERNEL);
6535 io_req_complete_failed(req, ret);
6539 spin_lock(&ctx->completion_lock);
6540 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6541 spin_unlock(&ctx->completion_lock);
6543 io_queue_async_work(req, NULL);
6547 trace_io_uring_defer(ctx, req, req->user_data);
6550 list_add_tail(&de->list, &ctx->defer_list);
6551 spin_unlock(&ctx->completion_lock);
6555 static void io_clean_op(struct io_kiocb *req)
6557 if (req->flags & REQ_F_BUFFER_SELECTED) {
6558 switch (req->opcode) {
6559 case IORING_OP_READV:
6560 case IORING_OP_READ_FIXED:
6561 case IORING_OP_READ:
6562 kfree((void *)(unsigned long)req->rw.addr);
6564 case IORING_OP_RECVMSG:
6565 case IORING_OP_RECV:
6566 kfree(req->sr_msg.kbuf);
6571 if (req->flags & REQ_F_NEED_CLEANUP) {
6572 switch (req->opcode) {
6573 case IORING_OP_READV:
6574 case IORING_OP_READ_FIXED:
6575 case IORING_OP_READ:
6576 case IORING_OP_WRITEV:
6577 case IORING_OP_WRITE_FIXED:
6578 case IORING_OP_WRITE: {
6579 struct io_async_rw *io = req->async_data;
6581 kfree(io->free_iovec);
6584 case IORING_OP_RECVMSG:
6585 case IORING_OP_SENDMSG: {
6586 struct io_async_msghdr *io = req->async_data;
6588 kfree(io->free_iov);
6591 case IORING_OP_SPLICE:
6593 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6594 io_put_file(req->splice.file_in);
6596 case IORING_OP_OPENAT:
6597 case IORING_OP_OPENAT2:
6598 if (req->open.filename)
6599 putname(req->open.filename);
6601 case IORING_OP_RENAMEAT:
6602 putname(req->rename.oldpath);
6603 putname(req->rename.newpath);
6605 case IORING_OP_UNLINKAT:
6606 putname(req->unlink.filename);
6608 case IORING_OP_MKDIRAT:
6609 putname(req->mkdir.filename);
6611 case IORING_OP_SYMLINKAT:
6612 putname(req->symlink.oldpath);
6613 putname(req->symlink.newpath);
6615 case IORING_OP_LINKAT:
6616 putname(req->hardlink.oldpath);
6617 putname(req->hardlink.newpath);
6621 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6622 kfree(req->apoll->double_poll);
6626 if (req->flags & REQ_F_INFLIGHT) {
6627 struct io_uring_task *tctx = req->task->io_uring;
6629 atomic_dec(&tctx->inflight_tracked);
6631 if (req->flags & REQ_F_CREDS)
6632 put_cred(req->creds);
6634 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6637 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6639 struct io_ring_ctx *ctx = req->ctx;
6640 const struct cred *creds = NULL;
6643 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6644 creds = override_creds(req->creds);
6646 switch (req->opcode) {
6648 ret = io_nop(req, issue_flags);
6650 case IORING_OP_READV:
6651 case IORING_OP_READ_FIXED:
6652 case IORING_OP_READ:
6653 ret = io_read(req, issue_flags);
6655 case IORING_OP_WRITEV:
6656 case IORING_OP_WRITE_FIXED:
6657 case IORING_OP_WRITE:
6658 ret = io_write(req, issue_flags);
6660 case IORING_OP_FSYNC:
6661 ret = io_fsync(req, issue_flags);
6663 case IORING_OP_POLL_ADD:
6664 ret = io_poll_add(req, issue_flags);
6666 case IORING_OP_POLL_REMOVE:
6667 ret = io_poll_update(req, issue_flags);
6669 case IORING_OP_SYNC_FILE_RANGE:
6670 ret = io_sync_file_range(req, issue_flags);
6672 case IORING_OP_SENDMSG:
6673 ret = io_sendmsg(req, issue_flags);
6675 case IORING_OP_SEND:
6676 ret = io_send(req, issue_flags);
6678 case IORING_OP_RECVMSG:
6679 ret = io_recvmsg(req, issue_flags);
6681 case IORING_OP_RECV:
6682 ret = io_recv(req, issue_flags);
6684 case IORING_OP_TIMEOUT:
6685 ret = io_timeout(req, issue_flags);
6687 case IORING_OP_TIMEOUT_REMOVE:
6688 ret = io_timeout_remove(req, issue_flags);
6690 case IORING_OP_ACCEPT:
6691 ret = io_accept(req, issue_flags);
6693 case IORING_OP_CONNECT:
6694 ret = io_connect(req, issue_flags);
6696 case IORING_OP_ASYNC_CANCEL:
6697 ret = io_async_cancel(req, issue_flags);
6699 case IORING_OP_FALLOCATE:
6700 ret = io_fallocate(req, issue_flags);
6702 case IORING_OP_OPENAT:
6703 ret = io_openat(req, issue_flags);
6705 case IORING_OP_CLOSE:
6706 ret = io_close(req, issue_flags);
6708 case IORING_OP_FILES_UPDATE:
6709 ret = io_files_update(req, issue_flags);
6711 case IORING_OP_STATX:
6712 ret = io_statx(req, issue_flags);
6714 case IORING_OP_FADVISE:
6715 ret = io_fadvise(req, issue_flags);
6717 case IORING_OP_MADVISE:
6718 ret = io_madvise(req, issue_flags);
6720 case IORING_OP_OPENAT2:
6721 ret = io_openat2(req, issue_flags);
6723 case IORING_OP_EPOLL_CTL:
6724 ret = io_epoll_ctl(req, issue_flags);
6726 case IORING_OP_SPLICE:
6727 ret = io_splice(req, issue_flags);
6729 case IORING_OP_PROVIDE_BUFFERS:
6730 ret = io_provide_buffers(req, issue_flags);
6732 case IORING_OP_REMOVE_BUFFERS:
6733 ret = io_remove_buffers(req, issue_flags);
6736 ret = io_tee(req, issue_flags);
6738 case IORING_OP_SHUTDOWN:
6739 ret = io_shutdown(req, issue_flags);
6741 case IORING_OP_RENAMEAT:
6742 ret = io_renameat(req, issue_flags);
6744 case IORING_OP_UNLINKAT:
6745 ret = io_unlinkat(req, issue_flags);
6747 case IORING_OP_MKDIRAT:
6748 ret = io_mkdirat(req, issue_flags);
6750 case IORING_OP_SYMLINKAT:
6751 ret = io_symlinkat(req, issue_flags);
6753 case IORING_OP_LINKAT:
6754 ret = io_linkat(req, issue_flags);
6762 revert_creds(creds);
6765 /* If the op doesn't have a file, we're not polling for it */
6766 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6767 io_iopoll_req_issued(req);
6772 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6774 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6776 req = io_put_req_find_next(req);
6777 return req ? &req->work : NULL;
6780 static void io_wq_submit_work(struct io_wq_work *work)
6782 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6783 struct io_kiocb *timeout;
6786 /* one will be dropped by ->io_free_work() after returning to io-wq */
6787 if (!(req->flags & REQ_F_REFCOUNT))
6788 __io_req_set_refcount(req, 2);
6792 timeout = io_prep_linked_timeout(req);
6794 io_queue_linked_timeout(timeout);
6796 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6797 if (work->flags & IO_WQ_WORK_CANCEL)
6802 ret = io_issue_sqe(req, 0);
6804 * We can get EAGAIN for polled IO even though we're
6805 * forcing a sync submission from here, since we can't
6806 * wait for request slots on the block side.
6814 /* avoid locking problems by failing it from a clean context */
6816 io_req_task_queue_fail(req, ret);
6819 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6822 return &table->files[i];
6825 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6828 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6830 return (struct file *) (slot->file_ptr & FFS_MASK);
6833 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6835 unsigned long file_ptr = (unsigned long) file;
6837 if (__io_file_supports_nowait(file, READ))
6838 file_ptr |= FFS_ASYNC_READ;
6839 if (__io_file_supports_nowait(file, WRITE))
6840 file_ptr |= FFS_ASYNC_WRITE;
6841 if (S_ISREG(file_inode(file)->i_mode))
6842 file_ptr |= FFS_ISREG;
6843 file_slot->file_ptr = file_ptr;
6846 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6847 struct io_kiocb *req, int fd)
6850 unsigned long file_ptr;
6852 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6854 fd = array_index_nospec(fd, ctx->nr_user_files);
6855 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6856 file = (struct file *) (file_ptr & FFS_MASK);
6857 file_ptr &= ~FFS_MASK;
6858 /* mask in overlapping REQ_F and FFS bits */
6859 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6860 io_req_set_rsrc_node(req);
6864 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6865 struct io_kiocb *req, int fd)
6867 struct file *file = fget(fd);
6869 trace_io_uring_file_get(ctx, fd);
6871 /* we don't allow fixed io_uring files */
6872 if (file && unlikely(file->f_op == &io_uring_fops))
6873 io_req_track_inflight(req);
6877 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6878 struct io_kiocb *req, int fd, bool fixed)
6881 return io_file_get_fixed(ctx, req, fd);
6883 return io_file_get_normal(ctx, req, fd);
6886 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6888 struct io_kiocb *prev = req->timeout.prev;
6892 ret = io_try_cancel_userdata(req, prev->user_data);
6893 io_req_complete_post(req, ret ?: -ETIME, 0);
6896 io_req_complete_post(req, -ETIME, 0);
6900 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6902 struct io_timeout_data *data = container_of(timer,
6903 struct io_timeout_data, timer);
6904 struct io_kiocb *prev, *req = data->req;
6905 struct io_ring_ctx *ctx = req->ctx;
6906 unsigned long flags;
6908 spin_lock_irqsave(&ctx->timeout_lock, flags);
6909 prev = req->timeout.head;
6910 req->timeout.head = NULL;
6913 * We don't expect the list to be empty, that will only happen if we
6914 * race with the completion of the linked work.
6917 io_remove_next_linked(prev);
6918 if (!req_ref_inc_not_zero(prev))
6921 list_del(&req->timeout.list);
6922 req->timeout.prev = prev;
6923 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6925 req->io_task_work.func = io_req_task_link_timeout;
6926 io_req_task_work_add(req);
6927 return HRTIMER_NORESTART;
6930 static void io_queue_linked_timeout(struct io_kiocb *req)
6932 struct io_ring_ctx *ctx = req->ctx;
6934 spin_lock_irq(&ctx->timeout_lock);
6936 * If the back reference is NULL, then our linked request finished
6937 * before we got a chance to setup the timer
6939 if (req->timeout.head) {
6940 struct io_timeout_data *data = req->async_data;
6942 data->timer.function = io_link_timeout_fn;
6943 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6945 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6947 spin_unlock_irq(&ctx->timeout_lock);
6948 /* drop submission reference */
6952 static void __io_queue_sqe(struct io_kiocb *req)
6953 __must_hold(&req->ctx->uring_lock)
6955 struct io_kiocb *linked_timeout;
6959 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6962 * We async punt it if the file wasn't marked NOWAIT, or if the file
6963 * doesn't support non-blocking read/write attempts
6966 if (req->flags & REQ_F_COMPLETE_INLINE) {
6967 struct io_ring_ctx *ctx = req->ctx;
6968 struct io_submit_state *state = &ctx->submit_state;
6970 state->compl_reqs[state->compl_nr++] = req;
6971 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6972 io_submit_flush_completions(ctx);
6976 linked_timeout = io_prep_linked_timeout(req);
6978 io_queue_linked_timeout(linked_timeout);
6979 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6980 linked_timeout = io_prep_linked_timeout(req);
6982 switch (io_arm_poll_handler(req)) {
6983 case IO_APOLL_READY:
6985 io_queue_linked_timeout(linked_timeout);
6987 case IO_APOLL_ABORTED:
6989 * Queued up for async execution, worker will release
6990 * submit reference when the iocb is actually submitted.
6992 io_queue_async_work(req, NULL);
6997 io_queue_linked_timeout(linked_timeout);
6999 io_req_complete_failed(req, ret);
7003 static inline void io_queue_sqe(struct io_kiocb *req)
7004 __must_hold(&req->ctx->uring_lock)
7006 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7009 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7010 __io_queue_sqe(req);
7011 } else if (req->flags & REQ_F_FAIL) {
7012 io_req_complete_fail_submit(req);
7014 int ret = io_req_prep_async(req);
7017 io_req_complete_failed(req, ret);
7019 io_queue_async_work(req, NULL);
7024 * Check SQE restrictions (opcode and flags).
7026 * Returns 'true' if SQE is allowed, 'false' otherwise.
7028 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7029 struct io_kiocb *req,
7030 unsigned int sqe_flags)
7032 if (likely(!ctx->restricted))
7035 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7038 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7039 ctx->restrictions.sqe_flags_required)
7042 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7043 ctx->restrictions.sqe_flags_required))
7049 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7050 const struct io_uring_sqe *sqe)
7051 __must_hold(&ctx->uring_lock)
7053 struct io_submit_state *state;
7054 unsigned int sqe_flags;
7055 int personality, ret = 0;
7057 /* req is partially pre-initialised, see io_preinit_req() */
7058 req->opcode = READ_ONCE(sqe->opcode);
7059 /* same numerical values with corresponding REQ_F_*, safe to copy */
7060 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7061 req->user_data = READ_ONCE(sqe->user_data);
7063 req->fixed_rsrc_refs = NULL;
7064 req->task = current;
7066 /* enforce forwards compatibility on users */
7067 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7069 if (unlikely(req->opcode >= IORING_OP_LAST))
7071 if (!io_check_restriction(ctx, req, sqe_flags))
7074 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7075 !io_op_defs[req->opcode].buffer_select)
7077 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7078 ctx->drain_active = true;
7080 personality = READ_ONCE(sqe->personality);
7082 req->creds = xa_load(&ctx->personalities, personality);
7085 get_cred(req->creds);
7086 req->flags |= REQ_F_CREDS;
7088 state = &ctx->submit_state;
7091 * Plug now if we have more than 1 IO left after this, and the target
7092 * is potentially a read/write to block based storage.
7094 if (!state->plug_started && state->ios_left > 1 &&
7095 io_op_defs[req->opcode].plug) {
7096 blk_start_plug(&state->plug);
7097 state->plug_started = true;
7100 if (io_op_defs[req->opcode].needs_file) {
7101 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7102 (sqe_flags & IOSQE_FIXED_FILE));
7103 if (unlikely(!req->file))
7111 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7112 const struct io_uring_sqe *sqe)
7113 __must_hold(&ctx->uring_lock)
7115 struct io_submit_link *link = &ctx->submit_state.link;
7118 ret = io_init_req(ctx, req, sqe);
7119 if (unlikely(ret)) {
7121 /* fail even hard links since we don't submit */
7124 * we can judge a link req is failed or cancelled by if
7125 * REQ_F_FAIL is set, but the head is an exception since
7126 * it may be set REQ_F_FAIL because of other req's failure
7127 * so let's leverage req->result to distinguish if a head
7128 * is set REQ_F_FAIL because of its failure or other req's
7129 * failure so that we can set the correct ret code for it.
7130 * init result here to avoid affecting the normal path.
7132 if (!(link->head->flags & REQ_F_FAIL))
7133 req_fail_link_node(link->head, -ECANCELED);
7134 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7136 * the current req is a normal req, we should return
7137 * error and thus break the submittion loop.
7139 io_req_complete_failed(req, ret);
7142 req_fail_link_node(req, ret);
7144 ret = io_req_prep(req, sqe);
7149 /* don't need @sqe from now on */
7150 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7152 ctx->flags & IORING_SETUP_SQPOLL);
7155 * If we already have a head request, queue this one for async
7156 * submittal once the head completes. If we don't have a head but
7157 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7158 * submitted sync once the chain is complete. If none of those
7159 * conditions are true (normal request), then just queue it.
7162 struct io_kiocb *head = link->head;
7164 if (!(req->flags & REQ_F_FAIL)) {
7165 ret = io_req_prep_async(req);
7166 if (unlikely(ret)) {
7167 req_fail_link_node(req, ret);
7168 if (!(head->flags & REQ_F_FAIL))
7169 req_fail_link_node(head, -ECANCELED);
7172 trace_io_uring_link(ctx, req, head);
7173 link->last->link = req;
7176 /* last request of a link, enqueue the link */
7177 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7182 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7194 * Batched submission is done, ensure local IO is flushed out.
7196 static void io_submit_state_end(struct io_submit_state *state,
7197 struct io_ring_ctx *ctx)
7199 if (state->link.head)
7200 io_queue_sqe(state->link.head);
7201 if (state->compl_nr)
7202 io_submit_flush_completions(ctx);
7203 if (state->plug_started)
7204 blk_finish_plug(&state->plug);
7208 * Start submission side cache.
7210 static void io_submit_state_start(struct io_submit_state *state,
7211 unsigned int max_ios)
7213 state->plug_started = false;
7214 state->ios_left = max_ios;
7215 /* set only head, no need to init link_last in advance */
7216 state->link.head = NULL;
7219 static void io_commit_sqring(struct io_ring_ctx *ctx)
7221 struct io_rings *rings = ctx->rings;
7224 * Ensure any loads from the SQEs are done at this point,
7225 * since once we write the new head, the application could
7226 * write new data to them.
7228 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7232 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7233 * that is mapped by userspace. This means that care needs to be taken to
7234 * ensure that reads are stable, as we cannot rely on userspace always
7235 * being a good citizen. If members of the sqe are validated and then later
7236 * used, it's important that those reads are done through READ_ONCE() to
7237 * prevent a re-load down the line.
7239 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7241 unsigned head, mask = ctx->sq_entries - 1;
7242 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7245 * The cached sq head (or cq tail) serves two purposes:
7247 * 1) allows us to batch the cost of updating the user visible
7249 * 2) allows the kernel side to track the head on its own, even
7250 * though the application is the one updating it.
7252 head = READ_ONCE(ctx->sq_array[sq_idx]);
7253 if (likely(head < ctx->sq_entries))
7254 return &ctx->sq_sqes[head];
7256 /* drop invalid entries */
7258 WRITE_ONCE(ctx->rings->sq_dropped,
7259 READ_ONCE(ctx->rings->sq_dropped) + 1);
7263 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7264 __must_hold(&ctx->uring_lock)
7268 /* make sure SQ entry isn't read before tail */
7269 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7270 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7272 io_get_task_refs(nr);
7274 io_submit_state_start(&ctx->submit_state, nr);
7275 while (submitted < nr) {
7276 const struct io_uring_sqe *sqe;
7277 struct io_kiocb *req;
7279 req = io_alloc_req(ctx);
7280 if (unlikely(!req)) {
7282 submitted = -EAGAIN;
7285 sqe = io_get_sqe(ctx);
7286 if (unlikely(!sqe)) {
7287 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7290 /* will complete beyond this point, count as submitted */
7292 if (io_submit_sqe(ctx, req, sqe))
7296 if (unlikely(submitted != nr)) {
7297 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7298 int unused = nr - ref_used;
7300 current->io_uring->cached_refs += unused;
7301 percpu_ref_put_many(&ctx->refs, unused);
7304 io_submit_state_end(&ctx->submit_state, ctx);
7305 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7306 io_commit_sqring(ctx);
7311 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7313 return READ_ONCE(sqd->state);
7316 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7318 /* Tell userspace we may need a wakeup call */
7319 spin_lock(&ctx->completion_lock);
7320 WRITE_ONCE(ctx->rings->sq_flags,
7321 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7322 spin_unlock(&ctx->completion_lock);
7325 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7327 spin_lock(&ctx->completion_lock);
7328 WRITE_ONCE(ctx->rings->sq_flags,
7329 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7330 spin_unlock(&ctx->completion_lock);
7333 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7335 unsigned int to_submit;
7338 to_submit = io_sqring_entries(ctx);
7339 /* if we're handling multiple rings, cap submit size for fairness */
7340 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7341 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7343 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7344 unsigned nr_events = 0;
7345 const struct cred *creds = NULL;
7347 if (ctx->sq_creds != current_cred())
7348 creds = override_creds(ctx->sq_creds);
7350 mutex_lock(&ctx->uring_lock);
7351 if (!list_empty(&ctx->iopoll_list))
7352 io_do_iopoll(ctx, &nr_events, 0);
7355 * Don't submit if refs are dying, good for io_uring_register(),
7356 * but also it is relied upon by io_ring_exit_work()
7358 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7359 !(ctx->flags & IORING_SETUP_R_DISABLED))
7360 ret = io_submit_sqes(ctx, to_submit);
7361 mutex_unlock(&ctx->uring_lock);
7363 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7364 wake_up(&ctx->sqo_sq_wait);
7366 revert_creds(creds);
7372 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7374 struct io_ring_ctx *ctx;
7375 unsigned sq_thread_idle = 0;
7377 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7378 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7379 sqd->sq_thread_idle = sq_thread_idle;
7382 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7384 bool did_sig = false;
7385 struct ksignal ksig;
7387 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7388 signal_pending(current)) {
7389 mutex_unlock(&sqd->lock);
7390 if (signal_pending(current))
7391 did_sig = get_signal(&ksig);
7393 mutex_lock(&sqd->lock);
7395 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7398 static int io_sq_thread(void *data)
7400 struct io_sq_data *sqd = data;
7401 struct io_ring_ctx *ctx;
7402 unsigned long timeout = 0;
7403 char buf[TASK_COMM_LEN];
7406 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7407 set_task_comm(current, buf);
7409 if (sqd->sq_cpu != -1)
7410 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7412 set_cpus_allowed_ptr(current, cpu_online_mask);
7413 current->flags |= PF_NO_SETAFFINITY;
7415 mutex_lock(&sqd->lock);
7417 bool cap_entries, sqt_spin = false;
7419 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7420 if (io_sqd_handle_event(sqd))
7422 timeout = jiffies + sqd->sq_thread_idle;
7425 cap_entries = !list_is_singular(&sqd->ctx_list);
7426 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7427 int ret = __io_sq_thread(ctx, cap_entries);
7429 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7432 if (io_run_task_work())
7435 if (sqt_spin || !time_after(jiffies, timeout)) {
7438 timeout = jiffies + sqd->sq_thread_idle;
7442 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7443 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7444 bool needs_sched = true;
7446 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7447 io_ring_set_wakeup_flag(ctx);
7449 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7450 !list_empty_careful(&ctx->iopoll_list)) {
7451 needs_sched = false;
7454 if (io_sqring_entries(ctx)) {
7455 needs_sched = false;
7461 mutex_unlock(&sqd->lock);
7463 mutex_lock(&sqd->lock);
7465 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7466 io_ring_clear_wakeup_flag(ctx);
7469 finish_wait(&sqd->wait, &wait);
7470 timeout = jiffies + sqd->sq_thread_idle;
7473 io_uring_cancel_generic(true, sqd);
7475 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7476 io_ring_set_wakeup_flag(ctx);
7478 mutex_unlock(&sqd->lock);
7480 complete(&sqd->exited);
7484 struct io_wait_queue {
7485 struct wait_queue_entry wq;
7486 struct io_ring_ctx *ctx;
7488 unsigned nr_timeouts;
7491 static inline bool io_should_wake(struct io_wait_queue *iowq)
7493 struct io_ring_ctx *ctx = iowq->ctx;
7494 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7497 * Wake up if we have enough events, or if a timeout occurred since we
7498 * started waiting. For timeouts, we always want to return to userspace,
7499 * regardless of event count.
7501 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7504 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7505 int wake_flags, void *key)
7507 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7511 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7512 * the task, and the next invocation will do it.
7514 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7515 return autoremove_wake_function(curr, mode, wake_flags, key);
7519 static int io_run_task_work_sig(void)
7521 if (io_run_task_work())
7523 if (!signal_pending(current))
7525 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7526 return -ERESTARTSYS;
7530 /* when returns >0, the caller should retry */
7531 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7532 struct io_wait_queue *iowq,
7533 signed long *timeout)
7537 /* make sure we run task_work before checking for signals */
7538 ret = io_run_task_work_sig();
7539 if (ret || io_should_wake(iowq))
7541 /* let the caller flush overflows, retry */
7542 if (test_bit(0, &ctx->check_cq_overflow))
7545 *timeout = schedule_timeout(*timeout);
7546 return !*timeout ? -ETIME : 1;
7550 * Wait until events become available, if we don't already have some. The
7551 * application must reap them itself, as they reside on the shared cq ring.
7553 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7554 const sigset_t __user *sig, size_t sigsz,
7555 struct __kernel_timespec __user *uts)
7557 struct io_wait_queue iowq;
7558 struct io_rings *rings = ctx->rings;
7559 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7563 io_cqring_overflow_flush(ctx);
7564 if (io_cqring_events(ctx) >= min_events)
7566 if (!io_run_task_work())
7571 struct timespec64 ts;
7573 if (get_timespec64(&ts, uts))
7575 timeout = timespec64_to_jiffies(&ts);
7579 #ifdef CONFIG_COMPAT
7580 if (in_compat_syscall())
7581 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7585 ret = set_user_sigmask(sig, sigsz);
7591 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7592 iowq.wq.private = current;
7593 INIT_LIST_HEAD(&iowq.wq.entry);
7595 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7596 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7598 trace_io_uring_cqring_wait(ctx, min_events);
7600 /* if we can't even flush overflow, don't wait for more */
7601 if (!io_cqring_overflow_flush(ctx)) {
7605 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7606 TASK_INTERRUPTIBLE);
7607 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7608 finish_wait(&ctx->cq_wait, &iowq.wq);
7612 restore_saved_sigmask_unless(ret == -EINTR);
7614 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7617 static void io_free_page_table(void **table, size_t size)
7619 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7621 for (i = 0; i < nr_tables; i++)
7626 static void **io_alloc_page_table(size_t size)
7628 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7629 size_t init_size = size;
7632 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7636 for (i = 0; i < nr_tables; i++) {
7637 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7639 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7641 io_free_page_table(table, init_size);
7649 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7651 percpu_ref_exit(&ref_node->refs);
7655 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7657 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7658 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7659 unsigned long flags;
7660 bool first_add = false;
7662 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7665 while (!list_empty(&ctx->rsrc_ref_list)) {
7666 node = list_first_entry(&ctx->rsrc_ref_list,
7667 struct io_rsrc_node, node);
7668 /* recycle ref nodes in order */
7671 list_del(&node->node);
7672 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7674 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7677 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7680 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7682 struct io_rsrc_node *ref_node;
7684 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7688 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7693 INIT_LIST_HEAD(&ref_node->node);
7694 INIT_LIST_HEAD(&ref_node->rsrc_list);
7695 ref_node->done = false;
7699 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7700 struct io_rsrc_data *data_to_kill)
7702 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7703 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7706 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7708 rsrc_node->rsrc_data = data_to_kill;
7709 spin_lock_irq(&ctx->rsrc_ref_lock);
7710 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7711 spin_unlock_irq(&ctx->rsrc_ref_lock);
7713 atomic_inc(&data_to_kill->refs);
7714 percpu_ref_kill(&rsrc_node->refs);
7715 ctx->rsrc_node = NULL;
7718 if (!ctx->rsrc_node) {
7719 ctx->rsrc_node = ctx->rsrc_backup_node;
7720 ctx->rsrc_backup_node = NULL;
7724 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7726 if (ctx->rsrc_backup_node)
7728 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7729 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7732 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7736 /* As we may drop ->uring_lock, other task may have started quiesce */
7740 data->quiesce = true;
7742 ret = io_rsrc_node_switch_start(ctx);
7745 io_rsrc_node_switch(ctx, data);
7747 /* kill initial ref, already quiesced if zero */
7748 if (atomic_dec_and_test(&data->refs))
7750 mutex_unlock(&ctx->uring_lock);
7751 flush_delayed_work(&ctx->rsrc_put_work);
7752 ret = wait_for_completion_interruptible(&data->done);
7754 mutex_lock(&ctx->uring_lock);
7758 atomic_inc(&data->refs);
7759 /* wait for all works potentially completing data->done */
7760 flush_delayed_work(&ctx->rsrc_put_work);
7761 reinit_completion(&data->done);
7763 ret = io_run_task_work_sig();
7764 mutex_lock(&ctx->uring_lock);
7766 data->quiesce = false;
7771 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7773 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7774 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7776 return &data->tags[table_idx][off];
7779 static void io_rsrc_data_free(struct io_rsrc_data *data)
7781 size_t size = data->nr * sizeof(data->tags[0][0]);
7784 io_free_page_table((void **)data->tags, size);
7788 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7789 u64 __user *utags, unsigned nr,
7790 struct io_rsrc_data **pdata)
7792 struct io_rsrc_data *data;
7796 data = kzalloc(sizeof(*data), GFP_KERNEL);
7799 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7807 data->do_put = do_put;
7810 for (i = 0; i < nr; i++) {
7811 u64 *tag_slot = io_get_tag_slot(data, i);
7813 if (copy_from_user(tag_slot, &utags[i],
7819 atomic_set(&data->refs, 1);
7820 init_completion(&data->done);
7824 io_rsrc_data_free(data);
7828 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7830 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7831 GFP_KERNEL_ACCOUNT);
7832 return !!table->files;
7835 static void io_free_file_tables(struct io_file_table *table)
7837 kvfree(table->files);
7838 table->files = NULL;
7841 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7843 #if defined(CONFIG_UNIX)
7844 if (ctx->ring_sock) {
7845 struct sock *sock = ctx->ring_sock->sk;
7846 struct sk_buff *skb;
7848 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7854 for (i = 0; i < ctx->nr_user_files; i++) {
7857 file = io_file_from_index(ctx, i);
7862 io_free_file_tables(&ctx->file_table);
7863 io_rsrc_data_free(ctx->file_data);
7864 ctx->file_data = NULL;
7865 ctx->nr_user_files = 0;
7868 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7872 if (!ctx->file_data)
7874 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7876 __io_sqe_files_unregister(ctx);
7880 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7881 __releases(&sqd->lock)
7883 WARN_ON_ONCE(sqd->thread == current);
7886 * Do the dance but not conditional clear_bit() because it'd race with
7887 * other threads incrementing park_pending and setting the bit.
7889 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7890 if (atomic_dec_return(&sqd->park_pending))
7891 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7892 mutex_unlock(&sqd->lock);
7895 static void io_sq_thread_park(struct io_sq_data *sqd)
7896 __acquires(&sqd->lock)
7898 WARN_ON_ONCE(sqd->thread == current);
7900 atomic_inc(&sqd->park_pending);
7901 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7902 mutex_lock(&sqd->lock);
7904 wake_up_process(sqd->thread);
7907 static void io_sq_thread_stop(struct io_sq_data *sqd)
7909 WARN_ON_ONCE(sqd->thread == current);
7910 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7912 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7913 mutex_lock(&sqd->lock);
7915 wake_up_process(sqd->thread);
7916 mutex_unlock(&sqd->lock);
7917 wait_for_completion(&sqd->exited);
7920 static void io_put_sq_data(struct io_sq_data *sqd)
7922 if (refcount_dec_and_test(&sqd->refs)) {
7923 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7925 io_sq_thread_stop(sqd);
7930 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7932 struct io_sq_data *sqd = ctx->sq_data;
7935 io_sq_thread_park(sqd);
7936 list_del_init(&ctx->sqd_list);
7937 io_sqd_update_thread_idle(sqd);
7938 io_sq_thread_unpark(sqd);
7940 io_put_sq_data(sqd);
7941 ctx->sq_data = NULL;
7945 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7947 struct io_ring_ctx *ctx_attach;
7948 struct io_sq_data *sqd;
7951 f = fdget(p->wq_fd);
7953 return ERR_PTR(-ENXIO);
7954 if (f.file->f_op != &io_uring_fops) {
7956 return ERR_PTR(-EINVAL);
7959 ctx_attach = f.file->private_data;
7960 sqd = ctx_attach->sq_data;
7963 return ERR_PTR(-EINVAL);
7965 if (sqd->task_tgid != current->tgid) {
7967 return ERR_PTR(-EPERM);
7970 refcount_inc(&sqd->refs);
7975 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7978 struct io_sq_data *sqd;
7981 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7982 sqd = io_attach_sq_data(p);
7987 /* fall through for EPERM case, setup new sqd/task */
7988 if (PTR_ERR(sqd) != -EPERM)
7992 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7994 return ERR_PTR(-ENOMEM);
7996 atomic_set(&sqd->park_pending, 0);
7997 refcount_set(&sqd->refs, 1);
7998 INIT_LIST_HEAD(&sqd->ctx_list);
7999 mutex_init(&sqd->lock);
8000 init_waitqueue_head(&sqd->wait);
8001 init_completion(&sqd->exited);
8005 #if defined(CONFIG_UNIX)
8007 * Ensure the UNIX gc is aware of our file set, so we are certain that
8008 * the io_uring can be safely unregistered on process exit, even if we have
8009 * loops in the file referencing.
8011 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8013 struct sock *sk = ctx->ring_sock->sk;
8014 struct scm_fp_list *fpl;
8015 struct sk_buff *skb;
8018 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8022 skb = alloc_skb(0, GFP_KERNEL);
8031 fpl->user = get_uid(current_user());
8032 for (i = 0; i < nr; i++) {
8033 struct file *file = io_file_from_index(ctx, i + offset);
8037 fpl->fp[nr_files] = get_file(file);
8038 unix_inflight(fpl->user, fpl->fp[nr_files]);
8043 fpl->max = SCM_MAX_FD;
8044 fpl->count = nr_files;
8045 UNIXCB(skb).fp = fpl;
8046 skb->destructor = unix_destruct_scm;
8047 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8048 skb_queue_head(&sk->sk_receive_queue, skb);
8050 for (i = 0; i < nr_files; i++)
8061 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8062 * causes regular reference counting to break down. We rely on the UNIX
8063 * garbage collection to take care of this problem for us.
8065 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8067 unsigned left, total;
8071 left = ctx->nr_user_files;
8073 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8075 ret = __io_sqe_files_scm(ctx, this_files, total);
8079 total += this_files;
8085 while (total < ctx->nr_user_files) {
8086 struct file *file = io_file_from_index(ctx, total);
8096 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8102 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8104 struct file *file = prsrc->file;
8105 #if defined(CONFIG_UNIX)
8106 struct sock *sock = ctx->ring_sock->sk;
8107 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8108 struct sk_buff *skb;
8111 __skb_queue_head_init(&list);
8114 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8115 * remove this entry and rearrange the file array.
8117 skb = skb_dequeue(head);
8119 struct scm_fp_list *fp;
8121 fp = UNIXCB(skb).fp;
8122 for (i = 0; i < fp->count; i++) {
8125 if (fp->fp[i] != file)
8128 unix_notinflight(fp->user, fp->fp[i]);
8129 left = fp->count - 1 - i;
8131 memmove(&fp->fp[i], &fp->fp[i + 1],
8132 left * sizeof(struct file *));
8139 __skb_queue_tail(&list, skb);
8149 __skb_queue_tail(&list, skb);
8151 skb = skb_dequeue(head);
8154 if (skb_peek(&list)) {
8155 spin_lock_irq(&head->lock);
8156 while ((skb = __skb_dequeue(&list)) != NULL)
8157 __skb_queue_tail(head, skb);
8158 spin_unlock_irq(&head->lock);
8165 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8167 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8168 struct io_ring_ctx *ctx = rsrc_data->ctx;
8169 struct io_rsrc_put *prsrc, *tmp;
8171 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8172 list_del(&prsrc->list);
8175 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8177 io_ring_submit_lock(ctx, lock_ring);
8178 spin_lock(&ctx->completion_lock);
8179 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8181 io_commit_cqring(ctx);
8182 spin_unlock(&ctx->completion_lock);
8183 io_cqring_ev_posted(ctx);
8184 io_ring_submit_unlock(ctx, lock_ring);
8187 rsrc_data->do_put(ctx, prsrc);
8191 io_rsrc_node_destroy(ref_node);
8192 if (atomic_dec_and_test(&rsrc_data->refs))
8193 complete(&rsrc_data->done);
8196 static void io_rsrc_put_work(struct work_struct *work)
8198 struct io_ring_ctx *ctx;
8199 struct llist_node *node;
8201 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8202 node = llist_del_all(&ctx->rsrc_put_llist);
8205 struct io_rsrc_node *ref_node;
8206 struct llist_node *next = node->next;
8208 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8209 __io_rsrc_put_work(ref_node);
8214 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8215 unsigned nr_args, u64 __user *tags)
8217 __s32 __user *fds = (__s32 __user *) arg;
8226 if (nr_args > IORING_MAX_FIXED_FILES)
8228 if (nr_args > rlimit(RLIMIT_NOFILE))
8230 ret = io_rsrc_node_switch_start(ctx);
8233 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8239 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8242 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8243 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8247 /* allow sparse sets */
8250 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8257 if (unlikely(!file))
8261 * Don't allow io_uring instances to be registered. If UNIX
8262 * isn't enabled, then this causes a reference cycle and this
8263 * instance can never get freed. If UNIX is enabled we'll
8264 * handle it just fine, but there's still no point in allowing
8265 * a ring fd as it doesn't support regular read/write anyway.
8267 if (file->f_op == &io_uring_fops) {
8271 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8274 ret = io_sqe_files_scm(ctx);
8276 __io_sqe_files_unregister(ctx);
8280 io_rsrc_node_switch(ctx, NULL);
8283 for (i = 0; i < ctx->nr_user_files; i++) {
8284 file = io_file_from_index(ctx, i);
8288 io_free_file_tables(&ctx->file_table);
8289 ctx->nr_user_files = 0;
8291 io_rsrc_data_free(ctx->file_data);
8292 ctx->file_data = NULL;
8296 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8299 #if defined(CONFIG_UNIX)
8300 struct sock *sock = ctx->ring_sock->sk;
8301 struct sk_buff_head *head = &sock->sk_receive_queue;
8302 struct sk_buff *skb;
8305 * See if we can merge this file into an existing skb SCM_RIGHTS
8306 * file set. If there's no room, fall back to allocating a new skb
8307 * and filling it in.
8309 spin_lock_irq(&head->lock);
8310 skb = skb_peek(head);
8312 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8314 if (fpl->count < SCM_MAX_FD) {
8315 __skb_unlink(skb, head);
8316 spin_unlock_irq(&head->lock);
8317 fpl->fp[fpl->count] = get_file(file);
8318 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8320 spin_lock_irq(&head->lock);
8321 __skb_queue_head(head, skb);
8326 spin_unlock_irq(&head->lock);
8333 return __io_sqe_files_scm(ctx, 1, index);
8339 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8340 struct io_rsrc_node *node, void *rsrc)
8342 struct io_rsrc_put *prsrc;
8344 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8348 prsrc->tag = *io_get_tag_slot(data, idx);
8350 list_add(&prsrc->list, &node->rsrc_list);
8354 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8355 unsigned int issue_flags, u32 slot_index)
8357 struct io_ring_ctx *ctx = req->ctx;
8358 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8359 bool needs_switch = false;
8360 struct io_fixed_file *file_slot;
8363 io_ring_submit_lock(ctx, !force_nonblock);
8364 if (file->f_op == &io_uring_fops)
8367 if (!ctx->file_data)
8370 if (slot_index >= ctx->nr_user_files)
8373 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8374 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8376 if (file_slot->file_ptr) {
8377 struct file *old_file;
8379 ret = io_rsrc_node_switch_start(ctx);
8383 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8384 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8385 ctx->rsrc_node, old_file);
8388 file_slot->file_ptr = 0;
8389 needs_switch = true;
8392 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8393 io_fixed_file_set(file_slot, file);
8394 ret = io_sqe_file_register(ctx, file, slot_index);
8396 file_slot->file_ptr = 0;
8403 io_rsrc_node_switch(ctx, ctx->file_data);
8404 io_ring_submit_unlock(ctx, !force_nonblock);
8410 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8412 unsigned int offset = req->close.file_slot - 1;
8413 struct io_ring_ctx *ctx = req->ctx;
8414 struct io_fixed_file *file_slot;
8418 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8420 if (unlikely(!ctx->file_data))
8423 if (offset >= ctx->nr_user_files)
8425 ret = io_rsrc_node_switch_start(ctx);
8429 i = array_index_nospec(offset, ctx->nr_user_files);
8430 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8432 if (!file_slot->file_ptr)
8435 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8436 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8440 file_slot->file_ptr = 0;
8441 io_rsrc_node_switch(ctx, ctx->file_data);
8444 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8448 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8449 struct io_uring_rsrc_update2 *up,
8452 u64 __user *tags = u64_to_user_ptr(up->tags);
8453 __s32 __user *fds = u64_to_user_ptr(up->data);
8454 struct io_rsrc_data *data = ctx->file_data;
8455 struct io_fixed_file *file_slot;
8459 bool needs_switch = false;
8461 if (!ctx->file_data)
8463 if (up->offset + nr_args > ctx->nr_user_files)
8466 for (done = 0; done < nr_args; done++) {
8469 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8470 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8474 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8478 if (fd == IORING_REGISTER_FILES_SKIP)
8481 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8482 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8484 if (file_slot->file_ptr) {
8485 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8486 err = io_queue_rsrc_removal(data, up->offset + done,
8487 ctx->rsrc_node, file);
8490 file_slot->file_ptr = 0;
8491 needs_switch = true;
8500 * Don't allow io_uring instances to be registered. If
8501 * UNIX isn't enabled, then this causes a reference
8502 * cycle and this instance can never get freed. If UNIX
8503 * is enabled we'll handle it just fine, but there's
8504 * still no point in allowing a ring fd as it doesn't
8505 * support regular read/write anyway.
8507 if (file->f_op == &io_uring_fops) {
8512 *io_get_tag_slot(data, up->offset + done) = tag;
8513 io_fixed_file_set(file_slot, file);
8514 err = io_sqe_file_register(ctx, file, i);
8516 file_slot->file_ptr = 0;
8524 io_rsrc_node_switch(ctx, data);
8525 return done ? done : err;
8528 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8529 struct task_struct *task)
8531 struct io_wq_hash *hash;
8532 struct io_wq_data data;
8533 unsigned int concurrency;
8535 mutex_lock(&ctx->uring_lock);
8536 hash = ctx->hash_map;
8538 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8540 mutex_unlock(&ctx->uring_lock);
8541 return ERR_PTR(-ENOMEM);
8543 refcount_set(&hash->refs, 1);
8544 init_waitqueue_head(&hash->wait);
8545 ctx->hash_map = hash;
8547 mutex_unlock(&ctx->uring_lock);
8551 data.free_work = io_wq_free_work;
8552 data.do_work = io_wq_submit_work;
8554 /* Do QD, or 4 * CPUS, whatever is smallest */
8555 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8557 return io_wq_create(concurrency, &data);
8560 static int io_uring_alloc_task_context(struct task_struct *task,
8561 struct io_ring_ctx *ctx)
8563 struct io_uring_task *tctx;
8566 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8567 if (unlikely(!tctx))
8570 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8571 if (unlikely(ret)) {
8576 tctx->io_wq = io_init_wq_offload(ctx, task);
8577 if (IS_ERR(tctx->io_wq)) {
8578 ret = PTR_ERR(tctx->io_wq);
8579 percpu_counter_destroy(&tctx->inflight);
8585 init_waitqueue_head(&tctx->wait);
8586 atomic_set(&tctx->in_idle, 0);
8587 atomic_set(&tctx->inflight_tracked, 0);
8588 task->io_uring = tctx;
8589 spin_lock_init(&tctx->task_lock);
8590 INIT_WQ_LIST(&tctx->task_list);
8591 init_task_work(&tctx->task_work, tctx_task_work);
8595 void __io_uring_free(struct task_struct *tsk)
8597 struct io_uring_task *tctx = tsk->io_uring;
8599 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8600 WARN_ON_ONCE(tctx->io_wq);
8601 WARN_ON_ONCE(tctx->cached_refs);
8603 percpu_counter_destroy(&tctx->inflight);
8605 tsk->io_uring = NULL;
8608 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8609 struct io_uring_params *p)
8613 /* Retain compatibility with failing for an invalid attach attempt */
8614 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8615 IORING_SETUP_ATTACH_WQ) {
8618 f = fdget(p->wq_fd);
8621 if (f.file->f_op != &io_uring_fops) {
8627 if (ctx->flags & IORING_SETUP_SQPOLL) {
8628 struct task_struct *tsk;
8629 struct io_sq_data *sqd;
8632 sqd = io_get_sq_data(p, &attached);
8638 ctx->sq_creds = get_current_cred();
8640 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8641 if (!ctx->sq_thread_idle)
8642 ctx->sq_thread_idle = HZ;
8644 io_sq_thread_park(sqd);
8645 list_add(&ctx->sqd_list, &sqd->ctx_list);
8646 io_sqd_update_thread_idle(sqd);
8647 /* don't attach to a dying SQPOLL thread, would be racy */
8648 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8649 io_sq_thread_unpark(sqd);
8656 if (p->flags & IORING_SETUP_SQ_AFF) {
8657 int cpu = p->sq_thread_cpu;
8660 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8667 sqd->task_pid = current->pid;
8668 sqd->task_tgid = current->tgid;
8669 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8676 ret = io_uring_alloc_task_context(tsk, ctx);
8677 wake_up_new_task(tsk);
8680 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8681 /* Can't have SQ_AFF without SQPOLL */
8688 complete(&ctx->sq_data->exited);
8690 io_sq_thread_finish(ctx);
8694 static inline void __io_unaccount_mem(struct user_struct *user,
8695 unsigned long nr_pages)
8697 atomic_long_sub(nr_pages, &user->locked_vm);
8700 static inline int __io_account_mem(struct user_struct *user,
8701 unsigned long nr_pages)
8703 unsigned long page_limit, cur_pages, new_pages;
8705 /* Don't allow more pages than we can safely lock */
8706 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8709 cur_pages = atomic_long_read(&user->locked_vm);
8710 new_pages = cur_pages + nr_pages;
8711 if (new_pages > page_limit)
8713 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8714 new_pages) != cur_pages);
8719 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8722 __io_unaccount_mem(ctx->user, nr_pages);
8724 if (ctx->mm_account)
8725 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8728 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8733 ret = __io_account_mem(ctx->user, nr_pages);
8738 if (ctx->mm_account)
8739 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8744 static void io_mem_free(void *ptr)
8751 page = virt_to_head_page(ptr);
8752 if (put_page_testzero(page))
8753 free_compound_page(page);
8756 static void *io_mem_alloc(size_t size)
8758 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8759 __GFP_NORETRY | __GFP_ACCOUNT;
8761 return (void *) __get_free_pages(gfp_flags, get_order(size));
8764 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8767 struct io_rings *rings;
8768 size_t off, sq_array_size;
8770 off = struct_size(rings, cqes, cq_entries);
8771 if (off == SIZE_MAX)
8775 off = ALIGN(off, SMP_CACHE_BYTES);
8783 sq_array_size = array_size(sizeof(u32), sq_entries);
8784 if (sq_array_size == SIZE_MAX)
8787 if (check_add_overflow(off, sq_array_size, &off))
8793 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8795 struct io_mapped_ubuf *imu = *slot;
8798 if (imu != ctx->dummy_ubuf) {
8799 for (i = 0; i < imu->nr_bvecs; i++)
8800 unpin_user_page(imu->bvec[i].bv_page);
8801 if (imu->acct_pages)
8802 io_unaccount_mem(ctx, imu->acct_pages);
8808 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8810 io_buffer_unmap(ctx, &prsrc->buf);
8814 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8818 for (i = 0; i < ctx->nr_user_bufs; i++)
8819 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8820 kfree(ctx->user_bufs);
8821 io_rsrc_data_free(ctx->buf_data);
8822 ctx->user_bufs = NULL;
8823 ctx->buf_data = NULL;
8824 ctx->nr_user_bufs = 0;
8827 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8834 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8836 __io_sqe_buffers_unregister(ctx);
8840 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8841 void __user *arg, unsigned index)
8843 struct iovec __user *src;
8845 #ifdef CONFIG_COMPAT
8847 struct compat_iovec __user *ciovs;
8848 struct compat_iovec ciov;
8850 ciovs = (struct compat_iovec __user *) arg;
8851 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8854 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8855 dst->iov_len = ciov.iov_len;
8859 src = (struct iovec __user *) arg;
8860 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8866 * Not super efficient, but this is just a registration time. And we do cache
8867 * the last compound head, so generally we'll only do a full search if we don't
8870 * We check if the given compound head page has already been accounted, to
8871 * avoid double accounting it. This allows us to account the full size of the
8872 * page, not just the constituent pages of a huge page.
8874 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8875 int nr_pages, struct page *hpage)
8879 /* check current page array */
8880 for (i = 0; i < nr_pages; i++) {
8881 if (!PageCompound(pages[i]))
8883 if (compound_head(pages[i]) == hpage)
8887 /* check previously registered pages */
8888 for (i = 0; i < ctx->nr_user_bufs; i++) {
8889 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8891 for (j = 0; j < imu->nr_bvecs; j++) {
8892 if (!PageCompound(imu->bvec[j].bv_page))
8894 if (compound_head(imu->bvec[j].bv_page) == hpage)
8902 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8903 int nr_pages, struct io_mapped_ubuf *imu,
8904 struct page **last_hpage)
8908 imu->acct_pages = 0;
8909 for (i = 0; i < nr_pages; i++) {
8910 if (!PageCompound(pages[i])) {
8915 hpage = compound_head(pages[i]);
8916 if (hpage == *last_hpage)
8918 *last_hpage = hpage;
8919 if (headpage_already_acct(ctx, pages, i, hpage))
8921 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8925 if (!imu->acct_pages)
8928 ret = io_account_mem(ctx, imu->acct_pages);
8930 imu->acct_pages = 0;
8934 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8935 struct io_mapped_ubuf **pimu,
8936 struct page **last_hpage)
8938 struct io_mapped_ubuf *imu = NULL;
8939 struct vm_area_struct **vmas = NULL;
8940 struct page **pages = NULL;
8941 unsigned long off, start, end, ubuf;
8943 int ret, pret, nr_pages, i;
8945 if (!iov->iov_base) {
8946 *pimu = ctx->dummy_ubuf;
8950 ubuf = (unsigned long) iov->iov_base;
8951 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8952 start = ubuf >> PAGE_SHIFT;
8953 nr_pages = end - start;
8958 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8962 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8967 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8972 mmap_read_lock(current->mm);
8973 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8975 if (pret == nr_pages) {
8976 /* don't support file backed memory */
8977 for (i = 0; i < nr_pages; i++) {
8978 struct vm_area_struct *vma = vmas[i];
8980 if (vma_is_shmem(vma))
8983 !is_file_hugepages(vma->vm_file)) {
8989 ret = pret < 0 ? pret : -EFAULT;
8991 mmap_read_unlock(current->mm);
8994 * if we did partial map, or found file backed vmas,
8995 * release any pages we did get
8998 unpin_user_pages(pages, pret);
9002 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9004 unpin_user_pages(pages, pret);
9008 off = ubuf & ~PAGE_MASK;
9009 size = iov->iov_len;
9010 for (i = 0; i < nr_pages; i++) {
9013 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9014 imu->bvec[i].bv_page = pages[i];
9015 imu->bvec[i].bv_len = vec_len;
9016 imu->bvec[i].bv_offset = off;
9020 /* store original address for later verification */
9022 imu->ubuf_end = ubuf + iov->iov_len;
9023 imu->nr_bvecs = nr_pages;
9034 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9036 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9037 return ctx->user_bufs ? 0 : -ENOMEM;
9040 static int io_buffer_validate(struct iovec *iov)
9042 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9045 * Don't impose further limits on the size and buffer
9046 * constraints here, we'll -EINVAL later when IO is
9047 * submitted if they are wrong.
9050 return iov->iov_len ? -EFAULT : 0;
9054 /* arbitrary limit, but we need something */
9055 if (iov->iov_len > SZ_1G)
9058 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9064 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9065 unsigned int nr_args, u64 __user *tags)
9067 struct page *last_hpage = NULL;
9068 struct io_rsrc_data *data;
9074 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9076 ret = io_rsrc_node_switch_start(ctx);
9079 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9082 ret = io_buffers_map_alloc(ctx, nr_args);
9084 io_rsrc_data_free(data);
9088 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9089 ret = io_copy_iov(ctx, &iov, arg, i);
9092 ret = io_buffer_validate(&iov);
9095 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9100 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9106 WARN_ON_ONCE(ctx->buf_data);
9108 ctx->buf_data = data;
9110 __io_sqe_buffers_unregister(ctx);
9112 io_rsrc_node_switch(ctx, NULL);
9116 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9117 struct io_uring_rsrc_update2 *up,
9118 unsigned int nr_args)
9120 u64 __user *tags = u64_to_user_ptr(up->tags);
9121 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9122 struct page *last_hpage = NULL;
9123 bool needs_switch = false;
9129 if (up->offset + nr_args > ctx->nr_user_bufs)
9132 for (done = 0; done < nr_args; done++) {
9133 struct io_mapped_ubuf *imu;
9134 int offset = up->offset + done;
9137 err = io_copy_iov(ctx, &iov, iovs, done);
9140 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9144 err = io_buffer_validate(&iov);
9147 if (!iov.iov_base && tag) {
9151 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9155 i = array_index_nospec(offset, ctx->nr_user_bufs);
9156 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9157 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9158 ctx->rsrc_node, ctx->user_bufs[i]);
9159 if (unlikely(err)) {
9160 io_buffer_unmap(ctx, &imu);
9163 ctx->user_bufs[i] = NULL;
9164 needs_switch = true;
9167 ctx->user_bufs[i] = imu;
9168 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9172 io_rsrc_node_switch(ctx, ctx->buf_data);
9173 return done ? done : err;
9176 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9178 __s32 __user *fds = arg;
9184 if (copy_from_user(&fd, fds, sizeof(*fds)))
9187 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9188 if (IS_ERR(ctx->cq_ev_fd)) {
9189 int ret = PTR_ERR(ctx->cq_ev_fd);
9191 ctx->cq_ev_fd = NULL;
9198 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9200 if (ctx->cq_ev_fd) {
9201 eventfd_ctx_put(ctx->cq_ev_fd);
9202 ctx->cq_ev_fd = NULL;
9209 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9211 struct io_buffer *buf;
9212 unsigned long index;
9214 xa_for_each(&ctx->io_buffers, index, buf) {
9215 __io_remove_buffers(ctx, buf, index, -1U);
9220 static void io_req_cache_free(struct list_head *list)
9222 struct io_kiocb *req, *nxt;
9224 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9225 list_del(&req->inflight_entry);
9226 kmem_cache_free(req_cachep, req);
9230 static void io_req_caches_free(struct io_ring_ctx *ctx)
9232 struct io_submit_state *state = &ctx->submit_state;
9234 mutex_lock(&ctx->uring_lock);
9236 if (state->free_reqs) {
9237 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9238 state->free_reqs = 0;
9241 io_flush_cached_locked_reqs(ctx, state);
9242 io_req_cache_free(&state->free_list);
9243 mutex_unlock(&ctx->uring_lock);
9246 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9248 if (data && !atomic_dec_and_test(&data->refs))
9249 wait_for_completion(&data->done);
9252 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9254 io_sq_thread_finish(ctx);
9256 if (ctx->mm_account) {
9257 mmdrop(ctx->mm_account);
9258 ctx->mm_account = NULL;
9261 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9262 io_wait_rsrc_data(ctx->buf_data);
9263 io_wait_rsrc_data(ctx->file_data);
9265 mutex_lock(&ctx->uring_lock);
9267 __io_sqe_buffers_unregister(ctx);
9269 __io_sqe_files_unregister(ctx);
9271 __io_cqring_overflow_flush(ctx, true);
9272 mutex_unlock(&ctx->uring_lock);
9273 io_eventfd_unregister(ctx);
9274 io_destroy_buffers(ctx);
9276 put_cred(ctx->sq_creds);
9278 /* there are no registered resources left, nobody uses it */
9280 io_rsrc_node_destroy(ctx->rsrc_node);
9281 if (ctx->rsrc_backup_node)
9282 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9283 flush_delayed_work(&ctx->rsrc_put_work);
9285 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9286 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9288 #if defined(CONFIG_UNIX)
9289 if (ctx->ring_sock) {
9290 ctx->ring_sock->file = NULL; /* so that iput() is called */
9291 sock_release(ctx->ring_sock);
9294 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9296 io_mem_free(ctx->rings);
9297 io_mem_free(ctx->sq_sqes);
9299 percpu_ref_exit(&ctx->refs);
9300 free_uid(ctx->user);
9301 io_req_caches_free(ctx);
9303 io_wq_put_hash(ctx->hash_map);
9304 kfree(ctx->cancel_hash);
9305 kfree(ctx->dummy_ubuf);
9309 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9311 struct io_ring_ctx *ctx = file->private_data;
9314 poll_wait(file, &ctx->poll_wait, wait);
9316 * synchronizes with barrier from wq_has_sleeper call in
9320 if (!io_sqring_full(ctx))
9321 mask |= EPOLLOUT | EPOLLWRNORM;
9324 * Don't flush cqring overflow list here, just do a simple check.
9325 * Otherwise there could possible be ABBA deadlock:
9328 * lock(&ctx->uring_lock);
9330 * lock(&ctx->uring_lock);
9333 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9334 * pushs them to do the flush.
9336 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9337 mask |= EPOLLIN | EPOLLRDNORM;
9342 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9344 const struct cred *creds;
9346 creds = xa_erase(&ctx->personalities, id);
9355 struct io_tctx_exit {
9356 struct callback_head task_work;
9357 struct completion completion;
9358 struct io_ring_ctx *ctx;
9361 static void io_tctx_exit_cb(struct callback_head *cb)
9363 struct io_uring_task *tctx = current->io_uring;
9364 struct io_tctx_exit *work;
9366 work = container_of(cb, struct io_tctx_exit, task_work);
9368 * When @in_idle, we're in cancellation and it's racy to remove the
9369 * node. It'll be removed by the end of cancellation, just ignore it.
9371 if (!atomic_read(&tctx->in_idle))
9372 io_uring_del_tctx_node((unsigned long)work->ctx);
9373 complete(&work->completion);
9376 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9378 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9380 return req->ctx == data;
9383 static void io_ring_exit_work(struct work_struct *work)
9385 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9386 unsigned long timeout = jiffies + HZ * 60 * 5;
9387 unsigned long interval = HZ / 20;
9388 struct io_tctx_exit exit;
9389 struct io_tctx_node *node;
9393 * If we're doing polled IO and end up having requests being
9394 * submitted async (out-of-line), then completions can come in while
9395 * we're waiting for refs to drop. We need to reap these manually,
9396 * as nobody else will be looking for them.
9399 io_uring_try_cancel_requests(ctx, NULL, true);
9401 struct io_sq_data *sqd = ctx->sq_data;
9402 struct task_struct *tsk;
9404 io_sq_thread_park(sqd);
9406 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9407 io_wq_cancel_cb(tsk->io_uring->io_wq,
9408 io_cancel_ctx_cb, ctx, true);
9409 io_sq_thread_unpark(sqd);
9412 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9413 /* there is little hope left, don't run it too often */
9416 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9418 init_completion(&exit.completion);
9419 init_task_work(&exit.task_work, io_tctx_exit_cb);
9422 * Some may use context even when all refs and requests have been put,
9423 * and they are free to do so while still holding uring_lock or
9424 * completion_lock, see io_req_task_submit(). Apart from other work,
9425 * this lock/unlock section also waits them to finish.
9427 mutex_lock(&ctx->uring_lock);
9428 while (!list_empty(&ctx->tctx_list)) {
9429 WARN_ON_ONCE(time_after(jiffies, timeout));
9431 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9433 /* don't spin on a single task if cancellation failed */
9434 list_rotate_left(&ctx->tctx_list);
9435 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9436 if (WARN_ON_ONCE(ret))
9438 wake_up_process(node->task);
9440 mutex_unlock(&ctx->uring_lock);
9441 wait_for_completion(&exit.completion);
9442 mutex_lock(&ctx->uring_lock);
9444 mutex_unlock(&ctx->uring_lock);
9445 spin_lock(&ctx->completion_lock);
9446 spin_unlock(&ctx->completion_lock);
9448 io_ring_ctx_free(ctx);
9451 /* Returns true if we found and killed one or more timeouts */
9452 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9455 struct io_kiocb *req, *tmp;
9458 spin_lock(&ctx->completion_lock);
9459 spin_lock_irq(&ctx->timeout_lock);
9460 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9461 if (io_match_task(req, tsk, cancel_all)) {
9462 io_kill_timeout(req, -ECANCELED);
9466 spin_unlock_irq(&ctx->timeout_lock);
9468 io_commit_cqring(ctx);
9469 spin_unlock(&ctx->completion_lock);
9471 io_cqring_ev_posted(ctx);
9472 return canceled != 0;
9475 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9477 unsigned long index;
9478 struct creds *creds;
9480 mutex_lock(&ctx->uring_lock);
9481 percpu_ref_kill(&ctx->refs);
9483 __io_cqring_overflow_flush(ctx, true);
9484 xa_for_each(&ctx->personalities, index, creds)
9485 io_unregister_personality(ctx, index);
9486 mutex_unlock(&ctx->uring_lock);
9488 io_kill_timeouts(ctx, NULL, true);
9489 io_poll_remove_all(ctx, NULL, true);
9491 /* if we failed setting up the ctx, we might not have any rings */
9492 io_iopoll_try_reap_events(ctx);
9494 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9496 * Use system_unbound_wq to avoid spawning tons of event kworkers
9497 * if we're exiting a ton of rings at the same time. It just adds
9498 * noise and overhead, there's no discernable change in runtime
9499 * over using system_wq.
9501 queue_work(system_unbound_wq, &ctx->exit_work);
9504 static int io_uring_release(struct inode *inode, struct file *file)
9506 struct io_ring_ctx *ctx = file->private_data;
9508 file->private_data = NULL;
9509 io_ring_ctx_wait_and_kill(ctx);
9513 struct io_task_cancel {
9514 struct task_struct *task;
9518 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9520 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9521 struct io_task_cancel *cancel = data;
9524 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9525 struct io_ring_ctx *ctx = req->ctx;
9527 /* protect against races with linked timeouts */
9528 spin_lock_irq(&ctx->timeout_lock);
9529 ret = io_match_task(req, cancel->task, cancel->all);
9530 spin_unlock_irq(&ctx->timeout_lock);
9532 ret = io_match_task(req, cancel->task, cancel->all);
9537 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9538 struct task_struct *task, bool cancel_all)
9540 struct io_defer_entry *de;
9543 spin_lock(&ctx->completion_lock);
9544 spin_lock_irq(&ctx->timeout_lock);
9545 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9546 if (io_match_task(de->req, task, cancel_all)) {
9547 list_cut_position(&list, &ctx->defer_list, &de->list);
9551 spin_unlock_irq(&ctx->timeout_lock);
9552 spin_unlock(&ctx->completion_lock);
9553 if (list_empty(&list))
9556 while (!list_empty(&list)) {
9557 de = list_first_entry(&list, struct io_defer_entry, list);
9558 list_del_init(&de->list);
9559 io_req_complete_failed(de->req, -ECANCELED);
9565 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9567 struct io_tctx_node *node;
9568 enum io_wq_cancel cret;
9571 mutex_lock(&ctx->uring_lock);
9572 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9573 struct io_uring_task *tctx = node->task->io_uring;
9576 * io_wq will stay alive while we hold uring_lock, because it's
9577 * killed after ctx nodes, which requires to take the lock.
9579 if (!tctx || !tctx->io_wq)
9581 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9582 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9584 mutex_unlock(&ctx->uring_lock);
9589 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9590 struct task_struct *task,
9593 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9594 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9597 enum io_wq_cancel cret;
9601 ret |= io_uring_try_cancel_iowq(ctx);
9602 } else if (tctx && tctx->io_wq) {
9604 * Cancels requests of all rings, not only @ctx, but
9605 * it's fine as the task is in exit/exec.
9607 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9609 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9612 /* SQPOLL thread does its own polling */
9613 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9614 (ctx->sq_data && ctx->sq_data->thread == current)) {
9615 while (!list_empty_careful(&ctx->iopoll_list)) {
9616 io_iopoll_try_reap_events(ctx);
9621 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9622 ret |= io_poll_remove_all(ctx, task, cancel_all);
9623 ret |= io_kill_timeouts(ctx, task, cancel_all);
9625 ret |= io_run_task_work();
9632 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9634 struct io_uring_task *tctx = current->io_uring;
9635 struct io_tctx_node *node;
9638 if (unlikely(!tctx)) {
9639 ret = io_uring_alloc_task_context(current, ctx);
9643 tctx = current->io_uring;
9644 if (ctx->iowq_limits_set) {
9645 unsigned int limits[2] = { ctx->iowq_limits[0],
9646 ctx->iowq_limits[1], };
9648 ret = io_wq_max_workers(tctx->io_wq, limits);
9653 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9654 node = kmalloc(sizeof(*node), GFP_KERNEL);
9658 node->task = current;
9660 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9667 mutex_lock(&ctx->uring_lock);
9668 list_add(&node->ctx_node, &ctx->tctx_list);
9669 mutex_unlock(&ctx->uring_lock);
9676 * Note that this task has used io_uring. We use it for cancelation purposes.
9678 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9680 struct io_uring_task *tctx = current->io_uring;
9682 if (likely(tctx && tctx->last == ctx))
9684 return __io_uring_add_tctx_node(ctx);
9688 * Remove this io_uring_file -> task mapping.
9690 static void io_uring_del_tctx_node(unsigned long index)
9692 struct io_uring_task *tctx = current->io_uring;
9693 struct io_tctx_node *node;
9697 node = xa_erase(&tctx->xa, index);
9701 WARN_ON_ONCE(current != node->task);
9702 WARN_ON_ONCE(list_empty(&node->ctx_node));
9704 mutex_lock(&node->ctx->uring_lock);
9705 list_del(&node->ctx_node);
9706 mutex_unlock(&node->ctx->uring_lock);
9708 if (tctx->last == node->ctx)
9713 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9715 struct io_wq *wq = tctx->io_wq;
9716 struct io_tctx_node *node;
9717 unsigned long index;
9719 xa_for_each(&tctx->xa, index, node) {
9720 io_uring_del_tctx_node(index);
9725 * Must be after io_uring_del_task_file() (removes nodes under
9726 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9728 io_wq_put_and_exit(wq);
9733 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9736 return atomic_read(&tctx->inflight_tracked);
9737 return percpu_counter_sum(&tctx->inflight);
9740 static void io_uring_drop_tctx_refs(struct task_struct *task)
9742 struct io_uring_task *tctx = task->io_uring;
9743 unsigned int refs = tctx->cached_refs;
9746 tctx->cached_refs = 0;
9747 percpu_counter_sub(&tctx->inflight, refs);
9748 put_task_struct_many(task, refs);
9753 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9754 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9756 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9758 struct io_uring_task *tctx = current->io_uring;
9759 struct io_ring_ctx *ctx;
9763 WARN_ON_ONCE(sqd && sqd->thread != current);
9765 if (!current->io_uring)
9768 io_wq_exit_start(tctx->io_wq);
9770 atomic_inc(&tctx->in_idle);
9772 io_uring_drop_tctx_refs(current);
9773 /* read completions before cancelations */
9774 inflight = tctx_inflight(tctx, !cancel_all);
9779 struct io_tctx_node *node;
9780 unsigned long index;
9782 xa_for_each(&tctx->xa, index, node) {
9783 /* sqpoll task will cancel all its requests */
9784 if (node->ctx->sq_data)
9786 io_uring_try_cancel_requests(node->ctx, current,
9790 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9791 io_uring_try_cancel_requests(ctx, current,
9795 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9796 io_uring_drop_tctx_refs(current);
9798 * If we've seen completions, retry without waiting. This
9799 * avoids a race where a completion comes in before we did
9800 * prepare_to_wait().
9802 if (inflight == tctx_inflight(tctx, !cancel_all))
9804 finish_wait(&tctx->wait, &wait);
9806 atomic_dec(&tctx->in_idle);
9808 io_uring_clean_tctx(tctx);
9810 /* for exec all current's requests should be gone, kill tctx */
9811 __io_uring_free(current);
9815 void __io_uring_cancel(bool cancel_all)
9817 io_uring_cancel_generic(cancel_all, NULL);
9820 static void *io_uring_validate_mmap_request(struct file *file,
9821 loff_t pgoff, size_t sz)
9823 struct io_ring_ctx *ctx = file->private_data;
9824 loff_t offset = pgoff << PAGE_SHIFT;
9829 case IORING_OFF_SQ_RING:
9830 case IORING_OFF_CQ_RING:
9833 case IORING_OFF_SQES:
9837 return ERR_PTR(-EINVAL);
9840 page = virt_to_head_page(ptr);
9841 if (sz > page_size(page))
9842 return ERR_PTR(-EINVAL);
9849 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9851 size_t sz = vma->vm_end - vma->vm_start;
9855 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9857 return PTR_ERR(ptr);
9859 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9860 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9863 #else /* !CONFIG_MMU */
9865 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9867 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9870 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9872 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9875 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9876 unsigned long addr, unsigned long len,
9877 unsigned long pgoff, unsigned long flags)
9881 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9883 return PTR_ERR(ptr);
9885 return (unsigned long) ptr;
9888 #endif /* !CONFIG_MMU */
9890 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9895 if (!io_sqring_full(ctx))
9897 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9899 if (!io_sqring_full(ctx))
9902 } while (!signal_pending(current));
9904 finish_wait(&ctx->sqo_sq_wait, &wait);
9908 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9909 struct __kernel_timespec __user **ts,
9910 const sigset_t __user **sig)
9912 struct io_uring_getevents_arg arg;
9915 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9916 * is just a pointer to the sigset_t.
9918 if (!(flags & IORING_ENTER_EXT_ARG)) {
9919 *sig = (const sigset_t __user *) argp;
9925 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9926 * timespec and sigset_t pointers if good.
9928 if (*argsz != sizeof(arg))
9930 if (copy_from_user(&arg, argp, sizeof(arg)))
9932 *sig = u64_to_user_ptr(arg.sigmask);
9933 *argsz = arg.sigmask_sz;
9934 *ts = u64_to_user_ptr(arg.ts);
9938 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9939 u32, min_complete, u32, flags, const void __user *, argp,
9942 struct io_ring_ctx *ctx;
9949 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9950 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9954 if (unlikely(!f.file))
9958 if (unlikely(f.file->f_op != &io_uring_fops))
9962 ctx = f.file->private_data;
9963 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9967 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9971 * For SQ polling, the thread will do all submissions and completions.
9972 * Just return the requested submit count, and wake the thread if
9976 if (ctx->flags & IORING_SETUP_SQPOLL) {
9977 io_cqring_overflow_flush(ctx);
9979 if (unlikely(ctx->sq_data->thread == NULL)) {
9983 if (flags & IORING_ENTER_SQ_WAKEUP)
9984 wake_up(&ctx->sq_data->wait);
9985 if (flags & IORING_ENTER_SQ_WAIT) {
9986 ret = io_sqpoll_wait_sq(ctx);
9990 submitted = to_submit;
9991 } else if (to_submit) {
9992 ret = io_uring_add_tctx_node(ctx);
9995 mutex_lock(&ctx->uring_lock);
9996 submitted = io_submit_sqes(ctx, to_submit);
9997 mutex_unlock(&ctx->uring_lock);
9999 if (submitted != to_submit)
10002 if (flags & IORING_ENTER_GETEVENTS) {
10003 const sigset_t __user *sig;
10004 struct __kernel_timespec __user *ts;
10006 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10010 min_complete = min(min_complete, ctx->cq_entries);
10013 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10014 * space applications don't need to do io completion events
10015 * polling again, they can rely on io_sq_thread to do polling
10016 * work, which can reduce cpu usage and uring_lock contention.
10018 if (ctx->flags & IORING_SETUP_IOPOLL &&
10019 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10020 ret = io_iopoll_check(ctx, min_complete);
10022 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10027 percpu_ref_put(&ctx->refs);
10030 return submitted ? submitted : ret;
10033 #ifdef CONFIG_PROC_FS
10034 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10035 const struct cred *cred)
10037 struct user_namespace *uns = seq_user_ns(m);
10038 struct group_info *gi;
10043 seq_printf(m, "%5d\n", id);
10044 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10045 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10046 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10047 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10048 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10049 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10050 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10051 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10052 seq_puts(m, "\n\tGroups:\t");
10053 gi = cred->group_info;
10054 for (g = 0; g < gi->ngroups; g++) {
10055 seq_put_decimal_ull(m, g ? " " : "",
10056 from_kgid_munged(uns, gi->gid[g]));
10058 seq_puts(m, "\n\tCapEff:\t");
10059 cap = cred->cap_effective;
10060 CAP_FOR_EACH_U32(__capi)
10061 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10066 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10068 struct io_sq_data *sq = NULL;
10073 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10074 * since fdinfo case grabs it in the opposite direction of normal use
10075 * cases. If we fail to get the lock, we just don't iterate any
10076 * structures that could be going away outside the io_uring mutex.
10078 has_lock = mutex_trylock(&ctx->uring_lock);
10080 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10086 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10087 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10088 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10089 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10090 struct file *f = io_file_from_index(ctx, i);
10093 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10095 seq_printf(m, "%5u: <none>\n", i);
10097 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10098 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10099 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10100 unsigned int len = buf->ubuf_end - buf->ubuf;
10102 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10104 if (has_lock && !xa_empty(&ctx->personalities)) {
10105 unsigned long index;
10106 const struct cred *cred;
10108 seq_printf(m, "Personalities:\n");
10109 xa_for_each(&ctx->personalities, index, cred)
10110 io_uring_show_cred(m, index, cred);
10112 seq_printf(m, "PollList:\n");
10113 spin_lock(&ctx->completion_lock);
10114 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10115 struct hlist_head *list = &ctx->cancel_hash[i];
10116 struct io_kiocb *req;
10118 hlist_for_each_entry(req, list, hash_node)
10119 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10120 req->task->task_works != NULL);
10122 spin_unlock(&ctx->completion_lock);
10124 mutex_unlock(&ctx->uring_lock);
10127 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10129 struct io_ring_ctx *ctx = f->private_data;
10131 if (percpu_ref_tryget(&ctx->refs)) {
10132 __io_uring_show_fdinfo(ctx, m);
10133 percpu_ref_put(&ctx->refs);
10138 static const struct file_operations io_uring_fops = {
10139 .release = io_uring_release,
10140 .mmap = io_uring_mmap,
10142 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10143 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10145 .poll = io_uring_poll,
10146 #ifdef CONFIG_PROC_FS
10147 .show_fdinfo = io_uring_show_fdinfo,
10151 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10152 struct io_uring_params *p)
10154 struct io_rings *rings;
10155 size_t size, sq_array_offset;
10157 /* make sure these are sane, as we already accounted them */
10158 ctx->sq_entries = p->sq_entries;
10159 ctx->cq_entries = p->cq_entries;
10161 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10162 if (size == SIZE_MAX)
10165 rings = io_mem_alloc(size);
10169 ctx->rings = rings;
10170 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10171 rings->sq_ring_mask = p->sq_entries - 1;
10172 rings->cq_ring_mask = p->cq_entries - 1;
10173 rings->sq_ring_entries = p->sq_entries;
10174 rings->cq_ring_entries = p->cq_entries;
10176 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10177 if (size == SIZE_MAX) {
10178 io_mem_free(ctx->rings);
10183 ctx->sq_sqes = io_mem_alloc(size);
10184 if (!ctx->sq_sqes) {
10185 io_mem_free(ctx->rings);
10193 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10197 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10201 ret = io_uring_add_tctx_node(ctx);
10206 fd_install(fd, file);
10211 * Allocate an anonymous fd, this is what constitutes the application
10212 * visible backing of an io_uring instance. The application mmaps this
10213 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10214 * we have to tie this fd to a socket for file garbage collection purposes.
10216 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10219 #if defined(CONFIG_UNIX)
10222 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10225 return ERR_PTR(ret);
10228 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10229 O_RDWR | O_CLOEXEC);
10230 #if defined(CONFIG_UNIX)
10231 if (IS_ERR(file)) {
10232 sock_release(ctx->ring_sock);
10233 ctx->ring_sock = NULL;
10235 ctx->ring_sock->file = file;
10241 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10242 struct io_uring_params __user *params)
10244 struct io_ring_ctx *ctx;
10250 if (entries > IORING_MAX_ENTRIES) {
10251 if (!(p->flags & IORING_SETUP_CLAMP))
10253 entries = IORING_MAX_ENTRIES;
10257 * Use twice as many entries for the CQ ring. It's possible for the
10258 * application to drive a higher depth than the size of the SQ ring,
10259 * since the sqes are only used at submission time. This allows for
10260 * some flexibility in overcommitting a bit. If the application has
10261 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10262 * of CQ ring entries manually.
10264 p->sq_entries = roundup_pow_of_two(entries);
10265 if (p->flags & IORING_SETUP_CQSIZE) {
10267 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10268 * to a power-of-two, if it isn't already. We do NOT impose
10269 * any cq vs sq ring sizing.
10271 if (!p->cq_entries)
10273 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10274 if (!(p->flags & IORING_SETUP_CLAMP))
10276 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10278 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10279 if (p->cq_entries < p->sq_entries)
10282 p->cq_entries = 2 * p->sq_entries;
10285 ctx = io_ring_ctx_alloc(p);
10288 ctx->compat = in_compat_syscall();
10289 if (!capable(CAP_IPC_LOCK))
10290 ctx->user = get_uid(current_user());
10293 * This is just grabbed for accounting purposes. When a process exits,
10294 * the mm is exited and dropped before the files, hence we need to hang
10295 * on to this mm purely for the purposes of being able to unaccount
10296 * memory (locked/pinned vm). It's not used for anything else.
10298 mmgrab(current->mm);
10299 ctx->mm_account = current->mm;
10301 ret = io_allocate_scq_urings(ctx, p);
10305 ret = io_sq_offload_create(ctx, p);
10308 /* always set a rsrc node */
10309 ret = io_rsrc_node_switch_start(ctx);
10312 io_rsrc_node_switch(ctx, NULL);
10314 memset(&p->sq_off, 0, sizeof(p->sq_off));
10315 p->sq_off.head = offsetof(struct io_rings, sq.head);
10316 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10317 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10318 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10319 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10320 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10321 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10323 memset(&p->cq_off, 0, sizeof(p->cq_off));
10324 p->cq_off.head = offsetof(struct io_rings, cq.head);
10325 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10326 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10327 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10328 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10329 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10330 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10332 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10333 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10334 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10335 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10336 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10337 IORING_FEAT_RSRC_TAGS;
10339 if (copy_to_user(params, p, sizeof(*p))) {
10344 file = io_uring_get_file(ctx);
10345 if (IS_ERR(file)) {
10346 ret = PTR_ERR(file);
10351 * Install ring fd as the very last thing, so we don't risk someone
10352 * having closed it before we finish setup
10354 ret = io_uring_install_fd(ctx, file);
10356 /* fput will clean it up */
10361 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10364 io_ring_ctx_wait_and_kill(ctx);
10369 * Sets up an aio uring context, and returns the fd. Applications asks for a
10370 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10371 * params structure passed in.
10373 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10375 struct io_uring_params p;
10378 if (copy_from_user(&p, params, sizeof(p)))
10380 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10385 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10386 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10387 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10388 IORING_SETUP_R_DISABLED))
10391 return io_uring_create(entries, &p, params);
10394 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10395 struct io_uring_params __user *, params)
10397 return io_uring_setup(entries, params);
10400 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10402 struct io_uring_probe *p;
10406 size = struct_size(p, ops, nr_args);
10407 if (size == SIZE_MAX)
10409 p = kzalloc(size, GFP_KERNEL);
10414 if (copy_from_user(p, arg, size))
10417 if (memchr_inv(p, 0, size))
10420 p->last_op = IORING_OP_LAST - 1;
10421 if (nr_args > IORING_OP_LAST)
10422 nr_args = IORING_OP_LAST;
10424 for (i = 0; i < nr_args; i++) {
10426 if (!io_op_defs[i].not_supported)
10427 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10432 if (copy_to_user(arg, p, size))
10439 static int io_register_personality(struct io_ring_ctx *ctx)
10441 const struct cred *creds;
10445 creds = get_current_cred();
10447 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10448 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10456 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10457 unsigned int nr_args)
10459 struct io_uring_restriction *res;
10463 /* Restrictions allowed only if rings started disabled */
10464 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10467 /* We allow only a single restrictions registration */
10468 if (ctx->restrictions.registered)
10471 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10474 size = array_size(nr_args, sizeof(*res));
10475 if (size == SIZE_MAX)
10478 res = memdup_user(arg, size);
10480 return PTR_ERR(res);
10484 for (i = 0; i < nr_args; i++) {
10485 switch (res[i].opcode) {
10486 case IORING_RESTRICTION_REGISTER_OP:
10487 if (res[i].register_op >= IORING_REGISTER_LAST) {
10492 __set_bit(res[i].register_op,
10493 ctx->restrictions.register_op);
10495 case IORING_RESTRICTION_SQE_OP:
10496 if (res[i].sqe_op >= IORING_OP_LAST) {
10501 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10503 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10504 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10506 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10507 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10516 /* Reset all restrictions if an error happened */
10518 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10520 ctx->restrictions.registered = true;
10526 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10528 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10531 if (ctx->restrictions.registered)
10532 ctx->restricted = 1;
10534 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10535 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10536 wake_up(&ctx->sq_data->wait);
10540 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10541 struct io_uring_rsrc_update2 *up,
10549 if (check_add_overflow(up->offset, nr_args, &tmp))
10551 err = io_rsrc_node_switch_start(ctx);
10556 case IORING_RSRC_FILE:
10557 return __io_sqe_files_update(ctx, up, nr_args);
10558 case IORING_RSRC_BUFFER:
10559 return __io_sqe_buffers_update(ctx, up, nr_args);
10564 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10567 struct io_uring_rsrc_update2 up;
10571 memset(&up, 0, sizeof(up));
10572 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10574 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10577 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10578 unsigned size, unsigned type)
10580 struct io_uring_rsrc_update2 up;
10582 if (size != sizeof(up))
10584 if (copy_from_user(&up, arg, sizeof(up)))
10586 if (!up.nr || up.resv)
10588 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10591 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10592 unsigned int size, unsigned int type)
10594 struct io_uring_rsrc_register rr;
10596 /* keep it extendible */
10597 if (size != sizeof(rr))
10600 memset(&rr, 0, sizeof(rr));
10601 if (copy_from_user(&rr, arg, size))
10603 if (!rr.nr || rr.resv || rr.resv2)
10607 case IORING_RSRC_FILE:
10608 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10609 rr.nr, u64_to_user_ptr(rr.tags));
10610 case IORING_RSRC_BUFFER:
10611 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10612 rr.nr, u64_to_user_ptr(rr.tags));
10617 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10620 struct io_uring_task *tctx = current->io_uring;
10621 cpumask_var_t new_mask;
10624 if (!tctx || !tctx->io_wq)
10627 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10630 cpumask_clear(new_mask);
10631 if (len > cpumask_size())
10632 len = cpumask_size();
10634 if (copy_from_user(new_mask, arg, len)) {
10635 free_cpumask_var(new_mask);
10639 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10640 free_cpumask_var(new_mask);
10644 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10646 struct io_uring_task *tctx = current->io_uring;
10648 if (!tctx || !tctx->io_wq)
10651 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10654 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10656 __must_hold(&ctx->uring_lock)
10658 struct io_tctx_node *node;
10659 struct io_uring_task *tctx = NULL;
10660 struct io_sq_data *sqd = NULL;
10661 __u32 new_count[2];
10664 if (copy_from_user(new_count, arg, sizeof(new_count)))
10666 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10667 if (new_count[i] > INT_MAX)
10670 if (ctx->flags & IORING_SETUP_SQPOLL) {
10671 sqd = ctx->sq_data;
10674 * Observe the correct sqd->lock -> ctx->uring_lock
10675 * ordering. Fine to drop uring_lock here, we hold
10676 * a ref to the ctx.
10678 refcount_inc(&sqd->refs);
10679 mutex_unlock(&ctx->uring_lock);
10680 mutex_lock(&sqd->lock);
10681 mutex_lock(&ctx->uring_lock);
10683 tctx = sqd->thread->io_uring;
10686 tctx = current->io_uring;
10689 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10691 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10693 ctx->iowq_limits[i] = new_count[i];
10694 ctx->iowq_limits_set = true;
10697 if (tctx && tctx->io_wq) {
10698 ret = io_wq_max_workers(tctx->io_wq, new_count);
10702 memset(new_count, 0, sizeof(new_count));
10706 mutex_unlock(&sqd->lock);
10707 io_put_sq_data(sqd);
10710 if (copy_to_user(arg, new_count, sizeof(new_count)))
10713 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10717 /* now propagate the restriction to all registered users */
10718 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10719 struct io_uring_task *tctx = node->task->io_uring;
10721 if (WARN_ON_ONCE(!tctx->io_wq))
10724 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10725 new_count[i] = ctx->iowq_limits[i];
10726 /* ignore errors, it always returns zero anyway */
10727 (void)io_wq_max_workers(tctx->io_wq, new_count);
10732 mutex_unlock(&sqd->lock);
10733 io_put_sq_data(sqd);
10738 static bool io_register_op_must_quiesce(int op)
10741 case IORING_REGISTER_BUFFERS:
10742 case IORING_UNREGISTER_BUFFERS:
10743 case IORING_REGISTER_FILES:
10744 case IORING_UNREGISTER_FILES:
10745 case IORING_REGISTER_FILES_UPDATE:
10746 case IORING_REGISTER_PROBE:
10747 case IORING_REGISTER_PERSONALITY:
10748 case IORING_UNREGISTER_PERSONALITY:
10749 case IORING_REGISTER_FILES2:
10750 case IORING_REGISTER_FILES_UPDATE2:
10751 case IORING_REGISTER_BUFFERS2:
10752 case IORING_REGISTER_BUFFERS_UPDATE:
10753 case IORING_REGISTER_IOWQ_AFF:
10754 case IORING_UNREGISTER_IOWQ_AFF:
10755 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10762 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10766 percpu_ref_kill(&ctx->refs);
10769 * Drop uring mutex before waiting for references to exit. If another
10770 * thread is currently inside io_uring_enter() it might need to grab the
10771 * uring_lock to make progress. If we hold it here across the drain
10772 * wait, then we can deadlock. It's safe to drop the mutex here, since
10773 * no new references will come in after we've killed the percpu ref.
10775 mutex_unlock(&ctx->uring_lock);
10777 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10780 ret = io_run_task_work_sig();
10781 } while (ret >= 0);
10782 mutex_lock(&ctx->uring_lock);
10785 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10789 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10790 void __user *arg, unsigned nr_args)
10791 __releases(ctx->uring_lock)
10792 __acquires(ctx->uring_lock)
10797 * We're inside the ring mutex, if the ref is already dying, then
10798 * someone else killed the ctx or is already going through
10799 * io_uring_register().
10801 if (percpu_ref_is_dying(&ctx->refs))
10804 if (ctx->restricted) {
10805 if (opcode >= IORING_REGISTER_LAST)
10807 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10808 if (!test_bit(opcode, ctx->restrictions.register_op))
10812 if (io_register_op_must_quiesce(opcode)) {
10813 ret = io_ctx_quiesce(ctx);
10819 case IORING_REGISTER_BUFFERS:
10820 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10822 case IORING_UNREGISTER_BUFFERS:
10824 if (arg || nr_args)
10826 ret = io_sqe_buffers_unregister(ctx);
10828 case IORING_REGISTER_FILES:
10829 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10831 case IORING_UNREGISTER_FILES:
10833 if (arg || nr_args)
10835 ret = io_sqe_files_unregister(ctx);
10837 case IORING_REGISTER_FILES_UPDATE:
10838 ret = io_register_files_update(ctx, arg, nr_args);
10840 case IORING_REGISTER_EVENTFD:
10841 case IORING_REGISTER_EVENTFD_ASYNC:
10845 ret = io_eventfd_register(ctx, arg);
10848 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10849 ctx->eventfd_async = 1;
10851 ctx->eventfd_async = 0;
10853 case IORING_UNREGISTER_EVENTFD:
10855 if (arg || nr_args)
10857 ret = io_eventfd_unregister(ctx);
10859 case IORING_REGISTER_PROBE:
10861 if (!arg || nr_args > 256)
10863 ret = io_probe(ctx, arg, nr_args);
10865 case IORING_REGISTER_PERSONALITY:
10867 if (arg || nr_args)
10869 ret = io_register_personality(ctx);
10871 case IORING_UNREGISTER_PERSONALITY:
10875 ret = io_unregister_personality(ctx, nr_args);
10877 case IORING_REGISTER_ENABLE_RINGS:
10879 if (arg || nr_args)
10881 ret = io_register_enable_rings(ctx);
10883 case IORING_REGISTER_RESTRICTIONS:
10884 ret = io_register_restrictions(ctx, arg, nr_args);
10886 case IORING_REGISTER_FILES2:
10887 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10889 case IORING_REGISTER_FILES_UPDATE2:
10890 ret = io_register_rsrc_update(ctx, arg, nr_args,
10893 case IORING_REGISTER_BUFFERS2:
10894 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10896 case IORING_REGISTER_BUFFERS_UPDATE:
10897 ret = io_register_rsrc_update(ctx, arg, nr_args,
10898 IORING_RSRC_BUFFER);
10900 case IORING_REGISTER_IOWQ_AFF:
10902 if (!arg || !nr_args)
10904 ret = io_register_iowq_aff(ctx, arg, nr_args);
10906 case IORING_UNREGISTER_IOWQ_AFF:
10908 if (arg || nr_args)
10910 ret = io_unregister_iowq_aff(ctx);
10912 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10914 if (!arg || nr_args != 2)
10916 ret = io_register_iowq_max_workers(ctx, arg);
10923 if (io_register_op_must_quiesce(opcode)) {
10924 /* bring the ctx back to life */
10925 percpu_ref_reinit(&ctx->refs);
10926 reinit_completion(&ctx->ref_comp);
10931 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10932 void __user *, arg, unsigned int, nr_args)
10934 struct io_ring_ctx *ctx;
10943 if (f.file->f_op != &io_uring_fops)
10946 ctx = f.file->private_data;
10948 io_run_task_work();
10950 mutex_lock(&ctx->uring_lock);
10951 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10952 mutex_unlock(&ctx->uring_lock);
10953 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10954 ctx->cq_ev_fd != NULL, ret);
10960 static int __init io_uring_init(void)
10962 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10963 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10964 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10967 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10968 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10969 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10970 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10971 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10972 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10973 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10974 BUILD_BUG_SQE_ELEM(8, __u64, off);
10975 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10976 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10977 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10978 BUILD_BUG_SQE_ELEM(24, __u32, len);
10979 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10980 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10981 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10982 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10983 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10984 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10985 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10986 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10987 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10988 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10989 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10990 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10991 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10992 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10993 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10994 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10995 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10996 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10997 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10998 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10999 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11001 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11002 sizeof(struct io_uring_rsrc_update));
11003 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11004 sizeof(struct io_uring_rsrc_update2));
11006 /* ->buf_index is u16 */
11007 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11009 /* should fit into one byte */
11010 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11012 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11013 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11015 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11019 __initcall(io_uring_init);