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(¤t->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);
3723 return ret ?: -EAGAIN;
3726 /* it's reportedly faster than delegating the null check to kfree() */
3732 static int io_renameat_prep(struct io_kiocb *req,
3733 const struct io_uring_sqe *sqe)
3735 struct io_rename *ren = &req->rename;
3736 const char __user *oldf, *newf;
3738 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3740 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3742 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3745 ren->old_dfd = READ_ONCE(sqe->fd);
3746 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3747 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3748 ren->new_dfd = READ_ONCE(sqe->len);
3749 ren->flags = READ_ONCE(sqe->rename_flags);
3751 ren->oldpath = getname(oldf);
3752 if (IS_ERR(ren->oldpath))
3753 return PTR_ERR(ren->oldpath);
3755 ren->newpath = getname(newf);
3756 if (IS_ERR(ren->newpath)) {
3757 putname(ren->oldpath);
3758 return PTR_ERR(ren->newpath);
3761 req->flags |= REQ_F_NEED_CLEANUP;
3765 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3767 struct io_rename *ren = &req->rename;
3770 if (issue_flags & IO_URING_F_NONBLOCK)
3773 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3774 ren->newpath, ren->flags);
3776 req->flags &= ~REQ_F_NEED_CLEANUP;
3779 io_req_complete(req, ret);
3783 static int io_unlinkat_prep(struct io_kiocb *req,
3784 const struct io_uring_sqe *sqe)
3786 struct io_unlink *un = &req->unlink;
3787 const char __user *fname;
3789 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3791 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3794 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3797 un->dfd = READ_ONCE(sqe->fd);
3799 un->flags = READ_ONCE(sqe->unlink_flags);
3800 if (un->flags & ~AT_REMOVEDIR)
3803 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3804 un->filename = getname(fname);
3805 if (IS_ERR(un->filename))
3806 return PTR_ERR(un->filename);
3808 req->flags |= REQ_F_NEED_CLEANUP;
3812 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3814 struct io_unlink *un = &req->unlink;
3817 if (issue_flags & IO_URING_F_NONBLOCK)
3820 if (un->flags & AT_REMOVEDIR)
3821 ret = do_rmdir(un->dfd, un->filename);
3823 ret = do_unlinkat(un->dfd, un->filename);
3825 req->flags &= ~REQ_F_NEED_CLEANUP;
3828 io_req_complete(req, ret);
3832 static int io_mkdirat_prep(struct io_kiocb *req,
3833 const struct io_uring_sqe *sqe)
3835 struct io_mkdir *mkd = &req->mkdir;
3836 const char __user *fname;
3838 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3840 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3843 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3846 mkd->dfd = READ_ONCE(sqe->fd);
3847 mkd->mode = READ_ONCE(sqe->len);
3849 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3850 mkd->filename = getname(fname);
3851 if (IS_ERR(mkd->filename))
3852 return PTR_ERR(mkd->filename);
3854 req->flags |= REQ_F_NEED_CLEANUP;
3858 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3860 struct io_mkdir *mkd = &req->mkdir;
3863 if (issue_flags & IO_URING_F_NONBLOCK)
3866 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3868 req->flags &= ~REQ_F_NEED_CLEANUP;
3871 io_req_complete(req, ret);
3875 static int io_symlinkat_prep(struct io_kiocb *req,
3876 const struct io_uring_sqe *sqe)
3878 struct io_symlink *sl = &req->symlink;
3879 const char __user *oldpath, *newpath;
3881 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3883 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3886 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3889 sl->new_dfd = READ_ONCE(sqe->fd);
3890 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3891 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3893 sl->oldpath = getname(oldpath);
3894 if (IS_ERR(sl->oldpath))
3895 return PTR_ERR(sl->oldpath);
3897 sl->newpath = getname(newpath);
3898 if (IS_ERR(sl->newpath)) {
3899 putname(sl->oldpath);
3900 return PTR_ERR(sl->newpath);
3903 req->flags |= REQ_F_NEED_CLEANUP;
3907 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3909 struct io_symlink *sl = &req->symlink;
3912 if (issue_flags & IO_URING_F_NONBLOCK)
3915 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3917 req->flags &= ~REQ_F_NEED_CLEANUP;
3920 io_req_complete(req, ret);
3924 static int io_linkat_prep(struct io_kiocb *req,
3925 const struct io_uring_sqe *sqe)
3927 struct io_hardlink *lnk = &req->hardlink;
3928 const char __user *oldf, *newf;
3930 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3932 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3934 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3937 lnk->old_dfd = READ_ONCE(sqe->fd);
3938 lnk->new_dfd = READ_ONCE(sqe->len);
3939 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3940 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3941 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3943 lnk->oldpath = getname(oldf);
3944 if (IS_ERR(lnk->oldpath))
3945 return PTR_ERR(lnk->oldpath);
3947 lnk->newpath = getname(newf);
3948 if (IS_ERR(lnk->newpath)) {
3949 putname(lnk->oldpath);
3950 return PTR_ERR(lnk->newpath);
3953 req->flags |= REQ_F_NEED_CLEANUP;
3957 static int io_linkat(struct io_kiocb *req, int issue_flags)
3959 struct io_hardlink *lnk = &req->hardlink;
3962 if (issue_flags & IO_URING_F_NONBLOCK)
3965 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3966 lnk->newpath, lnk->flags);
3968 req->flags &= ~REQ_F_NEED_CLEANUP;
3971 io_req_complete(req, ret);
3975 static int io_shutdown_prep(struct io_kiocb *req,
3976 const struct io_uring_sqe *sqe)
3978 #if defined(CONFIG_NET)
3979 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3981 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3982 sqe->buf_index || sqe->splice_fd_in))
3985 req->shutdown.how = READ_ONCE(sqe->len);
3992 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3994 #if defined(CONFIG_NET)
3995 struct socket *sock;
3998 if (issue_flags & IO_URING_F_NONBLOCK)
4001 sock = sock_from_file(req->file);
4002 if (unlikely(!sock))
4005 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4008 io_req_complete(req, ret);
4015 static int __io_splice_prep(struct io_kiocb *req,
4016 const struct io_uring_sqe *sqe)
4018 struct io_splice *sp = &req->splice;
4019 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4021 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4024 sp->len = READ_ONCE(sqe->len);
4025 sp->flags = READ_ONCE(sqe->splice_flags);
4026 if (unlikely(sp->flags & ~valid_flags))
4028 sp->splice_fd_in = READ_ONCE(sqe->splice_fd_in);
4032 static int io_tee_prep(struct io_kiocb *req,
4033 const struct io_uring_sqe *sqe)
4035 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4037 return __io_splice_prep(req, sqe);
4040 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4042 struct io_splice *sp = &req->splice;
4043 struct file *out = sp->file_out;
4044 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4048 if (issue_flags & IO_URING_F_NONBLOCK)
4051 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4052 (sp->flags & SPLICE_F_FD_IN_FIXED));
4059 ret = do_tee(in, out, sp->len, flags);
4061 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4066 io_req_complete(req, ret);
4070 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4072 struct io_splice *sp = &req->splice;
4074 sp->off_in = READ_ONCE(sqe->splice_off_in);
4075 sp->off_out = READ_ONCE(sqe->off);
4076 return __io_splice_prep(req, sqe);
4079 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4081 struct io_splice *sp = &req->splice;
4082 struct file *out = sp->file_out;
4083 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4084 loff_t *poff_in, *poff_out;
4088 if (issue_flags & IO_URING_F_NONBLOCK)
4091 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4092 (sp->flags & SPLICE_F_FD_IN_FIXED));
4098 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4099 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4102 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4104 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4109 io_req_complete(req, ret);
4114 * IORING_OP_NOP just posts a completion event, nothing else.
4116 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4118 struct io_ring_ctx *ctx = req->ctx;
4120 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4123 __io_req_complete(req, issue_flags, 0, 0);
4127 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4129 struct io_ring_ctx *ctx = req->ctx;
4131 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4133 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4137 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4138 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4141 req->sync.off = READ_ONCE(sqe->off);
4142 req->sync.len = READ_ONCE(sqe->len);
4146 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4148 loff_t end = req->sync.off + req->sync.len;
4151 /* fsync always requires a blocking context */
4152 if (issue_flags & IO_URING_F_NONBLOCK)
4155 ret = vfs_fsync_range(req->file, req->sync.off,
4156 end > 0 ? end : LLONG_MAX,
4157 req->sync.flags & IORING_FSYNC_DATASYNC);
4160 io_req_complete(req, ret);
4164 static int io_fallocate_prep(struct io_kiocb *req,
4165 const struct io_uring_sqe *sqe)
4167 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4170 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4173 req->sync.off = READ_ONCE(sqe->off);
4174 req->sync.len = READ_ONCE(sqe->addr);
4175 req->sync.mode = READ_ONCE(sqe->len);
4179 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4183 /* fallocate always requiring blocking context */
4184 if (issue_flags & IO_URING_F_NONBLOCK)
4186 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4191 fsnotify_modify(req->file);
4192 io_req_complete(req, ret);
4196 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4198 const char __user *fname;
4201 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4203 if (unlikely(sqe->ioprio || sqe->buf_index))
4205 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4208 /* open.how should be already initialised */
4209 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4210 req->open.how.flags |= O_LARGEFILE;
4212 req->open.dfd = READ_ONCE(sqe->fd);
4213 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4214 req->open.filename = getname(fname);
4215 if (IS_ERR(req->open.filename)) {
4216 ret = PTR_ERR(req->open.filename);
4217 req->open.filename = NULL;
4221 req->open.file_slot = READ_ONCE(sqe->file_index);
4222 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4225 req->open.nofile = rlimit(RLIMIT_NOFILE);
4226 req->flags |= REQ_F_NEED_CLEANUP;
4230 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4232 u64 mode = READ_ONCE(sqe->len);
4233 u64 flags = READ_ONCE(sqe->open_flags);
4235 req->open.how = build_open_how(flags, mode);
4236 return __io_openat_prep(req, sqe);
4239 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4241 struct open_how __user *how;
4245 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4246 len = READ_ONCE(sqe->len);
4247 if (len < OPEN_HOW_SIZE_VER0)
4250 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4255 return __io_openat_prep(req, sqe);
4258 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4260 struct open_flags op;
4262 bool resolve_nonblock, nonblock_set;
4263 bool fixed = !!req->open.file_slot;
4266 ret = build_open_flags(&req->open.how, &op);
4269 nonblock_set = op.open_flag & O_NONBLOCK;
4270 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4271 if (issue_flags & IO_URING_F_NONBLOCK) {
4273 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4274 * it'll always -EAGAIN
4276 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4278 op.lookup_flags |= LOOKUP_CACHED;
4279 op.open_flag |= O_NONBLOCK;
4283 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4288 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4291 * We could hang on to this 'fd' on retrying, but seems like
4292 * marginal gain for something that is now known to be a slower
4293 * path. So just put it, and we'll get a new one when we retry.
4298 ret = PTR_ERR(file);
4299 /* only retry if RESOLVE_CACHED wasn't already set by application */
4300 if (ret == -EAGAIN &&
4301 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4306 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4307 file->f_flags &= ~O_NONBLOCK;
4308 fsnotify_open(file);
4311 fd_install(ret, file);
4313 ret = io_install_fixed_file(req, file, issue_flags,
4314 req->open.file_slot - 1);
4316 putname(req->open.filename);
4317 req->flags &= ~REQ_F_NEED_CLEANUP;
4320 __io_req_complete(req, issue_flags, ret, 0);
4324 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4326 return io_openat2(req, issue_flags);
4329 static int io_remove_buffers_prep(struct io_kiocb *req,
4330 const struct io_uring_sqe *sqe)
4332 struct io_provide_buf *p = &req->pbuf;
4335 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4339 tmp = READ_ONCE(sqe->fd);
4340 if (!tmp || tmp > USHRT_MAX)
4343 memset(p, 0, sizeof(*p));
4345 p->bgid = READ_ONCE(sqe->buf_group);
4349 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4350 int bgid, unsigned nbufs)
4354 /* shouldn't happen */
4358 /* the head kbuf is the list itself */
4359 while (!list_empty(&buf->list)) {
4360 struct io_buffer *nxt;
4362 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4363 list_del(&nxt->list);
4371 xa_erase(&ctx->io_buffers, bgid);
4376 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4378 struct io_provide_buf *p = &req->pbuf;
4379 struct io_ring_ctx *ctx = req->ctx;
4380 struct io_buffer *head;
4382 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4384 io_ring_submit_lock(ctx, !force_nonblock);
4386 lockdep_assert_held(&ctx->uring_lock);
4389 head = xa_load(&ctx->io_buffers, p->bgid);
4391 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4395 /* complete before unlock, IOPOLL may need the lock */
4396 __io_req_complete(req, issue_flags, ret, 0);
4397 io_ring_submit_unlock(ctx, !force_nonblock);
4401 static int io_provide_buffers_prep(struct io_kiocb *req,
4402 const struct io_uring_sqe *sqe)
4404 unsigned long size, tmp_check;
4405 struct io_provide_buf *p = &req->pbuf;
4408 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4411 tmp = READ_ONCE(sqe->fd);
4412 if (!tmp || tmp > USHRT_MAX)
4415 p->addr = READ_ONCE(sqe->addr);
4416 p->len = READ_ONCE(sqe->len);
4418 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4421 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4424 size = (unsigned long)p->len * p->nbufs;
4425 if (!access_ok(u64_to_user_ptr(p->addr), size))
4428 p->bgid = READ_ONCE(sqe->buf_group);
4429 tmp = READ_ONCE(sqe->off);
4430 if (tmp > USHRT_MAX)
4436 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4438 struct io_buffer *buf;
4439 u64 addr = pbuf->addr;
4440 int i, bid = pbuf->bid;
4442 for (i = 0; i < pbuf->nbufs; i++) {
4443 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4448 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4453 INIT_LIST_HEAD(&buf->list);
4456 list_add_tail(&buf->list, &(*head)->list);
4461 return i ? i : -ENOMEM;
4464 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4466 struct io_provide_buf *p = &req->pbuf;
4467 struct io_ring_ctx *ctx = req->ctx;
4468 struct io_buffer *head, *list;
4470 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4472 io_ring_submit_lock(ctx, !force_nonblock);
4474 lockdep_assert_held(&ctx->uring_lock);
4476 list = head = xa_load(&ctx->io_buffers, p->bgid);
4478 ret = io_add_buffers(p, &head);
4479 if (ret >= 0 && !list) {
4480 ret = xa_insert(&ctx->io_buffers, p->bgid, head,
4481 GFP_KERNEL_ACCOUNT);
4483 __io_remove_buffers(ctx, head, p->bgid, -1U);
4487 /* complete before unlock, IOPOLL may need the lock */
4488 __io_req_complete(req, issue_flags, ret, 0);
4489 io_ring_submit_unlock(ctx, !force_nonblock);
4493 static int io_epoll_ctl_prep(struct io_kiocb *req,
4494 const struct io_uring_sqe *sqe)
4496 #if defined(CONFIG_EPOLL)
4497 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4499 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4502 req->epoll.epfd = READ_ONCE(sqe->fd);
4503 req->epoll.op = READ_ONCE(sqe->len);
4504 req->epoll.fd = READ_ONCE(sqe->off);
4506 if (ep_op_has_event(req->epoll.op)) {
4507 struct epoll_event __user *ev;
4509 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4510 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4520 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4522 #if defined(CONFIG_EPOLL)
4523 struct io_epoll *ie = &req->epoll;
4525 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4527 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4528 if (force_nonblock && ret == -EAGAIN)
4533 __io_req_complete(req, issue_flags, ret, 0);
4540 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4542 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4543 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4548 req->madvise.addr = READ_ONCE(sqe->addr);
4549 req->madvise.len = READ_ONCE(sqe->len);
4550 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4557 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4559 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4560 struct io_madvise *ma = &req->madvise;
4563 if (issue_flags & IO_URING_F_NONBLOCK)
4566 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4569 io_req_complete(req, ret);
4576 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4578 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4583 req->fadvise.offset = READ_ONCE(sqe->off);
4584 req->fadvise.len = READ_ONCE(sqe->len);
4585 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4589 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4591 struct io_fadvise *fa = &req->fadvise;
4594 if (issue_flags & IO_URING_F_NONBLOCK) {
4595 switch (fa->advice) {
4596 case POSIX_FADV_NORMAL:
4597 case POSIX_FADV_RANDOM:
4598 case POSIX_FADV_SEQUENTIAL:
4605 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4608 __io_req_complete(req, issue_flags, ret, 0);
4612 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4614 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4616 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4618 if (req->flags & REQ_F_FIXED_FILE)
4621 req->statx.dfd = READ_ONCE(sqe->fd);
4622 req->statx.mask = READ_ONCE(sqe->len);
4623 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4624 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4625 req->statx.flags = READ_ONCE(sqe->statx_flags);
4630 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4632 struct io_statx *ctx = &req->statx;
4635 if (issue_flags & IO_URING_F_NONBLOCK)
4638 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4643 io_req_complete(req, ret);
4647 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4649 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4651 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4652 sqe->rw_flags || sqe->buf_index)
4654 if (req->flags & REQ_F_FIXED_FILE)
4657 req->close.fd = READ_ONCE(sqe->fd);
4658 req->close.file_slot = READ_ONCE(sqe->file_index);
4659 if (req->close.file_slot && req->close.fd)
4665 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4667 struct files_struct *files = current->files;
4668 struct io_close *close = &req->close;
4669 struct fdtable *fdt;
4670 struct file *file = NULL;
4673 if (req->close.file_slot) {
4674 ret = io_close_fixed(req, issue_flags);
4678 spin_lock(&files->file_lock);
4679 fdt = files_fdtable(files);
4680 if (close->fd >= fdt->max_fds) {
4681 spin_unlock(&files->file_lock);
4684 file = fdt->fd[close->fd];
4685 if (!file || file->f_op == &io_uring_fops) {
4686 spin_unlock(&files->file_lock);
4691 /* if the file has a flush method, be safe and punt to async */
4692 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4693 spin_unlock(&files->file_lock);
4697 ret = __close_fd_get_file(close->fd, &file);
4698 spin_unlock(&files->file_lock);
4705 /* No ->flush() or already async, safely close from here */
4706 ret = filp_close(file, current->files);
4712 __io_req_complete(req, issue_flags, ret, 0);
4716 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718 struct io_ring_ctx *ctx = req->ctx;
4720 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4722 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4726 req->sync.off = READ_ONCE(sqe->off);
4727 req->sync.len = READ_ONCE(sqe->len);
4728 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4732 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4736 /* sync_file_range always requires a blocking context */
4737 if (issue_flags & IO_URING_F_NONBLOCK)
4740 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4744 io_req_complete(req, ret);
4748 #if defined(CONFIG_NET)
4749 static int io_setup_async_msg(struct io_kiocb *req,
4750 struct io_async_msghdr *kmsg)
4752 struct io_async_msghdr *async_msg = req->async_data;
4756 if (io_alloc_async_data(req)) {
4757 kfree(kmsg->free_iov);
4760 async_msg = req->async_data;
4761 req->flags |= REQ_F_NEED_CLEANUP;
4762 memcpy(async_msg, kmsg, sizeof(*kmsg));
4763 async_msg->msg.msg_name = &async_msg->addr;
4764 /* if were using fast_iov, set it to the new one */
4765 if (!async_msg->free_iov)
4766 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4771 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4772 struct io_async_msghdr *iomsg)
4774 iomsg->msg.msg_name = &iomsg->addr;
4775 iomsg->free_iov = iomsg->fast_iov;
4776 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4777 req->sr_msg.msg_flags, &iomsg->free_iov);
4780 static int io_sendmsg_prep_async(struct io_kiocb *req)
4784 ret = io_sendmsg_copy_hdr(req, req->async_data);
4786 req->flags |= REQ_F_NEED_CLEANUP;
4790 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4792 struct io_sr_msg *sr = &req->sr_msg;
4794 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4796 if (unlikely(sqe->addr2 || sqe->file_index))
4798 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
4801 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4802 sr->len = READ_ONCE(sqe->len);
4803 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4804 if (sr->msg_flags & MSG_DONTWAIT)
4805 req->flags |= REQ_F_NOWAIT;
4807 #ifdef CONFIG_COMPAT
4808 if (req->ctx->compat)
4809 sr->msg_flags |= MSG_CMSG_COMPAT;
4814 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4816 struct io_async_msghdr iomsg, *kmsg;
4817 struct socket *sock;
4822 sock = sock_from_file(req->file);
4823 if (unlikely(!sock))
4826 kmsg = req->async_data;
4828 ret = io_sendmsg_copy_hdr(req, &iomsg);
4834 flags = req->sr_msg.msg_flags;
4835 if (issue_flags & IO_URING_F_NONBLOCK)
4836 flags |= MSG_DONTWAIT;
4837 if (flags & MSG_WAITALL)
4838 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4840 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4841 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4842 return io_setup_async_msg(req, kmsg);
4843 if (ret == -ERESTARTSYS)
4846 /* fast path, check for non-NULL to avoid function call */
4848 kfree(kmsg->free_iov);
4849 req->flags &= ~REQ_F_NEED_CLEANUP;
4852 __io_req_complete(req, issue_flags, ret, 0);
4856 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4858 struct io_sr_msg *sr = &req->sr_msg;
4861 struct socket *sock;
4866 sock = sock_from_file(req->file);
4867 if (unlikely(!sock))
4870 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4874 msg.msg_name = NULL;
4875 msg.msg_control = NULL;
4876 msg.msg_controllen = 0;
4877 msg.msg_namelen = 0;
4879 flags = req->sr_msg.msg_flags;
4880 if (issue_flags & IO_URING_F_NONBLOCK)
4881 flags |= MSG_DONTWAIT;
4882 if (flags & MSG_WAITALL)
4883 min_ret = iov_iter_count(&msg.msg_iter);
4885 msg.msg_flags = flags;
4886 ret = sock_sendmsg(sock, &msg);
4887 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4889 if (ret == -ERESTARTSYS)
4894 __io_req_complete(req, issue_flags, ret, 0);
4898 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4899 struct io_async_msghdr *iomsg)
4901 struct io_sr_msg *sr = &req->sr_msg;
4902 struct iovec __user *uiov;
4906 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4907 &iomsg->uaddr, &uiov, &iov_len);
4911 if (req->flags & REQ_F_BUFFER_SELECT) {
4914 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4916 sr->len = iomsg->fast_iov[0].iov_len;
4917 iomsg->free_iov = NULL;
4919 iomsg->free_iov = iomsg->fast_iov;
4920 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4921 &iomsg->free_iov, &iomsg->msg.msg_iter,
4930 #ifdef CONFIG_COMPAT
4931 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4932 struct io_async_msghdr *iomsg)
4934 struct io_sr_msg *sr = &req->sr_msg;
4935 struct compat_iovec __user *uiov;
4940 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4945 uiov = compat_ptr(ptr);
4946 if (req->flags & REQ_F_BUFFER_SELECT) {
4947 compat_ssize_t clen;
4951 if (!access_ok(uiov, sizeof(*uiov)))
4953 if (__get_user(clen, &uiov->iov_len))
4958 iomsg->free_iov = NULL;
4960 iomsg->free_iov = iomsg->fast_iov;
4961 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4962 UIO_FASTIOV, &iomsg->free_iov,
4963 &iomsg->msg.msg_iter, true);
4972 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4973 struct io_async_msghdr *iomsg)
4975 iomsg->msg.msg_name = &iomsg->addr;
4977 #ifdef CONFIG_COMPAT
4978 if (req->ctx->compat)
4979 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4982 return __io_recvmsg_copy_hdr(req, iomsg);
4985 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4988 struct io_sr_msg *sr = &req->sr_msg;
4989 struct io_buffer *kbuf;
4991 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4996 req->flags |= REQ_F_BUFFER_SELECTED;
5000 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
5002 return io_put_kbuf(req, req->sr_msg.kbuf);
5005 static int io_recvmsg_prep_async(struct io_kiocb *req)
5009 ret = io_recvmsg_copy_hdr(req, req->async_data);
5011 req->flags |= REQ_F_NEED_CLEANUP;
5015 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5017 struct io_sr_msg *sr = &req->sr_msg;
5019 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5021 if (unlikely(sqe->addr2 || sqe->file_index))
5023 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
5026 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5027 sr->len = READ_ONCE(sqe->len);
5028 sr->bgid = READ_ONCE(sqe->buf_group);
5029 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5030 if (sr->msg_flags & MSG_DONTWAIT)
5031 req->flags |= REQ_F_NOWAIT;
5033 #ifdef CONFIG_COMPAT
5034 if (req->ctx->compat)
5035 sr->msg_flags |= MSG_CMSG_COMPAT;
5040 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5042 struct io_async_msghdr iomsg, *kmsg;
5043 struct socket *sock;
5044 struct io_buffer *kbuf;
5047 int ret, cflags = 0;
5048 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5050 sock = sock_from_file(req->file);
5051 if (unlikely(!sock))
5054 kmsg = req->async_data;
5056 ret = io_recvmsg_copy_hdr(req, &iomsg);
5062 if (req->flags & REQ_F_BUFFER_SELECT) {
5063 kbuf = io_recv_buffer_select(req, !force_nonblock);
5065 return PTR_ERR(kbuf);
5066 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5067 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5068 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5069 1, req->sr_msg.len);
5072 flags = req->sr_msg.msg_flags;
5074 flags |= MSG_DONTWAIT;
5075 if (flags & MSG_WAITALL)
5076 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5078 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5079 kmsg->uaddr, flags);
5080 if (force_nonblock && ret == -EAGAIN)
5081 return io_setup_async_msg(req, kmsg);
5082 if (ret == -ERESTARTSYS)
5085 if (req->flags & REQ_F_BUFFER_SELECTED)
5086 cflags = io_put_recv_kbuf(req);
5087 /* fast path, check for non-NULL to avoid function call */
5089 kfree(kmsg->free_iov);
5090 req->flags &= ~REQ_F_NEED_CLEANUP;
5091 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5093 __io_req_complete(req, issue_flags, ret, cflags);
5097 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5099 struct io_buffer *kbuf;
5100 struct io_sr_msg *sr = &req->sr_msg;
5102 void __user *buf = sr->buf;
5103 struct socket *sock;
5107 int ret, cflags = 0;
5108 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5110 sock = sock_from_file(req->file);
5111 if (unlikely(!sock))
5114 if (req->flags & REQ_F_BUFFER_SELECT) {
5115 kbuf = io_recv_buffer_select(req, !force_nonblock);
5117 return PTR_ERR(kbuf);
5118 buf = u64_to_user_ptr(kbuf->addr);
5121 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5125 msg.msg_name = NULL;
5126 msg.msg_control = NULL;
5127 msg.msg_controllen = 0;
5128 msg.msg_namelen = 0;
5129 msg.msg_iocb = NULL;
5132 flags = req->sr_msg.msg_flags;
5134 flags |= MSG_DONTWAIT;
5135 if (flags & MSG_WAITALL)
5136 min_ret = iov_iter_count(&msg.msg_iter);
5138 ret = sock_recvmsg(sock, &msg, flags);
5139 if (force_nonblock && ret == -EAGAIN)
5141 if (ret == -ERESTARTSYS)
5144 if (req->flags & REQ_F_BUFFER_SELECTED)
5145 cflags = io_put_recv_kbuf(req);
5146 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5148 __io_req_complete(req, issue_flags, ret, cflags);
5152 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5154 struct io_accept *accept = &req->accept;
5156 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5158 if (sqe->ioprio || sqe->len || sqe->buf_index)
5161 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5162 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5163 accept->flags = READ_ONCE(sqe->accept_flags);
5164 accept->nofile = rlimit(RLIMIT_NOFILE);
5166 accept->file_slot = READ_ONCE(sqe->file_index);
5167 if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))
5169 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5171 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5172 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5176 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5178 struct io_accept *accept = &req->accept;
5179 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5180 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5181 bool fixed = !!accept->file_slot;
5185 if (req->file->f_flags & O_NONBLOCK)
5186 req->flags |= REQ_F_NOWAIT;
5189 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5190 if (unlikely(fd < 0))
5193 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5198 ret = PTR_ERR(file);
5199 if (ret == -EAGAIN && force_nonblock)
5201 if (ret == -ERESTARTSYS)
5204 } else if (!fixed) {
5205 fd_install(fd, file);
5208 ret = io_install_fixed_file(req, file, issue_flags,
5209 accept->file_slot - 1);
5211 __io_req_complete(req, issue_flags, ret, 0);
5215 static int io_connect_prep_async(struct io_kiocb *req)
5217 struct io_async_connect *io = req->async_data;
5218 struct io_connect *conn = &req->connect;
5220 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5223 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5225 struct io_connect *conn = &req->connect;
5227 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5229 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5233 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5234 conn->addr_len = READ_ONCE(sqe->addr2);
5238 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5240 struct io_async_connect __io, *io;
5241 unsigned file_flags;
5243 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5245 if (req->async_data) {
5246 io = req->async_data;
5248 ret = move_addr_to_kernel(req->connect.addr,
5249 req->connect.addr_len,
5256 file_flags = force_nonblock ? O_NONBLOCK : 0;
5258 ret = __sys_connect_file(req->file, &io->address,
5259 req->connect.addr_len, file_flags);
5260 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5261 if (req->async_data)
5263 if (io_alloc_async_data(req)) {
5267 memcpy(req->async_data, &__io, sizeof(__io));
5270 if (ret == -ERESTARTSYS)
5275 __io_req_complete(req, issue_flags, ret, 0);
5278 #else /* !CONFIG_NET */
5279 #define IO_NETOP_FN(op) \
5280 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5282 return -EOPNOTSUPP; \
5285 #define IO_NETOP_PREP(op) \
5287 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5289 return -EOPNOTSUPP; \
5292 #define IO_NETOP_PREP_ASYNC(op) \
5294 static int io_##op##_prep_async(struct io_kiocb *req) \
5296 return -EOPNOTSUPP; \
5299 IO_NETOP_PREP_ASYNC(sendmsg);
5300 IO_NETOP_PREP_ASYNC(recvmsg);
5301 IO_NETOP_PREP_ASYNC(connect);
5302 IO_NETOP_PREP(accept);
5305 #endif /* CONFIG_NET */
5307 struct io_poll_table {
5308 struct poll_table_struct pt;
5309 struct io_kiocb *req;
5314 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5315 __poll_t mask, io_req_tw_func_t func)
5317 /* for instances that support it check for an event match first: */
5318 if (mask && !(mask & poll->events))
5321 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5323 list_del_init(&poll->wait.entry);
5326 req->io_task_work.func = func;
5329 * If this fails, then the task is exiting. When a task exits, the
5330 * work gets canceled, so just cancel this request as well instead
5331 * of executing it. We can't safely execute it anyway, as we may not
5332 * have the needed state needed for it anyway.
5334 io_req_task_work_add(req);
5338 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5339 __acquires(&req->ctx->completion_lock)
5341 struct io_ring_ctx *ctx = req->ctx;
5343 /* req->task == current here, checking PF_EXITING is safe */
5344 if (unlikely(req->task->flags & PF_EXITING))
5345 WRITE_ONCE(poll->canceled, true);
5347 if (!req->result && !READ_ONCE(poll->canceled)) {
5348 struct poll_table_struct pt = { ._key = poll->events };
5350 req->result = vfs_poll(req->file, &pt) & poll->events;
5353 spin_lock(&ctx->completion_lock);
5354 if (!req->result && !READ_ONCE(poll->canceled)) {
5355 add_wait_queue(poll->head, &poll->wait);
5362 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5364 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5365 if (req->opcode == IORING_OP_POLL_ADD)
5366 return req->async_data;
5367 return req->apoll->double_poll;
5370 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5372 if (req->opcode == IORING_OP_POLL_ADD)
5374 return &req->apoll->poll;
5377 static void io_poll_remove_double(struct io_kiocb *req)
5378 __must_hold(&req->ctx->completion_lock)
5380 struct io_poll_iocb *poll = io_poll_get_double(req);
5382 lockdep_assert_held(&req->ctx->completion_lock);
5384 if (poll && poll->head) {
5385 struct wait_queue_head *head = poll->head;
5387 spin_lock_irq(&head->lock);
5388 list_del_init(&poll->wait.entry);
5389 if (poll->wait.private)
5392 spin_unlock_irq(&head->lock);
5396 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5397 __must_hold(&req->ctx->completion_lock)
5399 struct io_ring_ctx *ctx = req->ctx;
5400 unsigned flags = IORING_CQE_F_MORE;
5403 if (READ_ONCE(req->poll.canceled)) {
5405 req->poll.events |= EPOLLONESHOT;
5407 error = mangle_poll(mask);
5409 if (req->poll.events & EPOLLONESHOT)
5411 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5412 req->poll.events |= EPOLLONESHOT;
5415 if (flags & IORING_CQE_F_MORE)
5418 return !(flags & IORING_CQE_F_MORE);
5421 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5422 __must_hold(&req->ctx->completion_lock)
5426 done = __io_poll_complete(req, mask);
5427 io_commit_cqring(req->ctx);
5431 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5433 struct io_ring_ctx *ctx = req->ctx;
5434 struct io_kiocb *nxt;
5436 if (io_poll_rewait(req, &req->poll)) {
5437 spin_unlock(&ctx->completion_lock);
5441 if (req->poll.done) {
5442 spin_unlock(&ctx->completion_lock);
5445 done = __io_poll_complete(req, req->result);
5447 io_poll_remove_double(req);
5448 hash_del(&req->hash_node);
5449 req->poll.done = true;
5452 add_wait_queue(req->poll.head, &req->poll.wait);
5454 io_commit_cqring(ctx);
5455 spin_unlock(&ctx->completion_lock);
5456 io_cqring_ev_posted(ctx);
5459 nxt = io_put_req_find_next(req);
5461 io_req_task_submit(nxt, locked);
5466 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5467 int sync, void *key)
5469 struct io_kiocb *req = wait->private;
5470 struct io_poll_iocb *poll = io_poll_get_single(req);
5471 __poll_t mask = key_to_poll(key);
5472 unsigned long flags;
5474 /* for instances that support it check for an event match first: */
5475 if (mask && !(mask & poll->events))
5477 if (!(poll->events & EPOLLONESHOT))
5478 return poll->wait.func(&poll->wait, mode, sync, key);
5480 list_del_init(&wait->entry);
5485 spin_lock_irqsave(&poll->head->lock, flags);
5486 done = list_empty(&poll->wait.entry);
5488 list_del_init(&poll->wait.entry);
5489 /* make sure double remove sees this as being gone */
5490 wait->private = NULL;
5491 spin_unlock_irqrestore(&poll->head->lock, flags);
5493 /* use wait func handler, so it matches the rq type */
5494 poll->wait.func(&poll->wait, mode, sync, key);
5501 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5502 wait_queue_func_t wake_func)
5506 poll->canceled = false;
5507 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5508 /* mask in events that we always want/need */
5509 poll->events = events | IO_POLL_UNMASK;
5510 INIT_LIST_HEAD(&poll->wait.entry);
5511 init_waitqueue_func_entry(&poll->wait, wake_func);
5514 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5515 struct wait_queue_head *head,
5516 struct io_poll_iocb **poll_ptr)
5518 struct io_kiocb *req = pt->req;
5521 * The file being polled uses multiple waitqueues for poll handling
5522 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5525 if (unlikely(pt->nr_entries)) {
5526 struct io_poll_iocb *poll_one = poll;
5528 /* double add on the same waitqueue head, ignore */
5529 if (poll_one->head == head)
5531 /* already have a 2nd entry, fail a third attempt */
5533 if ((*poll_ptr)->head == head)
5535 pt->error = -EINVAL;
5539 * Can't handle multishot for double wait for now, turn it
5540 * into one-shot mode.
5542 if (!(poll_one->events & EPOLLONESHOT))
5543 poll_one->events |= EPOLLONESHOT;
5544 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5546 pt->error = -ENOMEM;
5549 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5551 poll->wait.private = req;
5558 if (poll->events & EPOLLEXCLUSIVE)
5559 add_wait_queue_exclusive(head, &poll->wait);
5561 add_wait_queue(head, &poll->wait);
5564 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5565 struct poll_table_struct *p)
5567 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5568 struct async_poll *apoll = pt->req->apoll;
5570 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5573 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5575 struct async_poll *apoll = req->apoll;
5576 struct io_ring_ctx *ctx = req->ctx;
5578 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5580 if (io_poll_rewait(req, &apoll->poll)) {
5581 spin_unlock(&ctx->completion_lock);
5585 hash_del(&req->hash_node);
5586 io_poll_remove_double(req);
5587 apoll->poll.done = true;
5588 spin_unlock(&ctx->completion_lock);
5590 if (!READ_ONCE(apoll->poll.canceled))
5591 io_req_task_submit(req, locked);
5593 io_req_complete_failed(req, -ECANCELED);
5596 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5599 struct io_kiocb *req = wait->private;
5600 struct io_poll_iocb *poll = &req->apoll->poll;
5602 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5605 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5608 static void io_poll_req_insert(struct io_kiocb *req)
5610 struct io_ring_ctx *ctx = req->ctx;
5611 struct hlist_head *list;
5613 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5614 hlist_add_head(&req->hash_node, list);
5617 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5618 struct io_poll_iocb *poll,
5619 struct io_poll_table *ipt, __poll_t mask,
5620 wait_queue_func_t wake_func)
5621 __acquires(&ctx->completion_lock)
5623 struct io_ring_ctx *ctx = req->ctx;
5624 bool cancel = false;
5626 INIT_HLIST_NODE(&req->hash_node);
5627 io_init_poll_iocb(poll, mask, wake_func);
5628 poll->file = req->file;
5629 poll->wait.private = req;
5631 ipt->pt._key = mask;
5634 ipt->nr_entries = 0;
5636 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5637 if (unlikely(!ipt->nr_entries) && !ipt->error)
5638 ipt->error = -EINVAL;
5640 spin_lock(&ctx->completion_lock);
5641 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5642 io_poll_remove_double(req);
5643 if (likely(poll->head)) {
5644 spin_lock_irq(&poll->head->lock);
5645 if (unlikely(list_empty(&poll->wait.entry))) {
5651 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5652 list_del_init(&poll->wait.entry);
5654 WRITE_ONCE(poll->canceled, true);
5655 else if (!poll->done) /* actually waiting for an event */
5656 io_poll_req_insert(req);
5657 spin_unlock_irq(&poll->head->lock);
5669 static int io_arm_poll_handler(struct io_kiocb *req)
5671 const struct io_op_def *def = &io_op_defs[req->opcode];
5672 struct io_ring_ctx *ctx = req->ctx;
5673 struct async_poll *apoll;
5674 struct io_poll_table ipt;
5675 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5677 if (!req->file || !file_can_poll(req->file))
5678 return IO_APOLL_ABORTED;
5679 if (req->flags & REQ_F_POLLED)
5680 return IO_APOLL_ABORTED;
5681 if (!def->pollin && !def->pollout)
5682 return IO_APOLL_ABORTED;
5685 mask |= POLLIN | POLLRDNORM;
5687 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5688 if ((req->opcode == IORING_OP_RECVMSG) &&
5689 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5692 mask |= POLLOUT | POLLWRNORM;
5695 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5696 if (unlikely(!apoll))
5697 return IO_APOLL_ABORTED;
5698 apoll->double_poll = NULL;
5700 req->flags |= REQ_F_POLLED;
5701 ipt.pt._qproc = io_async_queue_proc;
5702 io_req_set_refcount(req);
5704 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5706 spin_unlock(&ctx->completion_lock);
5707 if (ret || ipt.error)
5708 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5710 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5711 mask, apoll->poll.events);
5715 static bool __io_poll_remove_one(struct io_kiocb *req,
5716 struct io_poll_iocb *poll, bool do_cancel)
5717 __must_hold(&req->ctx->completion_lock)
5719 bool do_complete = false;
5723 spin_lock_irq(&poll->head->lock);
5725 WRITE_ONCE(poll->canceled, true);
5726 if (!list_empty(&poll->wait.entry)) {
5727 list_del_init(&poll->wait.entry);
5730 spin_unlock_irq(&poll->head->lock);
5731 hash_del(&req->hash_node);
5735 static bool io_poll_remove_one(struct io_kiocb *req)
5736 __must_hold(&req->ctx->completion_lock)
5740 io_poll_remove_double(req);
5741 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5744 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5745 io_commit_cqring(req->ctx);
5747 io_put_req_deferred(req);
5753 * Returns true if we found and killed one or more poll requests
5755 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5758 struct hlist_node *tmp;
5759 struct io_kiocb *req;
5762 spin_lock(&ctx->completion_lock);
5763 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5764 struct hlist_head *list;
5766 list = &ctx->cancel_hash[i];
5767 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5768 if (io_match_task_safe(req, tsk, cancel_all))
5769 posted += io_poll_remove_one(req);
5772 spin_unlock(&ctx->completion_lock);
5775 io_cqring_ev_posted(ctx);
5780 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5782 __must_hold(&ctx->completion_lock)
5784 struct hlist_head *list;
5785 struct io_kiocb *req;
5787 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5788 hlist_for_each_entry(req, list, hash_node) {
5789 if (sqe_addr != req->user_data)
5791 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5798 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5800 __must_hold(&ctx->completion_lock)
5802 struct io_kiocb *req;
5804 req = io_poll_find(ctx, sqe_addr, poll_only);
5807 if (io_poll_remove_one(req))
5813 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5818 events = READ_ONCE(sqe->poll32_events);
5820 events = swahw32(events);
5822 if (!(flags & IORING_POLL_ADD_MULTI))
5823 events |= EPOLLONESHOT;
5824 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5827 static int io_poll_update_prep(struct io_kiocb *req,
5828 const struct io_uring_sqe *sqe)
5830 struct io_poll_update *upd = &req->poll_update;
5833 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5835 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5837 flags = READ_ONCE(sqe->len);
5838 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5839 IORING_POLL_ADD_MULTI))
5841 /* meaningless without update */
5842 if (flags == IORING_POLL_ADD_MULTI)
5845 upd->old_user_data = READ_ONCE(sqe->addr);
5846 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5847 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5849 upd->new_user_data = READ_ONCE(sqe->off);
5850 if (!upd->update_user_data && upd->new_user_data)
5852 if (upd->update_events)
5853 upd->events = io_poll_parse_events(sqe, flags);
5854 else if (sqe->poll32_events)
5860 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5863 struct io_kiocb *req = wait->private;
5864 struct io_poll_iocb *poll = &req->poll;
5866 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5869 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5870 struct poll_table_struct *p)
5872 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5874 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5877 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5879 struct io_poll_iocb *poll = &req->poll;
5882 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5884 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5886 flags = READ_ONCE(sqe->len);
5887 if (flags & ~IORING_POLL_ADD_MULTI)
5890 io_req_set_refcount(req);
5891 poll->events = io_poll_parse_events(sqe, flags);
5895 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5897 struct io_poll_iocb *poll = &req->poll;
5898 struct io_ring_ctx *ctx = req->ctx;
5899 struct io_poll_table ipt;
5903 ipt.pt._qproc = io_poll_queue_proc;
5905 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5908 if (mask) { /* no async, we'd stolen it */
5910 done = io_poll_complete(req, mask);
5912 spin_unlock(&ctx->completion_lock);
5915 io_cqring_ev_posted(ctx);
5922 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5924 struct io_ring_ctx *ctx = req->ctx;
5925 struct io_kiocb *preq;
5929 spin_lock(&ctx->completion_lock);
5930 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5936 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5938 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5943 * Don't allow racy completion with singleshot, as we cannot safely
5944 * update those. For multishot, if we're racing with completion, just
5945 * let completion re-add it.
5947 io_poll_remove_double(preq);
5948 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5949 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5953 /* we now have a detached poll request. reissue. */
5957 spin_unlock(&ctx->completion_lock);
5959 io_req_complete(req, ret);
5962 /* only mask one event flags, keep behavior flags */
5963 if (req->poll_update.update_events) {
5964 preq->poll.events &= ~0xffff;
5965 preq->poll.events |= req->poll_update.events & 0xffff;
5966 preq->poll.events |= IO_POLL_UNMASK;
5968 if (req->poll_update.update_user_data)
5969 preq->user_data = req->poll_update.new_user_data;
5970 spin_unlock(&ctx->completion_lock);
5972 /* complete update request, we're done with it */
5973 io_req_complete(req, ret);
5976 ret = io_poll_add(preq, issue_flags);
5979 io_req_complete(preq, ret);
5985 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5988 io_req_complete_post(req, -ETIME, 0);
5991 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5993 struct io_timeout_data *data = container_of(timer,
5994 struct io_timeout_data, timer);
5995 struct io_kiocb *req = data->req;
5996 struct io_ring_ctx *ctx = req->ctx;
5997 unsigned long flags;
5999 spin_lock_irqsave(&ctx->timeout_lock, flags);
6000 list_del_init(&req->timeout.list);
6001 atomic_set(&req->ctx->cq_timeouts,
6002 atomic_read(&req->ctx->cq_timeouts) + 1);
6003 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6005 req->io_task_work.func = io_req_task_timeout;
6006 io_req_task_work_add(req);
6007 return HRTIMER_NORESTART;
6010 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
6012 __must_hold(&ctx->timeout_lock)
6014 struct io_timeout_data *io;
6015 struct io_kiocb *req;
6018 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
6019 found = user_data == req->user_data;
6024 return ERR_PTR(-ENOENT);
6026 io = req->async_data;
6027 if (hrtimer_try_to_cancel(&io->timer) == -1)
6028 return ERR_PTR(-EALREADY);
6029 list_del_init(&req->timeout.list);
6033 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
6034 __must_hold(&ctx->completion_lock)
6035 __must_hold(&ctx->timeout_lock)
6037 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6040 return PTR_ERR(req);
6043 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
6044 io_put_req_deferred(req);
6048 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6050 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6051 case IORING_TIMEOUT_BOOTTIME:
6052 return CLOCK_BOOTTIME;
6053 case IORING_TIMEOUT_REALTIME:
6054 return CLOCK_REALTIME;
6056 /* can't happen, vetted at prep time */
6060 return CLOCK_MONOTONIC;
6064 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6065 struct timespec64 *ts, enum hrtimer_mode mode)
6066 __must_hold(&ctx->timeout_lock)
6068 struct io_timeout_data *io;
6069 struct io_kiocb *req;
6072 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6073 found = user_data == req->user_data;
6080 io = req->async_data;
6081 if (hrtimer_try_to_cancel(&io->timer) == -1)
6083 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6084 io->timer.function = io_link_timeout_fn;
6085 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6089 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6090 struct timespec64 *ts, enum hrtimer_mode mode)
6091 __must_hold(&ctx->timeout_lock)
6093 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6094 struct io_timeout_data *data;
6097 return PTR_ERR(req);
6099 req->timeout.off = 0; /* noseq */
6100 data = req->async_data;
6101 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6102 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6103 data->timer.function = io_timeout_fn;
6104 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6108 static int io_timeout_remove_prep(struct io_kiocb *req,
6109 const struct io_uring_sqe *sqe)
6111 struct io_timeout_rem *tr = &req->timeout_rem;
6113 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6115 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6117 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6120 tr->ltimeout = false;
6121 tr->addr = READ_ONCE(sqe->addr);
6122 tr->flags = READ_ONCE(sqe->timeout_flags);
6123 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6124 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6126 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6127 tr->ltimeout = true;
6128 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6130 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6132 } else if (tr->flags) {
6133 /* timeout removal doesn't support flags */
6140 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6142 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6147 * Remove or update an existing timeout command
6149 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6151 struct io_timeout_rem *tr = &req->timeout_rem;
6152 struct io_ring_ctx *ctx = req->ctx;
6155 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6156 spin_lock(&ctx->completion_lock);
6157 spin_lock_irq(&ctx->timeout_lock);
6158 ret = io_timeout_cancel(ctx, tr->addr);
6159 spin_unlock_irq(&ctx->timeout_lock);
6160 spin_unlock(&ctx->completion_lock);
6162 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6164 spin_lock_irq(&ctx->timeout_lock);
6166 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6168 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6169 spin_unlock_irq(&ctx->timeout_lock);
6174 io_req_complete_post(req, ret, 0);
6178 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6179 bool is_timeout_link)
6181 struct io_timeout_data *data;
6183 u32 off = READ_ONCE(sqe->off);
6185 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6187 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6190 if (off && is_timeout_link)
6192 flags = READ_ONCE(sqe->timeout_flags);
6193 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6195 /* more than one clock specified is invalid, obviously */
6196 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6199 INIT_LIST_HEAD(&req->timeout.list);
6200 req->timeout.off = off;
6201 if (unlikely(off && !req->ctx->off_timeout_used))
6202 req->ctx->off_timeout_used = true;
6204 if (!req->async_data && io_alloc_async_data(req))
6207 data = req->async_data;
6209 data->flags = flags;
6211 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6214 INIT_LIST_HEAD(&req->timeout.list);
6215 data->mode = io_translate_timeout_mode(flags);
6216 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6218 if (is_timeout_link) {
6219 struct io_submit_link *link = &req->ctx->submit_state.link;
6223 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6225 req->timeout.head = link->last;
6226 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6231 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6233 struct io_ring_ctx *ctx = req->ctx;
6234 struct io_timeout_data *data = req->async_data;
6235 struct list_head *entry;
6236 u32 tail, off = req->timeout.off;
6238 spin_lock_irq(&ctx->timeout_lock);
6241 * sqe->off holds how many events that need to occur for this
6242 * timeout event to be satisfied. If it isn't set, then this is
6243 * a pure timeout request, sequence isn't used.
6245 if (io_is_timeout_noseq(req)) {
6246 entry = ctx->timeout_list.prev;
6250 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6251 req->timeout.target_seq = tail + off;
6253 /* Update the last seq here in case io_flush_timeouts() hasn't.
6254 * This is safe because ->completion_lock is held, and submissions
6255 * and completions are never mixed in the same ->completion_lock section.
6257 ctx->cq_last_tm_flush = tail;
6260 * Insertion sort, ensuring the first entry in the list is always
6261 * the one we need first.
6263 list_for_each_prev(entry, &ctx->timeout_list) {
6264 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6267 if (io_is_timeout_noseq(nxt))
6269 /* nxt.seq is behind @tail, otherwise would've been completed */
6270 if (off >= nxt->timeout.target_seq - tail)
6274 list_add(&req->timeout.list, entry);
6275 data->timer.function = io_timeout_fn;
6276 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6277 spin_unlock_irq(&ctx->timeout_lock);
6281 struct io_cancel_data {
6282 struct io_ring_ctx *ctx;
6286 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6288 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6289 struct io_cancel_data *cd = data;
6291 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6294 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6295 struct io_ring_ctx *ctx)
6297 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6298 enum io_wq_cancel cancel_ret;
6301 if (!tctx || !tctx->io_wq)
6304 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6305 switch (cancel_ret) {
6306 case IO_WQ_CANCEL_OK:
6309 case IO_WQ_CANCEL_RUNNING:
6312 case IO_WQ_CANCEL_NOTFOUND:
6320 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6322 struct io_ring_ctx *ctx = req->ctx;
6325 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6327 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6331 spin_lock(&ctx->completion_lock);
6332 spin_lock_irq(&ctx->timeout_lock);
6333 ret = io_timeout_cancel(ctx, sqe_addr);
6334 spin_unlock_irq(&ctx->timeout_lock);
6337 ret = io_poll_cancel(ctx, sqe_addr, false);
6339 spin_unlock(&ctx->completion_lock);
6343 static int io_async_cancel_prep(struct io_kiocb *req,
6344 const struct io_uring_sqe *sqe)
6346 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6348 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6350 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6354 req->cancel.addr = READ_ONCE(sqe->addr);
6358 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6360 struct io_ring_ctx *ctx = req->ctx;
6361 u64 sqe_addr = req->cancel.addr;
6362 struct io_tctx_node *node;
6365 ret = io_try_cancel_userdata(req, sqe_addr);
6369 /* slow path, try all io-wq's */
6370 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6372 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6373 struct io_uring_task *tctx = node->task->io_uring;
6375 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6379 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6383 io_req_complete_post(req, ret, 0);
6387 static int io_rsrc_update_prep(struct io_kiocb *req,
6388 const struct io_uring_sqe *sqe)
6390 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6392 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6395 req->rsrc_update.offset = READ_ONCE(sqe->off);
6396 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6397 if (!req->rsrc_update.nr_args)
6399 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6403 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6405 struct io_ring_ctx *ctx = req->ctx;
6406 struct io_uring_rsrc_update2 up;
6409 up.offset = req->rsrc_update.offset;
6410 up.data = req->rsrc_update.arg;
6416 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6417 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6418 &up, req->rsrc_update.nr_args);
6419 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6423 __io_req_complete(req, issue_flags, ret, 0);
6427 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6429 switch (req->opcode) {
6432 case IORING_OP_READV:
6433 case IORING_OP_READ_FIXED:
6434 case IORING_OP_READ:
6435 return io_read_prep(req, sqe);
6436 case IORING_OP_WRITEV:
6437 case IORING_OP_WRITE_FIXED:
6438 case IORING_OP_WRITE:
6439 return io_write_prep(req, sqe);
6440 case IORING_OP_POLL_ADD:
6441 return io_poll_add_prep(req, sqe);
6442 case IORING_OP_POLL_REMOVE:
6443 return io_poll_update_prep(req, sqe);
6444 case IORING_OP_FSYNC:
6445 return io_fsync_prep(req, sqe);
6446 case IORING_OP_SYNC_FILE_RANGE:
6447 return io_sfr_prep(req, sqe);
6448 case IORING_OP_SENDMSG:
6449 case IORING_OP_SEND:
6450 return io_sendmsg_prep(req, sqe);
6451 case IORING_OP_RECVMSG:
6452 case IORING_OP_RECV:
6453 return io_recvmsg_prep(req, sqe);
6454 case IORING_OP_CONNECT:
6455 return io_connect_prep(req, sqe);
6456 case IORING_OP_TIMEOUT:
6457 return io_timeout_prep(req, sqe, false);
6458 case IORING_OP_TIMEOUT_REMOVE:
6459 return io_timeout_remove_prep(req, sqe);
6460 case IORING_OP_ASYNC_CANCEL:
6461 return io_async_cancel_prep(req, sqe);
6462 case IORING_OP_LINK_TIMEOUT:
6463 return io_timeout_prep(req, sqe, true);
6464 case IORING_OP_ACCEPT:
6465 return io_accept_prep(req, sqe);
6466 case IORING_OP_FALLOCATE:
6467 return io_fallocate_prep(req, sqe);
6468 case IORING_OP_OPENAT:
6469 return io_openat_prep(req, sqe);
6470 case IORING_OP_CLOSE:
6471 return io_close_prep(req, sqe);
6472 case IORING_OP_FILES_UPDATE:
6473 return io_rsrc_update_prep(req, sqe);
6474 case IORING_OP_STATX:
6475 return io_statx_prep(req, sqe);
6476 case IORING_OP_FADVISE:
6477 return io_fadvise_prep(req, sqe);
6478 case IORING_OP_MADVISE:
6479 return io_madvise_prep(req, sqe);
6480 case IORING_OP_OPENAT2:
6481 return io_openat2_prep(req, sqe);
6482 case IORING_OP_EPOLL_CTL:
6483 return io_epoll_ctl_prep(req, sqe);
6484 case IORING_OP_SPLICE:
6485 return io_splice_prep(req, sqe);
6486 case IORING_OP_PROVIDE_BUFFERS:
6487 return io_provide_buffers_prep(req, sqe);
6488 case IORING_OP_REMOVE_BUFFERS:
6489 return io_remove_buffers_prep(req, sqe);
6491 return io_tee_prep(req, sqe);
6492 case IORING_OP_SHUTDOWN:
6493 return io_shutdown_prep(req, sqe);
6494 case IORING_OP_RENAMEAT:
6495 return io_renameat_prep(req, sqe);
6496 case IORING_OP_UNLINKAT:
6497 return io_unlinkat_prep(req, sqe);
6498 case IORING_OP_MKDIRAT:
6499 return io_mkdirat_prep(req, sqe);
6500 case IORING_OP_SYMLINKAT:
6501 return io_symlinkat_prep(req, sqe);
6502 case IORING_OP_LINKAT:
6503 return io_linkat_prep(req, sqe);
6506 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6511 static int io_req_prep_async(struct io_kiocb *req)
6513 if (!io_op_defs[req->opcode].needs_async_setup)
6515 if (WARN_ON_ONCE(req->async_data))
6517 if (io_alloc_async_data(req))
6520 switch (req->opcode) {
6521 case IORING_OP_READV:
6522 return io_rw_prep_async(req, READ);
6523 case IORING_OP_WRITEV:
6524 return io_rw_prep_async(req, WRITE);
6525 case IORING_OP_SENDMSG:
6526 return io_sendmsg_prep_async(req);
6527 case IORING_OP_RECVMSG:
6528 return io_recvmsg_prep_async(req);
6529 case IORING_OP_CONNECT:
6530 return io_connect_prep_async(req);
6532 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6537 static u32 io_get_sequence(struct io_kiocb *req)
6539 u32 seq = req->ctx->cached_sq_head;
6541 /* need original cached_sq_head, but it was increased for each req */
6542 io_for_each_link(req, req)
6547 static bool io_drain_req(struct io_kiocb *req)
6549 struct io_kiocb *pos;
6550 struct io_ring_ctx *ctx = req->ctx;
6551 struct io_defer_entry *de;
6555 if (req->flags & REQ_F_FAIL) {
6556 io_req_complete_fail_submit(req);
6561 * If we need to drain a request in the middle of a link, drain the
6562 * head request and the next request/link after the current link.
6563 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6564 * maintained for every request of our link.
6566 if (ctx->drain_next) {
6567 req->flags |= REQ_F_IO_DRAIN;
6568 ctx->drain_next = false;
6570 /* not interested in head, start from the first linked */
6571 io_for_each_link(pos, req->link) {
6572 if (pos->flags & REQ_F_IO_DRAIN) {
6573 ctx->drain_next = true;
6574 req->flags |= REQ_F_IO_DRAIN;
6579 /* Still need defer if there is pending req in defer list. */
6580 spin_lock(&ctx->completion_lock);
6581 if (likely(list_empty_careful(&ctx->defer_list) &&
6582 !(req->flags & REQ_F_IO_DRAIN))) {
6583 spin_unlock(&ctx->completion_lock);
6584 ctx->drain_active = false;
6587 spin_unlock(&ctx->completion_lock);
6589 seq = io_get_sequence(req);
6590 /* Still a chance to pass the sequence check */
6591 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6594 ret = io_req_prep_async(req);
6597 io_prep_async_link(req);
6598 de = kmalloc(sizeof(*de), GFP_KERNEL);
6602 io_req_complete_failed(req, ret);
6606 spin_lock(&ctx->completion_lock);
6607 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6608 spin_unlock(&ctx->completion_lock);
6610 io_queue_async_work(req, NULL);
6614 trace_io_uring_defer(ctx, req, req->user_data);
6617 list_add_tail(&de->list, &ctx->defer_list);
6618 spin_unlock(&ctx->completion_lock);
6622 static void io_clean_op(struct io_kiocb *req)
6624 if (req->flags & REQ_F_BUFFER_SELECTED) {
6625 switch (req->opcode) {
6626 case IORING_OP_READV:
6627 case IORING_OP_READ_FIXED:
6628 case IORING_OP_READ:
6629 kfree((void *)(unsigned long)req->rw.addr);
6631 case IORING_OP_RECVMSG:
6632 case IORING_OP_RECV:
6633 kfree(req->sr_msg.kbuf);
6638 if (req->flags & REQ_F_NEED_CLEANUP) {
6639 switch (req->opcode) {
6640 case IORING_OP_READV:
6641 case IORING_OP_READ_FIXED:
6642 case IORING_OP_READ:
6643 case IORING_OP_WRITEV:
6644 case IORING_OP_WRITE_FIXED:
6645 case IORING_OP_WRITE: {
6646 struct io_async_rw *io = req->async_data;
6648 kfree(io->free_iovec);
6651 case IORING_OP_RECVMSG:
6652 case IORING_OP_SENDMSG: {
6653 struct io_async_msghdr *io = req->async_data;
6655 kfree(io->free_iov);
6658 case IORING_OP_OPENAT:
6659 case IORING_OP_OPENAT2:
6660 if (req->open.filename)
6661 putname(req->open.filename);
6663 case IORING_OP_RENAMEAT:
6664 putname(req->rename.oldpath);
6665 putname(req->rename.newpath);
6667 case IORING_OP_UNLINKAT:
6668 putname(req->unlink.filename);
6670 case IORING_OP_MKDIRAT:
6671 putname(req->mkdir.filename);
6673 case IORING_OP_SYMLINKAT:
6674 putname(req->symlink.oldpath);
6675 putname(req->symlink.newpath);
6677 case IORING_OP_LINKAT:
6678 putname(req->hardlink.oldpath);
6679 putname(req->hardlink.newpath);
6683 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6684 kfree(req->apoll->double_poll);
6688 if (req->flags & REQ_F_INFLIGHT) {
6689 struct io_uring_task *tctx = req->task->io_uring;
6691 atomic_dec(&tctx->inflight_tracked);
6693 if (req->flags & REQ_F_CREDS)
6694 put_cred(req->creds);
6696 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6699 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6701 struct io_ring_ctx *ctx = req->ctx;
6702 const struct cred *creds = NULL;
6705 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6706 creds = override_creds(req->creds);
6708 switch (req->opcode) {
6710 ret = io_nop(req, issue_flags);
6712 case IORING_OP_READV:
6713 case IORING_OP_READ_FIXED:
6714 case IORING_OP_READ:
6715 ret = io_read(req, issue_flags);
6717 case IORING_OP_WRITEV:
6718 case IORING_OP_WRITE_FIXED:
6719 case IORING_OP_WRITE:
6720 ret = io_write(req, issue_flags);
6722 case IORING_OP_FSYNC:
6723 ret = io_fsync(req, issue_flags);
6725 case IORING_OP_POLL_ADD:
6726 ret = io_poll_add(req, issue_flags);
6728 case IORING_OP_POLL_REMOVE:
6729 ret = io_poll_update(req, issue_flags);
6731 case IORING_OP_SYNC_FILE_RANGE:
6732 ret = io_sync_file_range(req, issue_flags);
6734 case IORING_OP_SENDMSG:
6735 ret = io_sendmsg(req, issue_flags);
6737 case IORING_OP_SEND:
6738 ret = io_send(req, issue_flags);
6740 case IORING_OP_RECVMSG:
6741 ret = io_recvmsg(req, issue_flags);
6743 case IORING_OP_RECV:
6744 ret = io_recv(req, issue_flags);
6746 case IORING_OP_TIMEOUT:
6747 ret = io_timeout(req, issue_flags);
6749 case IORING_OP_TIMEOUT_REMOVE:
6750 ret = io_timeout_remove(req, issue_flags);
6752 case IORING_OP_ACCEPT:
6753 ret = io_accept(req, issue_flags);
6755 case IORING_OP_CONNECT:
6756 ret = io_connect(req, issue_flags);
6758 case IORING_OP_ASYNC_CANCEL:
6759 ret = io_async_cancel(req, issue_flags);
6761 case IORING_OP_FALLOCATE:
6762 ret = io_fallocate(req, issue_flags);
6764 case IORING_OP_OPENAT:
6765 ret = io_openat(req, issue_flags);
6767 case IORING_OP_CLOSE:
6768 ret = io_close(req, issue_flags);
6770 case IORING_OP_FILES_UPDATE:
6771 ret = io_files_update(req, issue_flags);
6773 case IORING_OP_STATX:
6774 ret = io_statx(req, issue_flags);
6776 case IORING_OP_FADVISE:
6777 ret = io_fadvise(req, issue_flags);
6779 case IORING_OP_MADVISE:
6780 ret = io_madvise(req, issue_flags);
6782 case IORING_OP_OPENAT2:
6783 ret = io_openat2(req, issue_flags);
6785 case IORING_OP_EPOLL_CTL:
6786 ret = io_epoll_ctl(req, issue_flags);
6788 case IORING_OP_SPLICE:
6789 ret = io_splice(req, issue_flags);
6791 case IORING_OP_PROVIDE_BUFFERS:
6792 ret = io_provide_buffers(req, issue_flags);
6794 case IORING_OP_REMOVE_BUFFERS:
6795 ret = io_remove_buffers(req, issue_flags);
6798 ret = io_tee(req, issue_flags);
6800 case IORING_OP_SHUTDOWN:
6801 ret = io_shutdown(req, issue_flags);
6803 case IORING_OP_RENAMEAT:
6804 ret = io_renameat(req, issue_flags);
6806 case IORING_OP_UNLINKAT:
6807 ret = io_unlinkat(req, issue_flags);
6809 case IORING_OP_MKDIRAT:
6810 ret = io_mkdirat(req, issue_flags);
6812 case IORING_OP_SYMLINKAT:
6813 ret = io_symlinkat(req, issue_flags);
6815 case IORING_OP_LINKAT:
6816 ret = io_linkat(req, issue_flags);
6824 revert_creds(creds);
6827 /* If the op doesn't have a file, we're not polling for it */
6828 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6829 io_iopoll_req_issued(req);
6834 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6836 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6838 req = io_put_req_find_next(req);
6839 return req ? &req->work : NULL;
6842 static void io_wq_submit_work(struct io_wq_work *work)
6844 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6845 struct io_kiocb *timeout;
6848 /* one will be dropped by ->io_free_work() after returning to io-wq */
6849 if (!(req->flags & REQ_F_REFCOUNT))
6850 __io_req_set_refcount(req, 2);
6854 timeout = io_prep_linked_timeout(req);
6856 io_queue_linked_timeout(timeout);
6858 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6859 if (work->flags & IO_WQ_WORK_CANCEL)
6864 ret = io_issue_sqe(req, 0);
6866 * We can get EAGAIN for polled IO even though we're
6867 * forcing a sync submission from here, since we can't
6868 * wait for request slots on the block side.
6870 if (ret != -EAGAIN || !(req->ctx->flags & IORING_SETUP_IOPOLL))
6876 /* avoid locking problems by failing it from a clean context */
6878 io_req_task_queue_fail(req, ret);
6881 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6884 return &table->files[i];
6887 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6890 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6892 return (struct file *) (slot->file_ptr & FFS_MASK);
6895 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6897 unsigned long file_ptr = (unsigned long) file;
6899 if (__io_file_supports_nowait(file, READ))
6900 file_ptr |= FFS_ASYNC_READ;
6901 if (__io_file_supports_nowait(file, WRITE))
6902 file_ptr |= FFS_ASYNC_WRITE;
6903 if (S_ISREG(file_inode(file)->i_mode))
6904 file_ptr |= FFS_ISREG;
6905 file_slot->file_ptr = file_ptr;
6908 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6909 struct io_kiocb *req, int fd)
6912 unsigned long file_ptr;
6914 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6916 fd = array_index_nospec(fd, ctx->nr_user_files);
6917 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6918 file = (struct file *) (file_ptr & FFS_MASK);
6919 file_ptr &= ~FFS_MASK;
6920 /* mask in overlapping REQ_F and FFS bits */
6921 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6922 io_req_set_rsrc_node(req);
6926 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6927 struct io_kiocb *req, int fd)
6929 struct file *file = fget(fd);
6931 trace_io_uring_file_get(ctx, fd);
6933 /* we don't allow fixed io_uring files */
6934 if (file && unlikely(file->f_op == &io_uring_fops))
6935 io_req_track_inflight(req);
6939 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6940 struct io_kiocb *req, int fd, bool fixed)
6943 return io_file_get_fixed(ctx, req, fd);
6945 return io_file_get_normal(ctx, req, fd);
6948 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6950 struct io_kiocb *prev = req->timeout.prev;
6954 if (!(req->task->flags & PF_EXITING))
6955 ret = io_try_cancel_userdata(req, prev->user_data);
6956 io_req_complete_post(req, ret ?: -ETIME, 0);
6959 io_req_complete_post(req, -ETIME, 0);
6963 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6965 struct io_timeout_data *data = container_of(timer,
6966 struct io_timeout_data, timer);
6967 struct io_kiocb *prev, *req = data->req;
6968 struct io_ring_ctx *ctx = req->ctx;
6969 unsigned long flags;
6971 spin_lock_irqsave(&ctx->timeout_lock, flags);
6972 prev = req->timeout.head;
6973 req->timeout.head = NULL;
6976 * We don't expect the list to be empty, that will only happen if we
6977 * race with the completion of the linked work.
6980 io_remove_next_linked(prev);
6981 if (!req_ref_inc_not_zero(prev))
6984 list_del(&req->timeout.list);
6985 req->timeout.prev = prev;
6986 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6988 req->io_task_work.func = io_req_task_link_timeout;
6989 io_req_task_work_add(req);
6990 return HRTIMER_NORESTART;
6993 static void io_queue_linked_timeout(struct io_kiocb *req)
6995 struct io_ring_ctx *ctx = req->ctx;
6997 spin_lock_irq(&ctx->timeout_lock);
6999 * If the back reference is NULL, then our linked request finished
7000 * before we got a chance to setup the timer
7002 if (req->timeout.head) {
7003 struct io_timeout_data *data = req->async_data;
7005 data->timer.function = io_link_timeout_fn;
7006 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
7008 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
7010 spin_unlock_irq(&ctx->timeout_lock);
7011 /* drop submission reference */
7015 static void __io_queue_sqe(struct io_kiocb *req)
7016 __must_hold(&req->ctx->uring_lock)
7018 struct io_kiocb *linked_timeout;
7022 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
7025 * We async punt it if the file wasn't marked NOWAIT, or if the file
7026 * doesn't support non-blocking read/write attempts
7029 if (req->flags & REQ_F_COMPLETE_INLINE) {
7030 struct io_ring_ctx *ctx = req->ctx;
7031 struct io_submit_state *state = &ctx->submit_state;
7033 state->compl_reqs[state->compl_nr++] = req;
7034 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
7035 io_submit_flush_completions(ctx);
7039 linked_timeout = io_prep_linked_timeout(req);
7041 io_queue_linked_timeout(linked_timeout);
7042 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
7043 linked_timeout = io_prep_linked_timeout(req);
7045 switch (io_arm_poll_handler(req)) {
7046 case IO_APOLL_READY:
7048 io_queue_linked_timeout(linked_timeout);
7050 case IO_APOLL_ABORTED:
7052 * Queued up for async execution, worker will release
7053 * submit reference when the iocb is actually submitted.
7055 io_queue_async_work(req, NULL);
7060 io_queue_linked_timeout(linked_timeout);
7062 io_req_complete_failed(req, ret);
7066 static inline void io_queue_sqe(struct io_kiocb *req)
7067 __must_hold(&req->ctx->uring_lock)
7069 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7072 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7073 __io_queue_sqe(req);
7074 } else if (req->flags & REQ_F_FAIL) {
7075 io_req_complete_fail_submit(req);
7077 int ret = io_req_prep_async(req);
7080 io_req_complete_failed(req, ret);
7082 io_queue_async_work(req, NULL);
7087 * Check SQE restrictions (opcode and flags).
7089 * Returns 'true' if SQE is allowed, 'false' otherwise.
7091 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7092 struct io_kiocb *req,
7093 unsigned int sqe_flags)
7095 if (likely(!ctx->restricted))
7098 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7101 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7102 ctx->restrictions.sqe_flags_required)
7105 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7106 ctx->restrictions.sqe_flags_required))
7112 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7113 const struct io_uring_sqe *sqe)
7114 __must_hold(&ctx->uring_lock)
7116 struct io_submit_state *state;
7117 unsigned int sqe_flags;
7118 int personality, ret = 0;
7120 /* req is partially pre-initialised, see io_preinit_req() */
7121 req->opcode = READ_ONCE(sqe->opcode);
7122 /* same numerical values with corresponding REQ_F_*, safe to copy */
7123 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7124 req->user_data = READ_ONCE(sqe->user_data);
7126 req->fixed_rsrc_refs = NULL;
7127 req->task = current;
7129 /* enforce forwards compatibility on users */
7130 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7132 if (unlikely(req->opcode >= IORING_OP_LAST))
7134 if (!io_check_restriction(ctx, req, sqe_flags))
7137 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7138 !io_op_defs[req->opcode].buffer_select)
7140 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7141 ctx->drain_active = true;
7143 personality = READ_ONCE(sqe->personality);
7145 req->creds = xa_load(&ctx->personalities, personality);
7148 get_cred(req->creds);
7149 req->flags |= REQ_F_CREDS;
7151 state = &ctx->submit_state;
7154 * Plug now if we have more than 1 IO left after this, and the target
7155 * is potentially a read/write to block based storage.
7157 if (!state->plug_started && state->ios_left > 1 &&
7158 io_op_defs[req->opcode].plug) {
7159 blk_start_plug(&state->plug);
7160 state->plug_started = true;
7163 if (io_op_defs[req->opcode].needs_file) {
7164 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7165 (sqe_flags & IOSQE_FIXED_FILE));
7166 if (unlikely(!req->file))
7174 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7175 const struct io_uring_sqe *sqe)
7176 __must_hold(&ctx->uring_lock)
7178 struct io_submit_link *link = &ctx->submit_state.link;
7181 ret = io_init_req(ctx, req, sqe);
7182 if (unlikely(ret)) {
7184 /* fail even hard links since we don't submit */
7187 * we can judge a link req is failed or cancelled by if
7188 * REQ_F_FAIL is set, but the head is an exception since
7189 * it may be set REQ_F_FAIL because of other req's failure
7190 * so let's leverage req->result to distinguish if a head
7191 * is set REQ_F_FAIL because of its failure or other req's
7192 * failure so that we can set the correct ret code for it.
7193 * init result here to avoid affecting the normal path.
7195 if (!(link->head->flags & REQ_F_FAIL))
7196 req_fail_link_node(link->head, -ECANCELED);
7197 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7199 * the current req is a normal req, we should return
7200 * error and thus break the submittion loop.
7202 io_req_complete_failed(req, ret);
7205 req_fail_link_node(req, ret);
7207 ret = io_req_prep(req, sqe);
7212 /* don't need @sqe from now on */
7213 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7215 ctx->flags & IORING_SETUP_SQPOLL);
7218 * If we already have a head request, queue this one for async
7219 * submittal once the head completes. If we don't have a head but
7220 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7221 * submitted sync once the chain is complete. If none of those
7222 * conditions are true (normal request), then just queue it.
7225 struct io_kiocb *head = link->head;
7227 if (!(req->flags & REQ_F_FAIL)) {
7228 ret = io_req_prep_async(req);
7229 if (unlikely(ret)) {
7230 req_fail_link_node(req, ret);
7231 if (!(head->flags & REQ_F_FAIL))
7232 req_fail_link_node(head, -ECANCELED);
7235 trace_io_uring_link(ctx, req, head);
7236 link->last->link = req;
7239 /* last request of a link, enqueue the link */
7240 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7245 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7257 * Batched submission is done, ensure local IO is flushed out.
7259 static void io_submit_state_end(struct io_submit_state *state,
7260 struct io_ring_ctx *ctx)
7262 if (state->link.head)
7263 io_queue_sqe(state->link.head);
7264 if (state->compl_nr)
7265 io_submit_flush_completions(ctx);
7266 if (state->plug_started)
7267 blk_finish_plug(&state->plug);
7271 * Start submission side cache.
7273 static void io_submit_state_start(struct io_submit_state *state,
7274 unsigned int max_ios)
7276 state->plug_started = false;
7277 state->ios_left = max_ios;
7278 /* set only head, no need to init link_last in advance */
7279 state->link.head = NULL;
7282 static void io_commit_sqring(struct io_ring_ctx *ctx)
7284 struct io_rings *rings = ctx->rings;
7287 * Ensure any loads from the SQEs are done at this point,
7288 * since once we write the new head, the application could
7289 * write new data to them.
7291 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7295 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7296 * that is mapped by userspace. This means that care needs to be taken to
7297 * ensure that reads are stable, as we cannot rely on userspace always
7298 * being a good citizen. If members of the sqe are validated and then later
7299 * used, it's important that those reads are done through READ_ONCE() to
7300 * prevent a re-load down the line.
7302 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7304 unsigned head, mask = ctx->sq_entries - 1;
7305 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7308 * The cached sq head (or cq tail) serves two purposes:
7310 * 1) allows us to batch the cost of updating the user visible
7312 * 2) allows the kernel side to track the head on its own, even
7313 * though the application is the one updating it.
7315 head = READ_ONCE(ctx->sq_array[sq_idx]);
7316 if (likely(head < ctx->sq_entries))
7317 return &ctx->sq_sqes[head];
7319 /* drop invalid entries */
7321 WRITE_ONCE(ctx->rings->sq_dropped,
7322 READ_ONCE(ctx->rings->sq_dropped) + 1);
7326 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7327 __must_hold(&ctx->uring_lock)
7331 /* make sure SQ entry isn't read before tail */
7332 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7333 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7335 io_get_task_refs(nr);
7337 io_submit_state_start(&ctx->submit_state, nr);
7338 while (submitted < nr) {
7339 const struct io_uring_sqe *sqe;
7340 struct io_kiocb *req;
7342 req = io_alloc_req(ctx);
7343 if (unlikely(!req)) {
7345 submitted = -EAGAIN;
7348 sqe = io_get_sqe(ctx);
7349 if (unlikely(!sqe)) {
7350 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7353 /* will complete beyond this point, count as submitted */
7355 if (io_submit_sqe(ctx, req, sqe))
7359 if (unlikely(submitted != nr)) {
7360 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7361 int unused = nr - ref_used;
7363 current->io_uring->cached_refs += unused;
7364 percpu_ref_put_many(&ctx->refs, unused);
7367 io_submit_state_end(&ctx->submit_state, ctx);
7368 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7369 io_commit_sqring(ctx);
7374 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7376 return READ_ONCE(sqd->state);
7379 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7381 /* Tell userspace we may need a wakeup call */
7382 spin_lock(&ctx->completion_lock);
7383 WRITE_ONCE(ctx->rings->sq_flags,
7384 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7385 spin_unlock(&ctx->completion_lock);
7388 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7390 spin_lock(&ctx->completion_lock);
7391 WRITE_ONCE(ctx->rings->sq_flags,
7392 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7393 spin_unlock(&ctx->completion_lock);
7396 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7398 unsigned int to_submit;
7401 to_submit = io_sqring_entries(ctx);
7402 /* if we're handling multiple rings, cap submit size for fairness */
7403 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7404 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7406 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7407 unsigned nr_events = 0;
7408 const struct cred *creds = NULL;
7410 if (ctx->sq_creds != current_cred())
7411 creds = override_creds(ctx->sq_creds);
7413 mutex_lock(&ctx->uring_lock);
7414 if (!list_empty(&ctx->iopoll_list))
7415 io_do_iopoll(ctx, &nr_events, 0);
7418 * Don't submit if refs are dying, good for io_uring_register(),
7419 * but also it is relied upon by io_ring_exit_work()
7421 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7422 !(ctx->flags & IORING_SETUP_R_DISABLED))
7423 ret = io_submit_sqes(ctx, to_submit);
7424 mutex_unlock(&ctx->uring_lock);
7426 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7427 wake_up(&ctx->sqo_sq_wait);
7429 revert_creds(creds);
7435 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7437 struct io_ring_ctx *ctx;
7438 unsigned sq_thread_idle = 0;
7440 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7441 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7442 sqd->sq_thread_idle = sq_thread_idle;
7445 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7447 bool did_sig = false;
7448 struct ksignal ksig;
7450 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7451 signal_pending(current)) {
7452 mutex_unlock(&sqd->lock);
7453 if (signal_pending(current))
7454 did_sig = get_signal(&ksig);
7456 mutex_lock(&sqd->lock);
7458 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7461 static int io_sq_thread(void *data)
7463 struct io_sq_data *sqd = data;
7464 struct io_ring_ctx *ctx;
7465 unsigned long timeout = 0;
7466 char buf[TASK_COMM_LEN];
7469 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7470 set_task_comm(current, buf);
7472 if (sqd->sq_cpu != -1)
7473 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7475 set_cpus_allowed_ptr(current, cpu_online_mask);
7476 current->flags |= PF_NO_SETAFFINITY;
7478 mutex_lock(&sqd->lock);
7480 bool cap_entries, sqt_spin = false;
7482 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7483 if (io_sqd_handle_event(sqd))
7485 timeout = jiffies + sqd->sq_thread_idle;
7488 cap_entries = !list_is_singular(&sqd->ctx_list);
7489 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7490 int ret = __io_sq_thread(ctx, cap_entries);
7492 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7495 if (io_run_task_work())
7498 if (sqt_spin || !time_after(jiffies, timeout)) {
7501 timeout = jiffies + sqd->sq_thread_idle;
7505 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7506 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7507 bool needs_sched = true;
7509 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7510 io_ring_set_wakeup_flag(ctx);
7512 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7513 !list_empty_careful(&ctx->iopoll_list)) {
7514 needs_sched = false;
7517 if (io_sqring_entries(ctx)) {
7518 needs_sched = false;
7524 mutex_unlock(&sqd->lock);
7526 mutex_lock(&sqd->lock);
7528 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7529 io_ring_clear_wakeup_flag(ctx);
7532 finish_wait(&sqd->wait, &wait);
7533 timeout = jiffies + sqd->sq_thread_idle;
7536 io_uring_cancel_generic(true, sqd);
7538 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7539 io_ring_set_wakeup_flag(ctx);
7541 mutex_unlock(&sqd->lock);
7543 complete(&sqd->exited);
7547 struct io_wait_queue {
7548 struct wait_queue_entry wq;
7549 struct io_ring_ctx *ctx;
7551 unsigned nr_timeouts;
7554 static inline bool io_should_wake(struct io_wait_queue *iowq)
7556 struct io_ring_ctx *ctx = iowq->ctx;
7557 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7560 * Wake up if we have enough events, or if a timeout occurred since we
7561 * started waiting. For timeouts, we always want to return to userspace,
7562 * regardless of event count.
7564 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7567 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7568 int wake_flags, void *key)
7570 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7574 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7575 * the task, and the next invocation will do it.
7577 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7578 return autoremove_wake_function(curr, mode, wake_flags, key);
7582 static int io_run_task_work_sig(void)
7584 if (io_run_task_work())
7586 if (!signal_pending(current))
7588 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7589 return -ERESTARTSYS;
7593 /* when returns >0, the caller should retry */
7594 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7595 struct io_wait_queue *iowq,
7600 /* make sure we run task_work before checking for signals */
7601 ret = io_run_task_work_sig();
7602 if (ret || io_should_wake(iowq))
7604 /* let the caller flush overflows, retry */
7605 if (test_bit(0, &ctx->check_cq_overflow))
7608 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
7614 * Wait until events become available, if we don't already have some. The
7615 * application must reap them itself, as they reside on the shared cq ring.
7617 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7618 const sigset_t __user *sig, size_t sigsz,
7619 struct __kernel_timespec __user *uts)
7621 struct io_wait_queue iowq;
7622 struct io_rings *rings = ctx->rings;
7623 ktime_t timeout = KTIME_MAX;
7627 io_cqring_overflow_flush(ctx);
7628 if (io_cqring_events(ctx) >= min_events)
7630 if (!io_run_task_work())
7635 struct timespec64 ts;
7637 if (get_timespec64(&ts, uts))
7639 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
7643 #ifdef CONFIG_COMPAT
7644 if (in_compat_syscall())
7645 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7649 ret = set_user_sigmask(sig, sigsz);
7655 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7656 iowq.wq.private = current;
7657 INIT_LIST_HEAD(&iowq.wq.entry);
7659 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7660 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7662 trace_io_uring_cqring_wait(ctx, min_events);
7664 /* if we can't even flush overflow, don't wait for more */
7665 if (!io_cqring_overflow_flush(ctx)) {
7669 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7670 TASK_INTERRUPTIBLE);
7671 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
7672 finish_wait(&ctx->cq_wait, &iowq.wq);
7676 restore_saved_sigmask_unless(ret == -EINTR);
7678 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7681 static void io_free_page_table(void **table, size_t size)
7683 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7685 for (i = 0; i < nr_tables; i++)
7690 static void **io_alloc_page_table(size_t size)
7692 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7693 size_t init_size = size;
7696 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7700 for (i = 0; i < nr_tables; i++) {
7701 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7703 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7705 io_free_page_table(table, init_size);
7713 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7715 percpu_ref_exit(&ref_node->refs);
7719 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7721 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7722 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7723 unsigned long flags;
7724 bool first_add = false;
7725 unsigned long delay = HZ;
7727 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7730 /* if we are mid-quiesce then do not delay */
7731 if (node->rsrc_data->quiesce)
7734 while (!list_empty(&ctx->rsrc_ref_list)) {
7735 node = list_first_entry(&ctx->rsrc_ref_list,
7736 struct io_rsrc_node, node);
7737 /* recycle ref nodes in order */
7740 list_del(&node->node);
7741 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7743 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7746 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7749 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7751 struct io_rsrc_node *ref_node;
7753 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7757 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7762 INIT_LIST_HEAD(&ref_node->node);
7763 INIT_LIST_HEAD(&ref_node->rsrc_list);
7764 ref_node->done = false;
7768 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7769 struct io_rsrc_data *data_to_kill)
7771 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7772 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7775 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7777 rsrc_node->rsrc_data = data_to_kill;
7778 spin_lock_irq(&ctx->rsrc_ref_lock);
7779 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7780 spin_unlock_irq(&ctx->rsrc_ref_lock);
7782 atomic_inc(&data_to_kill->refs);
7783 percpu_ref_kill(&rsrc_node->refs);
7784 ctx->rsrc_node = NULL;
7787 if (!ctx->rsrc_node) {
7788 ctx->rsrc_node = ctx->rsrc_backup_node;
7789 ctx->rsrc_backup_node = NULL;
7793 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7795 if (ctx->rsrc_backup_node)
7797 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7798 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7801 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7805 /* As we may drop ->uring_lock, other task may have started quiesce */
7809 data->quiesce = true;
7811 ret = io_rsrc_node_switch_start(ctx);
7814 io_rsrc_node_switch(ctx, data);
7816 /* kill initial ref, already quiesced if zero */
7817 if (atomic_dec_and_test(&data->refs))
7819 mutex_unlock(&ctx->uring_lock);
7820 flush_delayed_work(&ctx->rsrc_put_work);
7821 ret = wait_for_completion_interruptible(&data->done);
7823 mutex_lock(&ctx->uring_lock);
7824 if (atomic_read(&data->refs) > 0) {
7826 * it has been revived by another thread while
7829 mutex_unlock(&ctx->uring_lock);
7835 atomic_inc(&data->refs);
7836 /* wait for all works potentially completing data->done */
7837 flush_delayed_work(&ctx->rsrc_put_work);
7838 reinit_completion(&data->done);
7840 ret = io_run_task_work_sig();
7841 mutex_lock(&ctx->uring_lock);
7843 data->quiesce = false;
7848 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7850 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7851 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7853 return &data->tags[table_idx][off];
7856 static void io_rsrc_data_free(struct io_rsrc_data *data)
7858 size_t size = data->nr * sizeof(data->tags[0][0]);
7861 io_free_page_table((void **)data->tags, size);
7865 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7866 u64 __user *utags, unsigned nr,
7867 struct io_rsrc_data **pdata)
7869 struct io_rsrc_data *data;
7873 data = kzalloc(sizeof(*data), GFP_KERNEL);
7876 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7884 data->do_put = do_put;
7887 for (i = 0; i < nr; i++) {
7888 u64 *tag_slot = io_get_tag_slot(data, i);
7890 if (copy_from_user(tag_slot, &utags[i],
7896 atomic_set(&data->refs, 1);
7897 init_completion(&data->done);
7901 io_rsrc_data_free(data);
7905 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7907 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7908 GFP_KERNEL_ACCOUNT);
7909 return !!table->files;
7912 static void io_free_file_tables(struct io_file_table *table)
7914 kvfree(table->files);
7915 table->files = NULL;
7918 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7920 #if defined(CONFIG_UNIX)
7921 if (ctx->ring_sock) {
7922 struct sock *sock = ctx->ring_sock->sk;
7923 struct sk_buff *skb;
7925 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7931 for (i = 0; i < ctx->nr_user_files; i++) {
7934 file = io_file_from_index(ctx, i);
7939 io_free_file_tables(&ctx->file_table);
7940 io_rsrc_data_free(ctx->file_data);
7941 ctx->file_data = NULL;
7942 ctx->nr_user_files = 0;
7945 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7947 unsigned nr = ctx->nr_user_files;
7950 if (!ctx->file_data)
7954 * Quiesce may unlock ->uring_lock, and while it's not held
7955 * prevent new requests using the table.
7957 ctx->nr_user_files = 0;
7958 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7959 ctx->nr_user_files = nr;
7961 __io_sqe_files_unregister(ctx);
7965 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7966 __releases(&sqd->lock)
7968 WARN_ON_ONCE(sqd->thread == current);
7971 * Do the dance but not conditional clear_bit() because it'd race with
7972 * other threads incrementing park_pending and setting the bit.
7974 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7975 if (atomic_dec_return(&sqd->park_pending))
7976 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7977 mutex_unlock(&sqd->lock);
7980 static void io_sq_thread_park(struct io_sq_data *sqd)
7981 __acquires(&sqd->lock)
7983 WARN_ON_ONCE(sqd->thread == current);
7985 atomic_inc(&sqd->park_pending);
7986 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7987 mutex_lock(&sqd->lock);
7989 wake_up_process(sqd->thread);
7992 static void io_sq_thread_stop(struct io_sq_data *sqd)
7994 WARN_ON_ONCE(sqd->thread == current);
7995 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7997 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7998 mutex_lock(&sqd->lock);
8000 wake_up_process(sqd->thread);
8001 mutex_unlock(&sqd->lock);
8002 wait_for_completion(&sqd->exited);
8005 static void io_put_sq_data(struct io_sq_data *sqd)
8007 if (refcount_dec_and_test(&sqd->refs)) {
8008 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
8010 io_sq_thread_stop(sqd);
8015 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
8017 struct io_sq_data *sqd = ctx->sq_data;
8020 io_sq_thread_park(sqd);
8021 list_del_init(&ctx->sqd_list);
8022 io_sqd_update_thread_idle(sqd);
8023 io_sq_thread_unpark(sqd);
8025 io_put_sq_data(sqd);
8026 ctx->sq_data = NULL;
8030 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
8032 struct io_ring_ctx *ctx_attach;
8033 struct io_sq_data *sqd;
8036 f = fdget(p->wq_fd);
8038 return ERR_PTR(-ENXIO);
8039 if (f.file->f_op != &io_uring_fops) {
8041 return ERR_PTR(-EINVAL);
8044 ctx_attach = f.file->private_data;
8045 sqd = ctx_attach->sq_data;
8048 return ERR_PTR(-EINVAL);
8050 if (sqd->task_tgid != current->tgid) {
8052 return ERR_PTR(-EPERM);
8055 refcount_inc(&sqd->refs);
8060 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
8063 struct io_sq_data *sqd;
8066 if (p->flags & IORING_SETUP_ATTACH_WQ) {
8067 sqd = io_attach_sq_data(p);
8072 /* fall through for EPERM case, setup new sqd/task */
8073 if (PTR_ERR(sqd) != -EPERM)
8077 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8079 return ERR_PTR(-ENOMEM);
8081 atomic_set(&sqd->park_pending, 0);
8082 refcount_set(&sqd->refs, 1);
8083 INIT_LIST_HEAD(&sqd->ctx_list);
8084 mutex_init(&sqd->lock);
8085 init_waitqueue_head(&sqd->wait);
8086 init_completion(&sqd->exited);
8090 #if defined(CONFIG_UNIX)
8092 * Ensure the UNIX gc is aware of our file set, so we are certain that
8093 * the io_uring can be safely unregistered on process exit, even if we have
8094 * loops in the file referencing.
8096 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8098 struct sock *sk = ctx->ring_sock->sk;
8099 struct scm_fp_list *fpl;
8100 struct sk_buff *skb;
8103 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8107 skb = alloc_skb(0, GFP_KERNEL);
8116 fpl->user = get_uid(current_user());
8117 for (i = 0; i < nr; i++) {
8118 struct file *file = io_file_from_index(ctx, i + offset);
8122 fpl->fp[nr_files] = get_file(file);
8123 unix_inflight(fpl->user, fpl->fp[nr_files]);
8128 fpl->max = SCM_MAX_FD;
8129 fpl->count = nr_files;
8130 UNIXCB(skb).fp = fpl;
8131 skb->destructor = unix_destruct_scm;
8132 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8133 skb_queue_head(&sk->sk_receive_queue, skb);
8135 for (i = 0; i < nr; i++) {
8136 struct file *file = io_file_from_index(ctx, i + offset);
8143 free_uid(fpl->user);
8151 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8152 * causes regular reference counting to break down. We rely on the UNIX
8153 * garbage collection to take care of this problem for us.
8155 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8157 unsigned left, total;
8161 left = ctx->nr_user_files;
8163 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8165 ret = __io_sqe_files_scm(ctx, this_files, total);
8169 total += this_files;
8175 while (total < ctx->nr_user_files) {
8176 struct file *file = io_file_from_index(ctx, total);
8186 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8192 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8194 struct file *file = prsrc->file;
8195 #if defined(CONFIG_UNIX)
8196 struct sock *sock = ctx->ring_sock->sk;
8197 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8198 struct sk_buff *skb;
8201 __skb_queue_head_init(&list);
8204 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8205 * remove this entry and rearrange the file array.
8207 skb = skb_dequeue(head);
8209 struct scm_fp_list *fp;
8211 fp = UNIXCB(skb).fp;
8212 for (i = 0; i < fp->count; i++) {
8215 if (fp->fp[i] != file)
8218 unix_notinflight(fp->user, fp->fp[i]);
8219 left = fp->count - 1 - i;
8221 memmove(&fp->fp[i], &fp->fp[i + 1],
8222 left * sizeof(struct file *));
8229 __skb_queue_tail(&list, skb);
8239 __skb_queue_tail(&list, skb);
8241 skb = skb_dequeue(head);
8244 if (skb_peek(&list)) {
8245 spin_lock_irq(&head->lock);
8246 while ((skb = __skb_dequeue(&list)) != NULL)
8247 __skb_queue_tail(head, skb);
8248 spin_unlock_irq(&head->lock);
8255 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8257 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8258 struct io_ring_ctx *ctx = rsrc_data->ctx;
8259 struct io_rsrc_put *prsrc, *tmp;
8261 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8262 list_del(&prsrc->list);
8265 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8267 io_ring_submit_lock(ctx, lock_ring);
8268 spin_lock(&ctx->completion_lock);
8269 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8271 io_commit_cqring(ctx);
8272 spin_unlock(&ctx->completion_lock);
8273 io_cqring_ev_posted(ctx);
8274 io_ring_submit_unlock(ctx, lock_ring);
8277 rsrc_data->do_put(ctx, prsrc);
8281 io_rsrc_node_destroy(ref_node);
8282 if (atomic_dec_and_test(&rsrc_data->refs))
8283 complete(&rsrc_data->done);
8286 static void io_rsrc_put_work(struct work_struct *work)
8288 struct io_ring_ctx *ctx;
8289 struct llist_node *node;
8291 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8292 node = llist_del_all(&ctx->rsrc_put_llist);
8295 struct io_rsrc_node *ref_node;
8296 struct llist_node *next = node->next;
8298 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8299 __io_rsrc_put_work(ref_node);
8304 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8305 unsigned nr_args, u64 __user *tags)
8307 __s32 __user *fds = (__s32 __user *) arg;
8316 if (nr_args > IORING_MAX_FIXED_FILES)
8318 if (nr_args > rlimit(RLIMIT_NOFILE))
8320 ret = io_rsrc_node_switch_start(ctx);
8323 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8329 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8332 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8333 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8337 /* allow sparse sets */
8340 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8347 if (unlikely(!file))
8351 * Don't allow io_uring instances to be registered. If UNIX
8352 * isn't enabled, then this causes a reference cycle and this
8353 * instance can never get freed. If UNIX is enabled we'll
8354 * handle it just fine, but there's still no point in allowing
8355 * a ring fd as it doesn't support regular read/write anyway.
8357 if (file->f_op == &io_uring_fops) {
8361 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8364 ret = io_sqe_files_scm(ctx);
8366 __io_sqe_files_unregister(ctx);
8370 io_rsrc_node_switch(ctx, NULL);
8373 for (i = 0; i < ctx->nr_user_files; i++) {
8374 file = io_file_from_index(ctx, i);
8378 io_free_file_tables(&ctx->file_table);
8379 ctx->nr_user_files = 0;
8381 io_rsrc_data_free(ctx->file_data);
8382 ctx->file_data = NULL;
8386 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8389 #if defined(CONFIG_UNIX)
8390 struct sock *sock = ctx->ring_sock->sk;
8391 struct sk_buff_head *head = &sock->sk_receive_queue;
8392 struct sk_buff *skb;
8395 * See if we can merge this file into an existing skb SCM_RIGHTS
8396 * file set. If there's no room, fall back to allocating a new skb
8397 * and filling it in.
8399 spin_lock_irq(&head->lock);
8400 skb = skb_peek(head);
8402 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8404 if (fpl->count < SCM_MAX_FD) {
8405 __skb_unlink(skb, head);
8406 spin_unlock_irq(&head->lock);
8407 fpl->fp[fpl->count] = get_file(file);
8408 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8410 spin_lock_irq(&head->lock);
8411 __skb_queue_head(head, skb);
8416 spin_unlock_irq(&head->lock);
8423 return __io_sqe_files_scm(ctx, 1, index);
8429 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8430 struct io_rsrc_node *node, void *rsrc)
8432 u64 *tag_slot = io_get_tag_slot(data, idx);
8433 struct io_rsrc_put *prsrc;
8435 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8439 prsrc->tag = *tag_slot;
8442 list_add(&prsrc->list, &node->rsrc_list);
8446 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8447 unsigned int issue_flags, u32 slot_index)
8449 struct io_ring_ctx *ctx = req->ctx;
8450 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8451 bool needs_switch = false;
8452 struct io_fixed_file *file_slot;
8455 io_ring_submit_lock(ctx, !force_nonblock);
8456 if (file->f_op == &io_uring_fops)
8459 if (!ctx->file_data)
8462 if (slot_index >= ctx->nr_user_files)
8465 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8466 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8468 if (file_slot->file_ptr) {
8469 struct file *old_file;
8471 ret = io_rsrc_node_switch_start(ctx);
8475 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8476 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8477 ctx->rsrc_node, old_file);
8480 file_slot->file_ptr = 0;
8481 needs_switch = true;
8484 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8485 io_fixed_file_set(file_slot, file);
8486 ret = io_sqe_file_register(ctx, file, slot_index);
8488 file_slot->file_ptr = 0;
8495 io_rsrc_node_switch(ctx, ctx->file_data);
8496 io_ring_submit_unlock(ctx, !force_nonblock);
8502 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8504 unsigned int offset = req->close.file_slot - 1;
8505 struct io_ring_ctx *ctx = req->ctx;
8506 struct io_fixed_file *file_slot;
8510 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8512 if (unlikely(!ctx->file_data))
8515 if (offset >= ctx->nr_user_files)
8517 ret = io_rsrc_node_switch_start(ctx);
8521 offset = array_index_nospec(offset, ctx->nr_user_files);
8522 file_slot = io_fixed_file_slot(&ctx->file_table, offset);
8524 if (!file_slot->file_ptr)
8527 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8528 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8532 file_slot->file_ptr = 0;
8533 io_rsrc_node_switch(ctx, ctx->file_data);
8536 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8540 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8541 struct io_uring_rsrc_update2 *up,
8544 u64 __user *tags = u64_to_user_ptr(up->tags);
8545 __s32 __user *fds = u64_to_user_ptr(up->data);
8546 struct io_rsrc_data *data = ctx->file_data;
8547 struct io_fixed_file *file_slot;
8551 bool needs_switch = false;
8553 if (!ctx->file_data)
8555 if (up->offset + nr_args > ctx->nr_user_files)
8558 for (done = 0; done < nr_args; done++) {
8561 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8562 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8566 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8570 if (fd == IORING_REGISTER_FILES_SKIP)
8573 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8574 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8576 if (file_slot->file_ptr) {
8577 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8578 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
8581 file_slot->file_ptr = 0;
8582 needs_switch = true;
8591 * Don't allow io_uring instances to be registered. If
8592 * UNIX isn't enabled, then this causes a reference
8593 * cycle and this instance can never get freed. If UNIX
8594 * is enabled we'll handle it just fine, but there's
8595 * still no point in allowing a ring fd as it doesn't
8596 * support regular read/write anyway.
8598 if (file->f_op == &io_uring_fops) {
8603 *io_get_tag_slot(data, i) = tag;
8604 io_fixed_file_set(file_slot, file);
8605 err = io_sqe_file_register(ctx, file, i);
8607 file_slot->file_ptr = 0;
8615 io_rsrc_node_switch(ctx, data);
8616 return done ? done : err;
8619 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8620 struct task_struct *task)
8622 struct io_wq_hash *hash;
8623 struct io_wq_data data;
8624 unsigned int concurrency;
8626 mutex_lock(&ctx->uring_lock);
8627 hash = ctx->hash_map;
8629 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8631 mutex_unlock(&ctx->uring_lock);
8632 return ERR_PTR(-ENOMEM);
8634 refcount_set(&hash->refs, 1);
8635 init_waitqueue_head(&hash->wait);
8636 ctx->hash_map = hash;
8638 mutex_unlock(&ctx->uring_lock);
8642 data.free_work = io_wq_free_work;
8643 data.do_work = io_wq_submit_work;
8645 /* Do QD, or 4 * CPUS, whatever is smallest */
8646 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8648 return io_wq_create(concurrency, &data);
8651 static int io_uring_alloc_task_context(struct task_struct *task,
8652 struct io_ring_ctx *ctx)
8654 struct io_uring_task *tctx;
8657 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8658 if (unlikely(!tctx))
8661 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8662 if (unlikely(ret)) {
8667 tctx->io_wq = io_init_wq_offload(ctx, task);
8668 if (IS_ERR(tctx->io_wq)) {
8669 ret = PTR_ERR(tctx->io_wq);
8670 percpu_counter_destroy(&tctx->inflight);
8676 init_waitqueue_head(&tctx->wait);
8677 atomic_set(&tctx->in_idle, 0);
8678 atomic_set(&tctx->inflight_tracked, 0);
8679 task->io_uring = tctx;
8680 spin_lock_init(&tctx->task_lock);
8681 INIT_WQ_LIST(&tctx->task_list);
8682 init_task_work(&tctx->task_work, tctx_task_work);
8686 void __io_uring_free(struct task_struct *tsk)
8688 struct io_uring_task *tctx = tsk->io_uring;
8690 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8691 WARN_ON_ONCE(tctx->io_wq);
8692 WARN_ON_ONCE(tctx->cached_refs);
8694 percpu_counter_destroy(&tctx->inflight);
8696 tsk->io_uring = NULL;
8699 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8700 struct io_uring_params *p)
8704 /* Retain compatibility with failing for an invalid attach attempt */
8705 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8706 IORING_SETUP_ATTACH_WQ) {
8709 f = fdget(p->wq_fd);
8712 if (f.file->f_op != &io_uring_fops) {
8718 if (ctx->flags & IORING_SETUP_SQPOLL) {
8719 struct task_struct *tsk;
8720 struct io_sq_data *sqd;
8723 sqd = io_get_sq_data(p, &attached);
8729 ctx->sq_creds = get_current_cred();
8731 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8732 if (!ctx->sq_thread_idle)
8733 ctx->sq_thread_idle = HZ;
8735 io_sq_thread_park(sqd);
8736 list_add(&ctx->sqd_list, &sqd->ctx_list);
8737 io_sqd_update_thread_idle(sqd);
8738 /* don't attach to a dying SQPOLL thread, would be racy */
8739 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8740 io_sq_thread_unpark(sqd);
8747 if (p->flags & IORING_SETUP_SQ_AFF) {
8748 int cpu = p->sq_thread_cpu;
8751 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8758 sqd->task_pid = current->pid;
8759 sqd->task_tgid = current->tgid;
8760 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8767 ret = io_uring_alloc_task_context(tsk, ctx);
8768 wake_up_new_task(tsk);
8771 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8772 /* Can't have SQ_AFF without SQPOLL */
8779 complete(&ctx->sq_data->exited);
8781 io_sq_thread_finish(ctx);
8785 static inline void __io_unaccount_mem(struct user_struct *user,
8786 unsigned long nr_pages)
8788 atomic_long_sub(nr_pages, &user->locked_vm);
8791 static inline int __io_account_mem(struct user_struct *user,
8792 unsigned long nr_pages)
8794 unsigned long page_limit, cur_pages, new_pages;
8796 /* Don't allow more pages than we can safely lock */
8797 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8800 cur_pages = atomic_long_read(&user->locked_vm);
8801 new_pages = cur_pages + nr_pages;
8802 if (new_pages > page_limit)
8804 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8805 new_pages) != cur_pages);
8810 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8813 __io_unaccount_mem(ctx->user, nr_pages);
8815 if (ctx->mm_account)
8816 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8819 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8824 ret = __io_account_mem(ctx->user, nr_pages);
8829 if (ctx->mm_account)
8830 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8835 static void io_mem_free(void *ptr)
8842 page = virt_to_head_page(ptr);
8843 if (put_page_testzero(page))
8844 free_compound_page(page);
8847 static void *io_mem_alloc(size_t size)
8849 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
8851 return (void *) __get_free_pages(gfp, get_order(size));
8854 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8857 struct io_rings *rings;
8858 size_t off, sq_array_size;
8860 off = struct_size(rings, cqes, cq_entries);
8861 if (off == SIZE_MAX)
8865 off = ALIGN(off, SMP_CACHE_BYTES);
8873 sq_array_size = array_size(sizeof(u32), sq_entries);
8874 if (sq_array_size == SIZE_MAX)
8877 if (check_add_overflow(off, sq_array_size, &off))
8883 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8885 struct io_mapped_ubuf *imu = *slot;
8888 if (imu != ctx->dummy_ubuf) {
8889 for (i = 0; i < imu->nr_bvecs; i++)
8890 unpin_user_page(imu->bvec[i].bv_page);
8891 if (imu->acct_pages)
8892 io_unaccount_mem(ctx, imu->acct_pages);
8898 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8900 io_buffer_unmap(ctx, &prsrc->buf);
8904 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8908 for (i = 0; i < ctx->nr_user_bufs; i++)
8909 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8910 kfree(ctx->user_bufs);
8911 io_rsrc_data_free(ctx->buf_data);
8912 ctx->user_bufs = NULL;
8913 ctx->buf_data = NULL;
8914 ctx->nr_user_bufs = 0;
8917 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8919 unsigned nr = ctx->nr_user_bufs;
8926 * Quiesce may unlock ->uring_lock, and while it's not held
8927 * prevent new requests using the table.
8929 ctx->nr_user_bufs = 0;
8930 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8931 ctx->nr_user_bufs = nr;
8933 __io_sqe_buffers_unregister(ctx);
8937 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8938 void __user *arg, unsigned index)
8940 struct iovec __user *src;
8942 #ifdef CONFIG_COMPAT
8944 struct compat_iovec __user *ciovs;
8945 struct compat_iovec ciov;
8947 ciovs = (struct compat_iovec __user *) arg;
8948 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8951 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8952 dst->iov_len = ciov.iov_len;
8956 src = (struct iovec __user *) arg;
8957 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8963 * Not super efficient, but this is just a registration time. And we do cache
8964 * the last compound head, so generally we'll only do a full search if we don't
8967 * We check if the given compound head page has already been accounted, to
8968 * avoid double accounting it. This allows us to account the full size of the
8969 * page, not just the constituent pages of a huge page.
8971 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8972 int nr_pages, struct page *hpage)
8976 /* check current page array */
8977 for (i = 0; i < nr_pages; i++) {
8978 if (!PageCompound(pages[i]))
8980 if (compound_head(pages[i]) == hpage)
8984 /* check previously registered pages */
8985 for (i = 0; i < ctx->nr_user_bufs; i++) {
8986 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8988 for (j = 0; j < imu->nr_bvecs; j++) {
8989 if (!PageCompound(imu->bvec[j].bv_page))
8991 if (compound_head(imu->bvec[j].bv_page) == hpage)
8999 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
9000 int nr_pages, struct io_mapped_ubuf *imu,
9001 struct page **last_hpage)
9005 imu->acct_pages = 0;
9006 for (i = 0; i < nr_pages; i++) {
9007 if (!PageCompound(pages[i])) {
9012 hpage = compound_head(pages[i]);
9013 if (hpage == *last_hpage)
9015 *last_hpage = hpage;
9016 if (headpage_already_acct(ctx, pages, i, hpage))
9018 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
9022 if (!imu->acct_pages)
9025 ret = io_account_mem(ctx, imu->acct_pages);
9027 imu->acct_pages = 0;
9031 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
9032 struct io_mapped_ubuf **pimu,
9033 struct page **last_hpage)
9035 struct io_mapped_ubuf *imu = NULL;
9036 struct vm_area_struct **vmas = NULL;
9037 struct page **pages = NULL;
9038 unsigned long off, start, end, ubuf;
9040 int ret, pret, nr_pages, i;
9042 if (!iov->iov_base) {
9043 *pimu = ctx->dummy_ubuf;
9047 ubuf = (unsigned long) iov->iov_base;
9048 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9049 start = ubuf >> PAGE_SHIFT;
9050 nr_pages = end - start;
9055 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9059 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9064 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9069 mmap_read_lock(current->mm);
9070 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9072 if (pret == nr_pages) {
9073 /* don't support file backed memory */
9074 for (i = 0; i < nr_pages; i++) {
9075 struct vm_area_struct *vma = vmas[i];
9077 if (vma_is_shmem(vma))
9080 !is_file_hugepages(vma->vm_file)) {
9086 ret = pret < 0 ? pret : -EFAULT;
9088 mmap_read_unlock(current->mm);
9091 * if we did partial map, or found file backed vmas,
9092 * release any pages we did get
9095 unpin_user_pages(pages, pret);
9099 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9101 unpin_user_pages(pages, pret);
9105 off = ubuf & ~PAGE_MASK;
9106 size = iov->iov_len;
9107 for (i = 0; i < nr_pages; i++) {
9110 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9111 imu->bvec[i].bv_page = pages[i];
9112 imu->bvec[i].bv_len = vec_len;
9113 imu->bvec[i].bv_offset = off;
9117 /* store original address for later verification */
9119 imu->ubuf_end = ubuf + iov->iov_len;
9120 imu->nr_bvecs = nr_pages;
9131 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9133 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9134 return ctx->user_bufs ? 0 : -ENOMEM;
9137 static int io_buffer_validate(struct iovec *iov)
9139 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9142 * Don't impose further limits on the size and buffer
9143 * constraints here, we'll -EINVAL later when IO is
9144 * submitted if they are wrong.
9147 return iov->iov_len ? -EFAULT : 0;
9151 /* arbitrary limit, but we need something */
9152 if (iov->iov_len > SZ_1G)
9155 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9161 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9162 unsigned int nr_args, u64 __user *tags)
9164 struct page *last_hpage = NULL;
9165 struct io_rsrc_data *data;
9171 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9173 ret = io_rsrc_node_switch_start(ctx);
9176 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9179 ret = io_buffers_map_alloc(ctx, nr_args);
9181 io_rsrc_data_free(data);
9185 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9186 ret = io_copy_iov(ctx, &iov, arg, i);
9189 ret = io_buffer_validate(&iov);
9192 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9197 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9203 WARN_ON_ONCE(ctx->buf_data);
9205 ctx->buf_data = data;
9207 __io_sqe_buffers_unregister(ctx);
9209 io_rsrc_node_switch(ctx, NULL);
9213 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9214 struct io_uring_rsrc_update2 *up,
9215 unsigned int nr_args)
9217 u64 __user *tags = u64_to_user_ptr(up->tags);
9218 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9219 struct page *last_hpage = NULL;
9220 bool needs_switch = false;
9226 if (up->offset + nr_args > ctx->nr_user_bufs)
9229 for (done = 0; done < nr_args; done++) {
9230 struct io_mapped_ubuf *imu;
9231 int offset = up->offset + done;
9234 err = io_copy_iov(ctx, &iov, iovs, done);
9237 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9241 err = io_buffer_validate(&iov);
9244 if (!iov.iov_base && tag) {
9248 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9252 i = array_index_nospec(offset, ctx->nr_user_bufs);
9253 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9254 err = io_queue_rsrc_removal(ctx->buf_data, i,
9255 ctx->rsrc_node, ctx->user_bufs[i]);
9256 if (unlikely(err)) {
9257 io_buffer_unmap(ctx, &imu);
9260 ctx->user_bufs[i] = NULL;
9261 needs_switch = true;
9264 ctx->user_bufs[i] = imu;
9265 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9269 io_rsrc_node_switch(ctx, ctx->buf_data);
9270 return done ? done : err;
9273 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9275 __s32 __user *fds = arg;
9281 if (copy_from_user(&fd, fds, sizeof(*fds)))
9284 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9285 if (IS_ERR(ctx->cq_ev_fd)) {
9286 int ret = PTR_ERR(ctx->cq_ev_fd);
9288 ctx->cq_ev_fd = NULL;
9295 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9297 if (ctx->cq_ev_fd) {
9298 eventfd_ctx_put(ctx->cq_ev_fd);
9299 ctx->cq_ev_fd = NULL;
9306 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9308 struct io_buffer *buf;
9309 unsigned long index;
9311 xa_for_each(&ctx->io_buffers, index, buf)
9312 __io_remove_buffers(ctx, buf, index, -1U);
9315 static void io_req_cache_free(struct list_head *list)
9317 struct io_kiocb *req, *nxt;
9319 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9320 list_del(&req->inflight_entry);
9321 kmem_cache_free(req_cachep, req);
9325 static void io_req_caches_free(struct io_ring_ctx *ctx)
9327 struct io_submit_state *state = &ctx->submit_state;
9329 mutex_lock(&ctx->uring_lock);
9331 if (state->free_reqs) {
9332 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9333 state->free_reqs = 0;
9336 io_flush_cached_locked_reqs(ctx, state);
9337 io_req_cache_free(&state->free_list);
9338 mutex_unlock(&ctx->uring_lock);
9341 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9343 if (data && !atomic_dec_and_test(&data->refs))
9344 wait_for_completion(&data->done);
9347 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9349 io_sq_thread_finish(ctx);
9351 if (ctx->mm_account) {
9352 mmdrop(ctx->mm_account);
9353 ctx->mm_account = NULL;
9356 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9357 io_wait_rsrc_data(ctx->buf_data);
9358 io_wait_rsrc_data(ctx->file_data);
9360 mutex_lock(&ctx->uring_lock);
9362 __io_sqe_buffers_unregister(ctx);
9364 __io_sqe_files_unregister(ctx);
9366 __io_cqring_overflow_flush(ctx, true);
9367 mutex_unlock(&ctx->uring_lock);
9368 io_eventfd_unregister(ctx);
9369 io_destroy_buffers(ctx);
9371 put_cred(ctx->sq_creds);
9373 /* there are no registered resources left, nobody uses it */
9375 io_rsrc_node_destroy(ctx->rsrc_node);
9376 if (ctx->rsrc_backup_node)
9377 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9378 flush_delayed_work(&ctx->rsrc_put_work);
9380 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9381 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9383 #if defined(CONFIG_UNIX)
9384 if (ctx->ring_sock) {
9385 ctx->ring_sock->file = NULL; /* so that iput() is called */
9386 sock_release(ctx->ring_sock);
9389 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9391 io_mem_free(ctx->rings);
9392 io_mem_free(ctx->sq_sqes);
9394 percpu_ref_exit(&ctx->refs);
9395 free_uid(ctx->user);
9396 io_req_caches_free(ctx);
9398 io_wq_put_hash(ctx->hash_map);
9399 kfree(ctx->cancel_hash);
9400 kfree(ctx->dummy_ubuf);
9404 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9406 struct io_ring_ctx *ctx = file->private_data;
9409 poll_wait(file, &ctx->poll_wait, wait);
9411 * synchronizes with barrier from wq_has_sleeper call in
9415 if (!io_sqring_full(ctx))
9416 mask |= EPOLLOUT | EPOLLWRNORM;
9419 * Don't flush cqring overflow list here, just do a simple check.
9420 * Otherwise there could possible be ABBA deadlock:
9423 * lock(&ctx->uring_lock);
9425 * lock(&ctx->uring_lock);
9428 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9429 * pushs them to do the flush.
9431 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9432 mask |= EPOLLIN | EPOLLRDNORM;
9437 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9439 const struct cred *creds;
9441 creds = xa_erase(&ctx->personalities, id);
9450 struct io_tctx_exit {
9451 struct callback_head task_work;
9452 struct completion completion;
9453 struct io_ring_ctx *ctx;
9456 static void io_tctx_exit_cb(struct callback_head *cb)
9458 struct io_uring_task *tctx = current->io_uring;
9459 struct io_tctx_exit *work;
9461 work = container_of(cb, struct io_tctx_exit, task_work);
9463 * When @in_idle, we're in cancellation and it's racy to remove the
9464 * node. It'll be removed by the end of cancellation, just ignore it.
9466 if (!atomic_read(&tctx->in_idle))
9467 io_uring_del_tctx_node((unsigned long)work->ctx);
9468 complete(&work->completion);
9471 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9473 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9475 return req->ctx == data;
9478 static void io_ring_exit_work(struct work_struct *work)
9480 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9481 unsigned long timeout = jiffies + HZ * 60 * 5;
9482 unsigned long interval = HZ / 20;
9483 struct io_tctx_exit exit;
9484 struct io_tctx_node *node;
9488 * If we're doing polled IO and end up having requests being
9489 * submitted async (out-of-line), then completions can come in while
9490 * we're waiting for refs to drop. We need to reap these manually,
9491 * as nobody else will be looking for them.
9494 io_uring_try_cancel_requests(ctx, NULL, true);
9496 struct io_sq_data *sqd = ctx->sq_data;
9497 struct task_struct *tsk;
9499 io_sq_thread_park(sqd);
9501 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9502 io_wq_cancel_cb(tsk->io_uring->io_wq,
9503 io_cancel_ctx_cb, ctx, true);
9504 io_sq_thread_unpark(sqd);
9507 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9508 /* there is little hope left, don't run it too often */
9511 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9513 init_completion(&exit.completion);
9514 init_task_work(&exit.task_work, io_tctx_exit_cb);
9517 * Some may use context even when all refs and requests have been put,
9518 * and they are free to do so while still holding uring_lock or
9519 * completion_lock, see io_req_task_submit(). Apart from other work,
9520 * this lock/unlock section also waits them to finish.
9522 mutex_lock(&ctx->uring_lock);
9523 while (!list_empty(&ctx->tctx_list)) {
9524 WARN_ON_ONCE(time_after(jiffies, timeout));
9526 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9528 /* don't spin on a single task if cancellation failed */
9529 list_rotate_left(&ctx->tctx_list);
9530 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9531 if (WARN_ON_ONCE(ret))
9533 wake_up_process(node->task);
9535 mutex_unlock(&ctx->uring_lock);
9536 wait_for_completion(&exit.completion);
9537 mutex_lock(&ctx->uring_lock);
9539 mutex_unlock(&ctx->uring_lock);
9540 spin_lock(&ctx->completion_lock);
9541 spin_unlock(&ctx->completion_lock);
9543 io_ring_ctx_free(ctx);
9546 /* Returns true if we found and killed one or more timeouts */
9547 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9550 struct io_kiocb *req, *tmp;
9553 spin_lock(&ctx->completion_lock);
9554 spin_lock_irq(&ctx->timeout_lock);
9555 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9556 if (io_match_task(req, tsk, cancel_all)) {
9557 io_kill_timeout(req, -ECANCELED);
9561 spin_unlock_irq(&ctx->timeout_lock);
9563 io_commit_cqring(ctx);
9564 spin_unlock(&ctx->completion_lock);
9566 io_cqring_ev_posted(ctx);
9567 return canceled != 0;
9570 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9572 unsigned long index;
9573 struct creds *creds;
9575 mutex_lock(&ctx->uring_lock);
9576 percpu_ref_kill(&ctx->refs);
9578 __io_cqring_overflow_flush(ctx, true);
9579 xa_for_each(&ctx->personalities, index, creds)
9580 io_unregister_personality(ctx, index);
9581 mutex_unlock(&ctx->uring_lock);
9583 io_kill_timeouts(ctx, NULL, true);
9584 io_poll_remove_all(ctx, NULL, true);
9586 /* if we failed setting up the ctx, we might not have any rings */
9587 io_iopoll_try_reap_events(ctx);
9589 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9591 * Use system_unbound_wq to avoid spawning tons of event kworkers
9592 * if we're exiting a ton of rings at the same time. It just adds
9593 * noise and overhead, there's no discernable change in runtime
9594 * over using system_wq.
9596 queue_work(system_unbound_wq, &ctx->exit_work);
9599 static int io_uring_release(struct inode *inode, struct file *file)
9601 struct io_ring_ctx *ctx = file->private_data;
9603 file->private_data = NULL;
9604 io_ring_ctx_wait_and_kill(ctx);
9608 struct io_task_cancel {
9609 struct task_struct *task;
9613 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9615 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9616 struct io_task_cancel *cancel = data;
9618 return io_match_task_safe(req, cancel->task, cancel->all);
9621 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9622 struct task_struct *task, bool cancel_all)
9624 struct io_defer_entry *de;
9627 spin_lock(&ctx->completion_lock);
9628 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9629 if (io_match_task_safe(de->req, task, cancel_all)) {
9630 list_cut_position(&list, &ctx->defer_list, &de->list);
9634 spin_unlock(&ctx->completion_lock);
9635 if (list_empty(&list))
9638 while (!list_empty(&list)) {
9639 de = list_first_entry(&list, struct io_defer_entry, list);
9640 list_del_init(&de->list);
9641 io_req_complete_failed(de->req, -ECANCELED);
9647 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9649 struct io_tctx_node *node;
9650 enum io_wq_cancel cret;
9653 mutex_lock(&ctx->uring_lock);
9654 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9655 struct io_uring_task *tctx = node->task->io_uring;
9658 * io_wq will stay alive while we hold uring_lock, because it's
9659 * killed after ctx nodes, which requires to take the lock.
9661 if (!tctx || !tctx->io_wq)
9663 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9664 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9666 mutex_unlock(&ctx->uring_lock);
9671 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9672 struct task_struct *task,
9675 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9676 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9679 enum io_wq_cancel cret;
9683 ret |= io_uring_try_cancel_iowq(ctx);
9684 } else if (tctx && tctx->io_wq) {
9686 * Cancels requests of all rings, not only @ctx, but
9687 * it's fine as the task is in exit/exec.
9689 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9691 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9694 /* SQPOLL thread does its own polling */
9695 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9696 (ctx->sq_data && ctx->sq_data->thread == current)) {
9697 while (!list_empty_careful(&ctx->iopoll_list)) {
9698 io_iopoll_try_reap_events(ctx);
9703 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9704 ret |= io_poll_remove_all(ctx, task, cancel_all);
9705 ret |= io_kill_timeouts(ctx, task, cancel_all);
9707 ret |= io_run_task_work();
9714 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9716 struct io_uring_task *tctx = current->io_uring;
9717 struct io_tctx_node *node;
9720 if (unlikely(!tctx)) {
9721 ret = io_uring_alloc_task_context(current, ctx);
9725 tctx = current->io_uring;
9726 if (ctx->iowq_limits_set) {
9727 unsigned int limits[2] = { ctx->iowq_limits[0],
9728 ctx->iowq_limits[1], };
9730 ret = io_wq_max_workers(tctx->io_wq, limits);
9735 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9736 node = kmalloc(sizeof(*node), GFP_KERNEL);
9740 node->task = current;
9742 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9749 mutex_lock(&ctx->uring_lock);
9750 list_add(&node->ctx_node, &ctx->tctx_list);
9751 mutex_unlock(&ctx->uring_lock);
9758 * Note that this task has used io_uring. We use it for cancelation purposes.
9760 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9762 struct io_uring_task *tctx = current->io_uring;
9764 if (likely(tctx && tctx->last == ctx))
9766 return __io_uring_add_tctx_node(ctx);
9770 * Remove this io_uring_file -> task mapping.
9772 static void io_uring_del_tctx_node(unsigned long index)
9774 struct io_uring_task *tctx = current->io_uring;
9775 struct io_tctx_node *node;
9779 node = xa_erase(&tctx->xa, index);
9783 WARN_ON_ONCE(current != node->task);
9784 WARN_ON_ONCE(list_empty(&node->ctx_node));
9786 mutex_lock(&node->ctx->uring_lock);
9787 list_del(&node->ctx_node);
9788 mutex_unlock(&node->ctx->uring_lock);
9790 if (tctx->last == node->ctx)
9795 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9797 struct io_wq *wq = tctx->io_wq;
9798 struct io_tctx_node *node;
9799 unsigned long index;
9801 xa_for_each(&tctx->xa, index, node) {
9802 io_uring_del_tctx_node(index);
9807 * Must be after io_uring_del_task_file() (removes nodes under
9808 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9810 io_wq_put_and_exit(wq);
9815 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9818 return atomic_read(&tctx->inflight_tracked);
9819 return percpu_counter_sum(&tctx->inflight);
9823 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9824 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9826 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9828 struct io_uring_task *tctx = current->io_uring;
9829 struct io_ring_ctx *ctx;
9833 WARN_ON_ONCE(sqd && sqd->thread != current);
9835 if (!current->io_uring)
9838 io_wq_exit_start(tctx->io_wq);
9840 atomic_inc(&tctx->in_idle);
9842 io_uring_drop_tctx_refs(current);
9843 /* read completions before cancelations */
9844 inflight = tctx_inflight(tctx, !cancel_all);
9849 struct io_tctx_node *node;
9850 unsigned long index;
9852 xa_for_each(&tctx->xa, index, node) {
9853 /* sqpoll task will cancel all its requests */
9854 if (node->ctx->sq_data)
9856 io_uring_try_cancel_requests(node->ctx, current,
9860 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9861 io_uring_try_cancel_requests(ctx, current,
9865 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9867 io_uring_drop_tctx_refs(current);
9870 * If we've seen completions, retry without waiting. This
9871 * avoids a race where a completion comes in before we did
9872 * prepare_to_wait().
9874 if (inflight == tctx_inflight(tctx, !cancel_all))
9876 finish_wait(&tctx->wait, &wait);
9879 io_uring_clean_tctx(tctx);
9882 * We shouldn't run task_works after cancel, so just leave
9883 * ->in_idle set for normal exit.
9885 atomic_dec(&tctx->in_idle);
9886 /* for exec all current's requests should be gone, kill tctx */
9887 __io_uring_free(current);
9891 void __io_uring_cancel(bool cancel_all)
9893 io_uring_cancel_generic(cancel_all, NULL);
9896 static void *io_uring_validate_mmap_request(struct file *file,
9897 loff_t pgoff, size_t sz)
9899 struct io_ring_ctx *ctx = file->private_data;
9900 loff_t offset = pgoff << PAGE_SHIFT;
9905 case IORING_OFF_SQ_RING:
9906 case IORING_OFF_CQ_RING:
9909 case IORING_OFF_SQES:
9913 return ERR_PTR(-EINVAL);
9916 page = virt_to_head_page(ptr);
9917 if (sz > page_size(page))
9918 return ERR_PTR(-EINVAL);
9925 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9927 size_t sz = vma->vm_end - vma->vm_start;
9931 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9933 return PTR_ERR(ptr);
9935 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9936 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9939 #else /* !CONFIG_MMU */
9941 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9943 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9946 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9948 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9951 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9952 unsigned long addr, unsigned long len,
9953 unsigned long pgoff, unsigned long flags)
9957 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9959 return PTR_ERR(ptr);
9961 return (unsigned long) ptr;
9964 #endif /* !CONFIG_MMU */
9966 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9971 if (!io_sqring_full(ctx))
9973 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9975 if (!io_sqring_full(ctx))
9978 } while (!signal_pending(current));
9980 finish_wait(&ctx->sqo_sq_wait, &wait);
9984 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9985 struct __kernel_timespec __user **ts,
9986 const sigset_t __user **sig)
9988 struct io_uring_getevents_arg arg;
9991 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9992 * is just a pointer to the sigset_t.
9994 if (!(flags & IORING_ENTER_EXT_ARG)) {
9995 *sig = (const sigset_t __user *) argp;
10001 * EXT_ARG is set - ensure we agree on the size of it and copy in our
10002 * timespec and sigset_t pointers if good.
10004 if (*argsz != sizeof(arg))
10006 if (copy_from_user(&arg, argp, sizeof(arg)))
10010 *sig = u64_to_user_ptr(arg.sigmask);
10011 *argsz = arg.sigmask_sz;
10012 *ts = u64_to_user_ptr(arg.ts);
10016 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
10017 u32, min_complete, u32, flags, const void __user *, argp,
10020 struct io_ring_ctx *ctx;
10025 io_run_task_work();
10027 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
10028 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
10032 if (unlikely(!f.file))
10036 if (unlikely(f.file->f_op != &io_uring_fops))
10040 ctx = f.file->private_data;
10041 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
10045 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10049 * For SQ polling, the thread will do all submissions and completions.
10050 * Just return the requested submit count, and wake the thread if
10051 * we were asked to.
10054 if (ctx->flags & IORING_SETUP_SQPOLL) {
10055 io_cqring_overflow_flush(ctx);
10057 if (unlikely(ctx->sq_data->thread == NULL)) {
10061 if (flags & IORING_ENTER_SQ_WAKEUP)
10062 wake_up(&ctx->sq_data->wait);
10063 if (flags & IORING_ENTER_SQ_WAIT) {
10064 ret = io_sqpoll_wait_sq(ctx);
10068 submitted = to_submit;
10069 } else if (to_submit) {
10070 ret = io_uring_add_tctx_node(ctx);
10073 mutex_lock(&ctx->uring_lock);
10074 submitted = io_submit_sqes(ctx, to_submit);
10075 mutex_unlock(&ctx->uring_lock);
10077 if (submitted != to_submit)
10080 if (flags & IORING_ENTER_GETEVENTS) {
10081 const sigset_t __user *sig;
10082 struct __kernel_timespec __user *ts;
10084 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10088 min_complete = min(min_complete, ctx->cq_entries);
10091 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10092 * space applications don't need to do io completion events
10093 * polling again, they can rely on io_sq_thread to do polling
10094 * work, which can reduce cpu usage and uring_lock contention.
10096 if (ctx->flags & IORING_SETUP_IOPOLL &&
10097 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10098 ret = io_iopoll_check(ctx, min_complete);
10100 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10105 percpu_ref_put(&ctx->refs);
10108 return submitted ? submitted : ret;
10111 #ifdef CONFIG_PROC_FS
10112 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10113 const struct cred *cred)
10115 struct user_namespace *uns = seq_user_ns(m);
10116 struct group_info *gi;
10121 seq_printf(m, "%5d\n", id);
10122 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10123 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10124 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10125 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10126 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10127 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10128 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10129 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10130 seq_puts(m, "\n\tGroups:\t");
10131 gi = cred->group_info;
10132 for (g = 0; g < gi->ngroups; g++) {
10133 seq_put_decimal_ull(m, g ? " " : "",
10134 from_kgid_munged(uns, gi->gid[g]));
10136 seq_puts(m, "\n\tCapEff:\t");
10137 cap = cred->cap_effective;
10138 CAP_FOR_EACH_U32(__capi)
10139 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10144 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10146 struct io_sq_data *sq = NULL;
10151 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10152 * since fdinfo case grabs it in the opposite direction of normal use
10153 * cases. If we fail to get the lock, we just don't iterate any
10154 * structures that could be going away outside the io_uring mutex.
10156 has_lock = mutex_trylock(&ctx->uring_lock);
10158 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10164 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10165 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10166 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10167 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10168 struct file *f = io_file_from_index(ctx, i);
10171 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10173 seq_printf(m, "%5u: <none>\n", i);
10175 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10176 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10177 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10178 unsigned int len = buf->ubuf_end - buf->ubuf;
10180 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10182 if (has_lock && !xa_empty(&ctx->personalities)) {
10183 unsigned long index;
10184 const struct cred *cred;
10186 seq_printf(m, "Personalities:\n");
10187 xa_for_each(&ctx->personalities, index, cred)
10188 io_uring_show_cred(m, index, cred);
10190 seq_printf(m, "PollList:\n");
10191 spin_lock(&ctx->completion_lock);
10192 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10193 struct hlist_head *list = &ctx->cancel_hash[i];
10194 struct io_kiocb *req;
10196 hlist_for_each_entry(req, list, hash_node)
10197 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10198 req->task->task_works != NULL);
10200 spin_unlock(&ctx->completion_lock);
10202 mutex_unlock(&ctx->uring_lock);
10205 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10207 struct io_ring_ctx *ctx = f->private_data;
10209 if (percpu_ref_tryget(&ctx->refs)) {
10210 __io_uring_show_fdinfo(ctx, m);
10211 percpu_ref_put(&ctx->refs);
10216 static const struct file_operations io_uring_fops = {
10217 .release = io_uring_release,
10218 .mmap = io_uring_mmap,
10220 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10221 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10223 .poll = io_uring_poll,
10224 #ifdef CONFIG_PROC_FS
10225 .show_fdinfo = io_uring_show_fdinfo,
10229 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10230 struct io_uring_params *p)
10232 struct io_rings *rings;
10233 size_t size, sq_array_offset;
10235 /* make sure these are sane, as we already accounted them */
10236 ctx->sq_entries = p->sq_entries;
10237 ctx->cq_entries = p->cq_entries;
10239 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10240 if (size == SIZE_MAX)
10243 rings = io_mem_alloc(size);
10247 ctx->rings = rings;
10248 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10249 rings->sq_ring_mask = p->sq_entries - 1;
10250 rings->cq_ring_mask = p->cq_entries - 1;
10251 rings->sq_ring_entries = p->sq_entries;
10252 rings->cq_ring_entries = p->cq_entries;
10254 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10255 if (size == SIZE_MAX) {
10256 io_mem_free(ctx->rings);
10261 ctx->sq_sqes = io_mem_alloc(size);
10262 if (!ctx->sq_sqes) {
10263 io_mem_free(ctx->rings);
10271 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10275 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10279 ret = io_uring_add_tctx_node(ctx);
10284 fd_install(fd, file);
10289 * Allocate an anonymous fd, this is what constitutes the application
10290 * visible backing of an io_uring instance. The application mmaps this
10291 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10292 * we have to tie this fd to a socket for file garbage collection purposes.
10294 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10297 #if defined(CONFIG_UNIX)
10300 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10303 return ERR_PTR(ret);
10306 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10307 O_RDWR | O_CLOEXEC);
10308 #if defined(CONFIG_UNIX)
10309 if (IS_ERR(file)) {
10310 sock_release(ctx->ring_sock);
10311 ctx->ring_sock = NULL;
10313 ctx->ring_sock->file = file;
10319 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10320 struct io_uring_params __user *params)
10322 struct io_ring_ctx *ctx;
10328 if (entries > IORING_MAX_ENTRIES) {
10329 if (!(p->flags & IORING_SETUP_CLAMP))
10331 entries = IORING_MAX_ENTRIES;
10335 * Use twice as many entries for the CQ ring. It's possible for the
10336 * application to drive a higher depth than the size of the SQ ring,
10337 * since the sqes are only used at submission time. This allows for
10338 * some flexibility in overcommitting a bit. If the application has
10339 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10340 * of CQ ring entries manually.
10342 p->sq_entries = roundup_pow_of_two(entries);
10343 if (p->flags & IORING_SETUP_CQSIZE) {
10345 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10346 * to a power-of-two, if it isn't already. We do NOT impose
10347 * any cq vs sq ring sizing.
10349 if (!p->cq_entries)
10351 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10352 if (!(p->flags & IORING_SETUP_CLAMP))
10354 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10356 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10357 if (p->cq_entries < p->sq_entries)
10360 p->cq_entries = 2 * p->sq_entries;
10363 ctx = io_ring_ctx_alloc(p);
10366 ctx->compat = in_compat_syscall();
10367 if (!capable(CAP_IPC_LOCK))
10368 ctx->user = get_uid(current_user());
10371 * This is just grabbed for accounting purposes. When a process exits,
10372 * the mm is exited and dropped before the files, hence we need to hang
10373 * on to this mm purely for the purposes of being able to unaccount
10374 * memory (locked/pinned vm). It's not used for anything else.
10376 mmgrab(current->mm);
10377 ctx->mm_account = current->mm;
10379 ret = io_allocate_scq_urings(ctx, p);
10383 ret = io_sq_offload_create(ctx, p);
10386 /* always set a rsrc node */
10387 ret = io_rsrc_node_switch_start(ctx);
10390 io_rsrc_node_switch(ctx, NULL);
10392 memset(&p->sq_off, 0, sizeof(p->sq_off));
10393 p->sq_off.head = offsetof(struct io_rings, sq.head);
10394 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10395 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10396 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10397 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10398 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10399 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10401 memset(&p->cq_off, 0, sizeof(p->cq_off));
10402 p->cq_off.head = offsetof(struct io_rings, cq.head);
10403 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10404 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10405 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10406 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10407 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10408 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10410 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10411 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10412 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10413 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10414 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10415 IORING_FEAT_RSRC_TAGS;
10417 if (copy_to_user(params, p, sizeof(*p))) {
10422 file = io_uring_get_file(ctx);
10423 if (IS_ERR(file)) {
10424 ret = PTR_ERR(file);
10429 * Install ring fd as the very last thing, so we don't risk someone
10430 * having closed it before we finish setup
10432 ret = io_uring_install_fd(ctx, file);
10434 /* fput will clean it up */
10439 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10442 io_ring_ctx_wait_and_kill(ctx);
10447 * Sets up an aio uring context, and returns the fd. Applications asks for a
10448 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10449 * params structure passed in.
10451 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10453 struct io_uring_params p;
10456 if (copy_from_user(&p, params, sizeof(p)))
10458 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10463 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10464 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10465 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10466 IORING_SETUP_R_DISABLED))
10469 return io_uring_create(entries, &p, params);
10472 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10473 struct io_uring_params __user *, params)
10475 return io_uring_setup(entries, params);
10478 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10480 struct io_uring_probe *p;
10484 size = struct_size(p, ops, nr_args);
10485 if (size == SIZE_MAX)
10487 p = kzalloc(size, GFP_KERNEL);
10492 if (copy_from_user(p, arg, size))
10495 if (memchr_inv(p, 0, size))
10498 p->last_op = IORING_OP_LAST - 1;
10499 if (nr_args > IORING_OP_LAST)
10500 nr_args = IORING_OP_LAST;
10502 for (i = 0; i < nr_args; i++) {
10504 if (!io_op_defs[i].not_supported)
10505 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10510 if (copy_to_user(arg, p, size))
10517 static int io_register_personality(struct io_ring_ctx *ctx)
10519 const struct cred *creds;
10523 creds = get_current_cred();
10525 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10526 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10534 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10535 unsigned int nr_args)
10537 struct io_uring_restriction *res;
10541 /* Restrictions allowed only if rings started disabled */
10542 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10545 /* We allow only a single restrictions registration */
10546 if (ctx->restrictions.registered)
10549 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10552 size = array_size(nr_args, sizeof(*res));
10553 if (size == SIZE_MAX)
10556 res = memdup_user(arg, size);
10558 return PTR_ERR(res);
10562 for (i = 0; i < nr_args; i++) {
10563 switch (res[i].opcode) {
10564 case IORING_RESTRICTION_REGISTER_OP:
10565 if (res[i].register_op >= IORING_REGISTER_LAST) {
10570 __set_bit(res[i].register_op,
10571 ctx->restrictions.register_op);
10573 case IORING_RESTRICTION_SQE_OP:
10574 if (res[i].sqe_op >= IORING_OP_LAST) {
10579 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10581 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10582 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10584 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10585 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10594 /* Reset all restrictions if an error happened */
10596 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10598 ctx->restrictions.registered = true;
10604 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10606 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10609 if (ctx->restrictions.registered)
10610 ctx->restricted = 1;
10612 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10613 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10614 wake_up(&ctx->sq_data->wait);
10618 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10619 struct io_uring_rsrc_update2 *up,
10625 if (check_add_overflow(up->offset, nr_args, &tmp))
10627 err = io_rsrc_node_switch_start(ctx);
10632 case IORING_RSRC_FILE:
10633 return __io_sqe_files_update(ctx, up, nr_args);
10634 case IORING_RSRC_BUFFER:
10635 return __io_sqe_buffers_update(ctx, up, nr_args);
10640 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10643 struct io_uring_rsrc_update2 up;
10647 memset(&up, 0, sizeof(up));
10648 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10650 if (up.resv || up.resv2)
10652 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10655 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10656 unsigned size, unsigned type)
10658 struct io_uring_rsrc_update2 up;
10660 if (size != sizeof(up))
10662 if (copy_from_user(&up, arg, sizeof(up)))
10664 if (!up.nr || up.resv || up.resv2)
10666 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10669 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10670 unsigned int size, unsigned int type)
10672 struct io_uring_rsrc_register rr;
10674 /* keep it extendible */
10675 if (size != sizeof(rr))
10678 memset(&rr, 0, sizeof(rr));
10679 if (copy_from_user(&rr, arg, size))
10681 if (!rr.nr || rr.resv || rr.resv2)
10685 case IORING_RSRC_FILE:
10686 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10687 rr.nr, u64_to_user_ptr(rr.tags));
10688 case IORING_RSRC_BUFFER:
10689 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10690 rr.nr, u64_to_user_ptr(rr.tags));
10695 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10698 struct io_uring_task *tctx = current->io_uring;
10699 cpumask_var_t new_mask;
10702 if (!tctx || !tctx->io_wq)
10705 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10708 cpumask_clear(new_mask);
10709 if (len > cpumask_size())
10710 len = cpumask_size();
10712 if (in_compat_syscall()) {
10713 ret = compat_get_bitmap(cpumask_bits(new_mask),
10714 (const compat_ulong_t __user *)arg,
10715 len * 8 /* CHAR_BIT */);
10717 ret = copy_from_user(new_mask, arg, len);
10721 free_cpumask_var(new_mask);
10725 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10726 free_cpumask_var(new_mask);
10730 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10732 struct io_uring_task *tctx = current->io_uring;
10734 if (!tctx || !tctx->io_wq)
10737 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10740 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10742 __must_hold(&ctx->uring_lock)
10744 struct io_tctx_node *node;
10745 struct io_uring_task *tctx = NULL;
10746 struct io_sq_data *sqd = NULL;
10747 __u32 new_count[2];
10750 if (copy_from_user(new_count, arg, sizeof(new_count)))
10752 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10753 if (new_count[i] > INT_MAX)
10756 if (ctx->flags & IORING_SETUP_SQPOLL) {
10757 sqd = ctx->sq_data;
10760 * Observe the correct sqd->lock -> ctx->uring_lock
10761 * ordering. Fine to drop uring_lock here, we hold
10762 * a ref to the ctx.
10764 refcount_inc(&sqd->refs);
10765 mutex_unlock(&ctx->uring_lock);
10766 mutex_lock(&sqd->lock);
10767 mutex_lock(&ctx->uring_lock);
10769 tctx = sqd->thread->io_uring;
10772 tctx = current->io_uring;
10775 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10777 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10779 ctx->iowq_limits[i] = new_count[i];
10780 ctx->iowq_limits_set = true;
10783 if (tctx && tctx->io_wq) {
10784 ret = io_wq_max_workers(tctx->io_wq, new_count);
10788 memset(new_count, 0, sizeof(new_count));
10792 mutex_unlock(&sqd->lock);
10793 io_put_sq_data(sqd);
10796 if (copy_to_user(arg, new_count, sizeof(new_count)))
10799 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10803 /* now propagate the restriction to all registered users */
10804 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10805 struct io_uring_task *tctx = node->task->io_uring;
10807 if (WARN_ON_ONCE(!tctx->io_wq))
10810 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10811 new_count[i] = ctx->iowq_limits[i];
10812 /* ignore errors, it always returns zero anyway */
10813 (void)io_wq_max_workers(tctx->io_wq, new_count);
10818 mutex_unlock(&sqd->lock);
10819 io_put_sq_data(sqd);
10824 static bool io_register_op_must_quiesce(int op)
10827 case IORING_REGISTER_BUFFERS:
10828 case IORING_UNREGISTER_BUFFERS:
10829 case IORING_REGISTER_FILES:
10830 case IORING_UNREGISTER_FILES:
10831 case IORING_REGISTER_FILES_UPDATE:
10832 case IORING_REGISTER_PROBE:
10833 case IORING_REGISTER_PERSONALITY:
10834 case IORING_UNREGISTER_PERSONALITY:
10835 case IORING_REGISTER_FILES2:
10836 case IORING_REGISTER_FILES_UPDATE2:
10837 case IORING_REGISTER_BUFFERS2:
10838 case IORING_REGISTER_BUFFERS_UPDATE:
10839 case IORING_REGISTER_IOWQ_AFF:
10840 case IORING_UNREGISTER_IOWQ_AFF:
10841 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10848 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10852 percpu_ref_kill(&ctx->refs);
10855 * Drop uring mutex before waiting for references to exit. If another
10856 * thread is currently inside io_uring_enter() it might need to grab the
10857 * uring_lock to make progress. If we hold it here across the drain
10858 * wait, then we can deadlock. It's safe to drop the mutex here, since
10859 * no new references will come in after we've killed the percpu ref.
10861 mutex_unlock(&ctx->uring_lock);
10863 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10866 ret = io_run_task_work_sig();
10867 } while (ret >= 0);
10868 mutex_lock(&ctx->uring_lock);
10871 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10875 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10876 void __user *arg, unsigned nr_args)
10877 __releases(ctx->uring_lock)
10878 __acquires(ctx->uring_lock)
10883 * We're inside the ring mutex, if the ref is already dying, then
10884 * someone else killed the ctx or is already going through
10885 * io_uring_register().
10887 if (percpu_ref_is_dying(&ctx->refs))
10890 if (ctx->restricted) {
10891 if (opcode >= IORING_REGISTER_LAST)
10893 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10894 if (!test_bit(opcode, ctx->restrictions.register_op))
10898 if (io_register_op_must_quiesce(opcode)) {
10899 ret = io_ctx_quiesce(ctx);
10905 case IORING_REGISTER_BUFFERS:
10906 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10908 case IORING_UNREGISTER_BUFFERS:
10910 if (arg || nr_args)
10912 ret = io_sqe_buffers_unregister(ctx);
10914 case IORING_REGISTER_FILES:
10915 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10917 case IORING_UNREGISTER_FILES:
10919 if (arg || nr_args)
10921 ret = io_sqe_files_unregister(ctx);
10923 case IORING_REGISTER_FILES_UPDATE:
10924 ret = io_register_files_update(ctx, arg, nr_args);
10926 case IORING_REGISTER_EVENTFD:
10927 case IORING_REGISTER_EVENTFD_ASYNC:
10931 ret = io_eventfd_register(ctx, arg);
10934 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10935 ctx->eventfd_async = 1;
10937 ctx->eventfd_async = 0;
10939 case IORING_UNREGISTER_EVENTFD:
10941 if (arg || nr_args)
10943 ret = io_eventfd_unregister(ctx);
10945 case IORING_REGISTER_PROBE:
10947 if (!arg || nr_args > 256)
10949 ret = io_probe(ctx, arg, nr_args);
10951 case IORING_REGISTER_PERSONALITY:
10953 if (arg || nr_args)
10955 ret = io_register_personality(ctx);
10957 case IORING_UNREGISTER_PERSONALITY:
10961 ret = io_unregister_personality(ctx, nr_args);
10963 case IORING_REGISTER_ENABLE_RINGS:
10965 if (arg || nr_args)
10967 ret = io_register_enable_rings(ctx);
10969 case IORING_REGISTER_RESTRICTIONS:
10970 ret = io_register_restrictions(ctx, arg, nr_args);
10972 case IORING_REGISTER_FILES2:
10973 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10975 case IORING_REGISTER_FILES_UPDATE2:
10976 ret = io_register_rsrc_update(ctx, arg, nr_args,
10979 case IORING_REGISTER_BUFFERS2:
10980 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10982 case IORING_REGISTER_BUFFERS_UPDATE:
10983 ret = io_register_rsrc_update(ctx, arg, nr_args,
10984 IORING_RSRC_BUFFER);
10986 case IORING_REGISTER_IOWQ_AFF:
10988 if (!arg || !nr_args)
10990 ret = io_register_iowq_aff(ctx, arg, nr_args);
10992 case IORING_UNREGISTER_IOWQ_AFF:
10994 if (arg || nr_args)
10996 ret = io_unregister_iowq_aff(ctx);
10998 case IORING_REGISTER_IOWQ_MAX_WORKERS:
11000 if (!arg || nr_args != 2)
11002 ret = io_register_iowq_max_workers(ctx, arg);
11009 if (io_register_op_must_quiesce(opcode)) {
11010 /* bring the ctx back to life */
11011 percpu_ref_reinit(&ctx->refs);
11012 reinit_completion(&ctx->ref_comp);
11017 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
11018 void __user *, arg, unsigned int, nr_args)
11020 struct io_ring_ctx *ctx;
11029 if (f.file->f_op != &io_uring_fops)
11032 ctx = f.file->private_data;
11034 io_run_task_work();
11036 mutex_lock(&ctx->uring_lock);
11037 ret = __io_uring_register(ctx, opcode, arg, nr_args);
11038 mutex_unlock(&ctx->uring_lock);
11039 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
11040 ctx->cq_ev_fd != NULL, ret);
11046 static int __init io_uring_init(void)
11048 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11049 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11050 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11053 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11054 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11055 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11056 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11057 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11058 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11059 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11060 BUILD_BUG_SQE_ELEM(8, __u64, off);
11061 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11062 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11063 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11064 BUILD_BUG_SQE_ELEM(24, __u32, len);
11065 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11066 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11067 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11068 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11069 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11070 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11071 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11072 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11073 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11074 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11075 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11076 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11077 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11078 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11079 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11080 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11081 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11082 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11083 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11084 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11085 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11087 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11088 sizeof(struct io_uring_rsrc_update));
11089 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11090 sizeof(struct io_uring_rsrc_update2));
11092 /* ->buf_index is u16 */
11093 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11095 /* should fit into one byte */
11096 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11098 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11099 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11101 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11105 __initcall(io_uring_init);