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;
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,
1207 __must_hold(&req->ctx->timeout_lock)
1209 struct io_kiocb *req;
1211 if (task && head->task != task)
1216 io_for_each_link(req, head) {
1217 if (req->flags & REQ_F_INFLIGHT)
1223 static bool io_match_linked(struct io_kiocb *head)
1225 struct io_kiocb *req;
1227 io_for_each_link(req, head) {
1228 if (req->flags & REQ_F_INFLIGHT)
1235 * As io_match_task() but protected against racing with linked timeouts.
1236 * User must not hold timeout_lock.
1238 static bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
1243 if (task && head->task != task)
1248 if (head->flags & REQ_F_LINK_TIMEOUT) {
1249 struct io_ring_ctx *ctx = head->ctx;
1251 /* protect against races with linked timeouts */
1252 spin_lock_irq(&ctx->timeout_lock);
1253 matched = io_match_linked(head);
1254 spin_unlock_irq(&ctx->timeout_lock);
1256 matched = io_match_linked(head);
1261 static inline void req_set_fail(struct io_kiocb *req)
1263 req->flags |= REQ_F_FAIL;
1266 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1272 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1274 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1276 complete(&ctx->ref_comp);
1279 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1281 return !req->timeout.off;
1284 static void io_fallback_req_func(struct work_struct *work)
1286 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1287 fallback_work.work);
1288 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1289 struct io_kiocb *req, *tmp;
1290 bool locked = false;
1292 percpu_ref_get(&ctx->refs);
1293 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1294 req->io_task_work.func(req, &locked);
1297 if (ctx->submit_state.compl_nr)
1298 io_submit_flush_completions(ctx);
1299 mutex_unlock(&ctx->uring_lock);
1301 percpu_ref_put(&ctx->refs);
1305 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1307 struct io_ring_ctx *ctx;
1310 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1315 * Use 5 bits less than the max cq entries, that should give us around
1316 * 32 entries per hash list if totally full and uniformly spread.
1318 hash_bits = ilog2(p->cq_entries);
1322 ctx->cancel_hash_bits = hash_bits;
1323 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1325 if (!ctx->cancel_hash)
1327 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1329 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1330 if (!ctx->dummy_ubuf)
1332 /* set invalid range, so io_import_fixed() fails meeting it */
1333 ctx->dummy_ubuf->ubuf = -1UL;
1335 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1336 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1339 ctx->flags = p->flags;
1340 init_waitqueue_head(&ctx->sqo_sq_wait);
1341 INIT_LIST_HEAD(&ctx->sqd_list);
1342 init_waitqueue_head(&ctx->poll_wait);
1343 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1344 init_completion(&ctx->ref_comp);
1345 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1346 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1347 mutex_init(&ctx->uring_lock);
1348 init_waitqueue_head(&ctx->cq_wait);
1349 spin_lock_init(&ctx->completion_lock);
1350 spin_lock_init(&ctx->timeout_lock);
1351 INIT_LIST_HEAD(&ctx->iopoll_list);
1352 INIT_LIST_HEAD(&ctx->defer_list);
1353 INIT_LIST_HEAD(&ctx->timeout_list);
1354 INIT_LIST_HEAD(&ctx->ltimeout_list);
1355 spin_lock_init(&ctx->rsrc_ref_lock);
1356 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1357 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1358 init_llist_head(&ctx->rsrc_put_llist);
1359 INIT_LIST_HEAD(&ctx->tctx_list);
1360 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1361 INIT_LIST_HEAD(&ctx->locked_free_list);
1362 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1365 kfree(ctx->dummy_ubuf);
1366 kfree(ctx->cancel_hash);
1371 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1373 struct io_rings *r = ctx->rings;
1375 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1379 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1381 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1382 struct io_ring_ctx *ctx = req->ctx;
1384 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1390 #define FFS_ASYNC_READ 0x1UL
1391 #define FFS_ASYNC_WRITE 0x2UL
1393 #define FFS_ISREG 0x4UL
1395 #define FFS_ISREG 0x0UL
1397 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1399 static inline bool io_req_ffs_set(struct io_kiocb *req)
1401 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1404 static void io_req_track_inflight(struct io_kiocb *req)
1406 if (!(req->flags & REQ_F_INFLIGHT)) {
1407 req->flags |= REQ_F_INFLIGHT;
1408 atomic_inc(&req->task->io_uring->inflight_tracked);
1412 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1414 if (WARN_ON_ONCE(!req->link))
1417 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1418 req->flags |= REQ_F_LINK_TIMEOUT;
1420 /* linked timeouts should have two refs once prep'ed */
1421 io_req_set_refcount(req);
1422 __io_req_set_refcount(req->link, 2);
1426 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1428 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1430 return __io_prep_linked_timeout(req);
1433 static void io_prep_async_work(struct io_kiocb *req)
1435 const struct io_op_def *def = &io_op_defs[req->opcode];
1436 struct io_ring_ctx *ctx = req->ctx;
1438 if (!(req->flags & REQ_F_CREDS)) {
1439 req->flags |= REQ_F_CREDS;
1440 req->creds = get_current_cred();
1443 req->work.list.next = NULL;
1444 req->work.flags = 0;
1445 if (req->flags & REQ_F_FORCE_ASYNC)
1446 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1448 if (req->flags & REQ_F_ISREG) {
1449 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1450 io_wq_hash_work(&req->work, file_inode(req->file));
1451 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1452 if (def->unbound_nonreg_file)
1453 req->work.flags |= IO_WQ_WORK_UNBOUND;
1457 static void io_prep_async_link(struct io_kiocb *req)
1459 struct io_kiocb *cur;
1461 if (req->flags & REQ_F_LINK_TIMEOUT) {
1462 struct io_ring_ctx *ctx = req->ctx;
1464 spin_lock_irq(&ctx->timeout_lock);
1465 io_for_each_link(cur, req)
1466 io_prep_async_work(cur);
1467 spin_unlock_irq(&ctx->timeout_lock);
1469 io_for_each_link(cur, req)
1470 io_prep_async_work(cur);
1474 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1476 struct io_ring_ctx *ctx = req->ctx;
1477 struct io_kiocb *link = io_prep_linked_timeout(req);
1478 struct io_uring_task *tctx = req->task->io_uring;
1480 /* must not take the lock, NULL it as a precaution */
1484 BUG_ON(!tctx->io_wq);
1486 /* init ->work of the whole link before punting */
1487 io_prep_async_link(req);
1490 * Not expected to happen, but if we do have a bug where this _can_
1491 * happen, catch it here and ensure the request is marked as
1492 * canceled. That will make io-wq go through the usual work cancel
1493 * procedure rather than attempt to run this request (or create a new
1496 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1497 req->work.flags |= IO_WQ_WORK_CANCEL;
1499 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1500 &req->work, req->flags);
1501 io_wq_enqueue(tctx->io_wq, &req->work);
1503 io_queue_linked_timeout(link);
1506 static void io_kill_timeout(struct io_kiocb *req, int status)
1507 __must_hold(&req->ctx->completion_lock)
1508 __must_hold(&req->ctx->timeout_lock)
1510 struct io_timeout_data *io = req->async_data;
1512 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1515 atomic_set(&req->ctx->cq_timeouts,
1516 atomic_read(&req->ctx->cq_timeouts) + 1);
1517 list_del_init(&req->timeout.list);
1518 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1519 io_put_req_deferred(req);
1523 static void io_queue_deferred(struct io_ring_ctx *ctx)
1525 while (!list_empty(&ctx->defer_list)) {
1526 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1527 struct io_defer_entry, list);
1529 if (req_need_defer(de->req, de->seq))
1531 list_del_init(&de->list);
1532 io_req_task_queue(de->req);
1537 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1538 __must_hold(&ctx->completion_lock)
1540 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1541 struct io_kiocb *req, *tmp;
1543 spin_lock_irq(&ctx->timeout_lock);
1544 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1545 u32 events_needed, events_got;
1547 if (io_is_timeout_noseq(req))
1551 * Since seq can easily wrap around over time, subtract
1552 * the last seq at which timeouts were flushed before comparing.
1553 * Assuming not more than 2^31-1 events have happened since,
1554 * these subtractions won't have wrapped, so we can check if
1555 * target is in [last_seq, current_seq] by comparing the two.
1557 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1558 events_got = seq - ctx->cq_last_tm_flush;
1559 if (events_got < events_needed)
1562 io_kill_timeout(req, 0);
1564 ctx->cq_last_tm_flush = seq;
1565 spin_unlock_irq(&ctx->timeout_lock);
1568 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1570 if (ctx->off_timeout_used)
1571 io_flush_timeouts(ctx);
1572 if (ctx->drain_active)
1573 io_queue_deferred(ctx);
1576 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1578 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1579 __io_commit_cqring_flush(ctx);
1580 /* order cqe stores with ring update */
1581 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1584 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1586 struct io_rings *r = ctx->rings;
1588 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1591 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1593 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1596 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1598 struct io_rings *rings = ctx->rings;
1599 unsigned tail, mask = ctx->cq_entries - 1;
1602 * writes to the cq entry need to come after reading head; the
1603 * control dependency is enough as we're using WRITE_ONCE to
1606 if (__io_cqring_events(ctx) == ctx->cq_entries)
1609 tail = ctx->cached_cq_tail++;
1610 return &rings->cqes[tail & mask];
1613 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1615 if (likely(!ctx->cq_ev_fd))
1617 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1619 return !ctx->eventfd_async || io_wq_current_is_worker();
1623 * This should only get called when at least one event has been posted.
1624 * Some applications rely on the eventfd notification count only changing
1625 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1626 * 1:1 relationship between how many times this function is called (and
1627 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1629 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1632 * wake_up_all() may seem excessive, but io_wake_function() and
1633 * io_should_wake() handle the termination of the loop and only
1634 * wake as many waiters as we need to.
1636 if (wq_has_sleeper(&ctx->cq_wait))
1637 wake_up_all(&ctx->cq_wait);
1638 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1639 wake_up(&ctx->sq_data->wait);
1640 if (io_should_trigger_evfd(ctx))
1641 eventfd_signal(ctx->cq_ev_fd, 1);
1642 if (waitqueue_active(&ctx->poll_wait))
1643 wake_up_interruptible(&ctx->poll_wait);
1646 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1648 /* see waitqueue_active() comment */
1651 if (ctx->flags & IORING_SETUP_SQPOLL) {
1652 if (waitqueue_active(&ctx->cq_wait))
1653 wake_up_all(&ctx->cq_wait);
1655 if (io_should_trigger_evfd(ctx))
1656 eventfd_signal(ctx->cq_ev_fd, 1);
1657 if (waitqueue_active(&ctx->poll_wait))
1658 wake_up_interruptible(&ctx->poll_wait);
1661 /* Returns true if there are no backlogged entries after the flush */
1662 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1664 bool all_flushed, posted;
1666 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1670 spin_lock(&ctx->completion_lock);
1671 while (!list_empty(&ctx->cq_overflow_list)) {
1672 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1673 struct io_overflow_cqe *ocqe;
1677 ocqe = list_first_entry(&ctx->cq_overflow_list,
1678 struct io_overflow_cqe, list);
1680 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1682 io_account_cq_overflow(ctx);
1685 list_del(&ocqe->list);
1689 all_flushed = list_empty(&ctx->cq_overflow_list);
1691 clear_bit(0, &ctx->check_cq_overflow);
1692 WRITE_ONCE(ctx->rings->sq_flags,
1693 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1697 io_commit_cqring(ctx);
1698 spin_unlock(&ctx->completion_lock);
1700 io_cqring_ev_posted(ctx);
1704 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1708 if (test_bit(0, &ctx->check_cq_overflow)) {
1709 /* iopoll syncs against uring_lock, not completion_lock */
1710 if (ctx->flags & IORING_SETUP_IOPOLL)
1711 mutex_lock(&ctx->uring_lock);
1712 ret = __io_cqring_overflow_flush(ctx, false);
1713 if (ctx->flags & IORING_SETUP_IOPOLL)
1714 mutex_unlock(&ctx->uring_lock);
1720 /* must to be called somewhat shortly after putting a request */
1721 static inline void io_put_task(struct task_struct *task, int nr)
1723 struct io_uring_task *tctx = task->io_uring;
1725 if (likely(task == current)) {
1726 tctx->cached_refs += nr;
1728 percpu_counter_sub(&tctx->inflight, nr);
1729 if (unlikely(atomic_read(&tctx->in_idle)))
1730 wake_up(&tctx->wait);
1731 put_task_struct_many(task, nr);
1735 static void io_task_refs_refill(struct io_uring_task *tctx)
1737 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1739 percpu_counter_add(&tctx->inflight, refill);
1740 refcount_add(refill, ¤t->usage);
1741 tctx->cached_refs += refill;
1744 static inline void io_get_task_refs(int nr)
1746 struct io_uring_task *tctx = current->io_uring;
1748 tctx->cached_refs -= nr;
1749 if (unlikely(tctx->cached_refs < 0))
1750 io_task_refs_refill(tctx);
1753 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
1755 struct io_uring_task *tctx = task->io_uring;
1756 unsigned int refs = tctx->cached_refs;
1759 tctx->cached_refs = 0;
1760 percpu_counter_sub(&tctx->inflight, refs);
1761 put_task_struct_many(task, refs);
1765 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1766 long res, unsigned int cflags)
1768 struct io_overflow_cqe *ocqe;
1770 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1773 * If we're in ring overflow flush mode, or in task cancel mode,
1774 * or cannot allocate an overflow entry, then we need to drop it
1777 io_account_cq_overflow(ctx);
1780 if (list_empty(&ctx->cq_overflow_list)) {
1781 set_bit(0, &ctx->check_cq_overflow);
1782 WRITE_ONCE(ctx->rings->sq_flags,
1783 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1786 ocqe->cqe.user_data = user_data;
1787 ocqe->cqe.res = res;
1788 ocqe->cqe.flags = cflags;
1789 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1793 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1794 long res, unsigned int cflags)
1796 struct io_uring_cqe *cqe;
1798 trace_io_uring_complete(ctx, user_data, res, cflags);
1801 * If we can't get a cq entry, userspace overflowed the
1802 * submission (by quite a lot). Increment the overflow count in
1805 cqe = io_get_cqe(ctx);
1807 WRITE_ONCE(cqe->user_data, user_data);
1808 WRITE_ONCE(cqe->res, res);
1809 WRITE_ONCE(cqe->flags, cflags);
1812 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1815 /* not as hot to bloat with inlining */
1816 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1817 long res, unsigned int cflags)
1819 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1822 static void io_req_complete_post(struct io_kiocb *req, long res,
1823 unsigned int cflags)
1825 struct io_ring_ctx *ctx = req->ctx;
1827 spin_lock(&ctx->completion_lock);
1828 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1830 * If we're the last reference to this request, add to our locked
1833 if (req_ref_put_and_test(req)) {
1834 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1835 if (req->flags & IO_DISARM_MASK)
1836 io_disarm_next(req);
1838 io_req_task_queue(req->link);
1842 io_dismantle_req(req);
1843 io_put_task(req->task, 1);
1844 list_add(&req->inflight_entry, &ctx->locked_free_list);
1845 ctx->locked_free_nr++;
1847 if (!percpu_ref_tryget(&ctx->refs))
1850 io_commit_cqring(ctx);
1851 spin_unlock(&ctx->completion_lock);
1854 io_cqring_ev_posted(ctx);
1855 percpu_ref_put(&ctx->refs);
1859 static inline bool io_req_needs_clean(struct io_kiocb *req)
1861 return req->flags & IO_REQ_CLEAN_FLAGS;
1864 static void io_req_complete_state(struct io_kiocb *req, long res,
1865 unsigned int cflags)
1867 if (io_req_needs_clean(req))
1870 req->compl.cflags = cflags;
1871 req->flags |= REQ_F_COMPLETE_INLINE;
1874 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1875 long res, unsigned cflags)
1877 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1878 io_req_complete_state(req, res, cflags);
1880 io_req_complete_post(req, res, cflags);
1883 static inline void io_req_complete(struct io_kiocb *req, long res)
1885 __io_req_complete(req, 0, res, 0);
1888 static void io_req_complete_failed(struct io_kiocb *req, long res)
1891 io_req_complete_post(req, res, 0);
1894 static void io_req_complete_fail_submit(struct io_kiocb *req)
1897 * We don't submit, fail them all, for that replace hardlinks with
1898 * normal links. Extra REQ_F_LINK is tolerated.
1900 req->flags &= ~REQ_F_HARDLINK;
1901 req->flags |= REQ_F_LINK;
1902 io_req_complete_failed(req, req->result);
1906 * Don't initialise the fields below on every allocation, but do that in
1907 * advance and keep them valid across allocations.
1909 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1913 req->async_data = NULL;
1914 /* not necessary, but safer to zero */
1918 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1919 struct io_submit_state *state)
1921 spin_lock(&ctx->completion_lock);
1922 list_splice_init(&ctx->locked_free_list, &state->free_list);
1923 ctx->locked_free_nr = 0;
1924 spin_unlock(&ctx->completion_lock);
1927 /* Returns true IFF there are requests in the cache */
1928 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1930 struct io_submit_state *state = &ctx->submit_state;
1934 * If we have more than a batch's worth of requests in our IRQ side
1935 * locked cache, grab the lock and move them over to our submission
1938 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1939 io_flush_cached_locked_reqs(ctx, state);
1941 nr = state->free_reqs;
1942 while (!list_empty(&state->free_list)) {
1943 struct io_kiocb *req = list_first_entry(&state->free_list,
1944 struct io_kiocb, inflight_entry);
1946 list_del(&req->inflight_entry);
1947 state->reqs[nr++] = req;
1948 if (nr == ARRAY_SIZE(state->reqs))
1952 state->free_reqs = nr;
1957 * A request might get retired back into the request caches even before opcode
1958 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1959 * Because of that, io_alloc_req() should be called only under ->uring_lock
1960 * and with extra caution to not get a request that is still worked on.
1962 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1963 __must_hold(&ctx->uring_lock)
1965 struct io_submit_state *state = &ctx->submit_state;
1966 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1969 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1971 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1974 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1978 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1979 * retry single alloc to be on the safe side.
1981 if (unlikely(ret <= 0)) {
1982 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1983 if (!state->reqs[0])
1988 for (i = 0; i < ret; i++)
1989 io_preinit_req(state->reqs[i], ctx);
1990 state->free_reqs = ret;
1993 return state->reqs[state->free_reqs];
1996 static inline void io_put_file(struct file *file)
2002 static void io_dismantle_req(struct io_kiocb *req)
2004 unsigned int flags = req->flags;
2006 if (io_req_needs_clean(req))
2008 if (!(flags & REQ_F_FIXED_FILE))
2009 io_put_file(req->file);
2010 if (req->fixed_rsrc_refs)
2011 percpu_ref_put(req->fixed_rsrc_refs);
2012 if (req->async_data) {
2013 kfree(req->async_data);
2014 req->async_data = NULL;
2018 static void __io_free_req(struct io_kiocb *req)
2020 struct io_ring_ctx *ctx = req->ctx;
2022 io_dismantle_req(req);
2023 io_put_task(req->task, 1);
2025 spin_lock(&ctx->completion_lock);
2026 list_add(&req->inflight_entry, &ctx->locked_free_list);
2027 ctx->locked_free_nr++;
2028 spin_unlock(&ctx->completion_lock);
2030 percpu_ref_put(&ctx->refs);
2033 static inline void io_remove_next_linked(struct io_kiocb *req)
2035 struct io_kiocb *nxt = req->link;
2037 req->link = nxt->link;
2041 static bool io_kill_linked_timeout(struct io_kiocb *req)
2042 __must_hold(&req->ctx->completion_lock)
2043 __must_hold(&req->ctx->timeout_lock)
2045 struct io_kiocb *link = req->link;
2047 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2048 struct io_timeout_data *io = link->async_data;
2050 io_remove_next_linked(req);
2051 link->timeout.head = NULL;
2052 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2053 list_del(&link->timeout.list);
2054 io_cqring_fill_event(link->ctx, link->user_data,
2056 io_put_req_deferred(link);
2063 static void io_fail_links(struct io_kiocb *req)
2064 __must_hold(&req->ctx->completion_lock)
2066 struct io_kiocb *nxt, *link = req->link;
2070 long res = -ECANCELED;
2072 if (link->flags & REQ_F_FAIL)
2078 trace_io_uring_fail_link(req, link);
2079 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2080 io_put_req_deferred(link);
2085 static bool io_disarm_next(struct io_kiocb *req)
2086 __must_hold(&req->ctx->completion_lock)
2088 bool posted = false;
2090 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2091 struct io_kiocb *link = req->link;
2093 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2094 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2095 io_remove_next_linked(req);
2096 io_cqring_fill_event(link->ctx, link->user_data,
2098 io_put_req_deferred(link);
2101 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2102 struct io_ring_ctx *ctx = req->ctx;
2104 spin_lock_irq(&ctx->timeout_lock);
2105 posted = io_kill_linked_timeout(req);
2106 spin_unlock_irq(&ctx->timeout_lock);
2108 if (unlikely((req->flags & REQ_F_FAIL) &&
2109 !(req->flags & REQ_F_HARDLINK))) {
2110 posted |= (req->link != NULL);
2116 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2118 struct io_kiocb *nxt;
2121 * If LINK is set, we have dependent requests in this chain. If we
2122 * didn't fail this request, queue the first one up, moving any other
2123 * dependencies to the next request. In case of failure, fail the rest
2126 if (req->flags & IO_DISARM_MASK) {
2127 struct io_ring_ctx *ctx = req->ctx;
2130 spin_lock(&ctx->completion_lock);
2131 posted = io_disarm_next(req);
2133 io_commit_cqring(req->ctx);
2134 spin_unlock(&ctx->completion_lock);
2136 io_cqring_ev_posted(ctx);
2143 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2145 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2147 return __io_req_find_next(req);
2150 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2155 if (ctx->submit_state.compl_nr)
2156 io_submit_flush_completions(ctx);
2157 mutex_unlock(&ctx->uring_lock);
2160 percpu_ref_put(&ctx->refs);
2163 static void tctx_task_work(struct callback_head *cb)
2165 bool locked = false;
2166 struct io_ring_ctx *ctx = NULL;
2167 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2171 struct io_wq_work_node *node;
2173 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2174 io_submit_flush_completions(ctx);
2176 spin_lock_irq(&tctx->task_lock);
2177 node = tctx->task_list.first;
2178 INIT_WQ_LIST(&tctx->task_list);
2180 tctx->task_running = false;
2181 spin_unlock_irq(&tctx->task_lock);
2186 struct io_wq_work_node *next = node->next;
2187 struct io_kiocb *req = container_of(node, struct io_kiocb,
2190 if (req->ctx != ctx) {
2191 ctx_flush_and_put(ctx, &locked);
2193 /* if not contended, grab and improve batching */
2194 locked = mutex_trylock(&ctx->uring_lock);
2195 percpu_ref_get(&ctx->refs);
2197 req->io_task_work.func(req, &locked);
2204 ctx_flush_and_put(ctx, &locked);
2206 /* relaxed read is enough as only the task itself sets ->in_idle */
2207 if (unlikely(atomic_read(&tctx->in_idle)))
2208 io_uring_drop_tctx_refs(current);
2211 static void io_req_task_work_add(struct io_kiocb *req)
2213 struct task_struct *tsk = req->task;
2214 struct io_uring_task *tctx = tsk->io_uring;
2215 enum task_work_notify_mode notify;
2216 struct io_wq_work_node *node;
2217 unsigned long flags;
2220 WARN_ON_ONCE(!tctx);
2222 spin_lock_irqsave(&tctx->task_lock, flags);
2223 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2224 running = tctx->task_running;
2226 tctx->task_running = true;
2227 spin_unlock_irqrestore(&tctx->task_lock, flags);
2229 /* task_work already pending, we're done */
2234 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2235 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2236 * processing task_work. There's no reliable way to tell if TWA_RESUME
2239 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2240 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2241 wake_up_process(tsk);
2245 spin_lock_irqsave(&tctx->task_lock, flags);
2246 tctx->task_running = false;
2247 node = tctx->task_list.first;
2248 INIT_WQ_LIST(&tctx->task_list);
2249 spin_unlock_irqrestore(&tctx->task_lock, flags);
2252 req = container_of(node, struct io_kiocb, io_task_work.node);
2254 if (llist_add(&req->io_task_work.fallback_node,
2255 &req->ctx->fallback_llist))
2256 schedule_delayed_work(&req->ctx->fallback_work, 1);
2260 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2262 struct io_ring_ctx *ctx = req->ctx;
2264 /* not needed for normal modes, but SQPOLL depends on it */
2265 io_tw_lock(ctx, locked);
2266 io_req_complete_failed(req, req->result);
2269 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2271 struct io_ring_ctx *ctx = req->ctx;
2273 io_tw_lock(ctx, locked);
2274 /* req->task == current here, checking PF_EXITING is safe */
2275 if (likely(!(req->task->flags & PF_EXITING)))
2276 __io_queue_sqe(req);
2278 io_req_complete_failed(req, -EFAULT);
2281 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2284 req->io_task_work.func = io_req_task_cancel;
2285 io_req_task_work_add(req);
2288 static void io_req_task_queue(struct io_kiocb *req)
2290 req->io_task_work.func = io_req_task_submit;
2291 io_req_task_work_add(req);
2294 static void io_req_task_queue_reissue(struct io_kiocb *req)
2296 req->io_task_work.func = io_queue_async_work;
2297 io_req_task_work_add(req);
2300 static inline void io_queue_next(struct io_kiocb *req)
2302 struct io_kiocb *nxt = io_req_find_next(req);
2305 io_req_task_queue(nxt);
2308 static void io_free_req(struct io_kiocb *req)
2314 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2320 struct task_struct *task;
2325 static inline void io_init_req_batch(struct req_batch *rb)
2332 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2333 struct req_batch *rb)
2336 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2338 io_put_task(rb->task, rb->task_refs);
2341 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2342 struct io_submit_state *state)
2345 io_dismantle_req(req);
2347 if (req->task != rb->task) {
2349 io_put_task(rb->task, rb->task_refs);
2350 rb->task = req->task;
2356 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2357 state->reqs[state->free_reqs++] = req;
2359 list_add(&req->inflight_entry, &state->free_list);
2362 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2363 __must_hold(&ctx->uring_lock)
2365 struct io_submit_state *state = &ctx->submit_state;
2366 int i, nr = state->compl_nr;
2367 struct req_batch rb;
2369 spin_lock(&ctx->completion_lock);
2370 for (i = 0; i < nr; i++) {
2371 struct io_kiocb *req = state->compl_reqs[i];
2373 __io_cqring_fill_event(ctx, req->user_data, req->result,
2376 io_commit_cqring(ctx);
2377 spin_unlock(&ctx->completion_lock);
2378 io_cqring_ev_posted(ctx);
2380 io_init_req_batch(&rb);
2381 for (i = 0; i < nr; i++) {
2382 struct io_kiocb *req = state->compl_reqs[i];
2384 if (req_ref_put_and_test(req))
2385 io_req_free_batch(&rb, req, &ctx->submit_state);
2388 io_req_free_batch_finish(ctx, &rb);
2389 state->compl_nr = 0;
2393 * Drop reference to request, return next in chain (if there is one) if this
2394 * was the last reference to this request.
2396 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2398 struct io_kiocb *nxt = NULL;
2400 if (req_ref_put_and_test(req)) {
2401 nxt = io_req_find_next(req);
2407 static inline void io_put_req(struct io_kiocb *req)
2409 if (req_ref_put_and_test(req))
2413 static inline void io_put_req_deferred(struct io_kiocb *req)
2415 if (req_ref_put_and_test(req)) {
2416 req->io_task_work.func = io_free_req_work;
2417 io_req_task_work_add(req);
2421 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2423 /* See comment at the top of this file */
2425 return __io_cqring_events(ctx);
2428 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2430 struct io_rings *rings = ctx->rings;
2432 /* make sure SQ entry isn't read before tail */
2433 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2436 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2438 unsigned int cflags;
2440 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2441 cflags |= IORING_CQE_F_BUFFER;
2442 req->flags &= ~REQ_F_BUFFER_SELECTED;
2447 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2449 struct io_buffer *kbuf;
2451 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2453 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2454 return io_put_kbuf(req, kbuf);
2457 static inline bool io_run_task_work(void)
2459 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2460 __set_current_state(TASK_RUNNING);
2461 tracehook_notify_signal();
2469 * Find and free completed poll iocbs
2471 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2472 struct list_head *done)
2474 struct req_batch rb;
2475 struct io_kiocb *req;
2477 /* order with ->result store in io_complete_rw_iopoll() */
2480 io_init_req_batch(&rb);
2481 while (!list_empty(done)) {
2482 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2483 list_del(&req->inflight_entry);
2485 __io_cqring_fill_event(ctx, req->user_data, req->result,
2486 io_put_rw_kbuf(req));
2489 if (req_ref_put_and_test(req))
2490 io_req_free_batch(&rb, req, &ctx->submit_state);
2493 io_commit_cqring(ctx);
2494 io_cqring_ev_posted_iopoll(ctx);
2495 io_req_free_batch_finish(ctx, &rb);
2498 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2501 struct io_kiocb *req, *tmp;
2506 * Only spin for completions if we don't have multiple devices hanging
2507 * off our complete list, and we're under the requested amount.
2509 spin = !ctx->poll_multi_queue && *nr_events < min;
2511 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2512 struct kiocb *kiocb = &req->rw.kiocb;
2516 * Move completed and retryable entries to our local lists.
2517 * If we find a request that requires polling, break out
2518 * and complete those lists first, if we have entries there.
2520 if (READ_ONCE(req->iopoll_completed)) {
2521 list_move_tail(&req->inflight_entry, &done);
2524 if (!list_empty(&done))
2527 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2528 if (unlikely(ret < 0))
2533 /* iopoll may have completed current req */
2534 if (READ_ONCE(req->iopoll_completed))
2535 list_move_tail(&req->inflight_entry, &done);
2538 if (!list_empty(&done))
2539 io_iopoll_complete(ctx, nr_events, &done);
2545 * We can't just wait for polled events to come to us, we have to actively
2546 * find and complete them.
2548 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2550 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2553 mutex_lock(&ctx->uring_lock);
2554 while (!list_empty(&ctx->iopoll_list)) {
2555 unsigned int nr_events = 0;
2557 io_do_iopoll(ctx, &nr_events, 0);
2559 /* let it sleep and repeat later if can't complete a request */
2563 * Ensure we allow local-to-the-cpu processing to take place,
2564 * in this case we need to ensure that we reap all events.
2565 * Also let task_work, etc. to progress by releasing the mutex
2567 if (need_resched()) {
2568 mutex_unlock(&ctx->uring_lock);
2570 mutex_lock(&ctx->uring_lock);
2573 mutex_unlock(&ctx->uring_lock);
2576 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2578 unsigned int nr_events = 0;
2582 * We disallow the app entering submit/complete with polling, but we
2583 * still need to lock the ring to prevent racing with polled issue
2584 * that got punted to a workqueue.
2586 mutex_lock(&ctx->uring_lock);
2588 * Don't enter poll loop if we already have events pending.
2589 * If we do, we can potentially be spinning for commands that
2590 * already triggered a CQE (eg in error).
2592 if (test_bit(0, &ctx->check_cq_overflow))
2593 __io_cqring_overflow_flush(ctx, false);
2594 if (io_cqring_events(ctx))
2598 * If a submit got punted to a workqueue, we can have the
2599 * application entering polling for a command before it gets
2600 * issued. That app will hold the uring_lock for the duration
2601 * of the poll right here, so we need to take a breather every
2602 * now and then to ensure that the issue has a chance to add
2603 * the poll to the issued list. Otherwise we can spin here
2604 * forever, while the workqueue is stuck trying to acquire the
2607 if (list_empty(&ctx->iopoll_list)) {
2608 u32 tail = ctx->cached_cq_tail;
2610 mutex_unlock(&ctx->uring_lock);
2612 mutex_lock(&ctx->uring_lock);
2614 /* some requests don't go through iopoll_list */
2615 if (tail != ctx->cached_cq_tail ||
2616 list_empty(&ctx->iopoll_list))
2619 ret = io_do_iopoll(ctx, &nr_events, min);
2620 } while (!ret && nr_events < min && !need_resched());
2622 mutex_unlock(&ctx->uring_lock);
2626 static void kiocb_end_write(struct io_kiocb *req)
2629 * Tell lockdep we inherited freeze protection from submission
2632 if (req->flags & REQ_F_ISREG) {
2633 struct super_block *sb = file_inode(req->file)->i_sb;
2635 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2641 static bool io_resubmit_prep(struct io_kiocb *req)
2643 struct io_async_rw *rw = req->async_data;
2646 return !io_req_prep_async(req);
2647 iov_iter_restore(&rw->iter, &rw->iter_state);
2651 static bool io_rw_should_reissue(struct io_kiocb *req)
2653 umode_t mode = file_inode(req->file)->i_mode;
2654 struct io_ring_ctx *ctx = req->ctx;
2656 if (!S_ISBLK(mode) && !S_ISREG(mode))
2658 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2659 !(ctx->flags & IORING_SETUP_IOPOLL)))
2662 * If ref is dying, we might be running poll reap from the exit work.
2663 * Don't attempt to reissue from that path, just let it fail with
2666 if (percpu_ref_is_dying(&ctx->refs))
2669 * Play it safe and assume not safe to re-import and reissue if we're
2670 * not in the original thread group (or in task context).
2672 if (!same_thread_group(req->task, current) || !in_task())
2677 static bool io_resubmit_prep(struct io_kiocb *req)
2681 static bool io_rw_should_reissue(struct io_kiocb *req)
2687 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2689 if (req->rw.kiocb.ki_flags & IOCB_WRITE) {
2690 kiocb_end_write(req);
2691 fsnotify_modify(req->file);
2693 fsnotify_access(req->file);
2695 if (res != req->result) {
2696 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2697 io_rw_should_reissue(req)) {
2698 req->flags |= REQ_F_REISSUE;
2707 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2709 unsigned int cflags = io_put_rw_kbuf(req);
2710 long res = req->result;
2713 struct io_ring_ctx *ctx = req->ctx;
2714 struct io_submit_state *state = &ctx->submit_state;
2716 io_req_complete_state(req, res, cflags);
2717 state->compl_reqs[state->compl_nr++] = req;
2718 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2719 io_submit_flush_completions(ctx);
2721 io_req_complete_post(req, res, cflags);
2725 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2726 unsigned int issue_flags)
2728 if (__io_complete_rw_common(req, res))
2730 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2733 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2735 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2737 if (__io_complete_rw_common(req, res))
2740 req->io_task_work.func = io_req_task_complete;
2741 io_req_task_work_add(req);
2744 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2746 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2748 if (kiocb->ki_flags & IOCB_WRITE)
2749 kiocb_end_write(req);
2750 if (unlikely(res != req->result)) {
2751 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2752 req->flags |= REQ_F_REISSUE;
2757 WRITE_ONCE(req->result, res);
2758 /* order with io_iopoll_complete() checking ->result */
2760 WRITE_ONCE(req->iopoll_completed, 1);
2764 * After the iocb has been issued, it's safe to be found on the poll list.
2765 * Adding the kiocb to the list AFTER submission ensures that we don't
2766 * find it from a io_do_iopoll() thread before the issuer is done
2767 * accessing the kiocb cookie.
2769 static void io_iopoll_req_issued(struct io_kiocb *req)
2771 struct io_ring_ctx *ctx = req->ctx;
2772 const bool in_async = io_wq_current_is_worker();
2774 /* workqueue context doesn't hold uring_lock, grab it now */
2775 if (unlikely(in_async))
2776 mutex_lock(&ctx->uring_lock);
2779 * Track whether we have multiple files in our lists. This will impact
2780 * how we do polling eventually, not spinning if we're on potentially
2781 * different devices.
2783 if (list_empty(&ctx->iopoll_list)) {
2784 ctx->poll_multi_queue = false;
2785 } else if (!ctx->poll_multi_queue) {
2786 struct io_kiocb *list_req;
2787 unsigned int queue_num0, queue_num1;
2789 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2792 if (list_req->file != req->file) {
2793 ctx->poll_multi_queue = true;
2795 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2796 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2797 if (queue_num0 != queue_num1)
2798 ctx->poll_multi_queue = true;
2803 * For fast devices, IO may have already completed. If it has, add
2804 * it to the front so we find it first.
2806 if (READ_ONCE(req->iopoll_completed))
2807 list_add(&req->inflight_entry, &ctx->iopoll_list);
2809 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2811 if (unlikely(in_async)) {
2813 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2814 * in sq thread task context or in io worker task context. If
2815 * current task context is sq thread, we don't need to check
2816 * whether should wake up sq thread.
2818 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2819 wq_has_sleeper(&ctx->sq_data->wait))
2820 wake_up(&ctx->sq_data->wait);
2822 mutex_unlock(&ctx->uring_lock);
2826 static bool io_bdev_nowait(struct block_device *bdev)
2828 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2832 * If we tracked the file through the SCM inflight mechanism, we could support
2833 * any file. For now, just ensure that anything potentially problematic is done
2836 static bool __io_file_supports_nowait(struct file *file, int rw)
2838 umode_t mode = file_inode(file)->i_mode;
2840 if (S_ISBLK(mode)) {
2841 if (IS_ENABLED(CONFIG_BLOCK) &&
2842 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2848 if (S_ISREG(mode)) {
2849 if (IS_ENABLED(CONFIG_BLOCK) &&
2850 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2851 file->f_op != &io_uring_fops)
2856 /* any ->read/write should understand O_NONBLOCK */
2857 if (file->f_flags & O_NONBLOCK)
2860 if (!(file->f_mode & FMODE_NOWAIT))
2864 return file->f_op->read_iter != NULL;
2866 return file->f_op->write_iter != NULL;
2869 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2871 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2873 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2876 return __io_file_supports_nowait(req->file, rw);
2879 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2882 struct io_ring_ctx *ctx = req->ctx;
2883 struct kiocb *kiocb = &req->rw.kiocb;
2884 struct file *file = req->file;
2888 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2889 req->flags |= REQ_F_ISREG;
2891 kiocb->ki_pos = READ_ONCE(sqe->off);
2892 if (kiocb->ki_pos == -1) {
2893 if (!(file->f_mode & FMODE_STREAM)) {
2894 req->flags |= REQ_F_CUR_POS;
2895 kiocb->ki_pos = file->f_pos;
2900 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2901 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2902 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2907 * If the file is marked O_NONBLOCK, still allow retry for it if it
2908 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2909 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2911 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2912 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2913 req->flags |= REQ_F_NOWAIT;
2915 ioprio = READ_ONCE(sqe->ioprio);
2917 ret = ioprio_check_cap(ioprio);
2921 kiocb->ki_ioprio = ioprio;
2923 kiocb->ki_ioprio = get_current_ioprio();
2925 if (ctx->flags & IORING_SETUP_IOPOLL) {
2926 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2927 !kiocb->ki_filp->f_op->iopoll)
2930 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2931 kiocb->ki_complete = io_complete_rw_iopoll;
2932 req->iopoll_completed = 0;
2934 if (kiocb->ki_flags & IOCB_HIPRI)
2936 kiocb->ki_complete = io_complete_rw;
2939 /* used for fixed read/write too - just read unconditionally */
2940 req->buf_index = READ_ONCE(sqe->buf_index);
2943 if (req->opcode == IORING_OP_READ_FIXED ||
2944 req->opcode == IORING_OP_WRITE_FIXED) {
2945 struct io_ring_ctx *ctx = req->ctx;
2948 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
2950 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
2951 req->imu = ctx->user_bufs[index];
2952 io_req_set_rsrc_node(req);
2955 req->rw.addr = READ_ONCE(sqe->addr);
2956 req->rw.len = READ_ONCE(sqe->len);
2960 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2966 case -ERESTARTNOINTR:
2967 case -ERESTARTNOHAND:
2968 case -ERESTART_RESTARTBLOCK:
2970 * We can't just restart the syscall, since previously
2971 * submitted sqes may already be in progress. Just fail this
2977 kiocb->ki_complete(kiocb, ret, 0);
2981 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2982 unsigned int issue_flags)
2984 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2985 struct io_async_rw *io = req->async_data;
2987 /* add previously done IO, if any */
2988 if (io && io->bytes_done > 0) {
2990 ret = io->bytes_done;
2992 ret += io->bytes_done;
2995 if (req->flags & REQ_F_CUR_POS)
2996 req->file->f_pos = kiocb->ki_pos;
2997 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2998 __io_complete_rw(req, ret, 0, issue_flags);
3000 io_rw_done(kiocb, ret);
3002 if (req->flags & REQ_F_REISSUE) {
3003 req->flags &= ~REQ_F_REISSUE;
3004 if (io_resubmit_prep(req)) {
3005 io_req_task_queue_reissue(req);
3007 unsigned int cflags = io_put_rw_kbuf(req);
3008 struct io_ring_ctx *ctx = req->ctx;
3011 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
3012 mutex_lock(&ctx->uring_lock);
3013 __io_req_complete(req, issue_flags, ret, cflags);
3014 mutex_unlock(&ctx->uring_lock);
3016 __io_req_complete(req, issue_flags, ret, cflags);
3022 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3023 struct io_mapped_ubuf *imu)
3025 size_t len = req->rw.len;
3026 u64 buf_end, buf_addr = req->rw.addr;
3029 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
3031 /* not inside the mapped region */
3032 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
3036 * May not be a start of buffer, set size appropriately
3037 * and advance us to the beginning.
3039 offset = buf_addr - imu->ubuf;
3040 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3044 * Don't use iov_iter_advance() here, as it's really slow for
3045 * using the latter parts of a big fixed buffer - it iterates
3046 * over each segment manually. We can cheat a bit here, because
3049 * 1) it's a BVEC iter, we set it up
3050 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3051 * first and last bvec
3053 * So just find our index, and adjust the iterator afterwards.
3054 * If the offset is within the first bvec (or the whole first
3055 * bvec, just use iov_iter_advance(). This makes it easier
3056 * since we can just skip the first segment, which may not
3057 * be PAGE_SIZE aligned.
3059 const struct bio_vec *bvec = imu->bvec;
3061 if (offset <= bvec->bv_len) {
3062 iov_iter_advance(iter, offset);
3064 unsigned long seg_skip;
3066 /* skip first vec */
3067 offset -= bvec->bv_len;
3068 seg_skip = 1 + (offset >> PAGE_SHIFT);
3070 iter->bvec = bvec + seg_skip;
3071 iter->nr_segs -= seg_skip;
3072 iter->count -= bvec->bv_len + offset;
3073 iter->iov_offset = offset & ~PAGE_MASK;
3080 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3082 if (WARN_ON_ONCE(!req->imu))
3084 return __io_import_fixed(req, rw, iter, req->imu);
3087 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3090 mutex_unlock(&ctx->uring_lock);
3093 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3096 * "Normal" inline submissions always hold the uring_lock, since we
3097 * grab it from the system call. Same is true for the SQPOLL offload.
3098 * The only exception is when we've detached the request and issue it
3099 * from an async worker thread, grab the lock for that case.
3102 mutex_lock(&ctx->uring_lock);
3105 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3106 int bgid, struct io_buffer *kbuf,
3109 struct io_buffer *head;
3111 if (req->flags & REQ_F_BUFFER_SELECTED)
3114 io_ring_submit_lock(req->ctx, needs_lock);
3116 lockdep_assert_held(&req->ctx->uring_lock);
3118 head = xa_load(&req->ctx->io_buffers, bgid);
3120 if (!list_empty(&head->list)) {
3121 kbuf = list_last_entry(&head->list, struct io_buffer,
3123 list_del(&kbuf->list);
3126 xa_erase(&req->ctx->io_buffers, bgid);
3128 if (*len > kbuf->len)
3131 kbuf = ERR_PTR(-ENOBUFS);
3134 io_ring_submit_unlock(req->ctx, needs_lock);
3139 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3142 struct io_buffer *kbuf;
3145 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3146 bgid = req->buf_index;
3147 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3150 req->rw.addr = (u64) (unsigned long) kbuf;
3151 req->flags |= REQ_F_BUFFER_SELECTED;
3152 return u64_to_user_ptr(kbuf->addr);
3155 #ifdef CONFIG_COMPAT
3156 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3159 struct compat_iovec __user *uiov;
3160 compat_ssize_t clen;
3164 uiov = u64_to_user_ptr(req->rw.addr);
3165 if (!access_ok(uiov, sizeof(*uiov)))
3167 if (__get_user(clen, &uiov->iov_len))
3173 buf = io_rw_buffer_select(req, &len, needs_lock);
3175 return PTR_ERR(buf);
3176 iov[0].iov_base = buf;
3177 iov[0].iov_len = (compat_size_t) len;
3182 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3185 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3189 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3192 len = iov[0].iov_len;
3195 buf = io_rw_buffer_select(req, &len, needs_lock);
3197 return PTR_ERR(buf);
3198 iov[0].iov_base = buf;
3199 iov[0].iov_len = len;
3203 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3206 if (req->flags & REQ_F_BUFFER_SELECTED) {
3207 struct io_buffer *kbuf;
3209 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3210 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3211 iov[0].iov_len = kbuf->len;
3214 if (req->rw.len != 1)
3217 #ifdef CONFIG_COMPAT
3218 if (req->ctx->compat)
3219 return io_compat_import(req, iov, needs_lock);
3222 return __io_iov_buffer_select(req, iov, needs_lock);
3225 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3226 struct iov_iter *iter, bool needs_lock)
3228 void __user *buf = u64_to_user_ptr(req->rw.addr);
3229 size_t sqe_len = req->rw.len;
3230 u8 opcode = req->opcode;
3233 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3235 return io_import_fixed(req, rw, iter);
3238 /* buffer index only valid with fixed read/write, or buffer select */
3239 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3242 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3243 if (req->flags & REQ_F_BUFFER_SELECT) {
3244 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3246 return PTR_ERR(buf);
3247 req->rw.len = sqe_len;
3250 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3255 if (req->flags & REQ_F_BUFFER_SELECT) {
3256 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3258 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3263 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3267 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3269 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3273 * For files that don't have ->read_iter() and ->write_iter(), handle them
3274 * by looping over ->read() or ->write() manually.
3276 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3278 struct kiocb *kiocb = &req->rw.kiocb;
3279 struct file *file = req->file;
3283 * Don't support polled IO through this interface, and we can't
3284 * support non-blocking either. For the latter, this just causes
3285 * the kiocb to be handled from an async context.
3287 if (kiocb->ki_flags & IOCB_HIPRI)
3289 if (kiocb->ki_flags & IOCB_NOWAIT)
3292 while (iov_iter_count(iter)) {
3296 if (!iov_iter_is_bvec(iter)) {
3297 iovec = iov_iter_iovec(iter);
3299 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3300 iovec.iov_len = req->rw.len;
3304 nr = file->f_op->read(file, iovec.iov_base,
3305 iovec.iov_len, io_kiocb_ppos(kiocb));
3307 nr = file->f_op->write(file, iovec.iov_base,
3308 iovec.iov_len, io_kiocb_ppos(kiocb));
3317 if (!iov_iter_is_bvec(iter)) {
3318 iov_iter_advance(iter, nr);
3325 if (nr != iovec.iov_len)
3332 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3333 const struct iovec *fast_iov, struct iov_iter *iter)
3335 struct io_async_rw *rw = req->async_data;
3337 memcpy(&rw->iter, iter, sizeof(*iter));
3338 rw->free_iovec = iovec;
3340 /* can only be fixed buffers, no need to do anything */
3341 if (iov_iter_is_bvec(iter))
3344 unsigned iov_off = 0;
3346 rw->iter.iov = rw->fast_iov;
3347 if (iter->iov != fast_iov) {
3348 iov_off = iter->iov - fast_iov;
3349 rw->iter.iov += iov_off;
3351 if (rw->fast_iov != fast_iov)
3352 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3353 sizeof(struct iovec) * iter->nr_segs);
3355 req->flags |= REQ_F_NEED_CLEANUP;
3359 static inline int io_alloc_async_data(struct io_kiocb *req)
3361 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3362 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3363 return req->async_data == NULL;
3366 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3367 const struct iovec *fast_iov,
3368 struct iov_iter *iter, bool force)
3370 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3372 if (!req->async_data) {
3373 struct io_async_rw *iorw;
3375 if (io_alloc_async_data(req)) {
3380 io_req_map_rw(req, iovec, fast_iov, iter);
3381 iorw = req->async_data;
3382 /* we've copied and mapped the iter, ensure state is saved */
3383 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3388 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3390 struct io_async_rw *iorw = req->async_data;
3391 struct iovec *iov = iorw->fast_iov;
3394 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3395 if (unlikely(ret < 0))
3398 iorw->bytes_done = 0;
3399 iorw->free_iovec = iov;
3401 req->flags |= REQ_F_NEED_CLEANUP;
3402 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3406 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3408 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3410 return io_prep_rw(req, sqe, READ);
3414 * This is our waitqueue callback handler, registered through lock_page_async()
3415 * when we initially tried to do the IO with the iocb armed our waitqueue.
3416 * This gets called when the page is unlocked, and we generally expect that to
3417 * happen when the page IO is completed and the page is now uptodate. This will
3418 * queue a task_work based retry of the operation, attempting to copy the data
3419 * again. If the latter fails because the page was NOT uptodate, then we will
3420 * do a thread based blocking retry of the operation. That's the unexpected
3423 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3424 int sync, void *arg)
3426 struct wait_page_queue *wpq;
3427 struct io_kiocb *req = wait->private;
3428 struct wait_page_key *key = arg;
3430 wpq = container_of(wait, struct wait_page_queue, wait);
3432 if (!wake_page_match(wpq, key))
3435 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3436 list_del_init(&wait->entry);
3437 io_req_task_queue(req);
3442 * This controls whether a given IO request should be armed for async page
3443 * based retry. If we return false here, the request is handed to the async
3444 * worker threads for retry. If we're doing buffered reads on a regular file,
3445 * we prepare a private wait_page_queue entry and retry the operation. This
3446 * will either succeed because the page is now uptodate and unlocked, or it
3447 * will register a callback when the page is unlocked at IO completion. Through
3448 * that callback, io_uring uses task_work to setup a retry of the operation.
3449 * That retry will attempt the buffered read again. The retry will generally
3450 * succeed, or in rare cases where it fails, we then fall back to using the
3451 * async worker threads for a blocking retry.
3453 static bool io_rw_should_retry(struct io_kiocb *req)
3455 struct io_async_rw *rw = req->async_data;
3456 struct wait_page_queue *wait = &rw->wpq;
3457 struct kiocb *kiocb = &req->rw.kiocb;
3459 /* never retry for NOWAIT, we just complete with -EAGAIN */
3460 if (req->flags & REQ_F_NOWAIT)
3463 /* Only for buffered IO */
3464 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3468 * just use poll if we can, and don't attempt if the fs doesn't
3469 * support callback based unlocks
3471 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3474 wait->wait.func = io_async_buf_func;
3475 wait->wait.private = req;
3476 wait->wait.flags = 0;
3477 INIT_LIST_HEAD(&wait->wait.entry);
3478 kiocb->ki_flags |= IOCB_WAITQ;
3479 kiocb->ki_flags &= ~IOCB_NOWAIT;
3480 kiocb->ki_waitq = wait;
3484 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3486 if (req->file->f_op->read_iter)
3487 return call_read_iter(req->file, &req->rw.kiocb, iter);
3488 else if (req->file->f_op->read)
3489 return loop_rw_iter(READ, req, iter);
3494 static bool need_read_all(struct io_kiocb *req)
3496 return req->flags & REQ_F_ISREG ||
3497 S_ISBLK(file_inode(req->file)->i_mode);
3500 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3502 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3503 struct kiocb *kiocb = &req->rw.kiocb;
3504 struct iov_iter __iter, *iter = &__iter;
3505 struct io_async_rw *rw = req->async_data;
3506 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3507 struct iov_iter_state __state, *state;
3512 state = &rw->iter_state;
3514 * We come here from an earlier attempt, restore our state to
3515 * match in case it doesn't. It's cheap enough that we don't
3516 * need to make this conditional.
3518 iov_iter_restore(iter, state);
3521 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3525 iov_iter_save_state(iter, state);
3527 req->result = iov_iter_count(iter);
3529 /* Ensure we clear previously set non-block flag */
3530 if (!force_nonblock)
3531 kiocb->ki_flags &= ~IOCB_NOWAIT;
3533 kiocb->ki_flags |= IOCB_NOWAIT;
3535 /* If the file doesn't support async, just async punt */
3536 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3537 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3538 return ret ?: -EAGAIN;
3541 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3542 if (unlikely(ret)) {
3547 ret = io_iter_do_read(req, iter);
3549 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3550 req->flags &= ~REQ_F_REISSUE;
3551 /* IOPOLL retry should happen for io-wq threads */
3552 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3554 /* no retry on NONBLOCK nor RWF_NOWAIT */
3555 if (req->flags & REQ_F_NOWAIT)
3558 } else if (ret == -EIOCBQUEUED) {
3560 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3561 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3562 /* read all, failed, already did sync or don't want to retry */
3567 * Don't depend on the iter state matching what was consumed, or being
3568 * untouched in case of error. Restore it and we'll advance it
3569 * manually if we need to.
3571 iov_iter_restore(iter, state);
3573 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3578 rw = req->async_data;
3580 * Now use our persistent iterator and state, if we aren't already.
3581 * We've restored and mapped the iter to match.
3583 if (iter != &rw->iter) {
3585 state = &rw->iter_state;
3590 * We end up here because of a partial read, either from
3591 * above or inside this loop. Advance the iter by the bytes
3592 * that were consumed.
3594 iov_iter_advance(iter, ret);
3595 if (!iov_iter_count(iter))
3597 rw->bytes_done += ret;
3598 iov_iter_save_state(iter, state);
3600 /* if we can retry, do so with the callbacks armed */
3601 if (!io_rw_should_retry(req)) {
3602 kiocb->ki_flags &= ~IOCB_WAITQ;
3607 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3608 * we get -EIOCBQUEUED, then we'll get a notification when the
3609 * desired page gets unlocked. We can also get a partial read
3610 * here, and if we do, then just retry at the new offset.
3612 ret = io_iter_do_read(req, iter);
3613 if (ret == -EIOCBQUEUED)
3615 /* we got some bytes, but not all. retry. */
3616 kiocb->ki_flags &= ~IOCB_WAITQ;
3617 iov_iter_restore(iter, state);
3620 kiocb_done(kiocb, ret, issue_flags);
3622 /* it's faster to check here then delegate to kfree */
3628 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3630 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3632 return io_prep_rw(req, sqe, WRITE);
3635 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3637 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3638 struct kiocb *kiocb = &req->rw.kiocb;
3639 struct iov_iter __iter, *iter = &__iter;
3640 struct io_async_rw *rw = req->async_data;
3641 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3642 struct iov_iter_state __state, *state;
3647 state = &rw->iter_state;
3648 iov_iter_restore(iter, state);
3651 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3655 iov_iter_save_state(iter, state);
3657 req->result = iov_iter_count(iter);
3659 /* Ensure we clear previously set non-block flag */
3660 if (!force_nonblock)
3661 kiocb->ki_flags &= ~IOCB_NOWAIT;
3663 kiocb->ki_flags |= IOCB_NOWAIT;
3665 /* If the file doesn't support async, just async punt */
3666 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3669 /* file path doesn't support NOWAIT for non-direct_IO */
3670 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3671 (req->flags & REQ_F_ISREG))
3674 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3679 * Open-code file_start_write here to grab freeze protection,
3680 * which will be released by another thread in
3681 * io_complete_rw(). Fool lockdep by telling it the lock got
3682 * released so that it doesn't complain about the held lock when
3683 * we return to userspace.
3685 if (req->flags & REQ_F_ISREG) {
3686 sb_start_write(file_inode(req->file)->i_sb);
3687 __sb_writers_release(file_inode(req->file)->i_sb,
3690 kiocb->ki_flags |= IOCB_WRITE;
3692 if (req->file->f_op->write_iter)
3693 ret2 = call_write_iter(req->file, kiocb, iter);
3694 else if (req->file->f_op->write)
3695 ret2 = loop_rw_iter(WRITE, req, iter);
3699 if (req->flags & REQ_F_REISSUE) {
3700 req->flags &= ~REQ_F_REISSUE;
3705 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3706 * retry them without IOCB_NOWAIT.
3708 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3710 /* no retry on NONBLOCK nor RWF_NOWAIT */
3711 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3713 if (!force_nonblock || ret2 != -EAGAIN) {
3714 /* IOPOLL retry should happen for io-wq threads */
3715 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3718 kiocb_done(kiocb, ret2, issue_flags);
3721 iov_iter_restore(iter, state);
3722 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3724 if (kiocb->ki_flags & IOCB_WRITE)
3725 kiocb_end_write(req);
3731 /* it's reportedly faster than delegating the null check to kfree() */
3737 static int io_renameat_prep(struct io_kiocb *req,
3738 const struct io_uring_sqe *sqe)
3740 struct io_rename *ren = &req->rename;
3741 const char __user *oldf, *newf;
3743 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3745 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3747 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3750 ren->old_dfd = READ_ONCE(sqe->fd);
3751 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3752 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3753 ren->new_dfd = READ_ONCE(sqe->len);
3754 ren->flags = READ_ONCE(sqe->rename_flags);
3756 ren->oldpath = getname(oldf);
3757 if (IS_ERR(ren->oldpath))
3758 return PTR_ERR(ren->oldpath);
3760 ren->newpath = getname(newf);
3761 if (IS_ERR(ren->newpath)) {
3762 putname(ren->oldpath);
3763 return PTR_ERR(ren->newpath);
3766 req->flags |= REQ_F_NEED_CLEANUP;
3770 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3772 struct io_rename *ren = &req->rename;
3775 if (issue_flags & IO_URING_F_NONBLOCK)
3778 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3779 ren->newpath, ren->flags);
3781 req->flags &= ~REQ_F_NEED_CLEANUP;
3784 io_req_complete(req, ret);
3788 static int io_unlinkat_prep(struct io_kiocb *req,
3789 const struct io_uring_sqe *sqe)
3791 struct io_unlink *un = &req->unlink;
3792 const char __user *fname;
3794 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3796 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3799 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3802 un->dfd = READ_ONCE(sqe->fd);
3804 un->flags = READ_ONCE(sqe->unlink_flags);
3805 if (un->flags & ~AT_REMOVEDIR)
3808 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3809 un->filename = getname(fname);
3810 if (IS_ERR(un->filename))
3811 return PTR_ERR(un->filename);
3813 req->flags |= REQ_F_NEED_CLEANUP;
3817 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3819 struct io_unlink *un = &req->unlink;
3822 if (issue_flags & IO_URING_F_NONBLOCK)
3825 if (un->flags & AT_REMOVEDIR)
3826 ret = do_rmdir(un->dfd, un->filename);
3828 ret = do_unlinkat(un->dfd, un->filename);
3830 req->flags &= ~REQ_F_NEED_CLEANUP;
3833 io_req_complete(req, ret);
3837 static int io_mkdirat_prep(struct io_kiocb *req,
3838 const struct io_uring_sqe *sqe)
3840 struct io_mkdir *mkd = &req->mkdir;
3841 const char __user *fname;
3843 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3845 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3848 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3851 mkd->dfd = READ_ONCE(sqe->fd);
3852 mkd->mode = READ_ONCE(sqe->len);
3854 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3855 mkd->filename = getname(fname);
3856 if (IS_ERR(mkd->filename))
3857 return PTR_ERR(mkd->filename);
3859 req->flags |= REQ_F_NEED_CLEANUP;
3863 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3865 struct io_mkdir *mkd = &req->mkdir;
3868 if (issue_flags & IO_URING_F_NONBLOCK)
3871 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3873 req->flags &= ~REQ_F_NEED_CLEANUP;
3876 io_req_complete(req, ret);
3880 static int io_symlinkat_prep(struct io_kiocb *req,
3881 const struct io_uring_sqe *sqe)
3883 struct io_symlink *sl = &req->symlink;
3884 const char __user *oldpath, *newpath;
3886 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3888 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3891 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3894 sl->new_dfd = READ_ONCE(sqe->fd);
3895 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3896 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3898 sl->oldpath = getname(oldpath);
3899 if (IS_ERR(sl->oldpath))
3900 return PTR_ERR(sl->oldpath);
3902 sl->newpath = getname(newpath);
3903 if (IS_ERR(sl->newpath)) {
3904 putname(sl->oldpath);
3905 return PTR_ERR(sl->newpath);
3908 req->flags |= REQ_F_NEED_CLEANUP;
3912 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3914 struct io_symlink *sl = &req->symlink;
3917 if (issue_flags & IO_URING_F_NONBLOCK)
3920 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3922 req->flags &= ~REQ_F_NEED_CLEANUP;
3925 io_req_complete(req, ret);
3929 static int io_linkat_prep(struct io_kiocb *req,
3930 const struct io_uring_sqe *sqe)
3932 struct io_hardlink *lnk = &req->hardlink;
3933 const char __user *oldf, *newf;
3935 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3937 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3939 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3942 lnk->old_dfd = READ_ONCE(sqe->fd);
3943 lnk->new_dfd = READ_ONCE(sqe->len);
3944 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3945 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3946 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3948 lnk->oldpath = getname(oldf);
3949 if (IS_ERR(lnk->oldpath))
3950 return PTR_ERR(lnk->oldpath);
3952 lnk->newpath = getname(newf);
3953 if (IS_ERR(lnk->newpath)) {
3954 putname(lnk->oldpath);
3955 return PTR_ERR(lnk->newpath);
3958 req->flags |= REQ_F_NEED_CLEANUP;
3962 static int io_linkat(struct io_kiocb *req, int issue_flags)
3964 struct io_hardlink *lnk = &req->hardlink;
3967 if (issue_flags & IO_URING_F_NONBLOCK)
3970 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3971 lnk->newpath, lnk->flags);
3973 req->flags &= ~REQ_F_NEED_CLEANUP;
3976 io_req_complete(req, ret);
3980 static int io_shutdown_prep(struct io_kiocb *req,
3981 const struct io_uring_sqe *sqe)
3983 #if defined(CONFIG_NET)
3984 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3986 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3987 sqe->buf_index || sqe->splice_fd_in))
3990 req->shutdown.how = READ_ONCE(sqe->len);
3997 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3999 #if defined(CONFIG_NET)
4000 struct socket *sock;
4003 if (issue_flags & IO_URING_F_NONBLOCK)
4006 sock = sock_from_file(req->file);
4007 if (unlikely(!sock))
4010 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4013 io_req_complete(req, ret);
4020 static int __io_splice_prep(struct io_kiocb *req,
4021 const struct io_uring_sqe *sqe)
4023 struct io_splice *sp = &req->splice;
4024 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4026 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4029 sp->len = READ_ONCE(sqe->len);
4030 sp->flags = READ_ONCE(sqe->splice_flags);
4031 if (unlikely(sp->flags & ~valid_flags))
4033 sp->splice_fd_in = READ_ONCE(sqe->splice_fd_in);
4037 static int io_tee_prep(struct io_kiocb *req,
4038 const struct io_uring_sqe *sqe)
4040 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4042 return __io_splice_prep(req, sqe);
4045 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4047 struct io_splice *sp = &req->splice;
4048 struct file *out = sp->file_out;
4049 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4053 if (issue_flags & IO_URING_F_NONBLOCK)
4056 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4057 (sp->flags & SPLICE_F_FD_IN_FIXED));
4064 ret = do_tee(in, out, sp->len, flags);
4066 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4071 io_req_complete(req, ret);
4075 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4077 struct io_splice *sp = &req->splice;
4079 sp->off_in = READ_ONCE(sqe->splice_off_in);
4080 sp->off_out = READ_ONCE(sqe->off);
4081 return __io_splice_prep(req, sqe);
4084 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4086 struct io_splice *sp = &req->splice;
4087 struct file *out = sp->file_out;
4088 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4089 loff_t *poff_in, *poff_out;
4093 if (issue_flags & IO_URING_F_NONBLOCK)
4096 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4097 (sp->flags & SPLICE_F_FD_IN_FIXED));
4103 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4104 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4107 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4109 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4114 io_req_complete(req, ret);
4119 * IORING_OP_NOP just posts a completion event, nothing else.
4121 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4123 struct io_ring_ctx *ctx = req->ctx;
4125 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4128 __io_req_complete(req, issue_flags, 0, 0);
4132 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4134 struct io_ring_ctx *ctx = req->ctx;
4136 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4138 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4142 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4143 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4146 req->sync.off = READ_ONCE(sqe->off);
4147 req->sync.len = READ_ONCE(sqe->len);
4151 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4153 loff_t end = req->sync.off + req->sync.len;
4156 /* fsync always requires a blocking context */
4157 if (issue_flags & IO_URING_F_NONBLOCK)
4160 ret = vfs_fsync_range(req->file, req->sync.off,
4161 end > 0 ? end : LLONG_MAX,
4162 req->sync.flags & IORING_FSYNC_DATASYNC);
4165 io_req_complete(req, ret);
4169 static int io_fallocate_prep(struct io_kiocb *req,
4170 const struct io_uring_sqe *sqe)
4172 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4175 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4178 req->sync.off = READ_ONCE(sqe->off);
4179 req->sync.len = READ_ONCE(sqe->addr);
4180 req->sync.mode = READ_ONCE(sqe->len);
4184 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4188 /* fallocate always requiring blocking context */
4189 if (issue_flags & IO_URING_F_NONBLOCK)
4191 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4196 fsnotify_modify(req->file);
4197 io_req_complete(req, ret);
4201 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4203 const char __user *fname;
4206 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4208 if (unlikely(sqe->ioprio || sqe->buf_index))
4210 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4213 /* open.how should be already initialised */
4214 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4215 req->open.how.flags |= O_LARGEFILE;
4217 req->open.dfd = READ_ONCE(sqe->fd);
4218 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4219 req->open.filename = getname(fname);
4220 if (IS_ERR(req->open.filename)) {
4221 ret = PTR_ERR(req->open.filename);
4222 req->open.filename = NULL;
4226 req->open.file_slot = READ_ONCE(sqe->file_index);
4227 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4230 req->open.nofile = rlimit(RLIMIT_NOFILE);
4231 req->flags |= REQ_F_NEED_CLEANUP;
4235 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4237 u64 mode = READ_ONCE(sqe->len);
4238 u64 flags = READ_ONCE(sqe->open_flags);
4240 req->open.how = build_open_how(flags, mode);
4241 return __io_openat_prep(req, sqe);
4244 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4246 struct open_how __user *how;
4250 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4251 len = READ_ONCE(sqe->len);
4252 if (len < OPEN_HOW_SIZE_VER0)
4255 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4260 return __io_openat_prep(req, sqe);
4263 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4265 struct open_flags op;
4267 bool resolve_nonblock, nonblock_set;
4268 bool fixed = !!req->open.file_slot;
4271 ret = build_open_flags(&req->open.how, &op);
4274 nonblock_set = op.open_flag & O_NONBLOCK;
4275 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4276 if (issue_flags & IO_URING_F_NONBLOCK) {
4278 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4279 * it'll always -EAGAIN
4281 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4283 op.lookup_flags |= LOOKUP_CACHED;
4284 op.open_flag |= O_NONBLOCK;
4288 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4293 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4296 * We could hang on to this 'fd' on retrying, but seems like
4297 * marginal gain for something that is now known to be a slower
4298 * path. So just put it, and we'll get a new one when we retry.
4303 ret = PTR_ERR(file);
4304 /* only retry if RESOLVE_CACHED wasn't already set by application */
4305 if (ret == -EAGAIN &&
4306 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4311 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4312 file->f_flags &= ~O_NONBLOCK;
4313 fsnotify_open(file);
4316 fd_install(ret, file);
4318 ret = io_install_fixed_file(req, file, issue_flags,
4319 req->open.file_slot - 1);
4321 putname(req->open.filename);
4322 req->flags &= ~REQ_F_NEED_CLEANUP;
4325 __io_req_complete(req, issue_flags, ret, 0);
4329 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4331 return io_openat2(req, issue_flags);
4334 static int io_remove_buffers_prep(struct io_kiocb *req,
4335 const struct io_uring_sqe *sqe)
4337 struct io_provide_buf *p = &req->pbuf;
4340 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4344 tmp = READ_ONCE(sqe->fd);
4345 if (!tmp || tmp > USHRT_MAX)
4348 memset(p, 0, sizeof(*p));
4350 p->bgid = READ_ONCE(sqe->buf_group);
4354 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4355 int bgid, unsigned nbufs)
4359 /* shouldn't happen */
4363 /* the head kbuf is the list itself */
4364 while (!list_empty(&buf->list)) {
4365 struct io_buffer *nxt;
4367 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4368 list_del(&nxt->list);
4376 xa_erase(&ctx->io_buffers, bgid);
4381 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4383 struct io_provide_buf *p = &req->pbuf;
4384 struct io_ring_ctx *ctx = req->ctx;
4385 struct io_buffer *head;
4387 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4389 io_ring_submit_lock(ctx, !force_nonblock);
4391 lockdep_assert_held(&ctx->uring_lock);
4394 head = xa_load(&ctx->io_buffers, p->bgid);
4396 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4400 /* complete before unlock, IOPOLL may need the lock */
4401 __io_req_complete(req, issue_flags, ret, 0);
4402 io_ring_submit_unlock(ctx, !force_nonblock);
4406 static int io_provide_buffers_prep(struct io_kiocb *req,
4407 const struct io_uring_sqe *sqe)
4409 unsigned long size, tmp_check;
4410 struct io_provide_buf *p = &req->pbuf;
4413 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4416 tmp = READ_ONCE(sqe->fd);
4417 if (!tmp || tmp > USHRT_MAX)
4420 p->addr = READ_ONCE(sqe->addr);
4421 p->len = READ_ONCE(sqe->len);
4423 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4426 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4429 size = (unsigned long)p->len * p->nbufs;
4430 if (!access_ok(u64_to_user_ptr(p->addr), size))
4433 p->bgid = READ_ONCE(sqe->buf_group);
4434 tmp = READ_ONCE(sqe->off);
4435 if (tmp > USHRT_MAX)
4441 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4443 struct io_buffer *buf;
4444 u64 addr = pbuf->addr;
4445 int i, bid = pbuf->bid;
4447 for (i = 0; i < pbuf->nbufs; i++) {
4448 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4453 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4458 INIT_LIST_HEAD(&buf->list);
4461 list_add_tail(&buf->list, &(*head)->list);
4466 return i ? i : -ENOMEM;
4469 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4471 struct io_provide_buf *p = &req->pbuf;
4472 struct io_ring_ctx *ctx = req->ctx;
4473 struct io_buffer *head, *list;
4475 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4477 io_ring_submit_lock(ctx, !force_nonblock);
4479 lockdep_assert_held(&ctx->uring_lock);
4481 list = head = xa_load(&ctx->io_buffers, p->bgid);
4483 ret = io_add_buffers(p, &head);
4484 if (ret >= 0 && !list) {
4485 ret = xa_insert(&ctx->io_buffers, p->bgid, head,
4486 GFP_KERNEL_ACCOUNT);
4488 __io_remove_buffers(ctx, head, p->bgid, -1U);
4492 /* complete before unlock, IOPOLL may need the lock */
4493 __io_req_complete(req, issue_flags, ret, 0);
4494 io_ring_submit_unlock(ctx, !force_nonblock);
4498 static int io_epoll_ctl_prep(struct io_kiocb *req,
4499 const struct io_uring_sqe *sqe)
4501 #if defined(CONFIG_EPOLL)
4502 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4504 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4507 req->epoll.epfd = READ_ONCE(sqe->fd);
4508 req->epoll.op = READ_ONCE(sqe->len);
4509 req->epoll.fd = READ_ONCE(sqe->off);
4511 if (ep_op_has_event(req->epoll.op)) {
4512 struct epoll_event __user *ev;
4514 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4515 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4525 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4527 #if defined(CONFIG_EPOLL)
4528 struct io_epoll *ie = &req->epoll;
4530 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4532 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4533 if (force_nonblock && ret == -EAGAIN)
4538 __io_req_complete(req, issue_flags, ret, 0);
4545 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4547 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4548 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4550 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4553 req->madvise.addr = READ_ONCE(sqe->addr);
4554 req->madvise.len = READ_ONCE(sqe->len);
4555 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4562 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4564 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4565 struct io_madvise *ma = &req->madvise;
4568 if (issue_flags & IO_URING_F_NONBLOCK)
4571 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4574 io_req_complete(req, ret);
4581 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4583 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4585 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4588 req->fadvise.offset = READ_ONCE(sqe->off);
4589 req->fadvise.len = READ_ONCE(sqe->len);
4590 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4594 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4596 struct io_fadvise *fa = &req->fadvise;
4599 if (issue_flags & IO_URING_F_NONBLOCK) {
4600 switch (fa->advice) {
4601 case POSIX_FADV_NORMAL:
4602 case POSIX_FADV_RANDOM:
4603 case POSIX_FADV_SEQUENTIAL:
4610 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4613 __io_req_complete(req, issue_flags, ret, 0);
4617 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4619 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4621 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4623 if (req->flags & REQ_F_FIXED_FILE)
4626 req->statx.dfd = READ_ONCE(sqe->fd);
4627 req->statx.mask = READ_ONCE(sqe->len);
4628 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4629 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4630 req->statx.flags = READ_ONCE(sqe->statx_flags);
4635 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4637 struct io_statx *ctx = &req->statx;
4640 if (issue_flags & IO_URING_F_NONBLOCK)
4643 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4648 io_req_complete(req, ret);
4652 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4654 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4656 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4657 sqe->rw_flags || sqe->buf_index)
4659 if (req->flags & REQ_F_FIXED_FILE)
4662 req->close.fd = READ_ONCE(sqe->fd);
4663 req->close.file_slot = READ_ONCE(sqe->file_index);
4664 if (req->close.file_slot && req->close.fd)
4670 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4672 struct files_struct *files = current->files;
4673 struct io_close *close = &req->close;
4674 struct fdtable *fdt;
4675 struct file *file = NULL;
4678 if (req->close.file_slot) {
4679 ret = io_close_fixed(req, issue_flags);
4683 spin_lock(&files->file_lock);
4684 fdt = files_fdtable(files);
4685 if (close->fd >= fdt->max_fds) {
4686 spin_unlock(&files->file_lock);
4689 file = fdt->fd[close->fd];
4690 if (!file || file->f_op == &io_uring_fops) {
4691 spin_unlock(&files->file_lock);
4696 /* if the file has a flush method, be safe and punt to async */
4697 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4698 spin_unlock(&files->file_lock);
4702 ret = __close_fd_get_file(close->fd, &file);
4703 spin_unlock(&files->file_lock);
4710 /* No ->flush() or already async, safely close from here */
4711 ret = filp_close(file, current->files);
4717 __io_req_complete(req, issue_flags, ret, 0);
4721 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4723 struct io_ring_ctx *ctx = req->ctx;
4725 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4727 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4731 req->sync.off = READ_ONCE(sqe->off);
4732 req->sync.len = READ_ONCE(sqe->len);
4733 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4737 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4741 /* sync_file_range always requires a blocking context */
4742 if (issue_flags & IO_URING_F_NONBLOCK)
4745 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4749 io_req_complete(req, ret);
4753 #if defined(CONFIG_NET)
4754 static int io_setup_async_msg(struct io_kiocb *req,
4755 struct io_async_msghdr *kmsg)
4757 struct io_async_msghdr *async_msg = req->async_data;
4761 if (io_alloc_async_data(req)) {
4762 kfree(kmsg->free_iov);
4765 async_msg = req->async_data;
4766 req->flags |= REQ_F_NEED_CLEANUP;
4767 memcpy(async_msg, kmsg, sizeof(*kmsg));
4768 async_msg->msg.msg_name = &async_msg->addr;
4769 /* if were using fast_iov, set it to the new one */
4770 if (!async_msg->free_iov)
4771 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4776 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4777 struct io_async_msghdr *iomsg)
4779 iomsg->msg.msg_name = &iomsg->addr;
4780 iomsg->free_iov = iomsg->fast_iov;
4781 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4782 req->sr_msg.msg_flags, &iomsg->free_iov);
4785 static int io_sendmsg_prep_async(struct io_kiocb *req)
4789 ret = io_sendmsg_copy_hdr(req, req->async_data);
4791 req->flags |= REQ_F_NEED_CLEANUP;
4795 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4797 struct io_sr_msg *sr = &req->sr_msg;
4799 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4801 if (unlikely(sqe->addr2 || sqe->file_index))
4803 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
4806 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4807 sr->len = READ_ONCE(sqe->len);
4808 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4809 if (sr->msg_flags & MSG_DONTWAIT)
4810 req->flags |= REQ_F_NOWAIT;
4812 #ifdef CONFIG_COMPAT
4813 if (req->ctx->compat)
4814 sr->msg_flags |= MSG_CMSG_COMPAT;
4819 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4821 struct io_async_msghdr iomsg, *kmsg;
4822 struct socket *sock;
4827 sock = sock_from_file(req->file);
4828 if (unlikely(!sock))
4831 kmsg = req->async_data;
4833 ret = io_sendmsg_copy_hdr(req, &iomsg);
4839 flags = req->sr_msg.msg_flags;
4840 if (issue_flags & IO_URING_F_NONBLOCK)
4841 flags |= MSG_DONTWAIT;
4842 if (flags & MSG_WAITALL)
4843 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4845 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4846 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4847 return io_setup_async_msg(req, kmsg);
4848 if (ret == -ERESTARTSYS)
4851 /* fast path, check for non-NULL to avoid function call */
4853 kfree(kmsg->free_iov);
4854 req->flags &= ~REQ_F_NEED_CLEANUP;
4857 __io_req_complete(req, issue_flags, ret, 0);
4861 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4863 struct io_sr_msg *sr = &req->sr_msg;
4866 struct socket *sock;
4871 sock = sock_from_file(req->file);
4872 if (unlikely(!sock))
4875 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4879 msg.msg_name = NULL;
4880 msg.msg_control = NULL;
4881 msg.msg_controllen = 0;
4882 msg.msg_namelen = 0;
4884 flags = req->sr_msg.msg_flags;
4885 if (issue_flags & IO_URING_F_NONBLOCK)
4886 flags |= MSG_DONTWAIT;
4887 if (flags & MSG_WAITALL)
4888 min_ret = iov_iter_count(&msg.msg_iter);
4890 msg.msg_flags = flags;
4891 ret = sock_sendmsg(sock, &msg);
4892 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4894 if (ret == -ERESTARTSYS)
4899 __io_req_complete(req, issue_flags, ret, 0);
4903 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4904 struct io_async_msghdr *iomsg)
4906 struct io_sr_msg *sr = &req->sr_msg;
4907 struct iovec __user *uiov;
4911 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4912 &iomsg->uaddr, &uiov, &iov_len);
4916 if (req->flags & REQ_F_BUFFER_SELECT) {
4919 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4921 sr->len = iomsg->fast_iov[0].iov_len;
4922 iomsg->free_iov = NULL;
4924 iomsg->free_iov = iomsg->fast_iov;
4925 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4926 &iomsg->free_iov, &iomsg->msg.msg_iter,
4935 #ifdef CONFIG_COMPAT
4936 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4937 struct io_async_msghdr *iomsg)
4939 struct io_sr_msg *sr = &req->sr_msg;
4940 struct compat_iovec __user *uiov;
4945 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4950 uiov = compat_ptr(ptr);
4951 if (req->flags & REQ_F_BUFFER_SELECT) {
4952 compat_ssize_t clen;
4956 if (!access_ok(uiov, sizeof(*uiov)))
4958 if (__get_user(clen, &uiov->iov_len))
4963 iomsg->free_iov = NULL;
4965 iomsg->free_iov = iomsg->fast_iov;
4966 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4967 UIO_FASTIOV, &iomsg->free_iov,
4968 &iomsg->msg.msg_iter, true);
4977 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4978 struct io_async_msghdr *iomsg)
4980 iomsg->msg.msg_name = &iomsg->addr;
4982 #ifdef CONFIG_COMPAT
4983 if (req->ctx->compat)
4984 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4987 return __io_recvmsg_copy_hdr(req, iomsg);
4990 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4993 struct io_sr_msg *sr = &req->sr_msg;
4994 struct io_buffer *kbuf;
4996 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
5001 req->flags |= REQ_F_BUFFER_SELECTED;
5005 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
5007 return io_put_kbuf(req, req->sr_msg.kbuf);
5010 static int io_recvmsg_prep_async(struct io_kiocb *req)
5014 ret = io_recvmsg_copy_hdr(req, req->async_data);
5016 req->flags |= REQ_F_NEED_CLEANUP;
5020 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5022 struct io_sr_msg *sr = &req->sr_msg;
5024 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5026 if (unlikely(sqe->addr2 || sqe->file_index))
5028 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
5031 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5032 sr->len = READ_ONCE(sqe->len);
5033 sr->bgid = READ_ONCE(sqe->buf_group);
5034 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5035 if (sr->msg_flags & MSG_DONTWAIT)
5036 req->flags |= REQ_F_NOWAIT;
5038 #ifdef CONFIG_COMPAT
5039 if (req->ctx->compat)
5040 sr->msg_flags |= MSG_CMSG_COMPAT;
5045 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5047 struct io_async_msghdr iomsg, *kmsg;
5048 struct socket *sock;
5049 struct io_buffer *kbuf;
5052 int ret, cflags = 0;
5053 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5055 sock = sock_from_file(req->file);
5056 if (unlikely(!sock))
5059 kmsg = req->async_data;
5061 ret = io_recvmsg_copy_hdr(req, &iomsg);
5067 if (req->flags & REQ_F_BUFFER_SELECT) {
5068 kbuf = io_recv_buffer_select(req, !force_nonblock);
5070 return PTR_ERR(kbuf);
5071 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5072 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5073 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5074 1, req->sr_msg.len);
5077 flags = req->sr_msg.msg_flags;
5079 flags |= MSG_DONTWAIT;
5080 if (flags & MSG_WAITALL)
5081 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5083 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5084 kmsg->uaddr, flags);
5085 if (force_nonblock && ret == -EAGAIN)
5086 return io_setup_async_msg(req, kmsg);
5087 if (ret == -ERESTARTSYS)
5090 if (req->flags & REQ_F_BUFFER_SELECTED)
5091 cflags = io_put_recv_kbuf(req);
5092 /* fast path, check for non-NULL to avoid function call */
5094 kfree(kmsg->free_iov);
5095 req->flags &= ~REQ_F_NEED_CLEANUP;
5096 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5098 __io_req_complete(req, issue_flags, ret, cflags);
5102 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5104 struct io_buffer *kbuf;
5105 struct io_sr_msg *sr = &req->sr_msg;
5107 void __user *buf = sr->buf;
5108 struct socket *sock;
5112 int ret, cflags = 0;
5113 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5115 sock = sock_from_file(req->file);
5116 if (unlikely(!sock))
5119 if (req->flags & REQ_F_BUFFER_SELECT) {
5120 kbuf = io_recv_buffer_select(req, !force_nonblock);
5122 return PTR_ERR(kbuf);
5123 buf = u64_to_user_ptr(kbuf->addr);
5126 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5130 msg.msg_name = NULL;
5131 msg.msg_control = NULL;
5132 msg.msg_controllen = 0;
5133 msg.msg_namelen = 0;
5134 msg.msg_iocb = NULL;
5137 flags = req->sr_msg.msg_flags;
5139 flags |= MSG_DONTWAIT;
5140 if (flags & MSG_WAITALL)
5141 min_ret = iov_iter_count(&msg.msg_iter);
5143 ret = sock_recvmsg(sock, &msg, flags);
5144 if (force_nonblock && ret == -EAGAIN)
5146 if (ret == -ERESTARTSYS)
5149 if (req->flags & REQ_F_BUFFER_SELECTED)
5150 cflags = io_put_recv_kbuf(req);
5151 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5153 __io_req_complete(req, issue_flags, ret, cflags);
5157 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5159 struct io_accept *accept = &req->accept;
5161 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5163 if (sqe->ioprio || sqe->len || sqe->buf_index)
5166 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5167 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5168 accept->flags = READ_ONCE(sqe->accept_flags);
5169 accept->nofile = rlimit(RLIMIT_NOFILE);
5171 accept->file_slot = READ_ONCE(sqe->file_index);
5172 if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))
5174 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5176 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5177 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5181 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5183 struct io_accept *accept = &req->accept;
5184 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5185 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5186 bool fixed = !!accept->file_slot;
5190 if (req->file->f_flags & O_NONBLOCK)
5191 req->flags |= REQ_F_NOWAIT;
5194 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5195 if (unlikely(fd < 0))
5198 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5203 ret = PTR_ERR(file);
5204 if (ret == -EAGAIN && force_nonblock)
5206 if (ret == -ERESTARTSYS)
5209 } else if (!fixed) {
5210 fd_install(fd, file);
5213 ret = io_install_fixed_file(req, file, issue_flags,
5214 accept->file_slot - 1);
5216 __io_req_complete(req, issue_flags, ret, 0);
5220 static int io_connect_prep_async(struct io_kiocb *req)
5222 struct io_async_connect *io = req->async_data;
5223 struct io_connect *conn = &req->connect;
5225 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5228 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5230 struct io_connect *conn = &req->connect;
5232 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5234 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5238 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5239 conn->addr_len = READ_ONCE(sqe->addr2);
5243 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5245 struct io_async_connect __io, *io;
5246 unsigned file_flags;
5248 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5250 if (req->async_data) {
5251 io = req->async_data;
5253 ret = move_addr_to_kernel(req->connect.addr,
5254 req->connect.addr_len,
5261 file_flags = force_nonblock ? O_NONBLOCK : 0;
5263 ret = __sys_connect_file(req->file, &io->address,
5264 req->connect.addr_len, file_flags);
5265 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5266 if (req->async_data)
5268 if (io_alloc_async_data(req)) {
5272 memcpy(req->async_data, &__io, sizeof(__io));
5275 if (ret == -ERESTARTSYS)
5280 __io_req_complete(req, issue_flags, ret, 0);
5283 #else /* !CONFIG_NET */
5284 #define IO_NETOP_FN(op) \
5285 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5287 return -EOPNOTSUPP; \
5290 #define IO_NETOP_PREP(op) \
5292 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5294 return -EOPNOTSUPP; \
5297 #define IO_NETOP_PREP_ASYNC(op) \
5299 static int io_##op##_prep_async(struct io_kiocb *req) \
5301 return -EOPNOTSUPP; \
5304 IO_NETOP_PREP_ASYNC(sendmsg);
5305 IO_NETOP_PREP_ASYNC(recvmsg);
5306 IO_NETOP_PREP_ASYNC(connect);
5307 IO_NETOP_PREP(accept);
5310 #endif /* CONFIG_NET */
5312 struct io_poll_table {
5313 struct poll_table_struct pt;
5314 struct io_kiocb *req;
5319 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5320 __poll_t mask, io_req_tw_func_t func)
5322 /* for instances that support it check for an event match first: */
5323 if (mask && !(mask & poll->events))
5326 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5328 list_del_init(&poll->wait.entry);
5331 req->io_task_work.func = func;
5334 * If this fails, then the task is exiting. When a task exits, the
5335 * work gets canceled, so just cancel this request as well instead
5336 * of executing it. We can't safely execute it anyway, as we may not
5337 * have the needed state needed for it anyway.
5339 io_req_task_work_add(req);
5343 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5344 __acquires(&req->ctx->completion_lock)
5346 struct io_ring_ctx *ctx = req->ctx;
5348 /* req->task == current here, checking PF_EXITING is safe */
5349 if (unlikely(req->task->flags & PF_EXITING))
5350 WRITE_ONCE(poll->canceled, true);
5352 if (!req->result && !READ_ONCE(poll->canceled)) {
5353 struct poll_table_struct pt = { ._key = poll->events };
5355 req->result = vfs_poll(req->file, &pt) & poll->events;
5358 spin_lock(&ctx->completion_lock);
5359 if (!req->result && !READ_ONCE(poll->canceled)) {
5360 add_wait_queue(poll->head, &poll->wait);
5367 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5369 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5370 if (req->opcode == IORING_OP_POLL_ADD)
5371 return req->async_data;
5372 return req->apoll->double_poll;
5375 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5377 if (req->opcode == IORING_OP_POLL_ADD)
5379 return &req->apoll->poll;
5382 static void io_poll_remove_double(struct io_kiocb *req)
5383 __must_hold(&req->ctx->completion_lock)
5385 struct io_poll_iocb *poll = io_poll_get_double(req);
5387 lockdep_assert_held(&req->ctx->completion_lock);
5389 if (poll && poll->head) {
5390 struct wait_queue_head *head = poll->head;
5392 spin_lock_irq(&head->lock);
5393 list_del_init(&poll->wait.entry);
5394 if (poll->wait.private)
5397 spin_unlock_irq(&head->lock);
5401 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5402 __must_hold(&req->ctx->completion_lock)
5404 struct io_ring_ctx *ctx = req->ctx;
5405 unsigned flags = IORING_CQE_F_MORE;
5408 if (READ_ONCE(req->poll.canceled)) {
5410 req->poll.events |= EPOLLONESHOT;
5412 error = mangle_poll(mask);
5414 if (req->poll.events & EPOLLONESHOT)
5416 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5417 req->poll.events |= EPOLLONESHOT;
5420 if (flags & IORING_CQE_F_MORE)
5423 return !(flags & IORING_CQE_F_MORE);
5426 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5427 __must_hold(&req->ctx->completion_lock)
5431 done = __io_poll_complete(req, mask);
5432 io_commit_cqring(req->ctx);
5436 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5438 struct io_ring_ctx *ctx = req->ctx;
5439 struct io_kiocb *nxt;
5441 if (io_poll_rewait(req, &req->poll)) {
5442 spin_unlock(&ctx->completion_lock);
5446 if (req->poll.done) {
5447 spin_unlock(&ctx->completion_lock);
5450 done = __io_poll_complete(req, req->result);
5452 io_poll_remove_double(req);
5453 hash_del(&req->hash_node);
5454 req->poll.done = true;
5457 add_wait_queue(req->poll.head, &req->poll.wait);
5459 io_commit_cqring(ctx);
5460 spin_unlock(&ctx->completion_lock);
5461 io_cqring_ev_posted(ctx);
5464 nxt = io_put_req_find_next(req);
5466 io_req_task_submit(nxt, locked);
5471 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5472 int sync, void *key)
5474 struct io_kiocb *req = wait->private;
5475 struct io_poll_iocb *poll = io_poll_get_single(req);
5476 __poll_t mask = key_to_poll(key);
5477 unsigned long flags;
5479 /* for instances that support it check for an event match first: */
5480 if (mask && !(mask & poll->events))
5482 if (!(poll->events & EPOLLONESHOT))
5483 return poll->wait.func(&poll->wait, mode, sync, key);
5485 list_del_init(&wait->entry);
5490 spin_lock_irqsave(&poll->head->lock, flags);
5491 done = list_empty(&poll->wait.entry);
5493 list_del_init(&poll->wait.entry);
5494 /* make sure double remove sees this as being gone */
5495 wait->private = NULL;
5496 spin_unlock_irqrestore(&poll->head->lock, flags);
5498 /* use wait func handler, so it matches the rq type */
5499 poll->wait.func(&poll->wait, mode, sync, key);
5506 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5507 wait_queue_func_t wake_func)
5511 poll->canceled = false;
5512 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5513 /* mask in events that we always want/need */
5514 poll->events = events | IO_POLL_UNMASK;
5515 INIT_LIST_HEAD(&poll->wait.entry);
5516 init_waitqueue_func_entry(&poll->wait, wake_func);
5519 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5520 struct wait_queue_head *head,
5521 struct io_poll_iocb **poll_ptr)
5523 struct io_kiocb *req = pt->req;
5526 * The file being polled uses multiple waitqueues for poll handling
5527 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5530 if (unlikely(pt->nr_entries)) {
5531 struct io_poll_iocb *poll_one = poll;
5533 /* double add on the same waitqueue head, ignore */
5534 if (poll_one->head == head)
5536 /* already have a 2nd entry, fail a third attempt */
5538 if ((*poll_ptr)->head == head)
5540 pt->error = -EINVAL;
5544 * Can't handle multishot for double wait for now, turn it
5545 * into one-shot mode.
5547 if (!(poll_one->events & EPOLLONESHOT))
5548 poll_one->events |= EPOLLONESHOT;
5549 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5551 pt->error = -ENOMEM;
5554 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5556 poll->wait.private = req;
5563 if (poll->events & EPOLLEXCLUSIVE)
5564 add_wait_queue_exclusive(head, &poll->wait);
5566 add_wait_queue(head, &poll->wait);
5569 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5570 struct poll_table_struct *p)
5572 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5573 struct async_poll *apoll = pt->req->apoll;
5575 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5578 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5580 struct async_poll *apoll = req->apoll;
5581 struct io_ring_ctx *ctx = req->ctx;
5583 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5585 if (io_poll_rewait(req, &apoll->poll)) {
5586 spin_unlock(&ctx->completion_lock);
5590 hash_del(&req->hash_node);
5591 io_poll_remove_double(req);
5592 apoll->poll.done = true;
5593 spin_unlock(&ctx->completion_lock);
5595 if (!READ_ONCE(apoll->poll.canceled))
5596 io_req_task_submit(req, locked);
5598 io_req_complete_failed(req, -ECANCELED);
5601 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5604 struct io_kiocb *req = wait->private;
5605 struct io_poll_iocb *poll = &req->apoll->poll;
5607 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5610 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5613 static void io_poll_req_insert(struct io_kiocb *req)
5615 struct io_ring_ctx *ctx = req->ctx;
5616 struct hlist_head *list;
5618 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5619 hlist_add_head(&req->hash_node, list);
5622 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5623 struct io_poll_iocb *poll,
5624 struct io_poll_table *ipt, __poll_t mask,
5625 wait_queue_func_t wake_func)
5626 __acquires(&ctx->completion_lock)
5628 struct io_ring_ctx *ctx = req->ctx;
5629 bool cancel = false;
5631 INIT_HLIST_NODE(&req->hash_node);
5632 io_init_poll_iocb(poll, mask, wake_func);
5633 poll->file = req->file;
5634 poll->wait.private = req;
5636 ipt->pt._key = mask;
5639 ipt->nr_entries = 0;
5641 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5642 if (unlikely(!ipt->nr_entries) && !ipt->error)
5643 ipt->error = -EINVAL;
5645 spin_lock(&ctx->completion_lock);
5646 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5647 io_poll_remove_double(req);
5648 if (likely(poll->head)) {
5649 spin_lock_irq(&poll->head->lock);
5650 if (unlikely(list_empty(&poll->wait.entry))) {
5656 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5657 list_del_init(&poll->wait.entry);
5659 WRITE_ONCE(poll->canceled, true);
5660 else if (!poll->done) /* actually waiting for an event */
5661 io_poll_req_insert(req);
5662 spin_unlock_irq(&poll->head->lock);
5674 static int io_arm_poll_handler(struct io_kiocb *req)
5676 const struct io_op_def *def = &io_op_defs[req->opcode];
5677 struct io_ring_ctx *ctx = req->ctx;
5678 struct async_poll *apoll;
5679 struct io_poll_table ipt;
5680 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5682 if (!req->file || !file_can_poll(req->file))
5683 return IO_APOLL_ABORTED;
5684 if (req->flags & REQ_F_POLLED)
5685 return IO_APOLL_ABORTED;
5686 if (!def->pollin && !def->pollout)
5687 return IO_APOLL_ABORTED;
5690 mask |= POLLIN | POLLRDNORM;
5692 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5693 if ((req->opcode == IORING_OP_RECVMSG) &&
5694 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5697 mask |= POLLOUT | POLLWRNORM;
5700 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5701 if (unlikely(!apoll))
5702 return IO_APOLL_ABORTED;
5703 apoll->double_poll = NULL;
5705 req->flags |= REQ_F_POLLED;
5706 ipt.pt._qproc = io_async_queue_proc;
5707 io_req_set_refcount(req);
5709 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5711 spin_unlock(&ctx->completion_lock);
5712 if (ret || ipt.error)
5713 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5715 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5716 mask, apoll->poll.events);
5720 static bool __io_poll_remove_one(struct io_kiocb *req,
5721 struct io_poll_iocb *poll, bool do_cancel)
5722 __must_hold(&req->ctx->completion_lock)
5724 bool do_complete = false;
5728 spin_lock_irq(&poll->head->lock);
5730 WRITE_ONCE(poll->canceled, true);
5731 if (!list_empty(&poll->wait.entry)) {
5732 list_del_init(&poll->wait.entry);
5735 spin_unlock_irq(&poll->head->lock);
5736 hash_del(&req->hash_node);
5740 static bool io_poll_remove_one(struct io_kiocb *req)
5741 __must_hold(&req->ctx->completion_lock)
5745 io_poll_remove_double(req);
5746 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5749 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5750 io_commit_cqring(req->ctx);
5752 io_put_req_deferred(req);
5758 * Returns true if we found and killed one or more poll requests
5760 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5763 struct hlist_node *tmp;
5764 struct io_kiocb *req;
5767 spin_lock(&ctx->completion_lock);
5768 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5769 struct hlist_head *list;
5771 list = &ctx->cancel_hash[i];
5772 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5773 if (io_match_task_safe(req, tsk, cancel_all))
5774 posted += io_poll_remove_one(req);
5777 spin_unlock(&ctx->completion_lock);
5780 io_cqring_ev_posted(ctx);
5785 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5787 __must_hold(&ctx->completion_lock)
5789 struct hlist_head *list;
5790 struct io_kiocb *req;
5792 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5793 hlist_for_each_entry(req, list, hash_node) {
5794 if (sqe_addr != req->user_data)
5796 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5803 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5805 __must_hold(&ctx->completion_lock)
5807 struct io_kiocb *req;
5809 req = io_poll_find(ctx, sqe_addr, poll_only);
5812 if (io_poll_remove_one(req))
5818 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5823 events = READ_ONCE(sqe->poll32_events);
5825 events = swahw32(events);
5827 if (!(flags & IORING_POLL_ADD_MULTI))
5828 events |= EPOLLONESHOT;
5829 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5832 static int io_poll_update_prep(struct io_kiocb *req,
5833 const struct io_uring_sqe *sqe)
5835 struct io_poll_update *upd = &req->poll_update;
5838 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5840 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5842 flags = READ_ONCE(sqe->len);
5843 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5844 IORING_POLL_ADD_MULTI))
5846 /* meaningless without update */
5847 if (flags == IORING_POLL_ADD_MULTI)
5850 upd->old_user_data = READ_ONCE(sqe->addr);
5851 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5852 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5854 upd->new_user_data = READ_ONCE(sqe->off);
5855 if (!upd->update_user_data && upd->new_user_data)
5857 if (upd->update_events)
5858 upd->events = io_poll_parse_events(sqe, flags);
5859 else if (sqe->poll32_events)
5865 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5868 struct io_kiocb *req = wait->private;
5869 struct io_poll_iocb *poll = &req->poll;
5871 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5874 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5875 struct poll_table_struct *p)
5877 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5879 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5882 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5884 struct io_poll_iocb *poll = &req->poll;
5887 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5889 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5891 flags = READ_ONCE(sqe->len);
5892 if (flags & ~IORING_POLL_ADD_MULTI)
5895 io_req_set_refcount(req);
5896 poll->events = io_poll_parse_events(sqe, flags);
5900 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5902 struct io_poll_iocb *poll = &req->poll;
5903 struct io_ring_ctx *ctx = req->ctx;
5904 struct io_poll_table ipt;
5908 ipt.pt._qproc = io_poll_queue_proc;
5910 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5913 if (mask) { /* no async, we'd stolen it */
5915 done = io_poll_complete(req, mask);
5917 spin_unlock(&ctx->completion_lock);
5920 io_cqring_ev_posted(ctx);
5927 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5929 struct io_ring_ctx *ctx = req->ctx;
5930 struct io_kiocb *preq;
5934 spin_lock(&ctx->completion_lock);
5935 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5941 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5943 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5948 * Don't allow racy completion with singleshot, as we cannot safely
5949 * update those. For multishot, if we're racing with completion, just
5950 * let completion re-add it.
5952 io_poll_remove_double(preq);
5953 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5954 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5958 /* we now have a detached poll request. reissue. */
5962 spin_unlock(&ctx->completion_lock);
5964 io_req_complete(req, ret);
5967 /* only mask one event flags, keep behavior flags */
5968 if (req->poll_update.update_events) {
5969 preq->poll.events &= ~0xffff;
5970 preq->poll.events |= req->poll_update.events & 0xffff;
5971 preq->poll.events |= IO_POLL_UNMASK;
5973 if (req->poll_update.update_user_data)
5974 preq->user_data = req->poll_update.new_user_data;
5975 spin_unlock(&ctx->completion_lock);
5977 /* complete update request, we're done with it */
5978 io_req_complete(req, ret);
5981 ret = io_poll_add(preq, issue_flags);
5984 io_req_complete(preq, ret);
5990 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5993 io_req_complete_post(req, -ETIME, 0);
5996 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5998 struct io_timeout_data *data = container_of(timer,
5999 struct io_timeout_data, timer);
6000 struct io_kiocb *req = data->req;
6001 struct io_ring_ctx *ctx = req->ctx;
6002 unsigned long flags;
6004 spin_lock_irqsave(&ctx->timeout_lock, flags);
6005 list_del_init(&req->timeout.list);
6006 atomic_set(&req->ctx->cq_timeouts,
6007 atomic_read(&req->ctx->cq_timeouts) + 1);
6008 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6010 req->io_task_work.func = io_req_task_timeout;
6011 io_req_task_work_add(req);
6012 return HRTIMER_NORESTART;
6015 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
6017 __must_hold(&ctx->timeout_lock)
6019 struct io_timeout_data *io;
6020 struct io_kiocb *req;
6023 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
6024 found = user_data == req->user_data;
6029 return ERR_PTR(-ENOENT);
6031 io = req->async_data;
6032 if (hrtimer_try_to_cancel(&io->timer) == -1)
6033 return ERR_PTR(-EALREADY);
6034 list_del_init(&req->timeout.list);
6038 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
6039 __must_hold(&ctx->completion_lock)
6040 __must_hold(&ctx->timeout_lock)
6042 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6045 return PTR_ERR(req);
6048 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
6049 io_put_req_deferred(req);
6053 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6055 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6056 case IORING_TIMEOUT_BOOTTIME:
6057 return CLOCK_BOOTTIME;
6058 case IORING_TIMEOUT_REALTIME:
6059 return CLOCK_REALTIME;
6061 /* can't happen, vetted at prep time */
6065 return CLOCK_MONOTONIC;
6069 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6070 struct timespec64 *ts, enum hrtimer_mode mode)
6071 __must_hold(&ctx->timeout_lock)
6073 struct io_timeout_data *io;
6074 struct io_kiocb *req;
6077 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6078 found = user_data == req->user_data;
6085 io = req->async_data;
6086 if (hrtimer_try_to_cancel(&io->timer) == -1)
6088 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6089 io->timer.function = io_link_timeout_fn;
6090 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6094 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6095 struct timespec64 *ts, enum hrtimer_mode mode)
6096 __must_hold(&ctx->timeout_lock)
6098 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6099 struct io_timeout_data *data;
6102 return PTR_ERR(req);
6104 req->timeout.off = 0; /* noseq */
6105 data = req->async_data;
6106 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6107 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6108 data->timer.function = io_timeout_fn;
6109 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6113 static int io_timeout_remove_prep(struct io_kiocb *req,
6114 const struct io_uring_sqe *sqe)
6116 struct io_timeout_rem *tr = &req->timeout_rem;
6118 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6120 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6122 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6125 tr->ltimeout = false;
6126 tr->addr = READ_ONCE(sqe->addr);
6127 tr->flags = READ_ONCE(sqe->timeout_flags);
6128 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6129 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6131 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6132 tr->ltimeout = true;
6133 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6135 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6137 } else if (tr->flags) {
6138 /* timeout removal doesn't support flags */
6145 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6147 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6152 * Remove or update an existing timeout command
6154 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6156 struct io_timeout_rem *tr = &req->timeout_rem;
6157 struct io_ring_ctx *ctx = req->ctx;
6160 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6161 spin_lock(&ctx->completion_lock);
6162 spin_lock_irq(&ctx->timeout_lock);
6163 ret = io_timeout_cancel(ctx, tr->addr);
6164 spin_unlock_irq(&ctx->timeout_lock);
6165 spin_unlock(&ctx->completion_lock);
6167 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6169 spin_lock_irq(&ctx->timeout_lock);
6171 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6173 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6174 spin_unlock_irq(&ctx->timeout_lock);
6179 io_req_complete_post(req, ret, 0);
6183 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6184 bool is_timeout_link)
6186 struct io_timeout_data *data;
6188 u32 off = READ_ONCE(sqe->off);
6190 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6192 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6195 if (off && is_timeout_link)
6197 flags = READ_ONCE(sqe->timeout_flags);
6198 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6200 /* more than one clock specified is invalid, obviously */
6201 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6204 INIT_LIST_HEAD(&req->timeout.list);
6205 req->timeout.off = off;
6206 if (unlikely(off && !req->ctx->off_timeout_used))
6207 req->ctx->off_timeout_used = true;
6209 if (!req->async_data && io_alloc_async_data(req))
6212 data = req->async_data;
6214 data->flags = flags;
6216 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6219 INIT_LIST_HEAD(&req->timeout.list);
6220 data->mode = io_translate_timeout_mode(flags);
6221 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6223 if (is_timeout_link) {
6224 struct io_submit_link *link = &req->ctx->submit_state.link;
6228 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6230 req->timeout.head = link->last;
6231 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6236 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6238 struct io_ring_ctx *ctx = req->ctx;
6239 struct io_timeout_data *data = req->async_data;
6240 struct list_head *entry;
6241 u32 tail, off = req->timeout.off;
6243 spin_lock_irq(&ctx->timeout_lock);
6246 * sqe->off holds how many events that need to occur for this
6247 * timeout event to be satisfied. If it isn't set, then this is
6248 * a pure timeout request, sequence isn't used.
6250 if (io_is_timeout_noseq(req)) {
6251 entry = ctx->timeout_list.prev;
6255 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6256 req->timeout.target_seq = tail + off;
6258 /* Update the last seq here in case io_flush_timeouts() hasn't.
6259 * This is safe because ->completion_lock is held, and submissions
6260 * and completions are never mixed in the same ->completion_lock section.
6262 ctx->cq_last_tm_flush = tail;
6265 * Insertion sort, ensuring the first entry in the list is always
6266 * the one we need first.
6268 list_for_each_prev(entry, &ctx->timeout_list) {
6269 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6272 if (io_is_timeout_noseq(nxt))
6274 /* nxt.seq is behind @tail, otherwise would've been completed */
6275 if (off >= nxt->timeout.target_seq - tail)
6279 list_add(&req->timeout.list, entry);
6280 data->timer.function = io_timeout_fn;
6281 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6282 spin_unlock_irq(&ctx->timeout_lock);
6286 struct io_cancel_data {
6287 struct io_ring_ctx *ctx;
6291 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6293 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6294 struct io_cancel_data *cd = data;
6296 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6299 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6300 struct io_ring_ctx *ctx)
6302 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6303 enum io_wq_cancel cancel_ret;
6306 if (!tctx || !tctx->io_wq)
6309 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6310 switch (cancel_ret) {
6311 case IO_WQ_CANCEL_OK:
6314 case IO_WQ_CANCEL_RUNNING:
6317 case IO_WQ_CANCEL_NOTFOUND:
6325 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6327 struct io_ring_ctx *ctx = req->ctx;
6330 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6332 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6336 spin_lock(&ctx->completion_lock);
6337 spin_lock_irq(&ctx->timeout_lock);
6338 ret = io_timeout_cancel(ctx, sqe_addr);
6339 spin_unlock_irq(&ctx->timeout_lock);
6342 ret = io_poll_cancel(ctx, sqe_addr, false);
6344 spin_unlock(&ctx->completion_lock);
6348 static int io_async_cancel_prep(struct io_kiocb *req,
6349 const struct io_uring_sqe *sqe)
6351 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6353 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6355 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6359 req->cancel.addr = READ_ONCE(sqe->addr);
6363 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6365 struct io_ring_ctx *ctx = req->ctx;
6366 u64 sqe_addr = req->cancel.addr;
6367 struct io_tctx_node *node;
6370 ret = io_try_cancel_userdata(req, sqe_addr);
6374 /* slow path, try all io-wq's */
6375 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6377 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6378 struct io_uring_task *tctx = node->task->io_uring;
6380 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6384 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6388 io_req_complete_post(req, ret, 0);
6392 static int io_rsrc_update_prep(struct io_kiocb *req,
6393 const struct io_uring_sqe *sqe)
6395 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6397 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6400 req->rsrc_update.offset = READ_ONCE(sqe->off);
6401 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6402 if (!req->rsrc_update.nr_args)
6404 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6408 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6410 struct io_ring_ctx *ctx = req->ctx;
6411 struct io_uring_rsrc_update2 up;
6414 up.offset = req->rsrc_update.offset;
6415 up.data = req->rsrc_update.arg;
6421 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6422 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6423 &up, req->rsrc_update.nr_args);
6424 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6428 __io_req_complete(req, issue_flags, ret, 0);
6432 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6434 switch (req->opcode) {
6437 case IORING_OP_READV:
6438 case IORING_OP_READ_FIXED:
6439 case IORING_OP_READ:
6440 return io_read_prep(req, sqe);
6441 case IORING_OP_WRITEV:
6442 case IORING_OP_WRITE_FIXED:
6443 case IORING_OP_WRITE:
6444 return io_write_prep(req, sqe);
6445 case IORING_OP_POLL_ADD:
6446 return io_poll_add_prep(req, sqe);
6447 case IORING_OP_POLL_REMOVE:
6448 return io_poll_update_prep(req, sqe);
6449 case IORING_OP_FSYNC:
6450 return io_fsync_prep(req, sqe);
6451 case IORING_OP_SYNC_FILE_RANGE:
6452 return io_sfr_prep(req, sqe);
6453 case IORING_OP_SENDMSG:
6454 case IORING_OP_SEND:
6455 return io_sendmsg_prep(req, sqe);
6456 case IORING_OP_RECVMSG:
6457 case IORING_OP_RECV:
6458 return io_recvmsg_prep(req, sqe);
6459 case IORING_OP_CONNECT:
6460 return io_connect_prep(req, sqe);
6461 case IORING_OP_TIMEOUT:
6462 return io_timeout_prep(req, sqe, false);
6463 case IORING_OP_TIMEOUT_REMOVE:
6464 return io_timeout_remove_prep(req, sqe);
6465 case IORING_OP_ASYNC_CANCEL:
6466 return io_async_cancel_prep(req, sqe);
6467 case IORING_OP_LINK_TIMEOUT:
6468 return io_timeout_prep(req, sqe, true);
6469 case IORING_OP_ACCEPT:
6470 return io_accept_prep(req, sqe);
6471 case IORING_OP_FALLOCATE:
6472 return io_fallocate_prep(req, sqe);
6473 case IORING_OP_OPENAT:
6474 return io_openat_prep(req, sqe);
6475 case IORING_OP_CLOSE:
6476 return io_close_prep(req, sqe);
6477 case IORING_OP_FILES_UPDATE:
6478 return io_rsrc_update_prep(req, sqe);
6479 case IORING_OP_STATX:
6480 return io_statx_prep(req, sqe);
6481 case IORING_OP_FADVISE:
6482 return io_fadvise_prep(req, sqe);
6483 case IORING_OP_MADVISE:
6484 return io_madvise_prep(req, sqe);
6485 case IORING_OP_OPENAT2:
6486 return io_openat2_prep(req, sqe);
6487 case IORING_OP_EPOLL_CTL:
6488 return io_epoll_ctl_prep(req, sqe);
6489 case IORING_OP_SPLICE:
6490 return io_splice_prep(req, sqe);
6491 case IORING_OP_PROVIDE_BUFFERS:
6492 return io_provide_buffers_prep(req, sqe);
6493 case IORING_OP_REMOVE_BUFFERS:
6494 return io_remove_buffers_prep(req, sqe);
6496 return io_tee_prep(req, sqe);
6497 case IORING_OP_SHUTDOWN:
6498 return io_shutdown_prep(req, sqe);
6499 case IORING_OP_RENAMEAT:
6500 return io_renameat_prep(req, sqe);
6501 case IORING_OP_UNLINKAT:
6502 return io_unlinkat_prep(req, sqe);
6503 case IORING_OP_MKDIRAT:
6504 return io_mkdirat_prep(req, sqe);
6505 case IORING_OP_SYMLINKAT:
6506 return io_symlinkat_prep(req, sqe);
6507 case IORING_OP_LINKAT:
6508 return io_linkat_prep(req, sqe);
6511 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6516 static int io_req_prep_async(struct io_kiocb *req)
6518 if (!io_op_defs[req->opcode].needs_async_setup)
6520 if (WARN_ON_ONCE(req->async_data))
6522 if (io_alloc_async_data(req))
6525 switch (req->opcode) {
6526 case IORING_OP_READV:
6527 return io_rw_prep_async(req, READ);
6528 case IORING_OP_WRITEV:
6529 return io_rw_prep_async(req, WRITE);
6530 case IORING_OP_SENDMSG:
6531 return io_sendmsg_prep_async(req);
6532 case IORING_OP_RECVMSG:
6533 return io_recvmsg_prep_async(req);
6534 case IORING_OP_CONNECT:
6535 return io_connect_prep_async(req);
6537 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6542 static u32 io_get_sequence(struct io_kiocb *req)
6544 u32 seq = req->ctx->cached_sq_head;
6546 /* need original cached_sq_head, but it was increased for each req */
6547 io_for_each_link(req, req)
6552 static bool io_drain_req(struct io_kiocb *req)
6554 struct io_kiocb *pos;
6555 struct io_ring_ctx *ctx = req->ctx;
6556 struct io_defer_entry *de;
6560 if (req->flags & REQ_F_FAIL) {
6561 io_req_complete_fail_submit(req);
6566 * If we need to drain a request in the middle of a link, drain the
6567 * head request and the next request/link after the current link.
6568 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6569 * maintained for every request of our link.
6571 if (ctx->drain_next) {
6572 req->flags |= REQ_F_IO_DRAIN;
6573 ctx->drain_next = false;
6575 /* not interested in head, start from the first linked */
6576 io_for_each_link(pos, req->link) {
6577 if (pos->flags & REQ_F_IO_DRAIN) {
6578 ctx->drain_next = true;
6579 req->flags |= REQ_F_IO_DRAIN;
6584 /* Still need defer if there is pending req in defer list. */
6585 spin_lock(&ctx->completion_lock);
6586 if (likely(list_empty_careful(&ctx->defer_list) &&
6587 !(req->flags & REQ_F_IO_DRAIN))) {
6588 spin_unlock(&ctx->completion_lock);
6589 ctx->drain_active = false;
6592 spin_unlock(&ctx->completion_lock);
6594 seq = io_get_sequence(req);
6595 /* Still a chance to pass the sequence check */
6596 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6599 ret = io_req_prep_async(req);
6602 io_prep_async_link(req);
6603 de = kmalloc(sizeof(*de), GFP_KERNEL);
6607 io_req_complete_failed(req, ret);
6611 spin_lock(&ctx->completion_lock);
6612 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6613 spin_unlock(&ctx->completion_lock);
6615 io_queue_async_work(req, NULL);
6619 trace_io_uring_defer(ctx, req, req->user_data);
6622 list_add_tail(&de->list, &ctx->defer_list);
6623 spin_unlock(&ctx->completion_lock);
6627 static void io_clean_op(struct io_kiocb *req)
6629 if (req->flags & REQ_F_BUFFER_SELECTED) {
6630 switch (req->opcode) {
6631 case IORING_OP_READV:
6632 case IORING_OP_READ_FIXED:
6633 case IORING_OP_READ:
6634 kfree((void *)(unsigned long)req->rw.addr);
6636 case IORING_OP_RECVMSG:
6637 case IORING_OP_RECV:
6638 kfree(req->sr_msg.kbuf);
6643 if (req->flags & REQ_F_NEED_CLEANUP) {
6644 switch (req->opcode) {
6645 case IORING_OP_READV:
6646 case IORING_OP_READ_FIXED:
6647 case IORING_OP_READ:
6648 case IORING_OP_WRITEV:
6649 case IORING_OP_WRITE_FIXED:
6650 case IORING_OP_WRITE: {
6651 struct io_async_rw *io = req->async_data;
6653 kfree(io->free_iovec);
6656 case IORING_OP_RECVMSG:
6657 case IORING_OP_SENDMSG: {
6658 struct io_async_msghdr *io = req->async_data;
6660 kfree(io->free_iov);
6663 case IORING_OP_OPENAT:
6664 case IORING_OP_OPENAT2:
6665 if (req->open.filename)
6666 putname(req->open.filename);
6668 case IORING_OP_RENAMEAT:
6669 putname(req->rename.oldpath);
6670 putname(req->rename.newpath);
6672 case IORING_OP_UNLINKAT:
6673 putname(req->unlink.filename);
6675 case IORING_OP_MKDIRAT:
6676 putname(req->mkdir.filename);
6678 case IORING_OP_SYMLINKAT:
6679 putname(req->symlink.oldpath);
6680 putname(req->symlink.newpath);
6682 case IORING_OP_LINKAT:
6683 putname(req->hardlink.oldpath);
6684 putname(req->hardlink.newpath);
6688 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6689 kfree(req->apoll->double_poll);
6693 if (req->flags & REQ_F_INFLIGHT) {
6694 struct io_uring_task *tctx = req->task->io_uring;
6696 atomic_dec(&tctx->inflight_tracked);
6698 if (req->flags & REQ_F_CREDS)
6699 put_cred(req->creds);
6701 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6704 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6706 struct io_ring_ctx *ctx = req->ctx;
6707 const struct cred *creds = NULL;
6710 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6711 creds = override_creds(req->creds);
6713 switch (req->opcode) {
6715 ret = io_nop(req, issue_flags);
6717 case IORING_OP_READV:
6718 case IORING_OP_READ_FIXED:
6719 case IORING_OP_READ:
6720 ret = io_read(req, issue_flags);
6722 case IORING_OP_WRITEV:
6723 case IORING_OP_WRITE_FIXED:
6724 case IORING_OP_WRITE:
6725 ret = io_write(req, issue_flags);
6727 case IORING_OP_FSYNC:
6728 ret = io_fsync(req, issue_flags);
6730 case IORING_OP_POLL_ADD:
6731 ret = io_poll_add(req, issue_flags);
6733 case IORING_OP_POLL_REMOVE:
6734 ret = io_poll_update(req, issue_flags);
6736 case IORING_OP_SYNC_FILE_RANGE:
6737 ret = io_sync_file_range(req, issue_flags);
6739 case IORING_OP_SENDMSG:
6740 ret = io_sendmsg(req, issue_flags);
6742 case IORING_OP_SEND:
6743 ret = io_send(req, issue_flags);
6745 case IORING_OP_RECVMSG:
6746 ret = io_recvmsg(req, issue_flags);
6748 case IORING_OP_RECV:
6749 ret = io_recv(req, issue_flags);
6751 case IORING_OP_TIMEOUT:
6752 ret = io_timeout(req, issue_flags);
6754 case IORING_OP_TIMEOUT_REMOVE:
6755 ret = io_timeout_remove(req, issue_flags);
6757 case IORING_OP_ACCEPT:
6758 ret = io_accept(req, issue_flags);
6760 case IORING_OP_CONNECT:
6761 ret = io_connect(req, issue_flags);
6763 case IORING_OP_ASYNC_CANCEL:
6764 ret = io_async_cancel(req, issue_flags);
6766 case IORING_OP_FALLOCATE:
6767 ret = io_fallocate(req, issue_flags);
6769 case IORING_OP_OPENAT:
6770 ret = io_openat(req, issue_flags);
6772 case IORING_OP_CLOSE:
6773 ret = io_close(req, issue_flags);
6775 case IORING_OP_FILES_UPDATE:
6776 ret = io_files_update(req, issue_flags);
6778 case IORING_OP_STATX:
6779 ret = io_statx(req, issue_flags);
6781 case IORING_OP_FADVISE:
6782 ret = io_fadvise(req, issue_flags);
6784 case IORING_OP_MADVISE:
6785 ret = io_madvise(req, issue_flags);
6787 case IORING_OP_OPENAT2:
6788 ret = io_openat2(req, issue_flags);
6790 case IORING_OP_EPOLL_CTL:
6791 ret = io_epoll_ctl(req, issue_flags);
6793 case IORING_OP_SPLICE:
6794 ret = io_splice(req, issue_flags);
6796 case IORING_OP_PROVIDE_BUFFERS:
6797 ret = io_provide_buffers(req, issue_flags);
6799 case IORING_OP_REMOVE_BUFFERS:
6800 ret = io_remove_buffers(req, issue_flags);
6803 ret = io_tee(req, issue_flags);
6805 case IORING_OP_SHUTDOWN:
6806 ret = io_shutdown(req, issue_flags);
6808 case IORING_OP_RENAMEAT:
6809 ret = io_renameat(req, issue_flags);
6811 case IORING_OP_UNLINKAT:
6812 ret = io_unlinkat(req, issue_flags);
6814 case IORING_OP_MKDIRAT:
6815 ret = io_mkdirat(req, issue_flags);
6817 case IORING_OP_SYMLINKAT:
6818 ret = io_symlinkat(req, issue_flags);
6820 case IORING_OP_LINKAT:
6821 ret = io_linkat(req, issue_flags);
6829 revert_creds(creds);
6832 /* If the op doesn't have a file, we're not polling for it */
6833 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6834 io_iopoll_req_issued(req);
6839 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6841 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6843 req = io_put_req_find_next(req);
6844 return req ? &req->work : NULL;
6847 static void io_wq_submit_work(struct io_wq_work *work)
6849 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6850 struct io_kiocb *timeout;
6853 /* one will be dropped by ->io_free_work() after returning to io-wq */
6854 if (!(req->flags & REQ_F_REFCOUNT))
6855 __io_req_set_refcount(req, 2);
6859 timeout = io_prep_linked_timeout(req);
6861 io_queue_linked_timeout(timeout);
6863 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6864 if (work->flags & IO_WQ_WORK_CANCEL)
6869 ret = io_issue_sqe(req, 0);
6871 * We can get EAGAIN for polled IO even though we're
6872 * forcing a sync submission from here, since we can't
6873 * wait for request slots on the block side.
6875 if (ret != -EAGAIN || !(req->ctx->flags & IORING_SETUP_IOPOLL))
6881 /* avoid locking problems by failing it from a clean context */
6883 io_req_task_queue_fail(req, ret);
6886 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6889 return &table->files[i];
6892 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6895 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6897 return (struct file *) (slot->file_ptr & FFS_MASK);
6900 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6902 unsigned long file_ptr = (unsigned long) file;
6904 if (__io_file_supports_nowait(file, READ))
6905 file_ptr |= FFS_ASYNC_READ;
6906 if (__io_file_supports_nowait(file, WRITE))
6907 file_ptr |= FFS_ASYNC_WRITE;
6908 if (S_ISREG(file_inode(file)->i_mode))
6909 file_ptr |= FFS_ISREG;
6910 file_slot->file_ptr = file_ptr;
6913 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6914 struct io_kiocb *req, int fd)
6917 unsigned long file_ptr;
6919 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6921 fd = array_index_nospec(fd, ctx->nr_user_files);
6922 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6923 file = (struct file *) (file_ptr & FFS_MASK);
6924 file_ptr &= ~FFS_MASK;
6925 /* mask in overlapping REQ_F and FFS bits */
6926 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6927 io_req_set_rsrc_node(req);
6931 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6932 struct io_kiocb *req, int fd)
6934 struct file *file = fget(fd);
6936 trace_io_uring_file_get(ctx, fd);
6938 /* we don't allow fixed io_uring files */
6939 if (file && unlikely(file->f_op == &io_uring_fops))
6940 io_req_track_inflight(req);
6944 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6945 struct io_kiocb *req, int fd, bool fixed)
6948 return io_file_get_fixed(ctx, req, fd);
6950 return io_file_get_normal(ctx, req, fd);
6953 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6955 struct io_kiocb *prev = req->timeout.prev;
6959 if (!(req->task->flags & PF_EXITING))
6960 ret = io_try_cancel_userdata(req, prev->user_data);
6961 io_req_complete_post(req, ret ?: -ETIME, 0);
6964 io_req_complete_post(req, -ETIME, 0);
6968 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6970 struct io_timeout_data *data = container_of(timer,
6971 struct io_timeout_data, timer);
6972 struct io_kiocb *prev, *req = data->req;
6973 struct io_ring_ctx *ctx = req->ctx;
6974 unsigned long flags;
6976 spin_lock_irqsave(&ctx->timeout_lock, flags);
6977 prev = req->timeout.head;
6978 req->timeout.head = NULL;
6981 * We don't expect the list to be empty, that will only happen if we
6982 * race with the completion of the linked work.
6985 io_remove_next_linked(prev);
6986 if (!req_ref_inc_not_zero(prev))
6989 list_del(&req->timeout.list);
6990 req->timeout.prev = prev;
6991 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6993 req->io_task_work.func = io_req_task_link_timeout;
6994 io_req_task_work_add(req);
6995 return HRTIMER_NORESTART;
6998 static void io_queue_linked_timeout(struct io_kiocb *req)
7000 struct io_ring_ctx *ctx = req->ctx;
7002 spin_lock_irq(&ctx->timeout_lock);
7004 * If the back reference is NULL, then our linked request finished
7005 * before we got a chance to setup the timer
7007 if (req->timeout.head) {
7008 struct io_timeout_data *data = req->async_data;
7010 data->timer.function = io_link_timeout_fn;
7011 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
7013 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
7015 spin_unlock_irq(&ctx->timeout_lock);
7016 /* drop submission reference */
7020 static void __io_queue_sqe(struct io_kiocb *req)
7021 __must_hold(&req->ctx->uring_lock)
7023 struct io_kiocb *linked_timeout;
7027 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
7030 * We async punt it if the file wasn't marked NOWAIT, or if the file
7031 * doesn't support non-blocking read/write attempts
7034 if (req->flags & REQ_F_COMPLETE_INLINE) {
7035 struct io_ring_ctx *ctx = req->ctx;
7036 struct io_submit_state *state = &ctx->submit_state;
7038 state->compl_reqs[state->compl_nr++] = req;
7039 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
7040 io_submit_flush_completions(ctx);
7044 linked_timeout = io_prep_linked_timeout(req);
7046 io_queue_linked_timeout(linked_timeout);
7047 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
7048 linked_timeout = io_prep_linked_timeout(req);
7050 switch (io_arm_poll_handler(req)) {
7051 case IO_APOLL_READY:
7053 io_queue_linked_timeout(linked_timeout);
7055 case IO_APOLL_ABORTED:
7057 * Queued up for async execution, worker will release
7058 * submit reference when the iocb is actually submitted.
7060 io_queue_async_work(req, NULL);
7065 io_queue_linked_timeout(linked_timeout);
7067 io_req_complete_failed(req, ret);
7071 static inline void io_queue_sqe(struct io_kiocb *req)
7072 __must_hold(&req->ctx->uring_lock)
7074 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7077 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7078 __io_queue_sqe(req);
7079 } else if (req->flags & REQ_F_FAIL) {
7080 io_req_complete_fail_submit(req);
7082 int ret = io_req_prep_async(req);
7085 io_req_complete_failed(req, ret);
7087 io_queue_async_work(req, NULL);
7092 * Check SQE restrictions (opcode and flags).
7094 * Returns 'true' if SQE is allowed, 'false' otherwise.
7096 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7097 struct io_kiocb *req,
7098 unsigned int sqe_flags)
7100 if (likely(!ctx->restricted))
7103 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7106 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7107 ctx->restrictions.sqe_flags_required)
7110 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7111 ctx->restrictions.sqe_flags_required))
7117 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7118 const struct io_uring_sqe *sqe)
7119 __must_hold(&ctx->uring_lock)
7121 struct io_submit_state *state;
7122 unsigned int sqe_flags;
7123 int personality, ret = 0;
7125 /* req is partially pre-initialised, see io_preinit_req() */
7126 req->opcode = READ_ONCE(sqe->opcode);
7127 /* same numerical values with corresponding REQ_F_*, safe to copy */
7128 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7129 req->user_data = READ_ONCE(sqe->user_data);
7131 req->fixed_rsrc_refs = NULL;
7132 req->task = current;
7134 /* enforce forwards compatibility on users */
7135 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7137 if (unlikely(req->opcode >= IORING_OP_LAST))
7139 if (!io_check_restriction(ctx, req, sqe_flags))
7142 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7143 !io_op_defs[req->opcode].buffer_select)
7145 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7146 ctx->drain_active = true;
7148 personality = READ_ONCE(sqe->personality);
7150 req->creds = xa_load(&ctx->personalities, personality);
7153 get_cred(req->creds);
7154 req->flags |= REQ_F_CREDS;
7156 state = &ctx->submit_state;
7159 * Plug now if we have more than 1 IO left after this, and the target
7160 * is potentially a read/write to block based storage.
7162 if (!state->plug_started && state->ios_left > 1 &&
7163 io_op_defs[req->opcode].plug) {
7164 blk_start_plug(&state->plug);
7165 state->plug_started = true;
7168 if (io_op_defs[req->opcode].needs_file) {
7169 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7170 (sqe_flags & IOSQE_FIXED_FILE));
7171 if (unlikely(!req->file))
7179 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7180 const struct io_uring_sqe *sqe)
7181 __must_hold(&ctx->uring_lock)
7183 struct io_submit_link *link = &ctx->submit_state.link;
7186 ret = io_init_req(ctx, req, sqe);
7187 if (unlikely(ret)) {
7189 /* fail even hard links since we don't submit */
7192 * we can judge a link req is failed or cancelled by if
7193 * REQ_F_FAIL is set, but the head is an exception since
7194 * it may be set REQ_F_FAIL because of other req's failure
7195 * so let's leverage req->result to distinguish if a head
7196 * is set REQ_F_FAIL because of its failure or other req's
7197 * failure so that we can set the correct ret code for it.
7198 * init result here to avoid affecting the normal path.
7200 if (!(link->head->flags & REQ_F_FAIL))
7201 req_fail_link_node(link->head, -ECANCELED);
7202 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7204 * the current req is a normal req, we should return
7205 * error and thus break the submittion loop.
7207 io_req_complete_failed(req, ret);
7210 req_fail_link_node(req, ret);
7212 ret = io_req_prep(req, sqe);
7217 /* don't need @sqe from now on */
7218 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7220 ctx->flags & IORING_SETUP_SQPOLL);
7223 * If we already have a head request, queue this one for async
7224 * submittal once the head completes. If we don't have a head but
7225 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7226 * submitted sync once the chain is complete. If none of those
7227 * conditions are true (normal request), then just queue it.
7230 struct io_kiocb *head = link->head;
7232 if (!(req->flags & REQ_F_FAIL)) {
7233 ret = io_req_prep_async(req);
7234 if (unlikely(ret)) {
7235 req_fail_link_node(req, ret);
7236 if (!(head->flags & REQ_F_FAIL))
7237 req_fail_link_node(head, -ECANCELED);
7240 trace_io_uring_link(ctx, req, head);
7241 link->last->link = req;
7244 /* last request of a link, enqueue the link */
7245 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7250 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7262 * Batched submission is done, ensure local IO is flushed out.
7264 static void io_submit_state_end(struct io_submit_state *state,
7265 struct io_ring_ctx *ctx)
7267 if (state->link.head)
7268 io_queue_sqe(state->link.head);
7269 if (state->compl_nr)
7270 io_submit_flush_completions(ctx);
7271 if (state->plug_started)
7272 blk_finish_plug(&state->plug);
7276 * Start submission side cache.
7278 static void io_submit_state_start(struct io_submit_state *state,
7279 unsigned int max_ios)
7281 state->plug_started = false;
7282 state->ios_left = max_ios;
7283 /* set only head, no need to init link_last in advance */
7284 state->link.head = NULL;
7287 static void io_commit_sqring(struct io_ring_ctx *ctx)
7289 struct io_rings *rings = ctx->rings;
7292 * Ensure any loads from the SQEs are done at this point,
7293 * since once we write the new head, the application could
7294 * write new data to them.
7296 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7300 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7301 * that is mapped by userspace. This means that care needs to be taken to
7302 * ensure that reads are stable, as we cannot rely on userspace always
7303 * being a good citizen. If members of the sqe are validated and then later
7304 * used, it's important that those reads are done through READ_ONCE() to
7305 * prevent a re-load down the line.
7307 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7309 unsigned head, mask = ctx->sq_entries - 1;
7310 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7313 * The cached sq head (or cq tail) serves two purposes:
7315 * 1) allows us to batch the cost of updating the user visible
7317 * 2) allows the kernel side to track the head on its own, even
7318 * though the application is the one updating it.
7320 head = READ_ONCE(ctx->sq_array[sq_idx]);
7321 if (likely(head < ctx->sq_entries))
7322 return &ctx->sq_sqes[head];
7324 /* drop invalid entries */
7326 WRITE_ONCE(ctx->rings->sq_dropped,
7327 READ_ONCE(ctx->rings->sq_dropped) + 1);
7331 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7332 __must_hold(&ctx->uring_lock)
7336 /* make sure SQ entry isn't read before tail */
7337 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7338 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7340 io_get_task_refs(nr);
7342 io_submit_state_start(&ctx->submit_state, nr);
7343 while (submitted < nr) {
7344 const struct io_uring_sqe *sqe;
7345 struct io_kiocb *req;
7347 req = io_alloc_req(ctx);
7348 if (unlikely(!req)) {
7350 submitted = -EAGAIN;
7353 sqe = io_get_sqe(ctx);
7354 if (unlikely(!sqe)) {
7355 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7358 /* will complete beyond this point, count as submitted */
7360 if (io_submit_sqe(ctx, req, sqe))
7364 if (unlikely(submitted != nr)) {
7365 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7366 int unused = nr - ref_used;
7368 current->io_uring->cached_refs += unused;
7369 percpu_ref_put_many(&ctx->refs, unused);
7372 io_submit_state_end(&ctx->submit_state, ctx);
7373 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7374 io_commit_sqring(ctx);
7379 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7381 return READ_ONCE(sqd->state);
7384 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7386 /* Tell userspace we may need a wakeup call */
7387 spin_lock(&ctx->completion_lock);
7388 WRITE_ONCE(ctx->rings->sq_flags,
7389 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7390 spin_unlock(&ctx->completion_lock);
7393 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7395 spin_lock(&ctx->completion_lock);
7396 WRITE_ONCE(ctx->rings->sq_flags,
7397 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7398 spin_unlock(&ctx->completion_lock);
7401 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7403 unsigned int to_submit;
7406 to_submit = io_sqring_entries(ctx);
7407 /* if we're handling multiple rings, cap submit size for fairness */
7408 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7409 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7411 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7412 unsigned nr_events = 0;
7413 const struct cred *creds = NULL;
7415 if (ctx->sq_creds != current_cred())
7416 creds = override_creds(ctx->sq_creds);
7418 mutex_lock(&ctx->uring_lock);
7419 if (!list_empty(&ctx->iopoll_list))
7420 io_do_iopoll(ctx, &nr_events, 0);
7423 * Don't submit if refs are dying, good for io_uring_register(),
7424 * but also it is relied upon by io_ring_exit_work()
7426 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7427 !(ctx->flags & IORING_SETUP_R_DISABLED))
7428 ret = io_submit_sqes(ctx, to_submit);
7429 mutex_unlock(&ctx->uring_lock);
7431 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7432 wake_up(&ctx->sqo_sq_wait);
7434 revert_creds(creds);
7440 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7442 struct io_ring_ctx *ctx;
7443 unsigned sq_thread_idle = 0;
7445 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7446 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7447 sqd->sq_thread_idle = sq_thread_idle;
7450 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7452 bool did_sig = false;
7453 struct ksignal ksig;
7455 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7456 signal_pending(current)) {
7457 mutex_unlock(&sqd->lock);
7458 if (signal_pending(current))
7459 did_sig = get_signal(&ksig);
7461 mutex_lock(&sqd->lock);
7463 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7466 static int io_sq_thread(void *data)
7468 struct io_sq_data *sqd = data;
7469 struct io_ring_ctx *ctx;
7470 unsigned long timeout = 0;
7471 char buf[TASK_COMM_LEN];
7474 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7475 set_task_comm(current, buf);
7477 if (sqd->sq_cpu != -1)
7478 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7480 set_cpus_allowed_ptr(current, cpu_online_mask);
7481 current->flags |= PF_NO_SETAFFINITY;
7483 mutex_lock(&sqd->lock);
7485 bool cap_entries, sqt_spin = false;
7487 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7488 if (io_sqd_handle_event(sqd))
7490 timeout = jiffies + sqd->sq_thread_idle;
7493 cap_entries = !list_is_singular(&sqd->ctx_list);
7494 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7495 int ret = __io_sq_thread(ctx, cap_entries);
7497 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7500 if (io_run_task_work())
7503 if (sqt_spin || !time_after(jiffies, timeout)) {
7506 timeout = jiffies + sqd->sq_thread_idle;
7510 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7511 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7512 bool needs_sched = true;
7514 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7515 io_ring_set_wakeup_flag(ctx);
7517 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7518 !list_empty_careful(&ctx->iopoll_list)) {
7519 needs_sched = false;
7522 if (io_sqring_entries(ctx)) {
7523 needs_sched = false;
7529 mutex_unlock(&sqd->lock);
7531 mutex_lock(&sqd->lock);
7533 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7534 io_ring_clear_wakeup_flag(ctx);
7537 finish_wait(&sqd->wait, &wait);
7538 timeout = jiffies + sqd->sq_thread_idle;
7541 io_uring_cancel_generic(true, sqd);
7543 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7544 io_ring_set_wakeup_flag(ctx);
7546 mutex_unlock(&sqd->lock);
7548 complete(&sqd->exited);
7552 struct io_wait_queue {
7553 struct wait_queue_entry wq;
7554 struct io_ring_ctx *ctx;
7556 unsigned nr_timeouts;
7559 static inline bool io_should_wake(struct io_wait_queue *iowq)
7561 struct io_ring_ctx *ctx = iowq->ctx;
7562 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7565 * Wake up if we have enough events, or if a timeout occurred since we
7566 * started waiting. For timeouts, we always want to return to userspace,
7567 * regardless of event count.
7569 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7572 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7573 int wake_flags, void *key)
7575 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7579 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7580 * the task, and the next invocation will do it.
7582 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7583 return autoremove_wake_function(curr, mode, wake_flags, key);
7587 static int io_run_task_work_sig(void)
7589 if (io_run_task_work())
7591 if (!signal_pending(current))
7593 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7594 return -ERESTARTSYS;
7598 /* when returns >0, the caller should retry */
7599 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7600 struct io_wait_queue *iowq,
7605 /* make sure we run task_work before checking for signals */
7606 ret = io_run_task_work_sig();
7607 if (ret || io_should_wake(iowq))
7609 /* let the caller flush overflows, retry */
7610 if (test_bit(0, &ctx->check_cq_overflow))
7613 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
7619 * Wait until events become available, if we don't already have some. The
7620 * application must reap them itself, as they reside on the shared cq ring.
7622 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7623 const sigset_t __user *sig, size_t sigsz,
7624 struct __kernel_timespec __user *uts)
7626 struct io_wait_queue iowq;
7627 struct io_rings *rings = ctx->rings;
7628 ktime_t timeout = KTIME_MAX;
7632 io_cqring_overflow_flush(ctx);
7633 if (io_cqring_events(ctx) >= min_events)
7635 if (!io_run_task_work())
7640 struct timespec64 ts;
7642 if (get_timespec64(&ts, uts))
7644 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
7648 #ifdef CONFIG_COMPAT
7649 if (in_compat_syscall())
7650 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7654 ret = set_user_sigmask(sig, sigsz);
7660 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7661 iowq.wq.private = current;
7662 INIT_LIST_HEAD(&iowq.wq.entry);
7664 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7665 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7667 trace_io_uring_cqring_wait(ctx, min_events);
7669 /* if we can't even flush overflow, don't wait for more */
7670 if (!io_cqring_overflow_flush(ctx)) {
7674 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7675 TASK_INTERRUPTIBLE);
7676 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
7677 finish_wait(&ctx->cq_wait, &iowq.wq);
7681 restore_saved_sigmask_unless(ret == -EINTR);
7683 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7686 static void io_free_page_table(void **table, size_t size)
7688 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7690 for (i = 0; i < nr_tables; i++)
7695 static void **io_alloc_page_table(size_t size)
7697 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7698 size_t init_size = size;
7701 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7705 for (i = 0; i < nr_tables; i++) {
7706 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7708 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7710 io_free_page_table(table, init_size);
7718 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7720 percpu_ref_exit(&ref_node->refs);
7724 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7726 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7727 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7728 unsigned long flags;
7729 bool first_add = false;
7730 unsigned long delay = HZ;
7732 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7735 /* if we are mid-quiesce then do not delay */
7736 if (node->rsrc_data->quiesce)
7739 while (!list_empty(&ctx->rsrc_ref_list)) {
7740 node = list_first_entry(&ctx->rsrc_ref_list,
7741 struct io_rsrc_node, node);
7742 /* recycle ref nodes in order */
7745 list_del(&node->node);
7746 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7748 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7751 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7754 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7756 struct io_rsrc_node *ref_node;
7758 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7762 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7767 INIT_LIST_HEAD(&ref_node->node);
7768 INIT_LIST_HEAD(&ref_node->rsrc_list);
7769 ref_node->done = false;
7773 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7774 struct io_rsrc_data *data_to_kill)
7776 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7777 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7780 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7782 rsrc_node->rsrc_data = data_to_kill;
7783 spin_lock_irq(&ctx->rsrc_ref_lock);
7784 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7785 spin_unlock_irq(&ctx->rsrc_ref_lock);
7787 atomic_inc(&data_to_kill->refs);
7788 percpu_ref_kill(&rsrc_node->refs);
7789 ctx->rsrc_node = NULL;
7792 if (!ctx->rsrc_node) {
7793 ctx->rsrc_node = ctx->rsrc_backup_node;
7794 ctx->rsrc_backup_node = NULL;
7798 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7800 if (ctx->rsrc_backup_node)
7802 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7803 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7806 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7810 /* As we may drop ->uring_lock, other task may have started quiesce */
7814 data->quiesce = true;
7816 ret = io_rsrc_node_switch_start(ctx);
7819 io_rsrc_node_switch(ctx, data);
7821 /* kill initial ref, already quiesced if zero */
7822 if (atomic_dec_and_test(&data->refs))
7824 mutex_unlock(&ctx->uring_lock);
7825 flush_delayed_work(&ctx->rsrc_put_work);
7826 ret = wait_for_completion_interruptible(&data->done);
7828 mutex_lock(&ctx->uring_lock);
7829 if (atomic_read(&data->refs) > 0) {
7831 * it has been revived by another thread while
7834 mutex_unlock(&ctx->uring_lock);
7840 atomic_inc(&data->refs);
7841 /* wait for all works potentially completing data->done */
7842 flush_delayed_work(&ctx->rsrc_put_work);
7843 reinit_completion(&data->done);
7845 ret = io_run_task_work_sig();
7846 mutex_lock(&ctx->uring_lock);
7848 data->quiesce = false;
7853 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7855 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7856 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7858 return &data->tags[table_idx][off];
7861 static void io_rsrc_data_free(struct io_rsrc_data *data)
7863 size_t size = data->nr * sizeof(data->tags[0][0]);
7866 io_free_page_table((void **)data->tags, size);
7870 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7871 u64 __user *utags, unsigned nr,
7872 struct io_rsrc_data **pdata)
7874 struct io_rsrc_data *data;
7878 data = kzalloc(sizeof(*data), GFP_KERNEL);
7881 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7889 data->do_put = do_put;
7892 for (i = 0; i < nr; i++) {
7893 u64 *tag_slot = io_get_tag_slot(data, i);
7895 if (copy_from_user(tag_slot, &utags[i],
7901 atomic_set(&data->refs, 1);
7902 init_completion(&data->done);
7906 io_rsrc_data_free(data);
7910 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7912 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7913 GFP_KERNEL_ACCOUNT);
7914 return !!table->files;
7917 static void io_free_file_tables(struct io_file_table *table)
7919 kvfree(table->files);
7920 table->files = NULL;
7923 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7925 #if defined(CONFIG_UNIX)
7926 if (ctx->ring_sock) {
7927 struct sock *sock = ctx->ring_sock->sk;
7928 struct sk_buff *skb;
7930 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7936 for (i = 0; i < ctx->nr_user_files; i++) {
7939 file = io_file_from_index(ctx, i);
7944 io_free_file_tables(&ctx->file_table);
7945 io_rsrc_data_free(ctx->file_data);
7946 ctx->file_data = NULL;
7947 ctx->nr_user_files = 0;
7950 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7952 unsigned nr = ctx->nr_user_files;
7955 if (!ctx->file_data)
7959 * Quiesce may unlock ->uring_lock, and while it's not held
7960 * prevent new requests using the table.
7962 ctx->nr_user_files = 0;
7963 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7964 ctx->nr_user_files = nr;
7966 __io_sqe_files_unregister(ctx);
7970 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7971 __releases(&sqd->lock)
7973 WARN_ON_ONCE(sqd->thread == current);
7976 * Do the dance but not conditional clear_bit() because it'd race with
7977 * other threads incrementing park_pending and setting the bit.
7979 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7980 if (atomic_dec_return(&sqd->park_pending))
7981 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7982 mutex_unlock(&sqd->lock);
7985 static void io_sq_thread_park(struct io_sq_data *sqd)
7986 __acquires(&sqd->lock)
7988 WARN_ON_ONCE(sqd->thread == current);
7990 atomic_inc(&sqd->park_pending);
7991 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7992 mutex_lock(&sqd->lock);
7994 wake_up_process(sqd->thread);
7997 static void io_sq_thread_stop(struct io_sq_data *sqd)
7999 WARN_ON_ONCE(sqd->thread == current);
8000 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
8002 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
8003 mutex_lock(&sqd->lock);
8005 wake_up_process(sqd->thread);
8006 mutex_unlock(&sqd->lock);
8007 wait_for_completion(&sqd->exited);
8010 static void io_put_sq_data(struct io_sq_data *sqd)
8012 if (refcount_dec_and_test(&sqd->refs)) {
8013 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
8015 io_sq_thread_stop(sqd);
8020 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
8022 struct io_sq_data *sqd = ctx->sq_data;
8025 io_sq_thread_park(sqd);
8026 list_del_init(&ctx->sqd_list);
8027 io_sqd_update_thread_idle(sqd);
8028 io_sq_thread_unpark(sqd);
8030 io_put_sq_data(sqd);
8031 ctx->sq_data = NULL;
8035 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
8037 struct io_ring_ctx *ctx_attach;
8038 struct io_sq_data *sqd;
8041 f = fdget(p->wq_fd);
8043 return ERR_PTR(-ENXIO);
8044 if (f.file->f_op != &io_uring_fops) {
8046 return ERR_PTR(-EINVAL);
8049 ctx_attach = f.file->private_data;
8050 sqd = ctx_attach->sq_data;
8053 return ERR_PTR(-EINVAL);
8055 if (sqd->task_tgid != current->tgid) {
8057 return ERR_PTR(-EPERM);
8060 refcount_inc(&sqd->refs);
8065 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
8068 struct io_sq_data *sqd;
8071 if (p->flags & IORING_SETUP_ATTACH_WQ) {
8072 sqd = io_attach_sq_data(p);
8077 /* fall through for EPERM case, setup new sqd/task */
8078 if (PTR_ERR(sqd) != -EPERM)
8082 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8084 return ERR_PTR(-ENOMEM);
8086 atomic_set(&sqd->park_pending, 0);
8087 refcount_set(&sqd->refs, 1);
8088 INIT_LIST_HEAD(&sqd->ctx_list);
8089 mutex_init(&sqd->lock);
8090 init_waitqueue_head(&sqd->wait);
8091 init_completion(&sqd->exited);
8095 #if defined(CONFIG_UNIX)
8097 * Ensure the UNIX gc is aware of our file set, so we are certain that
8098 * the io_uring can be safely unregistered on process exit, even if we have
8099 * loops in the file referencing.
8101 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8103 struct sock *sk = ctx->ring_sock->sk;
8104 struct scm_fp_list *fpl;
8105 struct sk_buff *skb;
8108 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8112 skb = alloc_skb(0, GFP_KERNEL);
8121 fpl->user = get_uid(current_user());
8122 for (i = 0; i < nr; i++) {
8123 struct file *file = io_file_from_index(ctx, i + offset);
8127 fpl->fp[nr_files] = get_file(file);
8128 unix_inflight(fpl->user, fpl->fp[nr_files]);
8133 fpl->max = SCM_MAX_FD;
8134 fpl->count = nr_files;
8135 UNIXCB(skb).fp = fpl;
8136 skb->destructor = unix_destruct_scm;
8137 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8138 skb_queue_head(&sk->sk_receive_queue, skb);
8140 for (i = 0; i < nr; i++) {
8141 struct file *file = io_file_from_index(ctx, i + offset);
8148 free_uid(fpl->user);
8156 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8157 * causes regular reference counting to break down. We rely on the UNIX
8158 * garbage collection to take care of this problem for us.
8160 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8162 unsigned left, total;
8166 left = ctx->nr_user_files;
8168 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8170 ret = __io_sqe_files_scm(ctx, this_files, total);
8174 total += this_files;
8180 while (total < ctx->nr_user_files) {
8181 struct file *file = io_file_from_index(ctx, total);
8191 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8197 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8199 struct file *file = prsrc->file;
8200 #if defined(CONFIG_UNIX)
8201 struct sock *sock = ctx->ring_sock->sk;
8202 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8203 struct sk_buff *skb;
8206 __skb_queue_head_init(&list);
8209 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8210 * remove this entry and rearrange the file array.
8212 skb = skb_dequeue(head);
8214 struct scm_fp_list *fp;
8216 fp = UNIXCB(skb).fp;
8217 for (i = 0; i < fp->count; i++) {
8220 if (fp->fp[i] != file)
8223 unix_notinflight(fp->user, fp->fp[i]);
8224 left = fp->count - 1 - i;
8226 memmove(&fp->fp[i], &fp->fp[i + 1],
8227 left * sizeof(struct file *));
8234 __skb_queue_tail(&list, skb);
8244 __skb_queue_tail(&list, skb);
8246 skb = skb_dequeue(head);
8249 if (skb_peek(&list)) {
8250 spin_lock_irq(&head->lock);
8251 while ((skb = __skb_dequeue(&list)) != NULL)
8252 __skb_queue_tail(head, skb);
8253 spin_unlock_irq(&head->lock);
8260 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8262 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8263 struct io_ring_ctx *ctx = rsrc_data->ctx;
8264 struct io_rsrc_put *prsrc, *tmp;
8266 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8267 list_del(&prsrc->list);
8270 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8272 io_ring_submit_lock(ctx, lock_ring);
8273 spin_lock(&ctx->completion_lock);
8274 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8276 io_commit_cqring(ctx);
8277 spin_unlock(&ctx->completion_lock);
8278 io_cqring_ev_posted(ctx);
8279 io_ring_submit_unlock(ctx, lock_ring);
8282 rsrc_data->do_put(ctx, prsrc);
8286 io_rsrc_node_destroy(ref_node);
8287 if (atomic_dec_and_test(&rsrc_data->refs))
8288 complete(&rsrc_data->done);
8291 static void io_rsrc_put_work(struct work_struct *work)
8293 struct io_ring_ctx *ctx;
8294 struct llist_node *node;
8296 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8297 node = llist_del_all(&ctx->rsrc_put_llist);
8300 struct io_rsrc_node *ref_node;
8301 struct llist_node *next = node->next;
8303 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8304 __io_rsrc_put_work(ref_node);
8309 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8310 unsigned nr_args, u64 __user *tags)
8312 __s32 __user *fds = (__s32 __user *) arg;
8321 if (nr_args > IORING_MAX_FIXED_FILES)
8323 if (nr_args > rlimit(RLIMIT_NOFILE))
8325 ret = io_rsrc_node_switch_start(ctx);
8328 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8334 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8337 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8338 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8342 /* allow sparse sets */
8345 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8352 if (unlikely(!file))
8356 * Don't allow io_uring instances to be registered. If UNIX
8357 * isn't enabled, then this causes a reference cycle and this
8358 * instance can never get freed. If UNIX is enabled we'll
8359 * handle it just fine, but there's still no point in allowing
8360 * a ring fd as it doesn't support regular read/write anyway.
8362 if (file->f_op == &io_uring_fops) {
8366 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8369 ret = io_sqe_files_scm(ctx);
8371 __io_sqe_files_unregister(ctx);
8375 io_rsrc_node_switch(ctx, NULL);
8378 for (i = 0; i < ctx->nr_user_files; i++) {
8379 file = io_file_from_index(ctx, i);
8383 io_free_file_tables(&ctx->file_table);
8384 ctx->nr_user_files = 0;
8386 io_rsrc_data_free(ctx->file_data);
8387 ctx->file_data = NULL;
8391 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8394 #if defined(CONFIG_UNIX)
8395 struct sock *sock = ctx->ring_sock->sk;
8396 struct sk_buff_head *head = &sock->sk_receive_queue;
8397 struct sk_buff *skb;
8400 * See if we can merge this file into an existing skb SCM_RIGHTS
8401 * file set. If there's no room, fall back to allocating a new skb
8402 * and filling it in.
8404 spin_lock_irq(&head->lock);
8405 skb = skb_peek(head);
8407 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8409 if (fpl->count < SCM_MAX_FD) {
8410 __skb_unlink(skb, head);
8411 spin_unlock_irq(&head->lock);
8412 fpl->fp[fpl->count] = get_file(file);
8413 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8415 spin_lock_irq(&head->lock);
8416 __skb_queue_head(head, skb);
8421 spin_unlock_irq(&head->lock);
8428 return __io_sqe_files_scm(ctx, 1, index);
8434 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8435 struct io_rsrc_node *node, void *rsrc)
8437 u64 *tag_slot = io_get_tag_slot(data, idx);
8438 struct io_rsrc_put *prsrc;
8440 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8444 prsrc->tag = *tag_slot;
8447 list_add(&prsrc->list, &node->rsrc_list);
8451 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8452 unsigned int issue_flags, u32 slot_index)
8454 struct io_ring_ctx *ctx = req->ctx;
8455 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8456 bool needs_switch = false;
8457 struct io_fixed_file *file_slot;
8460 io_ring_submit_lock(ctx, !force_nonblock);
8461 if (file->f_op == &io_uring_fops)
8464 if (!ctx->file_data)
8467 if (slot_index >= ctx->nr_user_files)
8470 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8471 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8473 if (file_slot->file_ptr) {
8474 struct file *old_file;
8476 ret = io_rsrc_node_switch_start(ctx);
8480 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8481 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8482 ctx->rsrc_node, old_file);
8485 file_slot->file_ptr = 0;
8486 needs_switch = true;
8489 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8490 io_fixed_file_set(file_slot, file);
8491 ret = io_sqe_file_register(ctx, file, slot_index);
8493 file_slot->file_ptr = 0;
8500 io_rsrc_node_switch(ctx, ctx->file_data);
8501 io_ring_submit_unlock(ctx, !force_nonblock);
8507 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8509 unsigned int offset = req->close.file_slot - 1;
8510 struct io_ring_ctx *ctx = req->ctx;
8511 struct io_fixed_file *file_slot;
8515 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8517 if (unlikely(!ctx->file_data))
8520 if (offset >= ctx->nr_user_files)
8522 ret = io_rsrc_node_switch_start(ctx);
8526 offset = array_index_nospec(offset, ctx->nr_user_files);
8527 file_slot = io_fixed_file_slot(&ctx->file_table, offset);
8529 if (!file_slot->file_ptr)
8532 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8533 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8537 file_slot->file_ptr = 0;
8538 io_rsrc_node_switch(ctx, ctx->file_data);
8541 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8545 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8546 struct io_uring_rsrc_update2 *up,
8549 u64 __user *tags = u64_to_user_ptr(up->tags);
8550 __s32 __user *fds = u64_to_user_ptr(up->data);
8551 struct io_rsrc_data *data = ctx->file_data;
8552 struct io_fixed_file *file_slot;
8556 bool needs_switch = false;
8558 if (!ctx->file_data)
8560 if (up->offset + nr_args > ctx->nr_user_files)
8563 for (done = 0; done < nr_args; done++) {
8566 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8567 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8571 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8575 if (fd == IORING_REGISTER_FILES_SKIP)
8578 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8579 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8581 if (file_slot->file_ptr) {
8582 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8583 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
8586 file_slot->file_ptr = 0;
8587 needs_switch = true;
8596 * Don't allow io_uring instances to be registered. If
8597 * UNIX isn't enabled, then this causes a reference
8598 * cycle and this instance can never get freed. If UNIX
8599 * is enabled we'll handle it just fine, but there's
8600 * still no point in allowing a ring fd as it doesn't
8601 * support regular read/write anyway.
8603 if (file->f_op == &io_uring_fops) {
8608 *io_get_tag_slot(data, i) = tag;
8609 io_fixed_file_set(file_slot, file);
8610 err = io_sqe_file_register(ctx, file, i);
8612 file_slot->file_ptr = 0;
8620 io_rsrc_node_switch(ctx, data);
8621 return done ? done : err;
8624 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8625 struct task_struct *task)
8627 struct io_wq_hash *hash;
8628 struct io_wq_data data;
8629 unsigned int concurrency;
8631 mutex_lock(&ctx->uring_lock);
8632 hash = ctx->hash_map;
8634 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8636 mutex_unlock(&ctx->uring_lock);
8637 return ERR_PTR(-ENOMEM);
8639 refcount_set(&hash->refs, 1);
8640 init_waitqueue_head(&hash->wait);
8641 ctx->hash_map = hash;
8643 mutex_unlock(&ctx->uring_lock);
8647 data.free_work = io_wq_free_work;
8648 data.do_work = io_wq_submit_work;
8650 /* Do QD, or 4 * CPUS, whatever is smallest */
8651 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8653 return io_wq_create(concurrency, &data);
8656 static int io_uring_alloc_task_context(struct task_struct *task,
8657 struct io_ring_ctx *ctx)
8659 struct io_uring_task *tctx;
8662 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8663 if (unlikely(!tctx))
8666 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8667 if (unlikely(ret)) {
8672 tctx->io_wq = io_init_wq_offload(ctx, task);
8673 if (IS_ERR(tctx->io_wq)) {
8674 ret = PTR_ERR(tctx->io_wq);
8675 percpu_counter_destroy(&tctx->inflight);
8681 init_waitqueue_head(&tctx->wait);
8682 atomic_set(&tctx->in_idle, 0);
8683 atomic_set(&tctx->inflight_tracked, 0);
8684 task->io_uring = tctx;
8685 spin_lock_init(&tctx->task_lock);
8686 INIT_WQ_LIST(&tctx->task_list);
8687 init_task_work(&tctx->task_work, tctx_task_work);
8691 void __io_uring_free(struct task_struct *tsk)
8693 struct io_uring_task *tctx = tsk->io_uring;
8695 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8696 WARN_ON_ONCE(tctx->io_wq);
8697 WARN_ON_ONCE(tctx->cached_refs);
8699 percpu_counter_destroy(&tctx->inflight);
8701 tsk->io_uring = NULL;
8704 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8705 struct io_uring_params *p)
8709 /* Retain compatibility with failing for an invalid attach attempt */
8710 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8711 IORING_SETUP_ATTACH_WQ) {
8714 f = fdget(p->wq_fd);
8717 if (f.file->f_op != &io_uring_fops) {
8723 if (ctx->flags & IORING_SETUP_SQPOLL) {
8724 struct task_struct *tsk;
8725 struct io_sq_data *sqd;
8728 sqd = io_get_sq_data(p, &attached);
8734 ctx->sq_creds = get_current_cred();
8736 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8737 if (!ctx->sq_thread_idle)
8738 ctx->sq_thread_idle = HZ;
8740 io_sq_thread_park(sqd);
8741 list_add(&ctx->sqd_list, &sqd->ctx_list);
8742 io_sqd_update_thread_idle(sqd);
8743 /* don't attach to a dying SQPOLL thread, would be racy */
8744 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8745 io_sq_thread_unpark(sqd);
8752 if (p->flags & IORING_SETUP_SQ_AFF) {
8753 int cpu = p->sq_thread_cpu;
8756 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8763 sqd->task_pid = current->pid;
8764 sqd->task_tgid = current->tgid;
8765 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8772 ret = io_uring_alloc_task_context(tsk, ctx);
8773 wake_up_new_task(tsk);
8776 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8777 /* Can't have SQ_AFF without SQPOLL */
8784 complete(&ctx->sq_data->exited);
8786 io_sq_thread_finish(ctx);
8790 static inline void __io_unaccount_mem(struct user_struct *user,
8791 unsigned long nr_pages)
8793 atomic_long_sub(nr_pages, &user->locked_vm);
8796 static inline int __io_account_mem(struct user_struct *user,
8797 unsigned long nr_pages)
8799 unsigned long page_limit, cur_pages, new_pages;
8801 /* Don't allow more pages than we can safely lock */
8802 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8805 cur_pages = atomic_long_read(&user->locked_vm);
8806 new_pages = cur_pages + nr_pages;
8807 if (new_pages > page_limit)
8809 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8810 new_pages) != cur_pages);
8815 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8818 __io_unaccount_mem(ctx->user, nr_pages);
8820 if (ctx->mm_account)
8821 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8824 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8829 ret = __io_account_mem(ctx->user, nr_pages);
8834 if (ctx->mm_account)
8835 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8840 static void io_mem_free(void *ptr)
8847 page = virt_to_head_page(ptr);
8848 if (put_page_testzero(page))
8849 free_compound_page(page);
8852 static void *io_mem_alloc(size_t size)
8854 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
8856 return (void *) __get_free_pages(gfp, get_order(size));
8859 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8862 struct io_rings *rings;
8863 size_t off, sq_array_size;
8865 off = struct_size(rings, cqes, cq_entries);
8866 if (off == SIZE_MAX)
8870 off = ALIGN(off, SMP_CACHE_BYTES);
8878 sq_array_size = array_size(sizeof(u32), sq_entries);
8879 if (sq_array_size == SIZE_MAX)
8882 if (check_add_overflow(off, sq_array_size, &off))
8888 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8890 struct io_mapped_ubuf *imu = *slot;
8893 if (imu != ctx->dummy_ubuf) {
8894 for (i = 0; i < imu->nr_bvecs; i++)
8895 unpin_user_page(imu->bvec[i].bv_page);
8896 if (imu->acct_pages)
8897 io_unaccount_mem(ctx, imu->acct_pages);
8903 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8905 io_buffer_unmap(ctx, &prsrc->buf);
8909 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8913 for (i = 0; i < ctx->nr_user_bufs; i++)
8914 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8915 kfree(ctx->user_bufs);
8916 io_rsrc_data_free(ctx->buf_data);
8917 ctx->user_bufs = NULL;
8918 ctx->buf_data = NULL;
8919 ctx->nr_user_bufs = 0;
8922 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8924 unsigned nr = ctx->nr_user_bufs;
8931 * Quiesce may unlock ->uring_lock, and while it's not held
8932 * prevent new requests using the table.
8934 ctx->nr_user_bufs = 0;
8935 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8936 ctx->nr_user_bufs = nr;
8938 __io_sqe_buffers_unregister(ctx);
8942 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8943 void __user *arg, unsigned index)
8945 struct iovec __user *src;
8947 #ifdef CONFIG_COMPAT
8949 struct compat_iovec __user *ciovs;
8950 struct compat_iovec ciov;
8952 ciovs = (struct compat_iovec __user *) arg;
8953 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8956 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8957 dst->iov_len = ciov.iov_len;
8961 src = (struct iovec __user *) arg;
8962 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8968 * Not super efficient, but this is just a registration time. And we do cache
8969 * the last compound head, so generally we'll only do a full search if we don't
8972 * We check if the given compound head page has already been accounted, to
8973 * avoid double accounting it. This allows us to account the full size of the
8974 * page, not just the constituent pages of a huge page.
8976 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8977 int nr_pages, struct page *hpage)
8981 /* check current page array */
8982 for (i = 0; i < nr_pages; i++) {
8983 if (!PageCompound(pages[i]))
8985 if (compound_head(pages[i]) == hpage)
8989 /* check previously registered pages */
8990 for (i = 0; i < ctx->nr_user_bufs; i++) {
8991 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8993 for (j = 0; j < imu->nr_bvecs; j++) {
8994 if (!PageCompound(imu->bvec[j].bv_page))
8996 if (compound_head(imu->bvec[j].bv_page) == hpage)
9004 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
9005 int nr_pages, struct io_mapped_ubuf *imu,
9006 struct page **last_hpage)
9010 imu->acct_pages = 0;
9011 for (i = 0; i < nr_pages; i++) {
9012 if (!PageCompound(pages[i])) {
9017 hpage = compound_head(pages[i]);
9018 if (hpage == *last_hpage)
9020 *last_hpage = hpage;
9021 if (headpage_already_acct(ctx, pages, i, hpage))
9023 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
9027 if (!imu->acct_pages)
9030 ret = io_account_mem(ctx, imu->acct_pages);
9032 imu->acct_pages = 0;
9036 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
9037 struct io_mapped_ubuf **pimu,
9038 struct page **last_hpage)
9040 struct io_mapped_ubuf *imu = NULL;
9041 struct vm_area_struct **vmas = NULL;
9042 struct page **pages = NULL;
9043 unsigned long off, start, end, ubuf;
9045 int ret, pret, nr_pages, i;
9047 if (!iov->iov_base) {
9048 *pimu = ctx->dummy_ubuf;
9052 ubuf = (unsigned long) iov->iov_base;
9053 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9054 start = ubuf >> PAGE_SHIFT;
9055 nr_pages = end - start;
9060 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9064 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9069 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9074 mmap_read_lock(current->mm);
9075 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9077 if (pret == nr_pages) {
9078 /* don't support file backed memory */
9079 for (i = 0; i < nr_pages; i++) {
9080 struct vm_area_struct *vma = vmas[i];
9082 if (vma_is_shmem(vma))
9085 !is_file_hugepages(vma->vm_file)) {
9091 ret = pret < 0 ? pret : -EFAULT;
9093 mmap_read_unlock(current->mm);
9096 * if we did partial map, or found file backed vmas,
9097 * release any pages we did get
9100 unpin_user_pages(pages, pret);
9104 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9106 unpin_user_pages(pages, pret);
9110 off = ubuf & ~PAGE_MASK;
9111 size = iov->iov_len;
9112 for (i = 0; i < nr_pages; i++) {
9115 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9116 imu->bvec[i].bv_page = pages[i];
9117 imu->bvec[i].bv_len = vec_len;
9118 imu->bvec[i].bv_offset = off;
9122 /* store original address for later verification */
9124 imu->ubuf_end = ubuf + iov->iov_len;
9125 imu->nr_bvecs = nr_pages;
9136 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9138 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9139 return ctx->user_bufs ? 0 : -ENOMEM;
9142 static int io_buffer_validate(struct iovec *iov)
9144 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9147 * Don't impose further limits on the size and buffer
9148 * constraints here, we'll -EINVAL later when IO is
9149 * submitted if they are wrong.
9152 return iov->iov_len ? -EFAULT : 0;
9156 /* arbitrary limit, but we need something */
9157 if (iov->iov_len > SZ_1G)
9160 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9166 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9167 unsigned int nr_args, u64 __user *tags)
9169 struct page *last_hpage = NULL;
9170 struct io_rsrc_data *data;
9176 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9178 ret = io_rsrc_node_switch_start(ctx);
9181 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9184 ret = io_buffers_map_alloc(ctx, nr_args);
9186 io_rsrc_data_free(data);
9190 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9191 ret = io_copy_iov(ctx, &iov, arg, i);
9194 ret = io_buffer_validate(&iov);
9197 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9202 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9208 WARN_ON_ONCE(ctx->buf_data);
9210 ctx->buf_data = data;
9212 __io_sqe_buffers_unregister(ctx);
9214 io_rsrc_node_switch(ctx, NULL);
9218 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9219 struct io_uring_rsrc_update2 *up,
9220 unsigned int nr_args)
9222 u64 __user *tags = u64_to_user_ptr(up->tags);
9223 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9224 struct page *last_hpage = NULL;
9225 bool needs_switch = false;
9231 if (up->offset + nr_args > ctx->nr_user_bufs)
9234 for (done = 0; done < nr_args; done++) {
9235 struct io_mapped_ubuf *imu;
9236 int offset = up->offset + done;
9239 err = io_copy_iov(ctx, &iov, iovs, done);
9242 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9246 err = io_buffer_validate(&iov);
9249 if (!iov.iov_base && tag) {
9253 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9257 i = array_index_nospec(offset, ctx->nr_user_bufs);
9258 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9259 err = io_queue_rsrc_removal(ctx->buf_data, i,
9260 ctx->rsrc_node, ctx->user_bufs[i]);
9261 if (unlikely(err)) {
9262 io_buffer_unmap(ctx, &imu);
9265 ctx->user_bufs[i] = NULL;
9266 needs_switch = true;
9269 ctx->user_bufs[i] = imu;
9270 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9274 io_rsrc_node_switch(ctx, ctx->buf_data);
9275 return done ? done : err;
9278 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9280 __s32 __user *fds = arg;
9286 if (copy_from_user(&fd, fds, sizeof(*fds)))
9289 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9290 if (IS_ERR(ctx->cq_ev_fd)) {
9291 int ret = PTR_ERR(ctx->cq_ev_fd);
9293 ctx->cq_ev_fd = NULL;
9300 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9302 if (ctx->cq_ev_fd) {
9303 eventfd_ctx_put(ctx->cq_ev_fd);
9304 ctx->cq_ev_fd = NULL;
9311 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9313 struct io_buffer *buf;
9314 unsigned long index;
9316 xa_for_each(&ctx->io_buffers, index, buf)
9317 __io_remove_buffers(ctx, buf, index, -1U);
9320 static void io_req_cache_free(struct list_head *list)
9322 struct io_kiocb *req, *nxt;
9324 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9325 list_del(&req->inflight_entry);
9326 kmem_cache_free(req_cachep, req);
9330 static void io_req_caches_free(struct io_ring_ctx *ctx)
9332 struct io_submit_state *state = &ctx->submit_state;
9334 mutex_lock(&ctx->uring_lock);
9336 if (state->free_reqs) {
9337 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9338 state->free_reqs = 0;
9341 io_flush_cached_locked_reqs(ctx, state);
9342 io_req_cache_free(&state->free_list);
9343 mutex_unlock(&ctx->uring_lock);
9346 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9348 if (data && !atomic_dec_and_test(&data->refs))
9349 wait_for_completion(&data->done);
9352 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9354 io_sq_thread_finish(ctx);
9356 if (ctx->mm_account) {
9357 mmdrop(ctx->mm_account);
9358 ctx->mm_account = NULL;
9361 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9362 io_wait_rsrc_data(ctx->buf_data);
9363 io_wait_rsrc_data(ctx->file_data);
9365 mutex_lock(&ctx->uring_lock);
9367 __io_sqe_buffers_unregister(ctx);
9369 __io_sqe_files_unregister(ctx);
9371 __io_cqring_overflow_flush(ctx, true);
9372 mutex_unlock(&ctx->uring_lock);
9373 io_eventfd_unregister(ctx);
9374 io_destroy_buffers(ctx);
9376 put_cred(ctx->sq_creds);
9378 /* there are no registered resources left, nobody uses it */
9380 io_rsrc_node_destroy(ctx->rsrc_node);
9381 if (ctx->rsrc_backup_node)
9382 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9383 flush_delayed_work(&ctx->rsrc_put_work);
9385 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9386 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9388 #if defined(CONFIG_UNIX)
9389 if (ctx->ring_sock) {
9390 ctx->ring_sock->file = NULL; /* so that iput() is called */
9391 sock_release(ctx->ring_sock);
9394 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9396 io_mem_free(ctx->rings);
9397 io_mem_free(ctx->sq_sqes);
9399 percpu_ref_exit(&ctx->refs);
9400 free_uid(ctx->user);
9401 io_req_caches_free(ctx);
9403 io_wq_put_hash(ctx->hash_map);
9404 kfree(ctx->cancel_hash);
9405 kfree(ctx->dummy_ubuf);
9409 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9411 struct io_ring_ctx *ctx = file->private_data;
9414 poll_wait(file, &ctx->poll_wait, wait);
9416 * synchronizes with barrier from wq_has_sleeper call in
9420 if (!io_sqring_full(ctx))
9421 mask |= EPOLLOUT | EPOLLWRNORM;
9424 * Don't flush cqring overflow list here, just do a simple check.
9425 * Otherwise there could possible be ABBA deadlock:
9428 * lock(&ctx->uring_lock);
9430 * lock(&ctx->uring_lock);
9433 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9434 * pushs them to do the flush.
9436 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9437 mask |= EPOLLIN | EPOLLRDNORM;
9442 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9444 const struct cred *creds;
9446 creds = xa_erase(&ctx->personalities, id);
9455 struct io_tctx_exit {
9456 struct callback_head task_work;
9457 struct completion completion;
9458 struct io_ring_ctx *ctx;
9461 static void io_tctx_exit_cb(struct callback_head *cb)
9463 struct io_uring_task *tctx = current->io_uring;
9464 struct io_tctx_exit *work;
9466 work = container_of(cb, struct io_tctx_exit, task_work);
9468 * When @in_idle, we're in cancellation and it's racy to remove the
9469 * node. It'll be removed by the end of cancellation, just ignore it.
9471 if (!atomic_read(&tctx->in_idle))
9472 io_uring_del_tctx_node((unsigned long)work->ctx);
9473 complete(&work->completion);
9476 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9478 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9480 return req->ctx == data;
9483 static void io_ring_exit_work(struct work_struct *work)
9485 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9486 unsigned long timeout = jiffies + HZ * 60 * 5;
9487 unsigned long interval = HZ / 20;
9488 struct io_tctx_exit exit;
9489 struct io_tctx_node *node;
9493 * If we're doing polled IO and end up having requests being
9494 * submitted async (out-of-line), then completions can come in while
9495 * we're waiting for refs to drop. We need to reap these manually,
9496 * as nobody else will be looking for them.
9499 io_uring_try_cancel_requests(ctx, NULL, true);
9501 struct io_sq_data *sqd = ctx->sq_data;
9502 struct task_struct *tsk;
9504 io_sq_thread_park(sqd);
9506 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9507 io_wq_cancel_cb(tsk->io_uring->io_wq,
9508 io_cancel_ctx_cb, ctx, true);
9509 io_sq_thread_unpark(sqd);
9512 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9513 /* there is little hope left, don't run it too often */
9516 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9518 init_completion(&exit.completion);
9519 init_task_work(&exit.task_work, io_tctx_exit_cb);
9522 * Some may use context even when all refs and requests have been put,
9523 * and they are free to do so while still holding uring_lock or
9524 * completion_lock, see io_req_task_submit(). Apart from other work,
9525 * this lock/unlock section also waits them to finish.
9527 mutex_lock(&ctx->uring_lock);
9528 while (!list_empty(&ctx->tctx_list)) {
9529 WARN_ON_ONCE(time_after(jiffies, timeout));
9531 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9533 /* don't spin on a single task if cancellation failed */
9534 list_rotate_left(&ctx->tctx_list);
9535 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9536 if (WARN_ON_ONCE(ret))
9538 wake_up_process(node->task);
9540 mutex_unlock(&ctx->uring_lock);
9541 wait_for_completion(&exit.completion);
9542 mutex_lock(&ctx->uring_lock);
9544 mutex_unlock(&ctx->uring_lock);
9545 spin_lock(&ctx->completion_lock);
9546 spin_unlock(&ctx->completion_lock);
9548 io_ring_ctx_free(ctx);
9551 /* Returns true if we found and killed one or more timeouts */
9552 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9555 struct io_kiocb *req, *tmp;
9558 spin_lock(&ctx->completion_lock);
9559 spin_lock_irq(&ctx->timeout_lock);
9560 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9561 if (io_match_task(req, tsk, cancel_all)) {
9562 io_kill_timeout(req, -ECANCELED);
9566 spin_unlock_irq(&ctx->timeout_lock);
9568 io_commit_cqring(ctx);
9569 spin_unlock(&ctx->completion_lock);
9571 io_cqring_ev_posted(ctx);
9572 return canceled != 0;
9575 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9577 unsigned long index;
9578 struct creds *creds;
9580 mutex_lock(&ctx->uring_lock);
9581 percpu_ref_kill(&ctx->refs);
9583 __io_cqring_overflow_flush(ctx, true);
9584 xa_for_each(&ctx->personalities, index, creds)
9585 io_unregister_personality(ctx, index);
9586 mutex_unlock(&ctx->uring_lock);
9588 io_kill_timeouts(ctx, NULL, true);
9589 io_poll_remove_all(ctx, NULL, true);
9591 /* if we failed setting up the ctx, we might not have any rings */
9592 io_iopoll_try_reap_events(ctx);
9594 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9596 * Use system_unbound_wq to avoid spawning tons of event kworkers
9597 * if we're exiting a ton of rings at the same time. It just adds
9598 * noise and overhead, there's no discernable change in runtime
9599 * over using system_wq.
9601 queue_work(system_unbound_wq, &ctx->exit_work);
9604 static int io_uring_release(struct inode *inode, struct file *file)
9606 struct io_ring_ctx *ctx = file->private_data;
9608 file->private_data = NULL;
9609 io_ring_ctx_wait_and_kill(ctx);
9613 struct io_task_cancel {
9614 struct task_struct *task;
9618 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9620 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9621 struct io_task_cancel *cancel = data;
9623 return io_match_task_safe(req, cancel->task, cancel->all);
9626 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9627 struct task_struct *task, bool cancel_all)
9629 struct io_defer_entry *de;
9632 spin_lock(&ctx->completion_lock);
9633 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9634 if (io_match_task_safe(de->req, task, cancel_all)) {
9635 list_cut_position(&list, &ctx->defer_list, &de->list);
9639 spin_unlock(&ctx->completion_lock);
9640 if (list_empty(&list))
9643 while (!list_empty(&list)) {
9644 de = list_first_entry(&list, struct io_defer_entry, list);
9645 list_del_init(&de->list);
9646 io_req_complete_failed(de->req, -ECANCELED);
9652 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9654 struct io_tctx_node *node;
9655 enum io_wq_cancel cret;
9658 mutex_lock(&ctx->uring_lock);
9659 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9660 struct io_uring_task *tctx = node->task->io_uring;
9663 * io_wq will stay alive while we hold uring_lock, because it's
9664 * killed after ctx nodes, which requires to take the lock.
9666 if (!tctx || !tctx->io_wq)
9668 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9669 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9671 mutex_unlock(&ctx->uring_lock);
9676 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9677 struct task_struct *task,
9680 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9681 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9684 enum io_wq_cancel cret;
9688 ret |= io_uring_try_cancel_iowq(ctx);
9689 } else if (tctx && tctx->io_wq) {
9691 * Cancels requests of all rings, not only @ctx, but
9692 * it's fine as the task is in exit/exec.
9694 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9696 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9699 /* SQPOLL thread does its own polling */
9700 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9701 (ctx->sq_data && ctx->sq_data->thread == current)) {
9702 while (!list_empty_careful(&ctx->iopoll_list)) {
9703 io_iopoll_try_reap_events(ctx);
9708 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9709 ret |= io_poll_remove_all(ctx, task, cancel_all);
9710 ret |= io_kill_timeouts(ctx, task, cancel_all);
9712 ret |= io_run_task_work();
9719 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9721 struct io_uring_task *tctx = current->io_uring;
9722 struct io_tctx_node *node;
9725 if (unlikely(!tctx)) {
9726 ret = io_uring_alloc_task_context(current, ctx);
9730 tctx = current->io_uring;
9731 if (ctx->iowq_limits_set) {
9732 unsigned int limits[2] = { ctx->iowq_limits[0],
9733 ctx->iowq_limits[1], };
9735 ret = io_wq_max_workers(tctx->io_wq, limits);
9740 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9741 node = kmalloc(sizeof(*node), GFP_KERNEL);
9745 node->task = current;
9747 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9754 mutex_lock(&ctx->uring_lock);
9755 list_add(&node->ctx_node, &ctx->tctx_list);
9756 mutex_unlock(&ctx->uring_lock);
9763 * Note that this task has used io_uring. We use it for cancelation purposes.
9765 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9767 struct io_uring_task *tctx = current->io_uring;
9769 if (likely(tctx && tctx->last == ctx))
9771 return __io_uring_add_tctx_node(ctx);
9775 * Remove this io_uring_file -> task mapping.
9777 static void io_uring_del_tctx_node(unsigned long index)
9779 struct io_uring_task *tctx = current->io_uring;
9780 struct io_tctx_node *node;
9784 node = xa_erase(&tctx->xa, index);
9788 WARN_ON_ONCE(current != node->task);
9789 WARN_ON_ONCE(list_empty(&node->ctx_node));
9791 mutex_lock(&node->ctx->uring_lock);
9792 list_del(&node->ctx_node);
9793 mutex_unlock(&node->ctx->uring_lock);
9795 if (tctx->last == node->ctx)
9800 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9802 struct io_wq *wq = tctx->io_wq;
9803 struct io_tctx_node *node;
9804 unsigned long index;
9806 xa_for_each(&tctx->xa, index, node) {
9807 io_uring_del_tctx_node(index);
9812 * Must be after io_uring_del_task_file() (removes nodes under
9813 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9815 io_wq_put_and_exit(wq);
9820 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9823 return atomic_read(&tctx->inflight_tracked);
9824 return percpu_counter_sum(&tctx->inflight);
9828 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9829 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9831 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9833 struct io_uring_task *tctx = current->io_uring;
9834 struct io_ring_ctx *ctx;
9838 WARN_ON_ONCE(sqd && sqd->thread != current);
9840 if (!current->io_uring)
9843 io_wq_exit_start(tctx->io_wq);
9845 atomic_inc(&tctx->in_idle);
9847 io_uring_drop_tctx_refs(current);
9848 /* read completions before cancelations */
9849 inflight = tctx_inflight(tctx, !cancel_all);
9854 struct io_tctx_node *node;
9855 unsigned long index;
9857 xa_for_each(&tctx->xa, index, node) {
9858 /* sqpoll task will cancel all its requests */
9859 if (node->ctx->sq_data)
9861 io_uring_try_cancel_requests(node->ctx, current,
9865 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9866 io_uring_try_cancel_requests(ctx, current,
9870 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9872 io_uring_drop_tctx_refs(current);
9875 * If we've seen completions, retry without waiting. This
9876 * avoids a race where a completion comes in before we did
9877 * prepare_to_wait().
9879 if (inflight == tctx_inflight(tctx, !cancel_all))
9881 finish_wait(&tctx->wait, &wait);
9884 io_uring_clean_tctx(tctx);
9887 * We shouldn't run task_works after cancel, so just leave
9888 * ->in_idle set for normal exit.
9890 atomic_dec(&tctx->in_idle);
9891 /* for exec all current's requests should be gone, kill tctx */
9892 __io_uring_free(current);
9896 void __io_uring_cancel(bool cancel_all)
9898 io_uring_cancel_generic(cancel_all, NULL);
9901 static void *io_uring_validate_mmap_request(struct file *file,
9902 loff_t pgoff, size_t sz)
9904 struct io_ring_ctx *ctx = file->private_data;
9905 loff_t offset = pgoff << PAGE_SHIFT;
9910 case IORING_OFF_SQ_RING:
9911 case IORING_OFF_CQ_RING:
9914 case IORING_OFF_SQES:
9918 return ERR_PTR(-EINVAL);
9921 page = virt_to_head_page(ptr);
9922 if (sz > page_size(page))
9923 return ERR_PTR(-EINVAL);
9930 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9932 size_t sz = vma->vm_end - vma->vm_start;
9936 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9938 return PTR_ERR(ptr);
9940 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9941 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9944 #else /* !CONFIG_MMU */
9946 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9948 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9951 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9953 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9956 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9957 unsigned long addr, unsigned long len,
9958 unsigned long pgoff, unsigned long flags)
9962 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9964 return PTR_ERR(ptr);
9966 return (unsigned long) ptr;
9969 #endif /* !CONFIG_MMU */
9971 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9976 if (!io_sqring_full(ctx))
9978 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9980 if (!io_sqring_full(ctx))
9983 } while (!signal_pending(current));
9985 finish_wait(&ctx->sqo_sq_wait, &wait);
9989 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9990 struct __kernel_timespec __user **ts,
9991 const sigset_t __user **sig)
9993 struct io_uring_getevents_arg arg;
9996 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9997 * is just a pointer to the sigset_t.
9999 if (!(flags & IORING_ENTER_EXT_ARG)) {
10000 *sig = (const sigset_t __user *) argp;
10006 * EXT_ARG is set - ensure we agree on the size of it and copy in our
10007 * timespec and sigset_t pointers if good.
10009 if (*argsz != sizeof(arg))
10011 if (copy_from_user(&arg, argp, sizeof(arg)))
10015 *sig = u64_to_user_ptr(arg.sigmask);
10016 *argsz = arg.sigmask_sz;
10017 *ts = u64_to_user_ptr(arg.ts);
10021 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
10022 u32, min_complete, u32, flags, const void __user *, argp,
10025 struct io_ring_ctx *ctx;
10030 io_run_task_work();
10032 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
10033 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
10037 if (unlikely(!f.file))
10041 if (unlikely(f.file->f_op != &io_uring_fops))
10045 ctx = f.file->private_data;
10046 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
10050 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10054 * For SQ polling, the thread will do all submissions and completions.
10055 * Just return the requested submit count, and wake the thread if
10056 * we were asked to.
10059 if (ctx->flags & IORING_SETUP_SQPOLL) {
10060 io_cqring_overflow_flush(ctx);
10062 if (unlikely(ctx->sq_data->thread == NULL)) {
10066 if (flags & IORING_ENTER_SQ_WAKEUP)
10067 wake_up(&ctx->sq_data->wait);
10068 if (flags & IORING_ENTER_SQ_WAIT) {
10069 ret = io_sqpoll_wait_sq(ctx);
10073 submitted = to_submit;
10074 } else if (to_submit) {
10075 ret = io_uring_add_tctx_node(ctx);
10078 mutex_lock(&ctx->uring_lock);
10079 submitted = io_submit_sqes(ctx, to_submit);
10080 mutex_unlock(&ctx->uring_lock);
10082 if (submitted != to_submit)
10085 if (flags & IORING_ENTER_GETEVENTS) {
10086 const sigset_t __user *sig;
10087 struct __kernel_timespec __user *ts;
10089 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10093 min_complete = min(min_complete, ctx->cq_entries);
10096 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10097 * space applications don't need to do io completion events
10098 * polling again, they can rely on io_sq_thread to do polling
10099 * work, which can reduce cpu usage and uring_lock contention.
10101 if (ctx->flags & IORING_SETUP_IOPOLL &&
10102 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10103 ret = io_iopoll_check(ctx, min_complete);
10105 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10110 percpu_ref_put(&ctx->refs);
10113 return submitted ? submitted : ret;
10116 #ifdef CONFIG_PROC_FS
10117 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10118 const struct cred *cred)
10120 struct user_namespace *uns = seq_user_ns(m);
10121 struct group_info *gi;
10126 seq_printf(m, "%5d\n", id);
10127 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10128 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10129 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10130 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10131 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10132 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10133 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10134 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10135 seq_puts(m, "\n\tGroups:\t");
10136 gi = cred->group_info;
10137 for (g = 0; g < gi->ngroups; g++) {
10138 seq_put_decimal_ull(m, g ? " " : "",
10139 from_kgid_munged(uns, gi->gid[g]));
10141 seq_puts(m, "\n\tCapEff:\t");
10142 cap = cred->cap_effective;
10143 CAP_FOR_EACH_U32(__capi)
10144 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10149 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10151 struct io_sq_data *sq = NULL;
10156 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10157 * since fdinfo case grabs it in the opposite direction of normal use
10158 * cases. If we fail to get the lock, we just don't iterate any
10159 * structures that could be going away outside the io_uring mutex.
10161 has_lock = mutex_trylock(&ctx->uring_lock);
10163 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10169 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10170 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10171 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10172 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10173 struct file *f = io_file_from_index(ctx, i);
10176 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10178 seq_printf(m, "%5u: <none>\n", i);
10180 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10181 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10182 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10183 unsigned int len = buf->ubuf_end - buf->ubuf;
10185 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10187 if (has_lock && !xa_empty(&ctx->personalities)) {
10188 unsigned long index;
10189 const struct cred *cred;
10191 seq_printf(m, "Personalities:\n");
10192 xa_for_each(&ctx->personalities, index, cred)
10193 io_uring_show_cred(m, index, cred);
10195 seq_printf(m, "PollList:\n");
10196 spin_lock(&ctx->completion_lock);
10197 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10198 struct hlist_head *list = &ctx->cancel_hash[i];
10199 struct io_kiocb *req;
10201 hlist_for_each_entry(req, list, hash_node)
10202 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10203 req->task->task_works != NULL);
10205 spin_unlock(&ctx->completion_lock);
10207 mutex_unlock(&ctx->uring_lock);
10210 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10212 struct io_ring_ctx *ctx = f->private_data;
10214 if (percpu_ref_tryget(&ctx->refs)) {
10215 __io_uring_show_fdinfo(ctx, m);
10216 percpu_ref_put(&ctx->refs);
10221 static const struct file_operations io_uring_fops = {
10222 .release = io_uring_release,
10223 .mmap = io_uring_mmap,
10225 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10226 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10228 .poll = io_uring_poll,
10229 #ifdef CONFIG_PROC_FS
10230 .show_fdinfo = io_uring_show_fdinfo,
10234 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10235 struct io_uring_params *p)
10237 struct io_rings *rings;
10238 size_t size, sq_array_offset;
10240 /* make sure these are sane, as we already accounted them */
10241 ctx->sq_entries = p->sq_entries;
10242 ctx->cq_entries = p->cq_entries;
10244 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10245 if (size == SIZE_MAX)
10248 rings = io_mem_alloc(size);
10252 ctx->rings = rings;
10253 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10254 rings->sq_ring_mask = p->sq_entries - 1;
10255 rings->cq_ring_mask = p->cq_entries - 1;
10256 rings->sq_ring_entries = p->sq_entries;
10257 rings->cq_ring_entries = p->cq_entries;
10259 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10260 if (size == SIZE_MAX) {
10261 io_mem_free(ctx->rings);
10266 ctx->sq_sqes = io_mem_alloc(size);
10267 if (!ctx->sq_sqes) {
10268 io_mem_free(ctx->rings);
10276 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10280 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10284 ret = io_uring_add_tctx_node(ctx);
10289 fd_install(fd, file);
10294 * Allocate an anonymous fd, this is what constitutes the application
10295 * visible backing of an io_uring instance. The application mmaps this
10296 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10297 * we have to tie this fd to a socket for file garbage collection purposes.
10299 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10302 #if defined(CONFIG_UNIX)
10305 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10308 return ERR_PTR(ret);
10311 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10312 O_RDWR | O_CLOEXEC);
10313 #if defined(CONFIG_UNIX)
10314 if (IS_ERR(file)) {
10315 sock_release(ctx->ring_sock);
10316 ctx->ring_sock = NULL;
10318 ctx->ring_sock->file = file;
10324 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10325 struct io_uring_params __user *params)
10327 struct io_ring_ctx *ctx;
10333 if (entries > IORING_MAX_ENTRIES) {
10334 if (!(p->flags & IORING_SETUP_CLAMP))
10336 entries = IORING_MAX_ENTRIES;
10340 * Use twice as many entries for the CQ ring. It's possible for the
10341 * application to drive a higher depth than the size of the SQ ring,
10342 * since the sqes are only used at submission time. This allows for
10343 * some flexibility in overcommitting a bit. If the application has
10344 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10345 * of CQ ring entries manually.
10347 p->sq_entries = roundup_pow_of_two(entries);
10348 if (p->flags & IORING_SETUP_CQSIZE) {
10350 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10351 * to a power-of-two, if it isn't already. We do NOT impose
10352 * any cq vs sq ring sizing.
10354 if (!p->cq_entries)
10356 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10357 if (!(p->flags & IORING_SETUP_CLAMP))
10359 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10361 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10362 if (p->cq_entries < p->sq_entries)
10365 p->cq_entries = 2 * p->sq_entries;
10368 ctx = io_ring_ctx_alloc(p);
10371 ctx->compat = in_compat_syscall();
10372 if (!capable(CAP_IPC_LOCK))
10373 ctx->user = get_uid(current_user());
10376 * This is just grabbed for accounting purposes. When a process exits,
10377 * the mm is exited and dropped before the files, hence we need to hang
10378 * on to this mm purely for the purposes of being able to unaccount
10379 * memory (locked/pinned vm). It's not used for anything else.
10381 mmgrab(current->mm);
10382 ctx->mm_account = current->mm;
10384 ret = io_allocate_scq_urings(ctx, p);
10388 ret = io_sq_offload_create(ctx, p);
10391 /* always set a rsrc node */
10392 ret = io_rsrc_node_switch_start(ctx);
10395 io_rsrc_node_switch(ctx, NULL);
10397 memset(&p->sq_off, 0, sizeof(p->sq_off));
10398 p->sq_off.head = offsetof(struct io_rings, sq.head);
10399 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10400 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10401 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10402 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10403 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10404 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10406 memset(&p->cq_off, 0, sizeof(p->cq_off));
10407 p->cq_off.head = offsetof(struct io_rings, cq.head);
10408 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10409 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10410 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10411 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10412 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10413 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10415 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10416 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10417 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10418 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10419 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10420 IORING_FEAT_RSRC_TAGS;
10422 if (copy_to_user(params, p, sizeof(*p))) {
10427 file = io_uring_get_file(ctx);
10428 if (IS_ERR(file)) {
10429 ret = PTR_ERR(file);
10434 * Install ring fd as the very last thing, so we don't risk someone
10435 * having closed it before we finish setup
10437 ret = io_uring_install_fd(ctx, file);
10439 /* fput will clean it up */
10444 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10447 io_ring_ctx_wait_and_kill(ctx);
10452 * Sets up an aio uring context, and returns the fd. Applications asks for a
10453 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10454 * params structure passed in.
10456 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10458 struct io_uring_params p;
10461 if (copy_from_user(&p, params, sizeof(p)))
10463 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10468 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10469 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10470 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10471 IORING_SETUP_R_DISABLED))
10474 return io_uring_create(entries, &p, params);
10477 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10478 struct io_uring_params __user *, params)
10480 return io_uring_setup(entries, params);
10483 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10485 struct io_uring_probe *p;
10489 size = struct_size(p, ops, nr_args);
10490 if (size == SIZE_MAX)
10492 p = kzalloc(size, GFP_KERNEL);
10497 if (copy_from_user(p, arg, size))
10500 if (memchr_inv(p, 0, size))
10503 p->last_op = IORING_OP_LAST - 1;
10504 if (nr_args > IORING_OP_LAST)
10505 nr_args = IORING_OP_LAST;
10507 for (i = 0; i < nr_args; i++) {
10509 if (!io_op_defs[i].not_supported)
10510 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10515 if (copy_to_user(arg, p, size))
10522 static int io_register_personality(struct io_ring_ctx *ctx)
10524 const struct cred *creds;
10528 creds = get_current_cred();
10530 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10531 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10539 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10540 unsigned int nr_args)
10542 struct io_uring_restriction *res;
10546 /* Restrictions allowed only if rings started disabled */
10547 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10550 /* We allow only a single restrictions registration */
10551 if (ctx->restrictions.registered)
10554 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10557 size = array_size(nr_args, sizeof(*res));
10558 if (size == SIZE_MAX)
10561 res = memdup_user(arg, size);
10563 return PTR_ERR(res);
10567 for (i = 0; i < nr_args; i++) {
10568 switch (res[i].opcode) {
10569 case IORING_RESTRICTION_REGISTER_OP:
10570 if (res[i].register_op >= IORING_REGISTER_LAST) {
10575 __set_bit(res[i].register_op,
10576 ctx->restrictions.register_op);
10578 case IORING_RESTRICTION_SQE_OP:
10579 if (res[i].sqe_op >= IORING_OP_LAST) {
10584 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10586 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10587 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10589 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10590 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10599 /* Reset all restrictions if an error happened */
10601 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10603 ctx->restrictions.registered = true;
10609 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10611 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10614 if (ctx->restrictions.registered)
10615 ctx->restricted = 1;
10617 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10618 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10619 wake_up(&ctx->sq_data->wait);
10623 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10624 struct io_uring_rsrc_update2 *up,
10630 if (check_add_overflow(up->offset, nr_args, &tmp))
10632 err = io_rsrc_node_switch_start(ctx);
10637 case IORING_RSRC_FILE:
10638 return __io_sqe_files_update(ctx, up, nr_args);
10639 case IORING_RSRC_BUFFER:
10640 return __io_sqe_buffers_update(ctx, up, nr_args);
10645 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10648 struct io_uring_rsrc_update2 up;
10652 memset(&up, 0, sizeof(up));
10653 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10655 if (up.resv || up.resv2)
10657 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10660 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10661 unsigned size, unsigned type)
10663 struct io_uring_rsrc_update2 up;
10665 if (size != sizeof(up))
10667 if (copy_from_user(&up, arg, sizeof(up)))
10669 if (!up.nr || up.resv || up.resv2)
10671 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10674 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10675 unsigned int size, unsigned int type)
10677 struct io_uring_rsrc_register rr;
10679 /* keep it extendible */
10680 if (size != sizeof(rr))
10683 memset(&rr, 0, sizeof(rr));
10684 if (copy_from_user(&rr, arg, size))
10686 if (!rr.nr || rr.resv || rr.resv2)
10690 case IORING_RSRC_FILE:
10691 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10692 rr.nr, u64_to_user_ptr(rr.tags));
10693 case IORING_RSRC_BUFFER:
10694 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10695 rr.nr, u64_to_user_ptr(rr.tags));
10700 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10703 struct io_uring_task *tctx = current->io_uring;
10704 cpumask_var_t new_mask;
10707 if (!tctx || !tctx->io_wq)
10710 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10713 cpumask_clear(new_mask);
10714 if (len > cpumask_size())
10715 len = cpumask_size();
10717 if (in_compat_syscall()) {
10718 ret = compat_get_bitmap(cpumask_bits(new_mask),
10719 (const compat_ulong_t __user *)arg,
10720 len * 8 /* CHAR_BIT */);
10722 ret = copy_from_user(new_mask, arg, len);
10726 free_cpumask_var(new_mask);
10730 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10731 free_cpumask_var(new_mask);
10735 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10737 struct io_uring_task *tctx = current->io_uring;
10739 if (!tctx || !tctx->io_wq)
10742 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10745 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10747 __must_hold(&ctx->uring_lock)
10749 struct io_tctx_node *node;
10750 struct io_uring_task *tctx = NULL;
10751 struct io_sq_data *sqd = NULL;
10752 __u32 new_count[2];
10755 if (copy_from_user(new_count, arg, sizeof(new_count)))
10757 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10758 if (new_count[i] > INT_MAX)
10761 if (ctx->flags & IORING_SETUP_SQPOLL) {
10762 sqd = ctx->sq_data;
10765 * Observe the correct sqd->lock -> ctx->uring_lock
10766 * ordering. Fine to drop uring_lock here, we hold
10767 * a ref to the ctx.
10769 refcount_inc(&sqd->refs);
10770 mutex_unlock(&ctx->uring_lock);
10771 mutex_lock(&sqd->lock);
10772 mutex_lock(&ctx->uring_lock);
10774 tctx = sqd->thread->io_uring;
10777 tctx = current->io_uring;
10780 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10782 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10784 ctx->iowq_limits[i] = new_count[i];
10785 ctx->iowq_limits_set = true;
10788 if (tctx && tctx->io_wq) {
10789 ret = io_wq_max_workers(tctx->io_wq, new_count);
10793 memset(new_count, 0, sizeof(new_count));
10797 mutex_unlock(&sqd->lock);
10798 io_put_sq_data(sqd);
10801 if (copy_to_user(arg, new_count, sizeof(new_count)))
10804 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10808 /* now propagate the restriction to all registered users */
10809 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10810 struct io_uring_task *tctx = node->task->io_uring;
10812 if (WARN_ON_ONCE(!tctx->io_wq))
10815 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10816 new_count[i] = ctx->iowq_limits[i];
10817 /* ignore errors, it always returns zero anyway */
10818 (void)io_wq_max_workers(tctx->io_wq, new_count);
10823 mutex_unlock(&sqd->lock);
10824 io_put_sq_data(sqd);
10829 static bool io_register_op_must_quiesce(int op)
10832 case IORING_REGISTER_BUFFERS:
10833 case IORING_UNREGISTER_BUFFERS:
10834 case IORING_REGISTER_FILES:
10835 case IORING_UNREGISTER_FILES:
10836 case IORING_REGISTER_FILES_UPDATE:
10837 case IORING_REGISTER_PROBE:
10838 case IORING_REGISTER_PERSONALITY:
10839 case IORING_UNREGISTER_PERSONALITY:
10840 case IORING_REGISTER_FILES2:
10841 case IORING_REGISTER_FILES_UPDATE2:
10842 case IORING_REGISTER_BUFFERS2:
10843 case IORING_REGISTER_BUFFERS_UPDATE:
10844 case IORING_REGISTER_IOWQ_AFF:
10845 case IORING_UNREGISTER_IOWQ_AFF:
10846 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10853 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10857 percpu_ref_kill(&ctx->refs);
10860 * Drop uring mutex before waiting for references to exit. If another
10861 * thread is currently inside io_uring_enter() it might need to grab the
10862 * uring_lock to make progress. If we hold it here across the drain
10863 * wait, then we can deadlock. It's safe to drop the mutex here, since
10864 * no new references will come in after we've killed the percpu ref.
10866 mutex_unlock(&ctx->uring_lock);
10868 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10871 ret = io_run_task_work_sig();
10872 } while (ret >= 0);
10873 mutex_lock(&ctx->uring_lock);
10876 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10880 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10881 void __user *arg, unsigned nr_args)
10882 __releases(ctx->uring_lock)
10883 __acquires(ctx->uring_lock)
10888 * We're inside the ring mutex, if the ref is already dying, then
10889 * someone else killed the ctx or is already going through
10890 * io_uring_register().
10892 if (percpu_ref_is_dying(&ctx->refs))
10895 if (ctx->restricted) {
10896 if (opcode >= IORING_REGISTER_LAST)
10898 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10899 if (!test_bit(opcode, ctx->restrictions.register_op))
10903 if (io_register_op_must_quiesce(opcode)) {
10904 ret = io_ctx_quiesce(ctx);
10910 case IORING_REGISTER_BUFFERS:
10911 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10913 case IORING_UNREGISTER_BUFFERS:
10915 if (arg || nr_args)
10917 ret = io_sqe_buffers_unregister(ctx);
10919 case IORING_REGISTER_FILES:
10920 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10922 case IORING_UNREGISTER_FILES:
10924 if (arg || nr_args)
10926 ret = io_sqe_files_unregister(ctx);
10928 case IORING_REGISTER_FILES_UPDATE:
10929 ret = io_register_files_update(ctx, arg, nr_args);
10931 case IORING_REGISTER_EVENTFD:
10932 case IORING_REGISTER_EVENTFD_ASYNC:
10936 ret = io_eventfd_register(ctx, arg);
10939 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10940 ctx->eventfd_async = 1;
10942 ctx->eventfd_async = 0;
10944 case IORING_UNREGISTER_EVENTFD:
10946 if (arg || nr_args)
10948 ret = io_eventfd_unregister(ctx);
10950 case IORING_REGISTER_PROBE:
10952 if (!arg || nr_args > 256)
10954 ret = io_probe(ctx, arg, nr_args);
10956 case IORING_REGISTER_PERSONALITY:
10958 if (arg || nr_args)
10960 ret = io_register_personality(ctx);
10962 case IORING_UNREGISTER_PERSONALITY:
10966 ret = io_unregister_personality(ctx, nr_args);
10968 case IORING_REGISTER_ENABLE_RINGS:
10970 if (arg || nr_args)
10972 ret = io_register_enable_rings(ctx);
10974 case IORING_REGISTER_RESTRICTIONS:
10975 ret = io_register_restrictions(ctx, arg, nr_args);
10977 case IORING_REGISTER_FILES2:
10978 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10980 case IORING_REGISTER_FILES_UPDATE2:
10981 ret = io_register_rsrc_update(ctx, arg, nr_args,
10984 case IORING_REGISTER_BUFFERS2:
10985 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10987 case IORING_REGISTER_BUFFERS_UPDATE:
10988 ret = io_register_rsrc_update(ctx, arg, nr_args,
10989 IORING_RSRC_BUFFER);
10991 case IORING_REGISTER_IOWQ_AFF:
10993 if (!arg || !nr_args)
10995 ret = io_register_iowq_aff(ctx, arg, nr_args);
10997 case IORING_UNREGISTER_IOWQ_AFF:
10999 if (arg || nr_args)
11001 ret = io_unregister_iowq_aff(ctx);
11003 case IORING_REGISTER_IOWQ_MAX_WORKERS:
11005 if (!arg || nr_args != 2)
11007 ret = io_register_iowq_max_workers(ctx, arg);
11014 if (io_register_op_must_quiesce(opcode)) {
11015 /* bring the ctx back to life */
11016 percpu_ref_reinit(&ctx->refs);
11017 reinit_completion(&ctx->ref_comp);
11022 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
11023 void __user *, arg, unsigned int, nr_args)
11025 struct io_ring_ctx *ctx;
11034 if (f.file->f_op != &io_uring_fops)
11037 ctx = f.file->private_data;
11039 io_run_task_work();
11041 mutex_lock(&ctx->uring_lock);
11042 ret = __io_uring_register(ctx, opcode, arg, nr_args);
11043 mutex_unlock(&ctx->uring_lock);
11044 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
11045 ctx->cq_ev_fd != NULL, ret);
11051 static int __init io_uring_init(void)
11053 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11054 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11055 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11058 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11059 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11060 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11061 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11062 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11063 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11064 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11065 BUILD_BUG_SQE_ELEM(8, __u64, off);
11066 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11067 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11068 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11069 BUILD_BUG_SQE_ELEM(24, __u32, len);
11070 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11071 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11072 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11073 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11074 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11075 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11076 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11077 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11078 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11079 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11080 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11081 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11082 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11083 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11084 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11085 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11086 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11087 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11088 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11089 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11090 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11092 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11093 sizeof(struct io_uring_rsrc_update));
11094 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11095 sizeof(struct io_uring_rsrc_update2));
11097 /* ->buf_index is u16 */
11098 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11100 /* should fit into one byte */
11101 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11103 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11104 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11106 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11110 __initcall(io_uring_init);