1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head ltimeout_list;
379 struct list_head cq_overflow_list;
380 struct xarray io_buffers;
381 struct xarray personalities;
383 unsigned sq_thread_idle;
384 } ____cacheline_aligned_in_smp;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list;
388 unsigned int locked_free_nr;
390 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
396 unsigned long check_cq_overflow;
399 unsigned cached_cq_tail;
401 struct eventfd_ctx *cq_ev_fd;
402 struct wait_queue_head poll_wait;
403 struct wait_queue_head cq_wait;
405 atomic_t cq_timeouts;
406 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
460 bool iowq_limits_set;
464 struct io_uring_task {
465 /* submission side */
468 struct wait_queue_head wait;
469 const struct io_ring_ctx *last;
471 struct percpu_counter inflight;
472 atomic_t inflight_tracked;
475 spinlock_t task_lock;
476 struct io_wq_work_list task_list;
477 struct callback_head task_work;
482 * First field must be the file pointer in all the
483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
485 struct io_poll_iocb {
487 struct wait_queue_head *head;
491 struct wait_queue_entry wait;
494 struct io_poll_update {
500 bool update_user_data;
509 struct io_timeout_data {
510 struct io_kiocb *req;
511 struct hrtimer timer;
512 struct timespec64 ts;
513 enum hrtimer_mode mode;
519 struct sockaddr __user *addr;
520 int __user *addr_len;
523 unsigned long nofile;
543 struct list_head list;
544 /* head of the link, used by linked timeouts only */
545 struct io_kiocb *head;
546 /* for linked completions */
547 struct io_kiocb *prev;
550 struct io_timeout_rem {
555 struct timespec64 ts;
561 /* NOTE: kiocb has the file as the first member, so don't do it here */
569 struct sockaddr __user *addr;
576 struct compat_msghdr __user *umsg_compat;
577 struct user_msghdr __user *umsg;
583 struct io_buffer *kbuf;
590 struct filename *filename;
592 unsigned long nofile;
595 struct io_rsrc_update {
621 struct epoll_event event;
625 struct file *file_out;
626 struct file *file_in;
633 struct io_provide_buf {
647 const char __user *filename;
648 struct statx __user *buffer;
660 struct filename *oldpath;
661 struct filename *newpath;
669 struct filename *filename;
676 struct filename *filename;
682 struct filename *oldpath;
683 struct filename *newpath;
690 struct filename *oldpath;
691 struct filename *newpath;
695 struct io_completion {
700 struct io_async_connect {
701 struct sockaddr_storage address;
704 struct io_async_msghdr {
705 struct iovec fast_iov[UIO_FASTIOV];
706 /* points to an allocated iov, if NULL we use fast_iov instead */
707 struct iovec *free_iov;
708 struct sockaddr __user *uaddr;
710 struct sockaddr_storage addr;
714 struct iovec fast_iov[UIO_FASTIOV];
715 const struct iovec *free_iovec;
716 struct iov_iter iter;
717 struct iov_iter_state iter_state;
719 struct wait_page_queue wpq;
723 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
724 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
725 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
726 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
727 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
728 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
730 /* first byte is taken by user flags, shift it to not overlap */
735 REQ_F_LINK_TIMEOUT_BIT,
736 REQ_F_NEED_CLEANUP_BIT,
738 REQ_F_BUFFER_SELECTED_BIT,
739 REQ_F_COMPLETE_INLINE_BIT,
743 REQ_F_ARM_LTIMEOUT_BIT,
744 /* keep async read/write and isreg together and in order */
745 REQ_F_NOWAIT_READ_BIT,
746 REQ_F_NOWAIT_WRITE_BIT,
749 /* not a real bit, just to check we're not overflowing the space */
755 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
756 /* drain existing IO first */
757 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
759 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
760 /* doesn't sever on completion < 0 */
761 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
763 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
764 /* IOSQE_BUFFER_SELECT */
765 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
767 /* fail rest of links */
768 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
769 /* on inflight list, should be cancelled and waited on exit reliably */
770 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
771 /* read/write uses file position */
772 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
773 /* must not punt to workers */
774 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
775 /* has or had linked timeout */
776 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
778 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
779 /* already went through poll handler */
780 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
781 /* buffer already selected */
782 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
783 /* completion is deferred through io_comp_state */
784 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
785 /* caller should reissue async */
786 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
787 /* supports async reads */
788 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
789 /* supports async writes */
790 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
792 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
793 /* has creds assigned */
794 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
795 /* skip refcounting if not set */
796 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
797 /* there is a linked timeout that has to be armed */
798 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
802 struct io_poll_iocb poll;
803 struct io_poll_iocb *double_poll;
806 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
808 struct io_task_work {
810 struct io_wq_work_node node;
811 struct llist_node fallback_node;
813 io_req_tw_func_t func;
817 IORING_RSRC_FILE = 0,
818 IORING_RSRC_BUFFER = 1,
822 * NOTE! Each of the iocb union members has the file pointer
823 * as the first entry in their struct definition. So you can
824 * access the file pointer through any of the sub-structs,
825 * or directly as just 'ki_filp' in this struct.
831 struct io_poll_iocb poll;
832 struct io_poll_update poll_update;
833 struct io_accept accept;
835 struct io_cancel cancel;
836 struct io_timeout timeout;
837 struct io_timeout_rem timeout_rem;
838 struct io_connect connect;
839 struct io_sr_msg sr_msg;
841 struct io_close close;
842 struct io_rsrc_update rsrc_update;
843 struct io_fadvise fadvise;
844 struct io_madvise madvise;
845 struct io_epoll epoll;
846 struct io_splice splice;
847 struct io_provide_buf pbuf;
848 struct io_statx statx;
849 struct io_shutdown shutdown;
850 struct io_rename rename;
851 struct io_unlink unlink;
852 struct io_mkdir mkdir;
853 struct io_symlink symlink;
854 struct io_hardlink hardlink;
855 /* use only after cleaning per-op data, see io_clean_op() */
856 struct io_completion compl;
859 /* opcode allocated if it needs to store data for async defer */
862 /* polled IO has completed */
868 struct io_ring_ctx *ctx;
871 struct task_struct *task;
874 struct io_kiocb *link;
875 struct percpu_ref *fixed_rsrc_refs;
877 /* used with ctx->iopoll_list with reads/writes */
878 struct list_head inflight_entry;
879 struct io_task_work io_task_work;
880 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
881 struct hlist_node hash_node;
882 struct async_poll *apoll;
883 struct io_wq_work work;
884 const struct cred *creds;
886 /* store used ubuf, so we can prevent reloading */
887 struct io_mapped_ubuf *imu;
890 struct io_tctx_node {
891 struct list_head ctx_node;
892 struct task_struct *task;
893 struct io_ring_ctx *ctx;
896 struct io_defer_entry {
897 struct list_head list;
898 struct io_kiocb *req;
903 /* needs req->file assigned */
904 unsigned needs_file : 1;
905 /* hash wq insertion if file is a regular file */
906 unsigned hash_reg_file : 1;
907 /* unbound wq insertion if file is a non-regular file */
908 unsigned unbound_nonreg_file : 1;
909 /* opcode is not supported by this kernel */
910 unsigned not_supported : 1;
911 /* set if opcode supports polled "wait" */
913 unsigned pollout : 1;
914 /* op supports buffer selection */
915 unsigned buffer_select : 1;
916 /* do prep async if is going to be punted */
917 unsigned needs_async_setup : 1;
918 /* should block plug */
920 /* size of async data needed, if any */
921 unsigned short async_size;
924 static const struct io_op_def io_op_defs[] = {
925 [IORING_OP_NOP] = {},
926 [IORING_OP_READV] = {
928 .unbound_nonreg_file = 1,
931 .needs_async_setup = 1,
933 .async_size = sizeof(struct io_async_rw),
935 [IORING_OP_WRITEV] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_setup = 1,
942 .async_size = sizeof(struct io_async_rw),
944 [IORING_OP_FSYNC] = {
947 [IORING_OP_READ_FIXED] = {
949 .unbound_nonreg_file = 1,
952 .async_size = sizeof(struct io_async_rw),
954 [IORING_OP_WRITE_FIXED] = {
957 .unbound_nonreg_file = 1,
960 .async_size = sizeof(struct io_async_rw),
962 [IORING_OP_POLL_ADD] = {
964 .unbound_nonreg_file = 1,
966 [IORING_OP_POLL_REMOVE] = {},
967 [IORING_OP_SYNC_FILE_RANGE] = {
970 [IORING_OP_SENDMSG] = {
972 .unbound_nonreg_file = 1,
974 .needs_async_setup = 1,
975 .async_size = sizeof(struct io_async_msghdr),
977 [IORING_OP_RECVMSG] = {
979 .unbound_nonreg_file = 1,
982 .needs_async_setup = 1,
983 .async_size = sizeof(struct io_async_msghdr),
985 [IORING_OP_TIMEOUT] = {
986 .async_size = sizeof(struct io_timeout_data),
988 [IORING_OP_TIMEOUT_REMOVE] = {
989 /* used by timeout updates' prep() */
991 [IORING_OP_ACCEPT] = {
993 .unbound_nonreg_file = 1,
996 [IORING_OP_ASYNC_CANCEL] = {},
997 [IORING_OP_LINK_TIMEOUT] = {
998 .async_size = sizeof(struct io_timeout_data),
1000 [IORING_OP_CONNECT] = {
1002 .unbound_nonreg_file = 1,
1004 .needs_async_setup = 1,
1005 .async_size = sizeof(struct io_async_connect),
1007 [IORING_OP_FALLOCATE] = {
1010 [IORING_OP_OPENAT] = {},
1011 [IORING_OP_CLOSE] = {},
1012 [IORING_OP_FILES_UPDATE] = {},
1013 [IORING_OP_STATX] = {},
1014 [IORING_OP_READ] = {
1016 .unbound_nonreg_file = 1,
1020 .async_size = sizeof(struct io_async_rw),
1022 [IORING_OP_WRITE] = {
1025 .unbound_nonreg_file = 1,
1028 .async_size = sizeof(struct io_async_rw),
1030 [IORING_OP_FADVISE] = {
1033 [IORING_OP_MADVISE] = {},
1034 [IORING_OP_SEND] = {
1036 .unbound_nonreg_file = 1,
1039 [IORING_OP_RECV] = {
1041 .unbound_nonreg_file = 1,
1045 [IORING_OP_OPENAT2] = {
1047 [IORING_OP_EPOLL_CTL] = {
1048 .unbound_nonreg_file = 1,
1050 [IORING_OP_SPLICE] = {
1053 .unbound_nonreg_file = 1,
1055 [IORING_OP_PROVIDE_BUFFERS] = {},
1056 [IORING_OP_REMOVE_BUFFERS] = {},
1060 .unbound_nonreg_file = 1,
1062 [IORING_OP_SHUTDOWN] = {
1065 [IORING_OP_RENAMEAT] = {},
1066 [IORING_OP_UNLINKAT] = {},
1067 [IORING_OP_MKDIRAT] = {},
1068 [IORING_OP_SYMLINKAT] = {},
1069 [IORING_OP_LINKAT] = {},
1072 /* requests with any of those set should undergo io_disarm_next() */
1073 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1075 static bool io_disarm_next(struct io_kiocb *req);
1076 static void io_uring_del_tctx_node(unsigned long index);
1077 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1078 struct task_struct *task,
1080 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1082 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1083 long res, unsigned int cflags);
1084 static void io_put_req(struct io_kiocb *req);
1085 static void io_put_req_deferred(struct io_kiocb *req);
1086 static void io_dismantle_req(struct io_kiocb *req);
1087 static void io_queue_linked_timeout(struct io_kiocb *req);
1088 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1089 struct io_uring_rsrc_update2 *up,
1091 static void io_clean_op(struct io_kiocb *req);
1092 static struct file *io_file_get(struct io_ring_ctx *ctx,
1093 struct io_kiocb *req, int fd, bool fixed);
1094 static void __io_queue_sqe(struct io_kiocb *req);
1095 static void io_rsrc_put_work(struct work_struct *work);
1097 static void io_req_task_queue(struct io_kiocb *req);
1098 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1099 static int io_req_prep_async(struct io_kiocb *req);
1101 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1102 unsigned int issue_flags, u32 slot_index);
1103 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1105 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1107 static struct kmem_cache *req_cachep;
1109 static const struct file_operations io_uring_fops;
1111 struct sock *io_uring_get_socket(struct file *file)
1113 #if defined(CONFIG_UNIX)
1114 if (file->f_op == &io_uring_fops) {
1115 struct io_ring_ctx *ctx = file->private_data;
1117 return ctx->ring_sock->sk;
1122 EXPORT_SYMBOL(io_uring_get_socket);
1124 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1127 mutex_lock(&ctx->uring_lock);
1132 #define io_for_each_link(pos, head) \
1133 for (pos = (head); pos; pos = pos->link)
1136 * Shamelessly stolen from the mm implementation of page reference checking,
1137 * see commit f958d7b528b1 for details.
1139 #define req_ref_zero_or_close_to_overflow(req) \
1140 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1142 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1144 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1145 return atomic_inc_not_zero(&req->refs);
1148 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1150 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1153 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1154 return atomic_dec_and_test(&req->refs);
1157 static inline void req_ref_put(struct io_kiocb *req)
1159 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1160 WARN_ON_ONCE(req_ref_put_and_test(req));
1163 static inline void req_ref_get(struct io_kiocb *req)
1165 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1166 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1167 atomic_inc(&req->refs);
1170 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1172 if (!(req->flags & REQ_F_REFCOUNT)) {
1173 req->flags |= REQ_F_REFCOUNT;
1174 atomic_set(&req->refs, nr);
1178 static inline void io_req_set_refcount(struct io_kiocb *req)
1180 __io_req_set_refcount(req, 1);
1183 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1185 struct io_ring_ctx *ctx = req->ctx;
1187 if (!req->fixed_rsrc_refs) {
1188 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1189 percpu_ref_get(req->fixed_rsrc_refs);
1193 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1195 bool got = percpu_ref_tryget(ref);
1197 /* already at zero, wait for ->release() */
1199 wait_for_completion(compl);
1200 percpu_ref_resurrect(ref);
1202 percpu_ref_put(ref);
1205 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
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;
1456 switch (req->opcode) {
1457 case IORING_OP_SPLICE:
1459 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1460 req->work.flags |= IO_WQ_WORK_UNBOUND;
1465 static void io_prep_async_link(struct io_kiocb *req)
1467 struct io_kiocb *cur;
1469 if (req->flags & REQ_F_LINK_TIMEOUT) {
1470 struct io_ring_ctx *ctx = req->ctx;
1472 spin_lock_irq(&ctx->timeout_lock);
1473 io_for_each_link(cur, req)
1474 io_prep_async_work(cur);
1475 spin_unlock_irq(&ctx->timeout_lock);
1477 io_for_each_link(cur, req)
1478 io_prep_async_work(cur);
1482 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1484 struct io_ring_ctx *ctx = req->ctx;
1485 struct io_kiocb *link = io_prep_linked_timeout(req);
1486 struct io_uring_task *tctx = req->task->io_uring;
1488 /* must not take the lock, NULL it as a precaution */
1492 BUG_ON(!tctx->io_wq);
1494 /* init ->work of the whole link before punting */
1495 io_prep_async_link(req);
1498 * Not expected to happen, but if we do have a bug where this _can_
1499 * happen, catch it here and ensure the request is marked as
1500 * canceled. That will make io-wq go through the usual work cancel
1501 * procedure rather than attempt to run this request (or create a new
1504 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1505 req->work.flags |= IO_WQ_WORK_CANCEL;
1507 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1508 &req->work, req->flags);
1509 io_wq_enqueue(tctx->io_wq, &req->work);
1511 io_queue_linked_timeout(link);
1514 static void io_kill_timeout(struct io_kiocb *req, int status)
1515 __must_hold(&req->ctx->completion_lock)
1516 __must_hold(&req->ctx->timeout_lock)
1518 struct io_timeout_data *io = req->async_data;
1520 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1523 atomic_set(&req->ctx->cq_timeouts,
1524 atomic_read(&req->ctx->cq_timeouts) + 1);
1525 list_del_init(&req->timeout.list);
1526 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1527 io_put_req_deferred(req);
1531 static void io_queue_deferred(struct io_ring_ctx *ctx)
1533 while (!list_empty(&ctx->defer_list)) {
1534 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1535 struct io_defer_entry, list);
1537 if (req_need_defer(de->req, de->seq))
1539 list_del_init(&de->list);
1540 io_req_task_queue(de->req);
1545 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1546 __must_hold(&ctx->completion_lock)
1548 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1550 spin_lock_irq(&ctx->timeout_lock);
1551 while (!list_empty(&ctx->timeout_list)) {
1552 u32 events_needed, events_got;
1553 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1554 struct io_kiocb, timeout.list);
1556 if (io_is_timeout_noseq(req))
1560 * Since seq can easily wrap around over time, subtract
1561 * the last seq at which timeouts were flushed before comparing.
1562 * Assuming not more than 2^31-1 events have happened since,
1563 * these subtractions won't have wrapped, so we can check if
1564 * target is in [last_seq, current_seq] by comparing the two.
1566 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1567 events_got = seq - ctx->cq_last_tm_flush;
1568 if (events_got < events_needed)
1571 list_del_init(&req->timeout.list);
1572 io_kill_timeout(req, 0);
1574 ctx->cq_last_tm_flush = seq;
1575 spin_unlock_irq(&ctx->timeout_lock);
1578 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1580 if (ctx->off_timeout_used)
1581 io_flush_timeouts(ctx);
1582 if (ctx->drain_active)
1583 io_queue_deferred(ctx);
1586 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1588 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1589 __io_commit_cqring_flush(ctx);
1590 /* order cqe stores with ring update */
1591 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1594 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1596 struct io_rings *r = ctx->rings;
1598 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1601 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1603 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1606 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1608 struct io_rings *rings = ctx->rings;
1609 unsigned tail, mask = ctx->cq_entries - 1;
1612 * writes to the cq entry need to come after reading head; the
1613 * control dependency is enough as we're using WRITE_ONCE to
1616 if (__io_cqring_events(ctx) == ctx->cq_entries)
1619 tail = ctx->cached_cq_tail++;
1620 return &rings->cqes[tail & mask];
1623 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1625 if (likely(!ctx->cq_ev_fd))
1627 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1629 return !ctx->eventfd_async || io_wq_current_is_worker();
1633 * This should only get called when at least one event has been posted.
1634 * Some applications rely on the eventfd notification count only changing
1635 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1636 * 1:1 relationship between how many times this function is called (and
1637 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1639 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1642 * wake_up_all() may seem excessive, but io_wake_function() and
1643 * io_should_wake() handle the termination of the loop and only
1644 * wake as many waiters as we need to.
1646 if (wq_has_sleeper(&ctx->cq_wait))
1647 wake_up_all(&ctx->cq_wait);
1648 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1649 wake_up(&ctx->sq_data->wait);
1650 if (io_should_trigger_evfd(ctx))
1651 eventfd_signal(ctx->cq_ev_fd, 1);
1652 if (waitqueue_active(&ctx->poll_wait))
1653 wake_up_interruptible(&ctx->poll_wait);
1656 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1658 /* see waitqueue_active() comment */
1661 if (ctx->flags & IORING_SETUP_SQPOLL) {
1662 if (waitqueue_active(&ctx->cq_wait))
1663 wake_up_all(&ctx->cq_wait);
1665 if (io_should_trigger_evfd(ctx))
1666 eventfd_signal(ctx->cq_ev_fd, 1);
1667 if (waitqueue_active(&ctx->poll_wait))
1668 wake_up_interruptible(&ctx->poll_wait);
1671 /* Returns true if there are no backlogged entries after the flush */
1672 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1674 bool all_flushed, posted;
1676 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1680 spin_lock(&ctx->completion_lock);
1681 while (!list_empty(&ctx->cq_overflow_list)) {
1682 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1683 struct io_overflow_cqe *ocqe;
1687 ocqe = list_first_entry(&ctx->cq_overflow_list,
1688 struct io_overflow_cqe, list);
1690 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1692 io_account_cq_overflow(ctx);
1695 list_del(&ocqe->list);
1699 all_flushed = list_empty(&ctx->cq_overflow_list);
1701 clear_bit(0, &ctx->check_cq_overflow);
1702 WRITE_ONCE(ctx->rings->sq_flags,
1703 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1707 io_commit_cqring(ctx);
1708 spin_unlock(&ctx->completion_lock);
1710 io_cqring_ev_posted(ctx);
1714 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1718 if (test_bit(0, &ctx->check_cq_overflow)) {
1719 /* iopoll syncs against uring_lock, not completion_lock */
1720 if (ctx->flags & IORING_SETUP_IOPOLL)
1721 mutex_lock(&ctx->uring_lock);
1722 ret = __io_cqring_overflow_flush(ctx, false);
1723 if (ctx->flags & IORING_SETUP_IOPOLL)
1724 mutex_unlock(&ctx->uring_lock);
1730 /* must to be called somewhat shortly after putting a request */
1731 static inline void io_put_task(struct task_struct *task, int nr)
1733 struct io_uring_task *tctx = task->io_uring;
1735 if (likely(task == current)) {
1736 tctx->cached_refs += nr;
1738 percpu_counter_sub(&tctx->inflight, nr);
1739 if (unlikely(atomic_read(&tctx->in_idle)))
1740 wake_up(&tctx->wait);
1741 put_task_struct_many(task, nr);
1745 static void io_task_refs_refill(struct io_uring_task *tctx)
1747 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1749 percpu_counter_add(&tctx->inflight, refill);
1750 refcount_add(refill, ¤t->usage);
1751 tctx->cached_refs += refill;
1754 static inline void io_get_task_refs(int nr)
1756 struct io_uring_task *tctx = current->io_uring;
1758 tctx->cached_refs -= nr;
1759 if (unlikely(tctx->cached_refs < 0))
1760 io_task_refs_refill(tctx);
1763 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1764 long res, unsigned int cflags)
1766 struct io_overflow_cqe *ocqe;
1768 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1771 * If we're in ring overflow flush mode, or in task cancel mode,
1772 * or cannot allocate an overflow entry, then we need to drop it
1775 io_account_cq_overflow(ctx);
1778 if (list_empty(&ctx->cq_overflow_list)) {
1779 set_bit(0, &ctx->check_cq_overflow);
1780 WRITE_ONCE(ctx->rings->sq_flags,
1781 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1784 ocqe->cqe.user_data = user_data;
1785 ocqe->cqe.res = res;
1786 ocqe->cqe.flags = cflags;
1787 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1791 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1792 long res, unsigned int cflags)
1794 struct io_uring_cqe *cqe;
1796 trace_io_uring_complete(ctx, user_data, res, cflags);
1799 * If we can't get a cq entry, userspace overflowed the
1800 * submission (by quite a lot). Increment the overflow count in
1803 cqe = io_get_cqe(ctx);
1805 WRITE_ONCE(cqe->user_data, user_data);
1806 WRITE_ONCE(cqe->res, res);
1807 WRITE_ONCE(cqe->flags, cflags);
1810 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1813 /* not as hot to bloat with inlining */
1814 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1815 long res, unsigned int cflags)
1817 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1820 static void io_req_complete_post(struct io_kiocb *req, long res,
1821 unsigned int cflags)
1823 struct io_ring_ctx *ctx = req->ctx;
1825 spin_lock(&ctx->completion_lock);
1826 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1828 * If we're the last reference to this request, add to our locked
1831 if (req_ref_put_and_test(req)) {
1832 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1833 if (req->flags & IO_DISARM_MASK)
1834 io_disarm_next(req);
1836 io_req_task_queue(req->link);
1840 io_dismantle_req(req);
1841 io_put_task(req->task, 1);
1842 list_add(&req->inflight_entry, &ctx->locked_free_list);
1843 ctx->locked_free_nr++;
1845 if (!percpu_ref_tryget(&ctx->refs))
1848 io_commit_cqring(ctx);
1849 spin_unlock(&ctx->completion_lock);
1852 io_cqring_ev_posted(ctx);
1853 percpu_ref_put(&ctx->refs);
1857 static inline bool io_req_needs_clean(struct io_kiocb *req)
1859 return req->flags & IO_REQ_CLEAN_FLAGS;
1862 static void io_req_complete_state(struct io_kiocb *req, long res,
1863 unsigned int cflags)
1865 if (io_req_needs_clean(req))
1868 req->compl.cflags = cflags;
1869 req->flags |= REQ_F_COMPLETE_INLINE;
1872 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1873 long res, unsigned cflags)
1875 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1876 io_req_complete_state(req, res, cflags);
1878 io_req_complete_post(req, res, cflags);
1881 static inline void io_req_complete(struct io_kiocb *req, long res)
1883 __io_req_complete(req, 0, res, 0);
1886 static void io_req_complete_failed(struct io_kiocb *req, long res)
1889 io_req_complete_post(req, res, 0);
1892 static void io_req_complete_fail_submit(struct io_kiocb *req)
1895 * We don't submit, fail them all, for that replace hardlinks with
1896 * normal links. Extra REQ_F_LINK is tolerated.
1898 req->flags &= ~REQ_F_HARDLINK;
1899 req->flags |= REQ_F_LINK;
1900 io_req_complete_failed(req, req->result);
1904 * Don't initialise the fields below on every allocation, but do that in
1905 * advance and keep them valid across allocations.
1907 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1911 req->async_data = NULL;
1912 /* not necessary, but safer to zero */
1916 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1917 struct io_submit_state *state)
1919 spin_lock(&ctx->completion_lock);
1920 list_splice_init(&ctx->locked_free_list, &state->free_list);
1921 ctx->locked_free_nr = 0;
1922 spin_unlock(&ctx->completion_lock);
1925 /* Returns true IFF there are requests in the cache */
1926 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1928 struct io_submit_state *state = &ctx->submit_state;
1932 * If we have more than a batch's worth of requests in our IRQ side
1933 * locked cache, grab the lock and move them over to our submission
1936 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1937 io_flush_cached_locked_reqs(ctx, state);
1939 nr = state->free_reqs;
1940 while (!list_empty(&state->free_list)) {
1941 struct io_kiocb *req = list_first_entry(&state->free_list,
1942 struct io_kiocb, inflight_entry);
1944 list_del(&req->inflight_entry);
1945 state->reqs[nr++] = req;
1946 if (nr == ARRAY_SIZE(state->reqs))
1950 state->free_reqs = nr;
1955 * A request might get retired back into the request caches even before opcode
1956 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1957 * Because of that, io_alloc_req() should be called only under ->uring_lock
1958 * and with extra caution to not get a request that is still worked on.
1960 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1961 __must_hold(&ctx->uring_lock)
1963 struct io_submit_state *state = &ctx->submit_state;
1964 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1967 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1969 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1972 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1976 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1977 * retry single alloc to be on the safe side.
1979 if (unlikely(ret <= 0)) {
1980 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1981 if (!state->reqs[0])
1986 for (i = 0; i < ret; i++)
1987 io_preinit_req(state->reqs[i], ctx);
1988 state->free_reqs = ret;
1991 return state->reqs[state->free_reqs];
1994 static inline void io_put_file(struct file *file)
2000 static void io_dismantle_req(struct io_kiocb *req)
2002 unsigned int flags = req->flags;
2004 if (io_req_needs_clean(req))
2006 if (!(flags & REQ_F_FIXED_FILE))
2007 io_put_file(req->file);
2008 if (req->fixed_rsrc_refs)
2009 percpu_ref_put(req->fixed_rsrc_refs);
2010 if (req->async_data) {
2011 kfree(req->async_data);
2012 req->async_data = NULL;
2016 static void __io_free_req(struct io_kiocb *req)
2018 struct io_ring_ctx *ctx = req->ctx;
2020 io_dismantle_req(req);
2021 io_put_task(req->task, 1);
2023 spin_lock(&ctx->completion_lock);
2024 list_add(&req->inflight_entry, &ctx->locked_free_list);
2025 ctx->locked_free_nr++;
2026 spin_unlock(&ctx->completion_lock);
2028 percpu_ref_put(&ctx->refs);
2031 static inline void io_remove_next_linked(struct io_kiocb *req)
2033 struct io_kiocb *nxt = req->link;
2035 req->link = nxt->link;
2039 static bool io_kill_linked_timeout(struct io_kiocb *req)
2040 __must_hold(&req->ctx->completion_lock)
2041 __must_hold(&req->ctx->timeout_lock)
2043 struct io_kiocb *link = req->link;
2045 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2046 struct io_timeout_data *io = link->async_data;
2048 io_remove_next_linked(req);
2049 link->timeout.head = NULL;
2050 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2051 list_del(&link->timeout.list);
2052 io_cqring_fill_event(link->ctx, link->user_data,
2054 io_put_req_deferred(link);
2061 static void io_fail_links(struct io_kiocb *req)
2062 __must_hold(&req->ctx->completion_lock)
2064 struct io_kiocb *nxt, *link = req->link;
2068 long res = -ECANCELED;
2070 if (link->flags & REQ_F_FAIL)
2076 trace_io_uring_fail_link(req, link);
2077 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2078 io_put_req_deferred(link);
2083 static bool io_disarm_next(struct io_kiocb *req)
2084 __must_hold(&req->ctx->completion_lock)
2086 bool posted = false;
2088 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2089 struct io_kiocb *link = req->link;
2091 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2092 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2093 io_remove_next_linked(req);
2094 io_cqring_fill_event(link->ctx, link->user_data,
2096 io_put_req_deferred(link);
2099 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2100 struct io_ring_ctx *ctx = req->ctx;
2102 spin_lock_irq(&ctx->timeout_lock);
2103 posted = io_kill_linked_timeout(req);
2104 spin_unlock_irq(&ctx->timeout_lock);
2106 if (unlikely((req->flags & REQ_F_FAIL) &&
2107 !(req->flags & REQ_F_HARDLINK))) {
2108 posted |= (req->link != NULL);
2114 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2116 struct io_kiocb *nxt;
2119 * If LINK is set, we have dependent requests in this chain. If we
2120 * didn't fail this request, queue the first one up, moving any other
2121 * dependencies to the next request. In case of failure, fail the rest
2124 if (req->flags & IO_DISARM_MASK) {
2125 struct io_ring_ctx *ctx = req->ctx;
2128 spin_lock(&ctx->completion_lock);
2129 posted = io_disarm_next(req);
2131 io_commit_cqring(req->ctx);
2132 spin_unlock(&ctx->completion_lock);
2134 io_cqring_ev_posted(ctx);
2141 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2143 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2145 return __io_req_find_next(req);
2148 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2153 if (ctx->submit_state.compl_nr)
2154 io_submit_flush_completions(ctx);
2155 mutex_unlock(&ctx->uring_lock);
2158 percpu_ref_put(&ctx->refs);
2161 static void tctx_task_work(struct callback_head *cb)
2163 bool locked = false;
2164 struct io_ring_ctx *ctx = NULL;
2165 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2169 struct io_wq_work_node *node;
2171 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2172 io_submit_flush_completions(ctx);
2174 spin_lock_irq(&tctx->task_lock);
2175 node = tctx->task_list.first;
2176 INIT_WQ_LIST(&tctx->task_list);
2178 tctx->task_running = false;
2179 spin_unlock_irq(&tctx->task_lock);
2184 struct io_wq_work_node *next = node->next;
2185 struct io_kiocb *req = container_of(node, struct io_kiocb,
2188 if (req->ctx != ctx) {
2189 ctx_flush_and_put(ctx, &locked);
2191 /* if not contended, grab and improve batching */
2192 locked = mutex_trylock(&ctx->uring_lock);
2193 percpu_ref_get(&ctx->refs);
2195 req->io_task_work.func(req, &locked);
2202 ctx_flush_and_put(ctx, &locked);
2205 static void io_req_task_work_add(struct io_kiocb *req)
2207 struct task_struct *tsk = req->task;
2208 struct io_uring_task *tctx = tsk->io_uring;
2209 enum task_work_notify_mode notify;
2210 struct io_wq_work_node *node;
2211 unsigned long flags;
2214 WARN_ON_ONCE(!tctx);
2216 spin_lock_irqsave(&tctx->task_lock, flags);
2217 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2218 running = tctx->task_running;
2220 tctx->task_running = true;
2221 spin_unlock_irqrestore(&tctx->task_lock, flags);
2223 /* task_work already pending, we're done */
2228 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2229 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2230 * processing task_work. There's no reliable way to tell if TWA_RESUME
2233 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2234 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2235 wake_up_process(tsk);
2239 spin_lock_irqsave(&tctx->task_lock, flags);
2240 tctx->task_running = false;
2241 node = tctx->task_list.first;
2242 INIT_WQ_LIST(&tctx->task_list);
2243 spin_unlock_irqrestore(&tctx->task_lock, flags);
2246 req = container_of(node, struct io_kiocb, io_task_work.node);
2248 if (llist_add(&req->io_task_work.fallback_node,
2249 &req->ctx->fallback_llist))
2250 schedule_delayed_work(&req->ctx->fallback_work, 1);
2254 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2256 struct io_ring_ctx *ctx = req->ctx;
2258 /* not needed for normal modes, but SQPOLL depends on it */
2259 io_tw_lock(ctx, locked);
2260 io_req_complete_failed(req, req->result);
2263 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2265 struct io_ring_ctx *ctx = req->ctx;
2267 io_tw_lock(ctx, locked);
2268 /* req->task == current here, checking PF_EXITING is safe */
2269 if (likely(!(req->task->flags & PF_EXITING)))
2270 __io_queue_sqe(req);
2272 io_req_complete_failed(req, -EFAULT);
2275 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2278 req->io_task_work.func = io_req_task_cancel;
2279 io_req_task_work_add(req);
2282 static void io_req_task_queue(struct io_kiocb *req)
2284 req->io_task_work.func = io_req_task_submit;
2285 io_req_task_work_add(req);
2288 static void io_req_task_queue_reissue(struct io_kiocb *req)
2290 req->io_task_work.func = io_queue_async_work;
2291 io_req_task_work_add(req);
2294 static inline void io_queue_next(struct io_kiocb *req)
2296 struct io_kiocb *nxt = io_req_find_next(req);
2299 io_req_task_queue(nxt);
2302 static void io_free_req(struct io_kiocb *req)
2308 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2314 struct task_struct *task;
2319 static inline void io_init_req_batch(struct req_batch *rb)
2326 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2327 struct req_batch *rb)
2330 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2332 io_put_task(rb->task, rb->task_refs);
2335 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2336 struct io_submit_state *state)
2339 io_dismantle_req(req);
2341 if (req->task != rb->task) {
2343 io_put_task(rb->task, rb->task_refs);
2344 rb->task = req->task;
2350 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2351 state->reqs[state->free_reqs++] = req;
2353 list_add(&req->inflight_entry, &state->free_list);
2356 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2357 __must_hold(&ctx->uring_lock)
2359 struct io_submit_state *state = &ctx->submit_state;
2360 int i, nr = state->compl_nr;
2361 struct req_batch rb;
2363 spin_lock(&ctx->completion_lock);
2364 for (i = 0; i < nr; i++) {
2365 struct io_kiocb *req = state->compl_reqs[i];
2367 __io_cqring_fill_event(ctx, req->user_data, req->result,
2370 io_commit_cqring(ctx);
2371 spin_unlock(&ctx->completion_lock);
2372 io_cqring_ev_posted(ctx);
2374 io_init_req_batch(&rb);
2375 for (i = 0; i < nr; i++) {
2376 struct io_kiocb *req = state->compl_reqs[i];
2378 if (req_ref_put_and_test(req))
2379 io_req_free_batch(&rb, req, &ctx->submit_state);
2382 io_req_free_batch_finish(ctx, &rb);
2383 state->compl_nr = 0;
2387 * Drop reference to request, return next in chain (if there is one) if this
2388 * was the last reference to this request.
2390 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2392 struct io_kiocb *nxt = NULL;
2394 if (req_ref_put_and_test(req)) {
2395 nxt = io_req_find_next(req);
2401 static inline void io_put_req(struct io_kiocb *req)
2403 if (req_ref_put_and_test(req))
2407 static inline void io_put_req_deferred(struct io_kiocb *req)
2409 if (req_ref_put_and_test(req)) {
2410 req->io_task_work.func = io_free_req_work;
2411 io_req_task_work_add(req);
2415 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2417 /* See comment at the top of this file */
2419 return __io_cqring_events(ctx);
2422 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2424 struct io_rings *rings = ctx->rings;
2426 /* make sure SQ entry isn't read before tail */
2427 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2430 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2432 unsigned int cflags;
2434 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2435 cflags |= IORING_CQE_F_BUFFER;
2436 req->flags &= ~REQ_F_BUFFER_SELECTED;
2441 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2443 struct io_buffer *kbuf;
2445 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2447 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2448 return io_put_kbuf(req, kbuf);
2451 static inline bool io_run_task_work(void)
2453 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2454 __set_current_state(TASK_RUNNING);
2455 tracehook_notify_signal();
2463 * Find and free completed poll iocbs
2465 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2466 struct list_head *done)
2468 struct req_batch rb;
2469 struct io_kiocb *req;
2471 /* order with ->result store in io_complete_rw_iopoll() */
2474 io_init_req_batch(&rb);
2475 while (!list_empty(done)) {
2476 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2477 list_del(&req->inflight_entry);
2479 __io_cqring_fill_event(ctx, req->user_data, req->result,
2480 io_put_rw_kbuf(req));
2483 if (req_ref_put_and_test(req))
2484 io_req_free_batch(&rb, req, &ctx->submit_state);
2487 io_commit_cqring(ctx);
2488 io_cqring_ev_posted_iopoll(ctx);
2489 io_req_free_batch_finish(ctx, &rb);
2492 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2495 struct io_kiocb *req, *tmp;
2500 * Only spin for completions if we don't have multiple devices hanging
2501 * off our complete list, and we're under the requested amount.
2503 spin = !ctx->poll_multi_queue && *nr_events < min;
2505 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2506 struct kiocb *kiocb = &req->rw.kiocb;
2510 * Move completed and retryable entries to our local lists.
2511 * If we find a request that requires polling, break out
2512 * and complete those lists first, if we have entries there.
2514 if (READ_ONCE(req->iopoll_completed)) {
2515 list_move_tail(&req->inflight_entry, &done);
2518 if (!list_empty(&done))
2521 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2522 if (unlikely(ret < 0))
2527 /* iopoll may have completed current req */
2528 if (READ_ONCE(req->iopoll_completed))
2529 list_move_tail(&req->inflight_entry, &done);
2532 if (!list_empty(&done))
2533 io_iopoll_complete(ctx, nr_events, &done);
2539 * We can't just wait for polled events to come to us, we have to actively
2540 * find and complete them.
2542 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2544 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2547 mutex_lock(&ctx->uring_lock);
2548 while (!list_empty(&ctx->iopoll_list)) {
2549 unsigned int nr_events = 0;
2551 io_do_iopoll(ctx, &nr_events, 0);
2553 /* let it sleep and repeat later if can't complete a request */
2557 * Ensure we allow local-to-the-cpu processing to take place,
2558 * in this case we need to ensure that we reap all events.
2559 * Also let task_work, etc. to progress by releasing the mutex
2561 if (need_resched()) {
2562 mutex_unlock(&ctx->uring_lock);
2564 mutex_lock(&ctx->uring_lock);
2567 mutex_unlock(&ctx->uring_lock);
2570 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2572 unsigned int nr_events = 0;
2576 * We disallow the app entering submit/complete with polling, but we
2577 * still need to lock the ring to prevent racing with polled issue
2578 * that got punted to a workqueue.
2580 mutex_lock(&ctx->uring_lock);
2582 * Don't enter poll loop if we already have events pending.
2583 * If we do, we can potentially be spinning for commands that
2584 * already triggered a CQE (eg in error).
2586 if (test_bit(0, &ctx->check_cq_overflow))
2587 __io_cqring_overflow_flush(ctx, false);
2588 if (io_cqring_events(ctx))
2592 * If a submit got punted to a workqueue, we can have the
2593 * application entering polling for a command before it gets
2594 * issued. That app will hold the uring_lock for the duration
2595 * of the poll right here, so we need to take a breather every
2596 * now and then to ensure that the issue has a chance to add
2597 * the poll to the issued list. Otherwise we can spin here
2598 * forever, while the workqueue is stuck trying to acquire the
2601 if (list_empty(&ctx->iopoll_list)) {
2602 u32 tail = ctx->cached_cq_tail;
2604 mutex_unlock(&ctx->uring_lock);
2606 mutex_lock(&ctx->uring_lock);
2608 /* some requests don't go through iopoll_list */
2609 if (tail != ctx->cached_cq_tail ||
2610 list_empty(&ctx->iopoll_list))
2613 ret = io_do_iopoll(ctx, &nr_events, min);
2614 } while (!ret && nr_events < min && !need_resched());
2616 mutex_unlock(&ctx->uring_lock);
2620 static void kiocb_end_write(struct io_kiocb *req)
2623 * Tell lockdep we inherited freeze protection from submission
2626 if (req->flags & REQ_F_ISREG) {
2627 struct super_block *sb = file_inode(req->file)->i_sb;
2629 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2635 static bool io_resubmit_prep(struct io_kiocb *req)
2637 struct io_async_rw *rw = req->async_data;
2640 return !io_req_prep_async(req);
2641 iov_iter_restore(&rw->iter, &rw->iter_state);
2645 static bool io_rw_should_reissue(struct io_kiocb *req)
2647 umode_t mode = file_inode(req->file)->i_mode;
2648 struct io_ring_ctx *ctx = req->ctx;
2650 if (!S_ISBLK(mode) && !S_ISREG(mode))
2652 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2653 !(ctx->flags & IORING_SETUP_IOPOLL)))
2656 * If ref is dying, we might be running poll reap from the exit work.
2657 * Don't attempt to reissue from that path, just let it fail with
2660 if (percpu_ref_is_dying(&ctx->refs))
2663 * Play it safe and assume not safe to re-import and reissue if we're
2664 * not in the original thread group (or in task context).
2666 if (!same_thread_group(req->task, current) || !in_task())
2671 static bool io_resubmit_prep(struct io_kiocb *req)
2675 static bool io_rw_should_reissue(struct io_kiocb *req)
2681 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2683 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2684 kiocb_end_write(req);
2685 if (res != req->result) {
2686 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2687 io_rw_should_reissue(req)) {
2688 req->flags |= REQ_F_REISSUE;
2697 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2699 unsigned int cflags = io_put_rw_kbuf(req);
2700 long res = req->result;
2703 struct io_ring_ctx *ctx = req->ctx;
2704 struct io_submit_state *state = &ctx->submit_state;
2706 io_req_complete_state(req, res, cflags);
2707 state->compl_reqs[state->compl_nr++] = req;
2708 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2709 io_submit_flush_completions(ctx);
2711 io_req_complete_post(req, res, cflags);
2715 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2716 unsigned int issue_flags)
2718 if (__io_complete_rw_common(req, res))
2720 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2723 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2725 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2727 if (__io_complete_rw_common(req, res))
2730 req->io_task_work.func = io_req_task_complete;
2731 io_req_task_work_add(req);
2734 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2736 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2738 if (kiocb->ki_flags & IOCB_WRITE)
2739 kiocb_end_write(req);
2740 if (unlikely(res != req->result)) {
2741 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2742 req->flags |= REQ_F_REISSUE;
2747 WRITE_ONCE(req->result, res);
2748 /* order with io_iopoll_complete() checking ->result */
2750 WRITE_ONCE(req->iopoll_completed, 1);
2754 * After the iocb has been issued, it's safe to be found on the poll list.
2755 * Adding the kiocb to the list AFTER submission ensures that we don't
2756 * find it from a io_do_iopoll() thread before the issuer is done
2757 * accessing the kiocb cookie.
2759 static void io_iopoll_req_issued(struct io_kiocb *req)
2761 struct io_ring_ctx *ctx = req->ctx;
2762 const bool in_async = io_wq_current_is_worker();
2764 /* workqueue context doesn't hold uring_lock, grab it now */
2765 if (unlikely(in_async))
2766 mutex_lock(&ctx->uring_lock);
2769 * Track whether we have multiple files in our lists. This will impact
2770 * how we do polling eventually, not spinning if we're on potentially
2771 * different devices.
2773 if (list_empty(&ctx->iopoll_list)) {
2774 ctx->poll_multi_queue = false;
2775 } else if (!ctx->poll_multi_queue) {
2776 struct io_kiocb *list_req;
2777 unsigned int queue_num0, queue_num1;
2779 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2782 if (list_req->file != req->file) {
2783 ctx->poll_multi_queue = true;
2785 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2786 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2787 if (queue_num0 != queue_num1)
2788 ctx->poll_multi_queue = true;
2793 * For fast devices, IO may have already completed. If it has, add
2794 * it to the front so we find it first.
2796 if (READ_ONCE(req->iopoll_completed))
2797 list_add(&req->inflight_entry, &ctx->iopoll_list);
2799 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2801 if (unlikely(in_async)) {
2803 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2804 * in sq thread task context or in io worker task context. If
2805 * current task context is sq thread, we don't need to check
2806 * whether should wake up sq thread.
2808 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2809 wq_has_sleeper(&ctx->sq_data->wait))
2810 wake_up(&ctx->sq_data->wait);
2812 mutex_unlock(&ctx->uring_lock);
2816 static bool io_bdev_nowait(struct block_device *bdev)
2818 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2822 * If we tracked the file through the SCM inflight mechanism, we could support
2823 * any file. For now, just ensure that anything potentially problematic is done
2826 static bool __io_file_supports_nowait(struct file *file, int rw)
2828 umode_t mode = file_inode(file)->i_mode;
2830 if (S_ISBLK(mode)) {
2831 if (IS_ENABLED(CONFIG_BLOCK) &&
2832 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2838 if (S_ISREG(mode)) {
2839 if (IS_ENABLED(CONFIG_BLOCK) &&
2840 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2841 file->f_op != &io_uring_fops)
2846 /* any ->read/write should understand O_NONBLOCK */
2847 if (file->f_flags & O_NONBLOCK)
2850 if (!(file->f_mode & FMODE_NOWAIT))
2854 return file->f_op->read_iter != NULL;
2856 return file->f_op->write_iter != NULL;
2859 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2861 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2863 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2866 return __io_file_supports_nowait(req->file, rw);
2869 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2872 struct io_ring_ctx *ctx = req->ctx;
2873 struct kiocb *kiocb = &req->rw.kiocb;
2874 struct file *file = req->file;
2878 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2879 req->flags |= REQ_F_ISREG;
2881 kiocb->ki_pos = READ_ONCE(sqe->off);
2882 if (kiocb->ki_pos == -1) {
2883 if (!(file->f_mode & FMODE_STREAM)) {
2884 req->flags |= REQ_F_CUR_POS;
2885 kiocb->ki_pos = file->f_pos;
2890 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2891 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2892 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2897 * If the file is marked O_NONBLOCK, still allow retry for it if it
2898 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2899 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2901 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2902 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2903 req->flags |= REQ_F_NOWAIT;
2905 ioprio = READ_ONCE(sqe->ioprio);
2907 ret = ioprio_check_cap(ioprio);
2911 kiocb->ki_ioprio = ioprio;
2913 kiocb->ki_ioprio = get_current_ioprio();
2915 if (ctx->flags & IORING_SETUP_IOPOLL) {
2916 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2917 !kiocb->ki_filp->f_op->iopoll)
2920 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2921 kiocb->ki_complete = io_complete_rw_iopoll;
2922 req->iopoll_completed = 0;
2924 if (kiocb->ki_flags & IOCB_HIPRI)
2926 kiocb->ki_complete = io_complete_rw;
2929 if (req->opcode == IORING_OP_READ_FIXED ||
2930 req->opcode == IORING_OP_WRITE_FIXED) {
2932 io_req_set_rsrc_node(req);
2935 req->rw.addr = READ_ONCE(sqe->addr);
2936 req->rw.len = READ_ONCE(sqe->len);
2937 req->buf_index = READ_ONCE(sqe->buf_index);
2941 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2947 case -ERESTARTNOINTR:
2948 case -ERESTARTNOHAND:
2949 case -ERESTART_RESTARTBLOCK:
2951 * We can't just restart the syscall, since previously
2952 * submitted sqes may already be in progress. Just fail this
2958 kiocb->ki_complete(kiocb, ret, 0);
2962 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2963 unsigned int issue_flags)
2965 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2966 struct io_async_rw *io = req->async_data;
2968 /* add previously done IO, if any */
2969 if (io && io->bytes_done > 0) {
2971 ret = io->bytes_done;
2973 ret += io->bytes_done;
2976 if (req->flags & REQ_F_CUR_POS)
2977 req->file->f_pos = kiocb->ki_pos;
2978 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2979 __io_complete_rw(req, ret, 0, issue_flags);
2981 io_rw_done(kiocb, ret);
2983 if (req->flags & REQ_F_REISSUE) {
2984 req->flags &= ~REQ_F_REISSUE;
2985 if (io_resubmit_prep(req)) {
2986 io_req_task_queue_reissue(req);
2988 unsigned int cflags = io_put_rw_kbuf(req);
2989 struct io_ring_ctx *ctx = req->ctx;
2992 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
2993 mutex_lock(&ctx->uring_lock);
2994 __io_req_complete(req, issue_flags, ret, cflags);
2995 mutex_unlock(&ctx->uring_lock);
2997 __io_req_complete(req, issue_flags, ret, cflags);
3003 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3004 struct io_mapped_ubuf *imu)
3006 size_t len = req->rw.len;
3007 u64 buf_end, buf_addr = req->rw.addr;
3010 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
3012 /* not inside the mapped region */
3013 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
3017 * May not be a start of buffer, set size appropriately
3018 * and advance us to the beginning.
3020 offset = buf_addr - imu->ubuf;
3021 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3025 * Don't use iov_iter_advance() here, as it's really slow for
3026 * using the latter parts of a big fixed buffer - it iterates
3027 * over each segment manually. We can cheat a bit here, because
3030 * 1) it's a BVEC iter, we set it up
3031 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3032 * first and last bvec
3034 * So just find our index, and adjust the iterator afterwards.
3035 * If the offset is within the first bvec (or the whole first
3036 * bvec, just use iov_iter_advance(). This makes it easier
3037 * since we can just skip the first segment, which may not
3038 * be PAGE_SIZE aligned.
3040 const struct bio_vec *bvec = imu->bvec;
3042 if (offset <= bvec->bv_len) {
3043 iov_iter_advance(iter, offset);
3045 unsigned long seg_skip;
3047 /* skip first vec */
3048 offset -= bvec->bv_len;
3049 seg_skip = 1 + (offset >> PAGE_SHIFT);
3051 iter->bvec = bvec + seg_skip;
3052 iter->nr_segs -= seg_skip;
3053 iter->count -= bvec->bv_len + offset;
3054 iter->iov_offset = offset & ~PAGE_MASK;
3061 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3063 struct io_ring_ctx *ctx = req->ctx;
3064 struct io_mapped_ubuf *imu = req->imu;
3065 u16 index, buf_index = req->buf_index;
3068 if (unlikely(buf_index >= ctx->nr_user_bufs))
3070 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3071 imu = READ_ONCE(ctx->user_bufs[index]);
3074 return __io_import_fixed(req, rw, iter, imu);
3077 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3080 mutex_unlock(&ctx->uring_lock);
3083 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3086 * "Normal" inline submissions always hold the uring_lock, since we
3087 * grab it from the system call. Same is true for the SQPOLL offload.
3088 * The only exception is when we've detached the request and issue it
3089 * from an async worker thread, grab the lock for that case.
3092 mutex_lock(&ctx->uring_lock);
3095 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3096 int bgid, struct io_buffer *kbuf,
3099 struct io_buffer *head;
3101 if (req->flags & REQ_F_BUFFER_SELECTED)
3104 io_ring_submit_lock(req->ctx, needs_lock);
3106 lockdep_assert_held(&req->ctx->uring_lock);
3108 head = xa_load(&req->ctx->io_buffers, bgid);
3110 if (!list_empty(&head->list)) {
3111 kbuf = list_last_entry(&head->list, struct io_buffer,
3113 list_del(&kbuf->list);
3116 xa_erase(&req->ctx->io_buffers, bgid);
3118 if (*len > kbuf->len)
3121 kbuf = ERR_PTR(-ENOBUFS);
3124 io_ring_submit_unlock(req->ctx, needs_lock);
3129 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3132 struct io_buffer *kbuf;
3135 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3136 bgid = req->buf_index;
3137 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3140 req->rw.addr = (u64) (unsigned long) kbuf;
3141 req->flags |= REQ_F_BUFFER_SELECTED;
3142 return u64_to_user_ptr(kbuf->addr);
3145 #ifdef CONFIG_COMPAT
3146 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3149 struct compat_iovec __user *uiov;
3150 compat_ssize_t clen;
3154 uiov = u64_to_user_ptr(req->rw.addr);
3155 if (!access_ok(uiov, sizeof(*uiov)))
3157 if (__get_user(clen, &uiov->iov_len))
3163 buf = io_rw_buffer_select(req, &len, needs_lock);
3165 return PTR_ERR(buf);
3166 iov[0].iov_base = buf;
3167 iov[0].iov_len = (compat_size_t) len;
3172 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3175 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3179 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3182 len = iov[0].iov_len;
3185 buf = io_rw_buffer_select(req, &len, needs_lock);
3187 return PTR_ERR(buf);
3188 iov[0].iov_base = buf;
3189 iov[0].iov_len = len;
3193 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3196 if (req->flags & REQ_F_BUFFER_SELECTED) {
3197 struct io_buffer *kbuf;
3199 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3200 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3201 iov[0].iov_len = kbuf->len;
3204 if (req->rw.len != 1)
3207 #ifdef CONFIG_COMPAT
3208 if (req->ctx->compat)
3209 return io_compat_import(req, iov, needs_lock);
3212 return __io_iov_buffer_select(req, iov, needs_lock);
3215 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3216 struct iov_iter *iter, bool needs_lock)
3218 void __user *buf = u64_to_user_ptr(req->rw.addr);
3219 size_t sqe_len = req->rw.len;
3220 u8 opcode = req->opcode;
3223 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3225 return io_import_fixed(req, rw, iter);
3228 /* buffer index only valid with fixed read/write, or buffer select */
3229 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3232 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3233 if (req->flags & REQ_F_BUFFER_SELECT) {
3234 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3236 return PTR_ERR(buf);
3237 req->rw.len = sqe_len;
3240 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3245 if (req->flags & REQ_F_BUFFER_SELECT) {
3246 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3248 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3253 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3257 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3259 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3263 * For files that don't have ->read_iter() and ->write_iter(), handle them
3264 * by looping over ->read() or ->write() manually.
3266 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3268 struct kiocb *kiocb = &req->rw.kiocb;
3269 struct file *file = req->file;
3273 * Don't support polled IO through this interface, and we can't
3274 * support non-blocking either. For the latter, this just causes
3275 * the kiocb to be handled from an async context.
3277 if (kiocb->ki_flags & IOCB_HIPRI)
3279 if (kiocb->ki_flags & IOCB_NOWAIT)
3282 while (iov_iter_count(iter)) {
3286 if (!iov_iter_is_bvec(iter)) {
3287 iovec = iov_iter_iovec(iter);
3289 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3290 iovec.iov_len = req->rw.len;
3294 nr = file->f_op->read(file, iovec.iov_base,
3295 iovec.iov_len, io_kiocb_ppos(kiocb));
3297 nr = file->f_op->write(file, iovec.iov_base,
3298 iovec.iov_len, io_kiocb_ppos(kiocb));
3306 if (!iov_iter_is_bvec(iter)) {
3307 iov_iter_advance(iter, nr);
3313 if (nr != iovec.iov_len)
3320 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3321 const struct iovec *fast_iov, struct iov_iter *iter)
3323 struct io_async_rw *rw = req->async_data;
3325 memcpy(&rw->iter, iter, sizeof(*iter));
3326 rw->free_iovec = iovec;
3328 /* can only be fixed buffers, no need to do anything */
3329 if (iov_iter_is_bvec(iter))
3332 unsigned iov_off = 0;
3334 rw->iter.iov = rw->fast_iov;
3335 if (iter->iov != fast_iov) {
3336 iov_off = iter->iov - fast_iov;
3337 rw->iter.iov += iov_off;
3339 if (rw->fast_iov != fast_iov)
3340 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3341 sizeof(struct iovec) * iter->nr_segs);
3343 req->flags |= REQ_F_NEED_CLEANUP;
3347 static inline int io_alloc_async_data(struct io_kiocb *req)
3349 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3350 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3351 return req->async_data == NULL;
3354 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3355 const struct iovec *fast_iov,
3356 struct iov_iter *iter, bool force)
3358 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3360 if (!req->async_data) {
3361 struct io_async_rw *iorw;
3363 if (io_alloc_async_data(req)) {
3368 io_req_map_rw(req, iovec, fast_iov, iter);
3369 iorw = req->async_data;
3370 /* we've copied and mapped the iter, ensure state is saved */
3371 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3376 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3378 struct io_async_rw *iorw = req->async_data;
3379 struct iovec *iov = iorw->fast_iov;
3382 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3383 if (unlikely(ret < 0))
3386 iorw->bytes_done = 0;
3387 iorw->free_iovec = iov;
3389 req->flags |= REQ_F_NEED_CLEANUP;
3390 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3394 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3396 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3398 return io_prep_rw(req, sqe, READ);
3402 * This is our waitqueue callback handler, registered through lock_page_async()
3403 * when we initially tried to do the IO with the iocb armed our waitqueue.
3404 * This gets called when the page is unlocked, and we generally expect that to
3405 * happen when the page IO is completed and the page is now uptodate. This will
3406 * queue a task_work based retry of the operation, attempting to copy the data
3407 * again. If the latter fails because the page was NOT uptodate, then we will
3408 * do a thread based blocking retry of the operation. That's the unexpected
3411 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3412 int sync, void *arg)
3414 struct wait_page_queue *wpq;
3415 struct io_kiocb *req = wait->private;
3416 struct wait_page_key *key = arg;
3418 wpq = container_of(wait, struct wait_page_queue, wait);
3420 if (!wake_page_match(wpq, key))
3423 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3424 list_del_init(&wait->entry);
3425 io_req_task_queue(req);
3430 * This controls whether a given IO request should be armed for async page
3431 * based retry. If we return false here, the request is handed to the async
3432 * worker threads for retry. If we're doing buffered reads on a regular file,
3433 * we prepare a private wait_page_queue entry and retry the operation. This
3434 * will either succeed because the page is now uptodate and unlocked, or it
3435 * will register a callback when the page is unlocked at IO completion. Through
3436 * that callback, io_uring uses task_work to setup a retry of the operation.
3437 * That retry will attempt the buffered read again. The retry will generally
3438 * succeed, or in rare cases where it fails, we then fall back to using the
3439 * async worker threads for a blocking retry.
3441 static bool io_rw_should_retry(struct io_kiocb *req)
3443 struct io_async_rw *rw = req->async_data;
3444 struct wait_page_queue *wait = &rw->wpq;
3445 struct kiocb *kiocb = &req->rw.kiocb;
3447 /* never retry for NOWAIT, we just complete with -EAGAIN */
3448 if (req->flags & REQ_F_NOWAIT)
3451 /* Only for buffered IO */
3452 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3456 * just use poll if we can, and don't attempt if the fs doesn't
3457 * support callback based unlocks
3459 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3462 wait->wait.func = io_async_buf_func;
3463 wait->wait.private = req;
3464 wait->wait.flags = 0;
3465 INIT_LIST_HEAD(&wait->wait.entry);
3466 kiocb->ki_flags |= IOCB_WAITQ;
3467 kiocb->ki_flags &= ~IOCB_NOWAIT;
3468 kiocb->ki_waitq = wait;
3472 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3474 if (req->file->f_op->read_iter)
3475 return call_read_iter(req->file, &req->rw.kiocb, iter);
3476 else if (req->file->f_op->read)
3477 return loop_rw_iter(READ, req, iter);
3482 static bool need_read_all(struct io_kiocb *req)
3484 return req->flags & REQ_F_ISREG ||
3485 S_ISBLK(file_inode(req->file)->i_mode);
3488 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3490 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3491 struct kiocb *kiocb = &req->rw.kiocb;
3492 struct iov_iter __iter, *iter = &__iter;
3493 struct io_async_rw *rw = req->async_data;
3494 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3495 struct iov_iter_state __state, *state;
3500 state = &rw->iter_state;
3502 * We come here from an earlier attempt, restore our state to
3503 * match in case it doesn't. It's cheap enough that we don't
3504 * need to make this conditional.
3506 iov_iter_restore(iter, state);
3509 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3513 iov_iter_save_state(iter, state);
3515 req->result = iov_iter_count(iter);
3517 /* Ensure we clear previously set non-block flag */
3518 if (!force_nonblock)
3519 kiocb->ki_flags &= ~IOCB_NOWAIT;
3521 kiocb->ki_flags |= IOCB_NOWAIT;
3523 /* If the file doesn't support async, just async punt */
3524 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3525 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3526 return ret ?: -EAGAIN;
3529 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3530 if (unlikely(ret)) {
3535 ret = io_iter_do_read(req, iter);
3537 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3538 req->flags &= ~REQ_F_REISSUE;
3539 /* IOPOLL retry should happen for io-wq threads */
3540 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3542 /* no retry on NONBLOCK nor RWF_NOWAIT */
3543 if (req->flags & REQ_F_NOWAIT)
3546 } else if (ret == -EIOCBQUEUED) {
3548 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3549 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3550 /* read all, failed, already did sync or don't want to retry */
3555 * Don't depend on the iter state matching what was consumed, or being
3556 * untouched in case of error. Restore it and we'll advance it
3557 * manually if we need to.
3559 iov_iter_restore(iter, state);
3561 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3566 rw = req->async_data;
3568 * Now use our persistent iterator and state, if we aren't already.
3569 * We've restored and mapped the iter to match.
3571 if (iter != &rw->iter) {
3573 state = &rw->iter_state;
3578 * We end up here because of a partial read, either from
3579 * above or inside this loop. Advance the iter by the bytes
3580 * that were consumed.
3582 iov_iter_advance(iter, ret);
3583 if (!iov_iter_count(iter))
3585 rw->bytes_done += ret;
3586 iov_iter_save_state(iter, state);
3588 /* if we can retry, do so with the callbacks armed */
3589 if (!io_rw_should_retry(req)) {
3590 kiocb->ki_flags &= ~IOCB_WAITQ;
3595 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3596 * we get -EIOCBQUEUED, then we'll get a notification when the
3597 * desired page gets unlocked. We can also get a partial read
3598 * here, and if we do, then just retry at the new offset.
3600 ret = io_iter_do_read(req, iter);
3601 if (ret == -EIOCBQUEUED)
3603 /* we got some bytes, but not all. retry. */
3604 kiocb->ki_flags &= ~IOCB_WAITQ;
3605 iov_iter_restore(iter, state);
3608 kiocb_done(kiocb, ret, issue_flags);
3610 /* it's faster to check here then delegate to kfree */
3616 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3618 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3620 return io_prep_rw(req, sqe, WRITE);
3623 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3625 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3626 struct kiocb *kiocb = &req->rw.kiocb;
3627 struct iov_iter __iter, *iter = &__iter;
3628 struct io_async_rw *rw = req->async_data;
3629 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3630 struct iov_iter_state __state, *state;
3635 state = &rw->iter_state;
3636 iov_iter_restore(iter, state);
3639 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3643 iov_iter_save_state(iter, state);
3645 req->result = iov_iter_count(iter);
3647 /* Ensure we clear previously set non-block flag */
3648 if (!force_nonblock)
3649 kiocb->ki_flags &= ~IOCB_NOWAIT;
3651 kiocb->ki_flags |= IOCB_NOWAIT;
3653 /* If the file doesn't support async, just async punt */
3654 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3657 /* file path doesn't support NOWAIT for non-direct_IO */
3658 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3659 (req->flags & REQ_F_ISREG))
3662 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3667 * Open-code file_start_write here to grab freeze protection,
3668 * which will be released by another thread in
3669 * io_complete_rw(). Fool lockdep by telling it the lock got
3670 * released so that it doesn't complain about the held lock when
3671 * we return to userspace.
3673 if (req->flags & REQ_F_ISREG) {
3674 sb_start_write(file_inode(req->file)->i_sb);
3675 __sb_writers_release(file_inode(req->file)->i_sb,
3678 kiocb->ki_flags |= IOCB_WRITE;
3680 if (req->file->f_op->write_iter)
3681 ret2 = call_write_iter(req->file, kiocb, iter);
3682 else if (req->file->f_op->write)
3683 ret2 = loop_rw_iter(WRITE, req, iter);
3687 if (req->flags & REQ_F_REISSUE) {
3688 req->flags &= ~REQ_F_REISSUE;
3693 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3694 * retry them without IOCB_NOWAIT.
3696 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3698 /* no retry on NONBLOCK nor RWF_NOWAIT */
3699 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3701 if (!force_nonblock || ret2 != -EAGAIN) {
3702 /* IOPOLL retry should happen for io-wq threads */
3703 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3706 kiocb_done(kiocb, ret2, issue_flags);
3709 iov_iter_restore(iter, state);
3710 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3711 return ret ?: -EAGAIN;
3714 /* it's reportedly faster than delegating the null check to kfree() */
3720 static int io_renameat_prep(struct io_kiocb *req,
3721 const struct io_uring_sqe *sqe)
3723 struct io_rename *ren = &req->rename;
3724 const char __user *oldf, *newf;
3726 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3728 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3730 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3733 ren->old_dfd = READ_ONCE(sqe->fd);
3734 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3735 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3736 ren->new_dfd = READ_ONCE(sqe->len);
3737 ren->flags = READ_ONCE(sqe->rename_flags);
3739 ren->oldpath = getname(oldf);
3740 if (IS_ERR(ren->oldpath))
3741 return PTR_ERR(ren->oldpath);
3743 ren->newpath = getname(newf);
3744 if (IS_ERR(ren->newpath)) {
3745 putname(ren->oldpath);
3746 return PTR_ERR(ren->newpath);
3749 req->flags |= REQ_F_NEED_CLEANUP;
3753 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3755 struct io_rename *ren = &req->rename;
3758 if (issue_flags & IO_URING_F_NONBLOCK)
3761 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3762 ren->newpath, ren->flags);
3764 req->flags &= ~REQ_F_NEED_CLEANUP;
3767 io_req_complete(req, ret);
3771 static int io_unlinkat_prep(struct io_kiocb *req,
3772 const struct io_uring_sqe *sqe)
3774 struct io_unlink *un = &req->unlink;
3775 const char __user *fname;
3777 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3779 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3782 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3785 un->dfd = READ_ONCE(sqe->fd);
3787 un->flags = READ_ONCE(sqe->unlink_flags);
3788 if (un->flags & ~AT_REMOVEDIR)
3791 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3792 un->filename = getname(fname);
3793 if (IS_ERR(un->filename))
3794 return PTR_ERR(un->filename);
3796 req->flags |= REQ_F_NEED_CLEANUP;
3800 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3802 struct io_unlink *un = &req->unlink;
3805 if (issue_flags & IO_URING_F_NONBLOCK)
3808 if (un->flags & AT_REMOVEDIR)
3809 ret = do_rmdir(un->dfd, un->filename);
3811 ret = do_unlinkat(un->dfd, un->filename);
3813 req->flags &= ~REQ_F_NEED_CLEANUP;
3816 io_req_complete(req, ret);
3820 static int io_mkdirat_prep(struct io_kiocb *req,
3821 const struct io_uring_sqe *sqe)
3823 struct io_mkdir *mkd = &req->mkdir;
3824 const char __user *fname;
3826 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3828 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3831 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3834 mkd->dfd = READ_ONCE(sqe->fd);
3835 mkd->mode = READ_ONCE(sqe->len);
3837 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3838 mkd->filename = getname(fname);
3839 if (IS_ERR(mkd->filename))
3840 return PTR_ERR(mkd->filename);
3842 req->flags |= REQ_F_NEED_CLEANUP;
3846 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3848 struct io_mkdir *mkd = &req->mkdir;
3851 if (issue_flags & IO_URING_F_NONBLOCK)
3854 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3856 req->flags &= ~REQ_F_NEED_CLEANUP;
3859 io_req_complete(req, ret);
3863 static int io_symlinkat_prep(struct io_kiocb *req,
3864 const struct io_uring_sqe *sqe)
3866 struct io_symlink *sl = &req->symlink;
3867 const char __user *oldpath, *newpath;
3869 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3871 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3874 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3877 sl->new_dfd = READ_ONCE(sqe->fd);
3878 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3879 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3881 sl->oldpath = getname(oldpath);
3882 if (IS_ERR(sl->oldpath))
3883 return PTR_ERR(sl->oldpath);
3885 sl->newpath = getname(newpath);
3886 if (IS_ERR(sl->newpath)) {
3887 putname(sl->oldpath);
3888 return PTR_ERR(sl->newpath);
3891 req->flags |= REQ_F_NEED_CLEANUP;
3895 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3897 struct io_symlink *sl = &req->symlink;
3900 if (issue_flags & IO_URING_F_NONBLOCK)
3903 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3905 req->flags &= ~REQ_F_NEED_CLEANUP;
3908 io_req_complete(req, ret);
3912 static int io_linkat_prep(struct io_kiocb *req,
3913 const struct io_uring_sqe *sqe)
3915 struct io_hardlink *lnk = &req->hardlink;
3916 const char __user *oldf, *newf;
3918 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3920 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3922 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3925 lnk->old_dfd = READ_ONCE(sqe->fd);
3926 lnk->new_dfd = READ_ONCE(sqe->len);
3927 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3928 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3929 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3931 lnk->oldpath = getname(oldf);
3932 if (IS_ERR(lnk->oldpath))
3933 return PTR_ERR(lnk->oldpath);
3935 lnk->newpath = getname(newf);
3936 if (IS_ERR(lnk->newpath)) {
3937 putname(lnk->oldpath);
3938 return PTR_ERR(lnk->newpath);
3941 req->flags |= REQ_F_NEED_CLEANUP;
3945 static int io_linkat(struct io_kiocb *req, int issue_flags)
3947 struct io_hardlink *lnk = &req->hardlink;
3950 if (issue_flags & IO_URING_F_NONBLOCK)
3953 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3954 lnk->newpath, lnk->flags);
3956 req->flags &= ~REQ_F_NEED_CLEANUP;
3959 io_req_complete(req, ret);
3963 static int io_shutdown_prep(struct io_kiocb *req,
3964 const struct io_uring_sqe *sqe)
3966 #if defined(CONFIG_NET)
3967 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3969 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3970 sqe->buf_index || sqe->splice_fd_in))
3973 req->shutdown.how = READ_ONCE(sqe->len);
3980 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3982 #if defined(CONFIG_NET)
3983 struct socket *sock;
3986 if (issue_flags & IO_URING_F_NONBLOCK)
3989 sock = sock_from_file(req->file);
3990 if (unlikely(!sock))
3993 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3996 io_req_complete(req, ret);
4003 static int __io_splice_prep(struct io_kiocb *req,
4004 const struct io_uring_sqe *sqe)
4006 struct io_splice *sp = &req->splice;
4007 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4009 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4013 sp->len = READ_ONCE(sqe->len);
4014 sp->flags = READ_ONCE(sqe->splice_flags);
4016 if (unlikely(sp->flags & ~valid_flags))
4019 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
4020 (sp->flags & SPLICE_F_FD_IN_FIXED));
4023 req->flags |= REQ_F_NEED_CLEANUP;
4027 static int io_tee_prep(struct io_kiocb *req,
4028 const struct io_uring_sqe *sqe)
4030 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4032 return __io_splice_prep(req, sqe);
4035 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4037 struct io_splice *sp = &req->splice;
4038 struct file *in = sp->file_in;
4039 struct file *out = sp->file_out;
4040 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4043 if (issue_flags & IO_URING_F_NONBLOCK)
4046 ret = do_tee(in, out, sp->len, flags);
4048 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4050 req->flags &= ~REQ_F_NEED_CLEANUP;
4054 io_req_complete(req, ret);
4058 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4060 struct io_splice *sp = &req->splice;
4062 sp->off_in = READ_ONCE(sqe->splice_off_in);
4063 sp->off_out = READ_ONCE(sqe->off);
4064 return __io_splice_prep(req, sqe);
4067 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4069 struct io_splice *sp = &req->splice;
4070 struct file *in = sp->file_in;
4071 struct file *out = sp->file_out;
4072 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4073 loff_t *poff_in, *poff_out;
4076 if (issue_flags & IO_URING_F_NONBLOCK)
4079 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4080 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4083 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4085 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4087 req->flags &= ~REQ_F_NEED_CLEANUP;
4091 io_req_complete(req, ret);
4096 * IORING_OP_NOP just posts a completion event, nothing else.
4098 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4100 struct io_ring_ctx *ctx = req->ctx;
4102 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4105 __io_req_complete(req, issue_flags, 0, 0);
4109 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4111 struct io_ring_ctx *ctx = req->ctx;
4116 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4118 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4122 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4123 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4126 req->sync.off = READ_ONCE(sqe->off);
4127 req->sync.len = READ_ONCE(sqe->len);
4131 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4133 loff_t end = req->sync.off + req->sync.len;
4136 /* fsync always requires a blocking context */
4137 if (issue_flags & IO_URING_F_NONBLOCK)
4140 ret = vfs_fsync_range(req->file, req->sync.off,
4141 end > 0 ? end : LLONG_MAX,
4142 req->sync.flags & IORING_FSYNC_DATASYNC);
4145 io_req_complete(req, ret);
4149 static int io_fallocate_prep(struct io_kiocb *req,
4150 const struct io_uring_sqe *sqe)
4152 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4155 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4158 req->sync.off = READ_ONCE(sqe->off);
4159 req->sync.len = READ_ONCE(sqe->addr);
4160 req->sync.mode = READ_ONCE(sqe->len);
4164 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4168 /* fallocate always requiring blocking context */
4169 if (issue_flags & IO_URING_F_NONBLOCK)
4171 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4175 io_req_complete(req, ret);
4179 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4181 const char __user *fname;
4184 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4186 if (unlikely(sqe->ioprio || sqe->buf_index))
4188 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4191 /* open.how should be already initialised */
4192 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4193 req->open.how.flags |= O_LARGEFILE;
4195 req->open.dfd = READ_ONCE(sqe->fd);
4196 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4197 req->open.filename = getname(fname);
4198 if (IS_ERR(req->open.filename)) {
4199 ret = PTR_ERR(req->open.filename);
4200 req->open.filename = NULL;
4204 req->open.file_slot = READ_ONCE(sqe->file_index);
4205 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4208 req->open.nofile = rlimit(RLIMIT_NOFILE);
4209 req->flags |= REQ_F_NEED_CLEANUP;
4213 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4215 u64 mode = READ_ONCE(sqe->len);
4216 u64 flags = READ_ONCE(sqe->open_flags);
4218 req->open.how = build_open_how(flags, mode);
4219 return __io_openat_prep(req, sqe);
4222 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4224 struct open_how __user *how;
4228 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4229 len = READ_ONCE(sqe->len);
4230 if (len < OPEN_HOW_SIZE_VER0)
4233 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4238 return __io_openat_prep(req, sqe);
4241 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4243 struct open_flags op;
4245 bool resolve_nonblock, nonblock_set;
4246 bool fixed = !!req->open.file_slot;
4249 ret = build_open_flags(&req->open.how, &op);
4252 nonblock_set = op.open_flag & O_NONBLOCK;
4253 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4254 if (issue_flags & IO_URING_F_NONBLOCK) {
4256 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4257 * it'll always -EAGAIN
4259 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4261 op.lookup_flags |= LOOKUP_CACHED;
4262 op.open_flag |= O_NONBLOCK;
4266 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4271 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4274 * We could hang on to this 'fd' on retrying, but seems like
4275 * marginal gain for something that is now known to be a slower
4276 * path. So just put it, and we'll get a new one when we retry.
4281 ret = PTR_ERR(file);
4282 /* only retry if RESOLVE_CACHED wasn't already set by application */
4283 if (ret == -EAGAIN &&
4284 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4289 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4290 file->f_flags &= ~O_NONBLOCK;
4291 fsnotify_open(file);
4294 fd_install(ret, file);
4296 ret = io_install_fixed_file(req, file, issue_flags,
4297 req->open.file_slot - 1);
4299 putname(req->open.filename);
4300 req->flags &= ~REQ_F_NEED_CLEANUP;
4303 __io_req_complete(req, issue_flags, ret, 0);
4307 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4309 return io_openat2(req, issue_flags);
4312 static int io_remove_buffers_prep(struct io_kiocb *req,
4313 const struct io_uring_sqe *sqe)
4315 struct io_provide_buf *p = &req->pbuf;
4318 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4322 tmp = READ_ONCE(sqe->fd);
4323 if (!tmp || tmp > USHRT_MAX)
4326 memset(p, 0, sizeof(*p));
4328 p->bgid = READ_ONCE(sqe->buf_group);
4332 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4333 int bgid, unsigned nbufs)
4337 /* shouldn't happen */
4341 /* the head kbuf is the list itself */
4342 while (!list_empty(&buf->list)) {
4343 struct io_buffer *nxt;
4345 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4346 list_del(&nxt->list);
4354 xa_erase(&ctx->io_buffers, bgid);
4359 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4361 struct io_provide_buf *p = &req->pbuf;
4362 struct io_ring_ctx *ctx = req->ctx;
4363 struct io_buffer *head;
4365 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4367 io_ring_submit_lock(ctx, !force_nonblock);
4369 lockdep_assert_held(&ctx->uring_lock);
4372 head = xa_load(&ctx->io_buffers, p->bgid);
4374 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4378 /* complete before unlock, IOPOLL may need the lock */
4379 __io_req_complete(req, issue_flags, ret, 0);
4380 io_ring_submit_unlock(ctx, !force_nonblock);
4384 static int io_provide_buffers_prep(struct io_kiocb *req,
4385 const struct io_uring_sqe *sqe)
4387 unsigned long size, tmp_check;
4388 struct io_provide_buf *p = &req->pbuf;
4391 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4394 tmp = READ_ONCE(sqe->fd);
4395 if (!tmp || tmp > USHRT_MAX)
4398 p->addr = READ_ONCE(sqe->addr);
4399 p->len = READ_ONCE(sqe->len);
4401 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4404 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4407 size = (unsigned long)p->len * p->nbufs;
4408 if (!access_ok(u64_to_user_ptr(p->addr), size))
4411 p->bgid = READ_ONCE(sqe->buf_group);
4412 tmp = READ_ONCE(sqe->off);
4413 if (tmp > USHRT_MAX)
4419 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4421 struct io_buffer *buf;
4422 u64 addr = pbuf->addr;
4423 int i, bid = pbuf->bid;
4425 for (i = 0; i < pbuf->nbufs; i++) {
4426 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4431 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4436 INIT_LIST_HEAD(&buf->list);
4439 list_add_tail(&buf->list, &(*head)->list);
4443 return i ? i : -ENOMEM;
4446 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4448 struct io_provide_buf *p = &req->pbuf;
4449 struct io_ring_ctx *ctx = req->ctx;
4450 struct io_buffer *head, *list;
4452 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4454 io_ring_submit_lock(ctx, !force_nonblock);
4456 lockdep_assert_held(&ctx->uring_lock);
4458 list = head = xa_load(&ctx->io_buffers, p->bgid);
4460 ret = io_add_buffers(p, &head);
4461 if (ret >= 0 && !list) {
4462 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4464 __io_remove_buffers(ctx, head, p->bgid, -1U);
4468 /* complete before unlock, IOPOLL may need the lock */
4469 __io_req_complete(req, issue_flags, ret, 0);
4470 io_ring_submit_unlock(ctx, !force_nonblock);
4474 static int io_epoll_ctl_prep(struct io_kiocb *req,
4475 const struct io_uring_sqe *sqe)
4477 #if defined(CONFIG_EPOLL)
4478 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4480 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4483 req->epoll.epfd = READ_ONCE(sqe->fd);
4484 req->epoll.op = READ_ONCE(sqe->len);
4485 req->epoll.fd = READ_ONCE(sqe->off);
4487 if (ep_op_has_event(req->epoll.op)) {
4488 struct epoll_event __user *ev;
4490 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4491 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4501 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4503 #if defined(CONFIG_EPOLL)
4504 struct io_epoll *ie = &req->epoll;
4506 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4508 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4509 if (force_nonblock && ret == -EAGAIN)
4514 __io_req_complete(req, issue_flags, ret, 0);
4521 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4523 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4524 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4526 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4529 req->madvise.addr = READ_ONCE(sqe->addr);
4530 req->madvise.len = READ_ONCE(sqe->len);
4531 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4538 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4540 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4541 struct io_madvise *ma = &req->madvise;
4544 if (issue_flags & IO_URING_F_NONBLOCK)
4547 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4550 io_req_complete(req, ret);
4557 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4559 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4561 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4564 req->fadvise.offset = READ_ONCE(sqe->off);
4565 req->fadvise.len = READ_ONCE(sqe->len);
4566 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4570 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4572 struct io_fadvise *fa = &req->fadvise;
4575 if (issue_flags & IO_URING_F_NONBLOCK) {
4576 switch (fa->advice) {
4577 case POSIX_FADV_NORMAL:
4578 case POSIX_FADV_RANDOM:
4579 case POSIX_FADV_SEQUENTIAL:
4586 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4589 __io_req_complete(req, issue_flags, ret, 0);
4593 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4595 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4597 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4599 if (req->flags & REQ_F_FIXED_FILE)
4602 req->statx.dfd = READ_ONCE(sqe->fd);
4603 req->statx.mask = READ_ONCE(sqe->len);
4604 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4605 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4606 req->statx.flags = READ_ONCE(sqe->statx_flags);
4611 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4613 struct io_statx *ctx = &req->statx;
4616 if (issue_flags & IO_URING_F_NONBLOCK)
4619 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4624 io_req_complete(req, ret);
4628 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4630 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4632 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4633 sqe->rw_flags || sqe->buf_index)
4635 if (req->flags & REQ_F_FIXED_FILE)
4638 req->close.fd = READ_ONCE(sqe->fd);
4639 req->close.file_slot = READ_ONCE(sqe->file_index);
4640 if (req->close.file_slot && req->close.fd)
4646 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4648 struct files_struct *files = current->files;
4649 struct io_close *close = &req->close;
4650 struct fdtable *fdt;
4651 struct file *file = NULL;
4654 if (req->close.file_slot) {
4655 ret = io_close_fixed(req, issue_flags);
4659 spin_lock(&files->file_lock);
4660 fdt = files_fdtable(files);
4661 if (close->fd >= fdt->max_fds) {
4662 spin_unlock(&files->file_lock);
4665 file = fdt->fd[close->fd];
4666 if (!file || file->f_op == &io_uring_fops) {
4667 spin_unlock(&files->file_lock);
4672 /* if the file has a flush method, be safe and punt to async */
4673 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4674 spin_unlock(&files->file_lock);
4678 ret = __close_fd_get_file(close->fd, &file);
4679 spin_unlock(&files->file_lock);
4686 /* No ->flush() or already async, safely close from here */
4687 ret = filp_close(file, current->files);
4693 __io_req_complete(req, issue_flags, ret, 0);
4697 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4699 struct io_ring_ctx *ctx = req->ctx;
4701 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4703 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4707 req->sync.off = READ_ONCE(sqe->off);
4708 req->sync.len = READ_ONCE(sqe->len);
4709 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4713 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4717 /* sync_file_range always requires a blocking context */
4718 if (issue_flags & IO_URING_F_NONBLOCK)
4721 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4725 io_req_complete(req, ret);
4729 #if defined(CONFIG_NET)
4730 static int io_setup_async_msg(struct io_kiocb *req,
4731 struct io_async_msghdr *kmsg)
4733 struct io_async_msghdr *async_msg = req->async_data;
4737 if (io_alloc_async_data(req)) {
4738 kfree(kmsg->free_iov);
4741 async_msg = req->async_data;
4742 req->flags |= REQ_F_NEED_CLEANUP;
4743 memcpy(async_msg, kmsg, sizeof(*kmsg));
4744 async_msg->msg.msg_name = &async_msg->addr;
4745 /* if were using fast_iov, set it to the new one */
4746 if (!async_msg->free_iov)
4747 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4752 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4753 struct io_async_msghdr *iomsg)
4755 iomsg->msg.msg_name = &iomsg->addr;
4756 iomsg->free_iov = iomsg->fast_iov;
4757 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4758 req->sr_msg.msg_flags, &iomsg->free_iov);
4761 static int io_sendmsg_prep_async(struct io_kiocb *req)
4765 ret = io_sendmsg_copy_hdr(req, req->async_data);
4767 req->flags |= REQ_F_NEED_CLEANUP;
4771 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4773 struct io_sr_msg *sr = &req->sr_msg;
4775 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4778 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4779 sr->len = READ_ONCE(sqe->len);
4780 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4781 if (sr->msg_flags & MSG_DONTWAIT)
4782 req->flags |= REQ_F_NOWAIT;
4784 #ifdef CONFIG_COMPAT
4785 if (req->ctx->compat)
4786 sr->msg_flags |= MSG_CMSG_COMPAT;
4791 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4793 struct io_async_msghdr iomsg, *kmsg;
4794 struct socket *sock;
4799 sock = sock_from_file(req->file);
4800 if (unlikely(!sock))
4803 kmsg = req->async_data;
4805 ret = io_sendmsg_copy_hdr(req, &iomsg);
4811 flags = req->sr_msg.msg_flags;
4812 if (issue_flags & IO_URING_F_NONBLOCK)
4813 flags |= MSG_DONTWAIT;
4814 if (flags & MSG_WAITALL)
4815 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4817 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4818 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4819 return io_setup_async_msg(req, kmsg);
4820 if (ret == -ERESTARTSYS)
4823 /* fast path, check for non-NULL to avoid function call */
4825 kfree(kmsg->free_iov);
4826 req->flags &= ~REQ_F_NEED_CLEANUP;
4829 __io_req_complete(req, issue_flags, ret, 0);
4833 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4835 struct io_sr_msg *sr = &req->sr_msg;
4838 struct socket *sock;
4843 sock = sock_from_file(req->file);
4844 if (unlikely(!sock))
4847 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4851 msg.msg_name = NULL;
4852 msg.msg_control = NULL;
4853 msg.msg_controllen = 0;
4854 msg.msg_namelen = 0;
4856 flags = req->sr_msg.msg_flags;
4857 if (issue_flags & IO_URING_F_NONBLOCK)
4858 flags |= MSG_DONTWAIT;
4859 if (flags & MSG_WAITALL)
4860 min_ret = iov_iter_count(&msg.msg_iter);
4862 msg.msg_flags = flags;
4863 ret = sock_sendmsg(sock, &msg);
4864 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4866 if (ret == -ERESTARTSYS)
4871 __io_req_complete(req, issue_flags, ret, 0);
4875 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4876 struct io_async_msghdr *iomsg)
4878 struct io_sr_msg *sr = &req->sr_msg;
4879 struct iovec __user *uiov;
4883 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4884 &iomsg->uaddr, &uiov, &iov_len);
4888 if (req->flags & REQ_F_BUFFER_SELECT) {
4891 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4893 sr->len = iomsg->fast_iov[0].iov_len;
4894 iomsg->free_iov = NULL;
4896 iomsg->free_iov = iomsg->fast_iov;
4897 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4898 &iomsg->free_iov, &iomsg->msg.msg_iter,
4907 #ifdef CONFIG_COMPAT
4908 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4909 struct io_async_msghdr *iomsg)
4911 struct io_sr_msg *sr = &req->sr_msg;
4912 struct compat_iovec __user *uiov;
4917 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4922 uiov = compat_ptr(ptr);
4923 if (req->flags & REQ_F_BUFFER_SELECT) {
4924 compat_ssize_t clen;
4928 if (!access_ok(uiov, sizeof(*uiov)))
4930 if (__get_user(clen, &uiov->iov_len))
4935 iomsg->free_iov = NULL;
4937 iomsg->free_iov = iomsg->fast_iov;
4938 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4939 UIO_FASTIOV, &iomsg->free_iov,
4940 &iomsg->msg.msg_iter, true);
4949 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4950 struct io_async_msghdr *iomsg)
4952 iomsg->msg.msg_name = &iomsg->addr;
4954 #ifdef CONFIG_COMPAT
4955 if (req->ctx->compat)
4956 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4959 return __io_recvmsg_copy_hdr(req, iomsg);
4962 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4965 struct io_sr_msg *sr = &req->sr_msg;
4966 struct io_buffer *kbuf;
4968 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4973 req->flags |= REQ_F_BUFFER_SELECTED;
4977 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4979 return io_put_kbuf(req, req->sr_msg.kbuf);
4982 static int io_recvmsg_prep_async(struct io_kiocb *req)
4986 ret = io_recvmsg_copy_hdr(req, req->async_data);
4988 req->flags |= REQ_F_NEED_CLEANUP;
4992 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4994 struct io_sr_msg *sr = &req->sr_msg;
4996 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4999 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5000 sr->len = READ_ONCE(sqe->len);
5001 sr->bgid = READ_ONCE(sqe->buf_group);
5002 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5003 if (sr->msg_flags & MSG_DONTWAIT)
5004 req->flags |= REQ_F_NOWAIT;
5006 #ifdef CONFIG_COMPAT
5007 if (req->ctx->compat)
5008 sr->msg_flags |= MSG_CMSG_COMPAT;
5013 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5015 struct io_async_msghdr iomsg, *kmsg;
5016 struct socket *sock;
5017 struct io_buffer *kbuf;
5020 int ret, cflags = 0;
5021 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5023 sock = sock_from_file(req->file);
5024 if (unlikely(!sock))
5027 kmsg = req->async_data;
5029 ret = io_recvmsg_copy_hdr(req, &iomsg);
5035 if (req->flags & REQ_F_BUFFER_SELECT) {
5036 kbuf = io_recv_buffer_select(req, !force_nonblock);
5038 return PTR_ERR(kbuf);
5039 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5040 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5041 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5042 1, req->sr_msg.len);
5045 flags = req->sr_msg.msg_flags;
5047 flags |= MSG_DONTWAIT;
5048 if (flags & MSG_WAITALL)
5049 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5051 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5052 kmsg->uaddr, flags);
5053 if (force_nonblock && ret == -EAGAIN)
5054 return io_setup_async_msg(req, kmsg);
5055 if (ret == -ERESTARTSYS)
5058 if (req->flags & REQ_F_BUFFER_SELECTED)
5059 cflags = io_put_recv_kbuf(req);
5060 /* fast path, check for non-NULL to avoid function call */
5062 kfree(kmsg->free_iov);
5063 req->flags &= ~REQ_F_NEED_CLEANUP;
5064 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5066 __io_req_complete(req, issue_flags, ret, cflags);
5070 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5072 struct io_buffer *kbuf;
5073 struct io_sr_msg *sr = &req->sr_msg;
5075 void __user *buf = sr->buf;
5076 struct socket *sock;
5080 int ret, cflags = 0;
5081 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5083 sock = sock_from_file(req->file);
5084 if (unlikely(!sock))
5087 if (req->flags & REQ_F_BUFFER_SELECT) {
5088 kbuf = io_recv_buffer_select(req, !force_nonblock);
5090 return PTR_ERR(kbuf);
5091 buf = u64_to_user_ptr(kbuf->addr);
5094 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5098 msg.msg_name = NULL;
5099 msg.msg_control = NULL;
5100 msg.msg_controllen = 0;
5101 msg.msg_namelen = 0;
5102 msg.msg_iocb = NULL;
5105 flags = req->sr_msg.msg_flags;
5107 flags |= MSG_DONTWAIT;
5108 if (flags & MSG_WAITALL)
5109 min_ret = iov_iter_count(&msg.msg_iter);
5111 ret = sock_recvmsg(sock, &msg, flags);
5112 if (force_nonblock && ret == -EAGAIN)
5114 if (ret == -ERESTARTSYS)
5117 if (req->flags & REQ_F_BUFFER_SELECTED)
5118 cflags = io_put_recv_kbuf(req);
5119 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5121 __io_req_complete(req, issue_flags, ret, cflags);
5125 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5127 struct io_accept *accept = &req->accept;
5129 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5131 if (sqe->ioprio || sqe->len || sqe->buf_index)
5134 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5135 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5136 accept->flags = READ_ONCE(sqe->accept_flags);
5137 accept->nofile = rlimit(RLIMIT_NOFILE);
5139 accept->file_slot = READ_ONCE(sqe->file_index);
5140 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5141 (accept->flags & SOCK_CLOEXEC)))
5143 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5145 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5146 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5150 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5152 struct io_accept *accept = &req->accept;
5153 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5154 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5155 bool fixed = !!accept->file_slot;
5159 if (req->file->f_flags & O_NONBLOCK)
5160 req->flags |= REQ_F_NOWAIT;
5163 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5164 if (unlikely(fd < 0))
5167 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5172 ret = PTR_ERR(file);
5173 if (ret == -EAGAIN && force_nonblock)
5175 if (ret == -ERESTARTSYS)
5178 } else if (!fixed) {
5179 fd_install(fd, file);
5182 ret = io_install_fixed_file(req, file, issue_flags,
5183 accept->file_slot - 1);
5185 __io_req_complete(req, issue_flags, ret, 0);
5189 static int io_connect_prep_async(struct io_kiocb *req)
5191 struct io_async_connect *io = req->async_data;
5192 struct io_connect *conn = &req->connect;
5194 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5197 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5199 struct io_connect *conn = &req->connect;
5201 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5203 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5207 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5208 conn->addr_len = READ_ONCE(sqe->addr2);
5212 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5214 struct io_async_connect __io, *io;
5215 unsigned file_flags;
5217 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5219 if (req->async_data) {
5220 io = req->async_data;
5222 ret = move_addr_to_kernel(req->connect.addr,
5223 req->connect.addr_len,
5230 file_flags = force_nonblock ? O_NONBLOCK : 0;
5232 ret = __sys_connect_file(req->file, &io->address,
5233 req->connect.addr_len, file_flags);
5234 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5235 if (req->async_data)
5237 if (io_alloc_async_data(req)) {
5241 memcpy(req->async_data, &__io, sizeof(__io));
5244 if (ret == -ERESTARTSYS)
5249 __io_req_complete(req, issue_flags, ret, 0);
5252 #else /* !CONFIG_NET */
5253 #define IO_NETOP_FN(op) \
5254 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5256 return -EOPNOTSUPP; \
5259 #define IO_NETOP_PREP(op) \
5261 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5263 return -EOPNOTSUPP; \
5266 #define IO_NETOP_PREP_ASYNC(op) \
5268 static int io_##op##_prep_async(struct io_kiocb *req) \
5270 return -EOPNOTSUPP; \
5273 IO_NETOP_PREP_ASYNC(sendmsg);
5274 IO_NETOP_PREP_ASYNC(recvmsg);
5275 IO_NETOP_PREP_ASYNC(connect);
5276 IO_NETOP_PREP(accept);
5279 #endif /* CONFIG_NET */
5281 struct io_poll_table {
5282 struct poll_table_struct pt;
5283 struct io_kiocb *req;
5288 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5289 __poll_t mask, io_req_tw_func_t func)
5291 /* for instances that support it check for an event match first: */
5292 if (mask && !(mask & poll->events))
5295 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5297 list_del_init(&poll->wait.entry);
5300 req->io_task_work.func = func;
5303 * If this fails, then the task is exiting. When a task exits, the
5304 * work gets canceled, so just cancel this request as well instead
5305 * of executing it. We can't safely execute it anyway, as we may not
5306 * have the needed state needed for it anyway.
5308 io_req_task_work_add(req);
5312 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5313 __acquires(&req->ctx->completion_lock)
5315 struct io_ring_ctx *ctx = req->ctx;
5317 /* req->task == current here, checking PF_EXITING is safe */
5318 if (unlikely(req->task->flags & PF_EXITING))
5319 WRITE_ONCE(poll->canceled, true);
5321 if (!req->result && !READ_ONCE(poll->canceled)) {
5322 struct poll_table_struct pt = { ._key = poll->events };
5324 req->result = vfs_poll(req->file, &pt) & poll->events;
5327 spin_lock(&ctx->completion_lock);
5328 if (!req->result && !READ_ONCE(poll->canceled)) {
5329 add_wait_queue(poll->head, &poll->wait);
5336 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5338 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5339 if (req->opcode == IORING_OP_POLL_ADD)
5340 return req->async_data;
5341 return req->apoll->double_poll;
5344 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5346 if (req->opcode == IORING_OP_POLL_ADD)
5348 return &req->apoll->poll;
5351 static void io_poll_remove_double(struct io_kiocb *req)
5352 __must_hold(&req->ctx->completion_lock)
5354 struct io_poll_iocb *poll = io_poll_get_double(req);
5356 lockdep_assert_held(&req->ctx->completion_lock);
5358 if (poll && poll->head) {
5359 struct wait_queue_head *head = poll->head;
5361 spin_lock_irq(&head->lock);
5362 list_del_init(&poll->wait.entry);
5363 if (poll->wait.private)
5366 spin_unlock_irq(&head->lock);
5370 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5371 __must_hold(&req->ctx->completion_lock)
5373 struct io_ring_ctx *ctx = req->ctx;
5374 unsigned flags = IORING_CQE_F_MORE;
5377 if (READ_ONCE(req->poll.canceled)) {
5379 req->poll.events |= EPOLLONESHOT;
5381 error = mangle_poll(mask);
5383 if (req->poll.events & EPOLLONESHOT)
5385 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5386 req->poll.events |= EPOLLONESHOT;
5389 if (flags & IORING_CQE_F_MORE)
5392 return !(flags & IORING_CQE_F_MORE);
5395 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5396 __must_hold(&req->ctx->completion_lock)
5400 done = __io_poll_complete(req, mask);
5401 io_commit_cqring(req->ctx);
5405 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5407 struct io_ring_ctx *ctx = req->ctx;
5408 struct io_kiocb *nxt;
5410 if (io_poll_rewait(req, &req->poll)) {
5411 spin_unlock(&ctx->completion_lock);
5415 if (req->poll.done) {
5416 spin_unlock(&ctx->completion_lock);
5419 done = __io_poll_complete(req, req->result);
5421 io_poll_remove_double(req);
5422 hash_del(&req->hash_node);
5423 req->poll.done = true;
5426 add_wait_queue(req->poll.head, &req->poll.wait);
5428 io_commit_cqring(ctx);
5429 spin_unlock(&ctx->completion_lock);
5430 io_cqring_ev_posted(ctx);
5433 nxt = io_put_req_find_next(req);
5435 io_req_task_submit(nxt, locked);
5440 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5441 int sync, void *key)
5443 struct io_kiocb *req = wait->private;
5444 struct io_poll_iocb *poll = io_poll_get_single(req);
5445 __poll_t mask = key_to_poll(key);
5446 unsigned long flags;
5448 /* for instances that support it check for an event match first: */
5449 if (mask && !(mask & poll->events))
5451 if (!(poll->events & EPOLLONESHOT))
5452 return poll->wait.func(&poll->wait, mode, sync, key);
5454 list_del_init(&wait->entry);
5459 spin_lock_irqsave(&poll->head->lock, flags);
5460 done = list_empty(&poll->wait.entry);
5462 list_del_init(&poll->wait.entry);
5463 /* make sure double remove sees this as being gone */
5464 wait->private = NULL;
5465 spin_unlock_irqrestore(&poll->head->lock, flags);
5467 /* use wait func handler, so it matches the rq type */
5468 poll->wait.func(&poll->wait, mode, sync, key);
5475 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5476 wait_queue_func_t wake_func)
5480 poll->canceled = false;
5481 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5482 /* mask in events that we always want/need */
5483 poll->events = events | IO_POLL_UNMASK;
5484 INIT_LIST_HEAD(&poll->wait.entry);
5485 init_waitqueue_func_entry(&poll->wait, wake_func);
5488 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5489 struct wait_queue_head *head,
5490 struct io_poll_iocb **poll_ptr)
5492 struct io_kiocb *req = pt->req;
5495 * The file being polled uses multiple waitqueues for poll handling
5496 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5499 if (unlikely(pt->nr_entries)) {
5500 struct io_poll_iocb *poll_one = poll;
5502 /* double add on the same waitqueue head, ignore */
5503 if (poll_one->head == head)
5505 /* already have a 2nd entry, fail a third attempt */
5507 if ((*poll_ptr)->head == head)
5509 pt->error = -EINVAL;
5513 * Can't handle multishot for double wait for now, turn it
5514 * into one-shot mode.
5516 if (!(poll_one->events & EPOLLONESHOT))
5517 poll_one->events |= EPOLLONESHOT;
5518 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5520 pt->error = -ENOMEM;
5523 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5525 poll->wait.private = req;
5532 if (poll->events & EPOLLEXCLUSIVE)
5533 add_wait_queue_exclusive(head, &poll->wait);
5535 add_wait_queue(head, &poll->wait);
5538 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5539 struct poll_table_struct *p)
5541 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5542 struct async_poll *apoll = pt->req->apoll;
5544 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5547 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5549 struct async_poll *apoll = req->apoll;
5550 struct io_ring_ctx *ctx = req->ctx;
5552 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5554 if (io_poll_rewait(req, &apoll->poll)) {
5555 spin_unlock(&ctx->completion_lock);
5559 hash_del(&req->hash_node);
5560 io_poll_remove_double(req);
5561 apoll->poll.done = true;
5562 spin_unlock(&ctx->completion_lock);
5564 if (!READ_ONCE(apoll->poll.canceled))
5565 io_req_task_submit(req, locked);
5567 io_req_complete_failed(req, -ECANCELED);
5570 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5573 struct io_kiocb *req = wait->private;
5574 struct io_poll_iocb *poll = &req->apoll->poll;
5576 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5579 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5582 static void io_poll_req_insert(struct io_kiocb *req)
5584 struct io_ring_ctx *ctx = req->ctx;
5585 struct hlist_head *list;
5587 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5588 hlist_add_head(&req->hash_node, list);
5591 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5592 struct io_poll_iocb *poll,
5593 struct io_poll_table *ipt, __poll_t mask,
5594 wait_queue_func_t wake_func)
5595 __acquires(&ctx->completion_lock)
5597 struct io_ring_ctx *ctx = req->ctx;
5598 bool cancel = false;
5600 INIT_HLIST_NODE(&req->hash_node);
5601 io_init_poll_iocb(poll, mask, wake_func);
5602 poll->file = req->file;
5603 poll->wait.private = req;
5605 ipt->pt._key = mask;
5608 ipt->nr_entries = 0;
5610 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5611 if (unlikely(!ipt->nr_entries) && !ipt->error)
5612 ipt->error = -EINVAL;
5614 spin_lock(&ctx->completion_lock);
5615 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5616 io_poll_remove_double(req);
5617 if (likely(poll->head)) {
5618 spin_lock_irq(&poll->head->lock);
5619 if (unlikely(list_empty(&poll->wait.entry))) {
5625 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5626 list_del_init(&poll->wait.entry);
5628 WRITE_ONCE(poll->canceled, true);
5629 else if (!poll->done) /* actually waiting for an event */
5630 io_poll_req_insert(req);
5631 spin_unlock_irq(&poll->head->lock);
5643 static int io_arm_poll_handler(struct io_kiocb *req)
5645 const struct io_op_def *def = &io_op_defs[req->opcode];
5646 struct io_ring_ctx *ctx = req->ctx;
5647 struct async_poll *apoll;
5648 struct io_poll_table ipt;
5649 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5652 if (!req->file || !file_can_poll(req->file))
5653 return IO_APOLL_ABORTED;
5654 if (req->flags & REQ_F_POLLED)
5655 return IO_APOLL_ABORTED;
5656 if (!def->pollin && !def->pollout)
5657 return IO_APOLL_ABORTED;
5661 mask |= POLLIN | POLLRDNORM;
5663 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5664 if ((req->opcode == IORING_OP_RECVMSG) &&
5665 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5669 mask |= POLLOUT | POLLWRNORM;
5672 /* if we can't nonblock try, then no point in arming a poll handler */
5673 if (!io_file_supports_nowait(req, rw))
5674 return IO_APOLL_ABORTED;
5676 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5677 if (unlikely(!apoll))
5678 return IO_APOLL_ABORTED;
5679 apoll->double_poll = NULL;
5681 req->flags |= REQ_F_POLLED;
5682 ipt.pt._qproc = io_async_queue_proc;
5683 io_req_set_refcount(req);
5685 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5687 spin_unlock(&ctx->completion_lock);
5688 if (ret || ipt.error)
5689 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5691 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5692 mask, apoll->poll.events);
5696 static bool __io_poll_remove_one(struct io_kiocb *req,
5697 struct io_poll_iocb *poll, bool do_cancel)
5698 __must_hold(&req->ctx->completion_lock)
5700 bool do_complete = false;
5704 spin_lock_irq(&poll->head->lock);
5706 WRITE_ONCE(poll->canceled, true);
5707 if (!list_empty(&poll->wait.entry)) {
5708 list_del_init(&poll->wait.entry);
5711 spin_unlock_irq(&poll->head->lock);
5712 hash_del(&req->hash_node);
5716 static bool io_poll_remove_one(struct io_kiocb *req)
5717 __must_hold(&req->ctx->completion_lock)
5721 io_poll_remove_double(req);
5722 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5725 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5726 io_commit_cqring(req->ctx);
5728 io_put_req_deferred(req);
5734 * Returns true if we found and killed one or more poll requests
5736 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5739 struct hlist_node *tmp;
5740 struct io_kiocb *req;
5743 spin_lock(&ctx->completion_lock);
5744 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5745 struct hlist_head *list;
5747 list = &ctx->cancel_hash[i];
5748 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5749 if (io_match_task_safe(req, tsk, cancel_all))
5750 posted += io_poll_remove_one(req);
5753 spin_unlock(&ctx->completion_lock);
5756 io_cqring_ev_posted(ctx);
5761 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5763 __must_hold(&ctx->completion_lock)
5765 struct hlist_head *list;
5766 struct io_kiocb *req;
5768 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5769 hlist_for_each_entry(req, list, hash_node) {
5770 if (sqe_addr != req->user_data)
5772 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5779 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5781 __must_hold(&ctx->completion_lock)
5783 struct io_kiocb *req;
5785 req = io_poll_find(ctx, sqe_addr, poll_only);
5788 if (io_poll_remove_one(req))
5794 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5799 events = READ_ONCE(sqe->poll32_events);
5801 events = swahw32(events);
5803 if (!(flags & IORING_POLL_ADD_MULTI))
5804 events |= EPOLLONESHOT;
5805 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5808 static int io_poll_update_prep(struct io_kiocb *req,
5809 const struct io_uring_sqe *sqe)
5811 struct io_poll_update *upd = &req->poll_update;
5814 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5816 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5818 flags = READ_ONCE(sqe->len);
5819 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5820 IORING_POLL_ADD_MULTI))
5822 /* meaningless without update */
5823 if (flags == IORING_POLL_ADD_MULTI)
5826 upd->old_user_data = READ_ONCE(sqe->addr);
5827 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5828 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5830 upd->new_user_data = READ_ONCE(sqe->off);
5831 if (!upd->update_user_data && upd->new_user_data)
5833 if (upd->update_events)
5834 upd->events = io_poll_parse_events(sqe, flags);
5835 else if (sqe->poll32_events)
5841 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5844 struct io_kiocb *req = wait->private;
5845 struct io_poll_iocb *poll = &req->poll;
5847 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5850 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5851 struct poll_table_struct *p)
5853 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5855 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5858 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5860 struct io_poll_iocb *poll = &req->poll;
5863 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5865 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5867 flags = READ_ONCE(sqe->len);
5868 if (flags & ~IORING_POLL_ADD_MULTI)
5871 io_req_set_refcount(req);
5872 poll->events = io_poll_parse_events(sqe, flags);
5876 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5878 struct io_poll_iocb *poll = &req->poll;
5879 struct io_ring_ctx *ctx = req->ctx;
5880 struct io_poll_table ipt;
5884 ipt.pt._qproc = io_poll_queue_proc;
5886 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5889 if (mask) { /* no async, we'd stolen it */
5891 done = io_poll_complete(req, mask);
5893 spin_unlock(&ctx->completion_lock);
5896 io_cqring_ev_posted(ctx);
5903 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5905 struct io_ring_ctx *ctx = req->ctx;
5906 struct io_kiocb *preq;
5910 spin_lock(&ctx->completion_lock);
5911 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5917 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5919 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5924 * Don't allow racy completion with singleshot, as we cannot safely
5925 * update those. For multishot, if we're racing with completion, just
5926 * let completion re-add it.
5928 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5929 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5933 /* we now have a detached poll request. reissue. */
5937 spin_unlock(&ctx->completion_lock);
5939 io_req_complete(req, ret);
5942 /* only mask one event flags, keep behavior flags */
5943 if (req->poll_update.update_events) {
5944 preq->poll.events &= ~0xffff;
5945 preq->poll.events |= req->poll_update.events & 0xffff;
5946 preq->poll.events |= IO_POLL_UNMASK;
5948 if (req->poll_update.update_user_data)
5949 preq->user_data = req->poll_update.new_user_data;
5950 spin_unlock(&ctx->completion_lock);
5952 /* complete update request, we're done with it */
5953 io_req_complete(req, ret);
5956 ret = io_poll_add(preq, issue_flags);
5959 io_req_complete(preq, ret);
5965 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5968 io_req_complete_post(req, -ETIME, 0);
5971 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5973 struct io_timeout_data *data = container_of(timer,
5974 struct io_timeout_data, timer);
5975 struct io_kiocb *req = data->req;
5976 struct io_ring_ctx *ctx = req->ctx;
5977 unsigned long flags;
5979 spin_lock_irqsave(&ctx->timeout_lock, flags);
5980 list_del_init(&req->timeout.list);
5981 atomic_set(&req->ctx->cq_timeouts,
5982 atomic_read(&req->ctx->cq_timeouts) + 1);
5983 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5985 req->io_task_work.func = io_req_task_timeout;
5986 io_req_task_work_add(req);
5987 return HRTIMER_NORESTART;
5990 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5992 __must_hold(&ctx->timeout_lock)
5994 struct io_timeout_data *io;
5995 struct io_kiocb *req;
5998 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5999 found = user_data == req->user_data;
6004 return ERR_PTR(-ENOENT);
6006 io = req->async_data;
6007 if (hrtimer_try_to_cancel(&io->timer) == -1)
6008 return ERR_PTR(-EALREADY);
6009 list_del_init(&req->timeout.list);
6013 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
6014 __must_hold(&ctx->completion_lock)
6015 __must_hold(&ctx->timeout_lock)
6017 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6020 return PTR_ERR(req);
6023 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
6024 io_put_req_deferred(req);
6028 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6030 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6031 case IORING_TIMEOUT_BOOTTIME:
6032 return CLOCK_BOOTTIME;
6033 case IORING_TIMEOUT_REALTIME:
6034 return CLOCK_REALTIME;
6036 /* can't happen, vetted at prep time */
6040 return CLOCK_MONOTONIC;
6044 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6045 struct timespec64 *ts, enum hrtimer_mode mode)
6046 __must_hold(&ctx->timeout_lock)
6048 struct io_timeout_data *io;
6049 struct io_kiocb *req;
6052 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6053 found = user_data == req->user_data;
6060 io = req->async_data;
6061 if (hrtimer_try_to_cancel(&io->timer) == -1)
6063 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6064 io->timer.function = io_link_timeout_fn;
6065 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6069 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6070 struct timespec64 *ts, enum hrtimer_mode mode)
6071 __must_hold(&ctx->timeout_lock)
6073 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6074 struct io_timeout_data *data;
6077 return PTR_ERR(req);
6079 req->timeout.off = 0; /* noseq */
6080 data = req->async_data;
6081 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6082 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6083 data->timer.function = io_timeout_fn;
6084 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6088 static int io_timeout_remove_prep(struct io_kiocb *req,
6089 const struct io_uring_sqe *sqe)
6091 struct io_timeout_rem *tr = &req->timeout_rem;
6093 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6095 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6097 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6100 tr->ltimeout = false;
6101 tr->addr = READ_ONCE(sqe->addr);
6102 tr->flags = READ_ONCE(sqe->timeout_flags);
6103 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6104 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6106 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6107 tr->ltimeout = true;
6108 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6110 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6112 } else if (tr->flags) {
6113 /* timeout removal doesn't support flags */
6120 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6122 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6127 * Remove or update an existing timeout command
6129 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6131 struct io_timeout_rem *tr = &req->timeout_rem;
6132 struct io_ring_ctx *ctx = req->ctx;
6135 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6136 spin_lock(&ctx->completion_lock);
6137 spin_lock_irq(&ctx->timeout_lock);
6138 ret = io_timeout_cancel(ctx, tr->addr);
6139 spin_unlock_irq(&ctx->timeout_lock);
6140 spin_unlock(&ctx->completion_lock);
6142 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6144 spin_lock_irq(&ctx->timeout_lock);
6146 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6148 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6149 spin_unlock_irq(&ctx->timeout_lock);
6154 io_req_complete_post(req, ret, 0);
6158 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6159 bool is_timeout_link)
6161 struct io_timeout_data *data;
6163 u32 off = READ_ONCE(sqe->off);
6165 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6167 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6170 if (off && is_timeout_link)
6172 flags = READ_ONCE(sqe->timeout_flags);
6173 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6175 /* more than one clock specified is invalid, obviously */
6176 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6179 INIT_LIST_HEAD(&req->timeout.list);
6180 req->timeout.off = off;
6181 if (unlikely(off && !req->ctx->off_timeout_used))
6182 req->ctx->off_timeout_used = true;
6184 if (!req->async_data && io_alloc_async_data(req))
6187 data = req->async_data;
6189 data->flags = flags;
6191 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6194 data->mode = io_translate_timeout_mode(flags);
6195 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6197 if (is_timeout_link) {
6198 struct io_submit_link *link = &req->ctx->submit_state.link;
6202 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6204 req->timeout.head = link->last;
6205 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6210 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6212 struct io_ring_ctx *ctx = req->ctx;
6213 struct io_timeout_data *data = req->async_data;
6214 struct list_head *entry;
6215 u32 tail, off = req->timeout.off;
6217 spin_lock_irq(&ctx->timeout_lock);
6220 * sqe->off holds how many events that need to occur for this
6221 * timeout event to be satisfied. If it isn't set, then this is
6222 * a pure timeout request, sequence isn't used.
6224 if (io_is_timeout_noseq(req)) {
6225 entry = ctx->timeout_list.prev;
6229 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6230 req->timeout.target_seq = tail + off;
6232 /* Update the last seq here in case io_flush_timeouts() hasn't.
6233 * This is safe because ->completion_lock is held, and submissions
6234 * and completions are never mixed in the same ->completion_lock section.
6236 ctx->cq_last_tm_flush = tail;
6239 * Insertion sort, ensuring the first entry in the list is always
6240 * the one we need first.
6242 list_for_each_prev(entry, &ctx->timeout_list) {
6243 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6246 if (io_is_timeout_noseq(nxt))
6248 /* nxt.seq is behind @tail, otherwise would've been completed */
6249 if (off >= nxt->timeout.target_seq - tail)
6253 list_add(&req->timeout.list, entry);
6254 data->timer.function = io_timeout_fn;
6255 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6256 spin_unlock_irq(&ctx->timeout_lock);
6260 struct io_cancel_data {
6261 struct io_ring_ctx *ctx;
6265 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6267 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6268 struct io_cancel_data *cd = data;
6270 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6273 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6274 struct io_ring_ctx *ctx)
6276 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6277 enum io_wq_cancel cancel_ret;
6280 if (!tctx || !tctx->io_wq)
6283 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6284 switch (cancel_ret) {
6285 case IO_WQ_CANCEL_OK:
6288 case IO_WQ_CANCEL_RUNNING:
6291 case IO_WQ_CANCEL_NOTFOUND:
6299 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6301 struct io_ring_ctx *ctx = req->ctx;
6304 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6306 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6310 spin_lock(&ctx->completion_lock);
6311 spin_lock_irq(&ctx->timeout_lock);
6312 ret = io_timeout_cancel(ctx, sqe_addr);
6313 spin_unlock_irq(&ctx->timeout_lock);
6316 ret = io_poll_cancel(ctx, sqe_addr, false);
6318 spin_unlock(&ctx->completion_lock);
6322 static int io_async_cancel_prep(struct io_kiocb *req,
6323 const struct io_uring_sqe *sqe)
6325 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6327 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6329 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6333 req->cancel.addr = READ_ONCE(sqe->addr);
6337 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6339 struct io_ring_ctx *ctx = req->ctx;
6340 u64 sqe_addr = req->cancel.addr;
6341 struct io_tctx_node *node;
6344 ret = io_try_cancel_userdata(req, sqe_addr);
6348 /* slow path, try all io-wq's */
6349 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6351 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6352 struct io_uring_task *tctx = node->task->io_uring;
6354 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6358 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6362 io_req_complete_post(req, ret, 0);
6366 static int io_rsrc_update_prep(struct io_kiocb *req,
6367 const struct io_uring_sqe *sqe)
6369 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6371 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6374 req->rsrc_update.offset = READ_ONCE(sqe->off);
6375 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6376 if (!req->rsrc_update.nr_args)
6378 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6382 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6384 struct io_ring_ctx *ctx = req->ctx;
6385 struct io_uring_rsrc_update2 up;
6388 up.offset = req->rsrc_update.offset;
6389 up.data = req->rsrc_update.arg;
6394 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6395 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6396 &up, req->rsrc_update.nr_args);
6397 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6401 __io_req_complete(req, issue_flags, ret, 0);
6405 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6407 switch (req->opcode) {
6410 case IORING_OP_READV:
6411 case IORING_OP_READ_FIXED:
6412 case IORING_OP_READ:
6413 return io_read_prep(req, sqe);
6414 case IORING_OP_WRITEV:
6415 case IORING_OP_WRITE_FIXED:
6416 case IORING_OP_WRITE:
6417 return io_write_prep(req, sqe);
6418 case IORING_OP_POLL_ADD:
6419 return io_poll_add_prep(req, sqe);
6420 case IORING_OP_POLL_REMOVE:
6421 return io_poll_update_prep(req, sqe);
6422 case IORING_OP_FSYNC:
6423 return io_fsync_prep(req, sqe);
6424 case IORING_OP_SYNC_FILE_RANGE:
6425 return io_sfr_prep(req, sqe);
6426 case IORING_OP_SENDMSG:
6427 case IORING_OP_SEND:
6428 return io_sendmsg_prep(req, sqe);
6429 case IORING_OP_RECVMSG:
6430 case IORING_OP_RECV:
6431 return io_recvmsg_prep(req, sqe);
6432 case IORING_OP_CONNECT:
6433 return io_connect_prep(req, sqe);
6434 case IORING_OP_TIMEOUT:
6435 return io_timeout_prep(req, sqe, false);
6436 case IORING_OP_TIMEOUT_REMOVE:
6437 return io_timeout_remove_prep(req, sqe);
6438 case IORING_OP_ASYNC_CANCEL:
6439 return io_async_cancel_prep(req, sqe);
6440 case IORING_OP_LINK_TIMEOUT:
6441 return io_timeout_prep(req, sqe, true);
6442 case IORING_OP_ACCEPT:
6443 return io_accept_prep(req, sqe);
6444 case IORING_OP_FALLOCATE:
6445 return io_fallocate_prep(req, sqe);
6446 case IORING_OP_OPENAT:
6447 return io_openat_prep(req, sqe);
6448 case IORING_OP_CLOSE:
6449 return io_close_prep(req, sqe);
6450 case IORING_OP_FILES_UPDATE:
6451 return io_rsrc_update_prep(req, sqe);
6452 case IORING_OP_STATX:
6453 return io_statx_prep(req, sqe);
6454 case IORING_OP_FADVISE:
6455 return io_fadvise_prep(req, sqe);
6456 case IORING_OP_MADVISE:
6457 return io_madvise_prep(req, sqe);
6458 case IORING_OP_OPENAT2:
6459 return io_openat2_prep(req, sqe);
6460 case IORING_OP_EPOLL_CTL:
6461 return io_epoll_ctl_prep(req, sqe);
6462 case IORING_OP_SPLICE:
6463 return io_splice_prep(req, sqe);
6464 case IORING_OP_PROVIDE_BUFFERS:
6465 return io_provide_buffers_prep(req, sqe);
6466 case IORING_OP_REMOVE_BUFFERS:
6467 return io_remove_buffers_prep(req, sqe);
6469 return io_tee_prep(req, sqe);
6470 case IORING_OP_SHUTDOWN:
6471 return io_shutdown_prep(req, sqe);
6472 case IORING_OP_RENAMEAT:
6473 return io_renameat_prep(req, sqe);
6474 case IORING_OP_UNLINKAT:
6475 return io_unlinkat_prep(req, sqe);
6476 case IORING_OP_MKDIRAT:
6477 return io_mkdirat_prep(req, sqe);
6478 case IORING_OP_SYMLINKAT:
6479 return io_symlinkat_prep(req, sqe);
6480 case IORING_OP_LINKAT:
6481 return io_linkat_prep(req, sqe);
6484 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6489 static int io_req_prep_async(struct io_kiocb *req)
6491 if (!io_op_defs[req->opcode].needs_async_setup)
6493 if (WARN_ON_ONCE(req->async_data))
6495 if (io_alloc_async_data(req))
6498 switch (req->opcode) {
6499 case IORING_OP_READV:
6500 return io_rw_prep_async(req, READ);
6501 case IORING_OP_WRITEV:
6502 return io_rw_prep_async(req, WRITE);
6503 case IORING_OP_SENDMSG:
6504 return io_sendmsg_prep_async(req);
6505 case IORING_OP_RECVMSG:
6506 return io_recvmsg_prep_async(req);
6507 case IORING_OP_CONNECT:
6508 return io_connect_prep_async(req);
6510 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6515 static u32 io_get_sequence(struct io_kiocb *req)
6517 u32 seq = req->ctx->cached_sq_head;
6519 /* need original cached_sq_head, but it was increased for each req */
6520 io_for_each_link(req, req)
6525 static bool io_drain_req(struct io_kiocb *req)
6527 struct io_kiocb *pos;
6528 struct io_ring_ctx *ctx = req->ctx;
6529 struct io_defer_entry *de;
6533 if (req->flags & REQ_F_FAIL) {
6534 io_req_complete_fail_submit(req);
6539 * If we need to drain a request in the middle of a link, drain the
6540 * head request and the next request/link after the current link.
6541 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6542 * maintained for every request of our link.
6544 if (ctx->drain_next) {
6545 req->flags |= REQ_F_IO_DRAIN;
6546 ctx->drain_next = false;
6548 /* not interested in head, start from the first linked */
6549 io_for_each_link(pos, req->link) {
6550 if (pos->flags & REQ_F_IO_DRAIN) {
6551 ctx->drain_next = true;
6552 req->flags |= REQ_F_IO_DRAIN;
6557 /* Still need defer if there is pending req in defer list. */
6558 if (likely(list_empty_careful(&ctx->defer_list) &&
6559 !(req->flags & REQ_F_IO_DRAIN))) {
6560 ctx->drain_active = false;
6564 seq = io_get_sequence(req);
6565 /* Still a chance to pass the sequence check */
6566 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6569 ret = io_req_prep_async(req);
6572 io_prep_async_link(req);
6573 de = kmalloc(sizeof(*de), GFP_KERNEL);
6577 io_req_complete_failed(req, ret);
6581 spin_lock(&ctx->completion_lock);
6582 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6583 spin_unlock(&ctx->completion_lock);
6585 io_queue_async_work(req, NULL);
6589 trace_io_uring_defer(ctx, req, req->user_data);
6592 list_add_tail(&de->list, &ctx->defer_list);
6593 spin_unlock(&ctx->completion_lock);
6597 static void io_clean_op(struct io_kiocb *req)
6599 if (req->flags & REQ_F_BUFFER_SELECTED) {
6600 switch (req->opcode) {
6601 case IORING_OP_READV:
6602 case IORING_OP_READ_FIXED:
6603 case IORING_OP_READ:
6604 kfree((void *)(unsigned long)req->rw.addr);
6606 case IORING_OP_RECVMSG:
6607 case IORING_OP_RECV:
6608 kfree(req->sr_msg.kbuf);
6613 if (req->flags & REQ_F_NEED_CLEANUP) {
6614 switch (req->opcode) {
6615 case IORING_OP_READV:
6616 case IORING_OP_READ_FIXED:
6617 case IORING_OP_READ:
6618 case IORING_OP_WRITEV:
6619 case IORING_OP_WRITE_FIXED:
6620 case IORING_OP_WRITE: {
6621 struct io_async_rw *io = req->async_data;
6623 kfree(io->free_iovec);
6626 case IORING_OP_RECVMSG:
6627 case IORING_OP_SENDMSG: {
6628 struct io_async_msghdr *io = req->async_data;
6630 kfree(io->free_iov);
6633 case IORING_OP_SPLICE:
6635 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6636 io_put_file(req->splice.file_in);
6638 case IORING_OP_OPENAT:
6639 case IORING_OP_OPENAT2:
6640 if (req->open.filename)
6641 putname(req->open.filename);
6643 case IORING_OP_RENAMEAT:
6644 putname(req->rename.oldpath);
6645 putname(req->rename.newpath);
6647 case IORING_OP_UNLINKAT:
6648 putname(req->unlink.filename);
6650 case IORING_OP_MKDIRAT:
6651 putname(req->mkdir.filename);
6653 case IORING_OP_SYMLINKAT:
6654 putname(req->symlink.oldpath);
6655 putname(req->symlink.newpath);
6657 case IORING_OP_LINKAT:
6658 putname(req->hardlink.oldpath);
6659 putname(req->hardlink.newpath);
6663 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6664 kfree(req->apoll->double_poll);
6668 if (req->flags & REQ_F_INFLIGHT) {
6669 struct io_uring_task *tctx = req->task->io_uring;
6671 atomic_dec(&tctx->inflight_tracked);
6673 if (req->flags & REQ_F_CREDS)
6674 put_cred(req->creds);
6676 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6679 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6681 struct io_ring_ctx *ctx = req->ctx;
6682 const struct cred *creds = NULL;
6685 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6686 creds = override_creds(req->creds);
6688 switch (req->opcode) {
6690 ret = io_nop(req, issue_flags);
6692 case IORING_OP_READV:
6693 case IORING_OP_READ_FIXED:
6694 case IORING_OP_READ:
6695 ret = io_read(req, issue_flags);
6697 case IORING_OP_WRITEV:
6698 case IORING_OP_WRITE_FIXED:
6699 case IORING_OP_WRITE:
6700 ret = io_write(req, issue_flags);
6702 case IORING_OP_FSYNC:
6703 ret = io_fsync(req, issue_flags);
6705 case IORING_OP_POLL_ADD:
6706 ret = io_poll_add(req, issue_flags);
6708 case IORING_OP_POLL_REMOVE:
6709 ret = io_poll_update(req, issue_flags);
6711 case IORING_OP_SYNC_FILE_RANGE:
6712 ret = io_sync_file_range(req, issue_flags);
6714 case IORING_OP_SENDMSG:
6715 ret = io_sendmsg(req, issue_flags);
6717 case IORING_OP_SEND:
6718 ret = io_send(req, issue_flags);
6720 case IORING_OP_RECVMSG:
6721 ret = io_recvmsg(req, issue_flags);
6723 case IORING_OP_RECV:
6724 ret = io_recv(req, issue_flags);
6726 case IORING_OP_TIMEOUT:
6727 ret = io_timeout(req, issue_flags);
6729 case IORING_OP_TIMEOUT_REMOVE:
6730 ret = io_timeout_remove(req, issue_flags);
6732 case IORING_OP_ACCEPT:
6733 ret = io_accept(req, issue_flags);
6735 case IORING_OP_CONNECT:
6736 ret = io_connect(req, issue_flags);
6738 case IORING_OP_ASYNC_CANCEL:
6739 ret = io_async_cancel(req, issue_flags);
6741 case IORING_OP_FALLOCATE:
6742 ret = io_fallocate(req, issue_flags);
6744 case IORING_OP_OPENAT:
6745 ret = io_openat(req, issue_flags);
6747 case IORING_OP_CLOSE:
6748 ret = io_close(req, issue_flags);
6750 case IORING_OP_FILES_UPDATE:
6751 ret = io_files_update(req, issue_flags);
6753 case IORING_OP_STATX:
6754 ret = io_statx(req, issue_flags);
6756 case IORING_OP_FADVISE:
6757 ret = io_fadvise(req, issue_flags);
6759 case IORING_OP_MADVISE:
6760 ret = io_madvise(req, issue_flags);
6762 case IORING_OP_OPENAT2:
6763 ret = io_openat2(req, issue_flags);
6765 case IORING_OP_EPOLL_CTL:
6766 ret = io_epoll_ctl(req, issue_flags);
6768 case IORING_OP_SPLICE:
6769 ret = io_splice(req, issue_flags);
6771 case IORING_OP_PROVIDE_BUFFERS:
6772 ret = io_provide_buffers(req, issue_flags);
6774 case IORING_OP_REMOVE_BUFFERS:
6775 ret = io_remove_buffers(req, issue_flags);
6778 ret = io_tee(req, issue_flags);
6780 case IORING_OP_SHUTDOWN:
6781 ret = io_shutdown(req, issue_flags);
6783 case IORING_OP_RENAMEAT:
6784 ret = io_renameat(req, issue_flags);
6786 case IORING_OP_UNLINKAT:
6787 ret = io_unlinkat(req, issue_flags);
6789 case IORING_OP_MKDIRAT:
6790 ret = io_mkdirat(req, issue_flags);
6792 case IORING_OP_SYMLINKAT:
6793 ret = io_symlinkat(req, issue_flags);
6795 case IORING_OP_LINKAT:
6796 ret = io_linkat(req, issue_flags);
6804 revert_creds(creds);
6807 /* If the op doesn't have a file, we're not polling for it */
6808 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6809 io_iopoll_req_issued(req);
6814 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6816 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6818 req = io_put_req_find_next(req);
6819 return req ? &req->work : NULL;
6822 static void io_wq_submit_work(struct io_wq_work *work)
6824 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6825 struct io_kiocb *timeout;
6828 /* one will be dropped by ->io_free_work() after returning to io-wq */
6829 if (!(req->flags & REQ_F_REFCOUNT))
6830 __io_req_set_refcount(req, 2);
6834 timeout = io_prep_linked_timeout(req);
6836 io_queue_linked_timeout(timeout);
6838 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6839 if (work->flags & IO_WQ_WORK_CANCEL)
6844 ret = io_issue_sqe(req, 0);
6846 * We can get EAGAIN for polled IO even though we're
6847 * forcing a sync submission from here, since we can't
6848 * wait for request slots on the block side.
6856 /* avoid locking problems by failing it from a clean context */
6858 io_req_task_queue_fail(req, ret);
6861 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6864 return &table->files[i];
6867 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6870 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6872 return (struct file *) (slot->file_ptr & FFS_MASK);
6875 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6877 unsigned long file_ptr = (unsigned long) file;
6879 if (__io_file_supports_nowait(file, READ))
6880 file_ptr |= FFS_ASYNC_READ;
6881 if (__io_file_supports_nowait(file, WRITE))
6882 file_ptr |= FFS_ASYNC_WRITE;
6883 if (S_ISREG(file_inode(file)->i_mode))
6884 file_ptr |= FFS_ISREG;
6885 file_slot->file_ptr = file_ptr;
6888 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6889 struct io_kiocb *req, int fd)
6892 unsigned long file_ptr;
6894 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6896 fd = array_index_nospec(fd, ctx->nr_user_files);
6897 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6898 file = (struct file *) (file_ptr & FFS_MASK);
6899 file_ptr &= ~FFS_MASK;
6900 /* mask in overlapping REQ_F and FFS bits */
6901 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6902 io_req_set_rsrc_node(req);
6906 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6907 struct io_kiocb *req, int fd)
6909 struct file *file = fget(fd);
6911 trace_io_uring_file_get(ctx, fd);
6913 /* we don't allow fixed io_uring files */
6914 if (file && unlikely(file->f_op == &io_uring_fops))
6915 io_req_track_inflight(req);
6919 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6920 struct io_kiocb *req, int fd, bool fixed)
6923 return io_file_get_fixed(ctx, req, fd);
6925 return io_file_get_normal(ctx, req, fd);
6928 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6930 struct io_kiocb *prev = req->timeout.prev;
6934 if (!(req->task->flags & PF_EXITING))
6935 ret = io_try_cancel_userdata(req, prev->user_data);
6936 io_req_complete_post(req, ret ?: -ETIME, 0);
6939 io_req_complete_post(req, -ETIME, 0);
6943 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6945 struct io_timeout_data *data = container_of(timer,
6946 struct io_timeout_data, timer);
6947 struct io_kiocb *prev, *req = data->req;
6948 struct io_ring_ctx *ctx = req->ctx;
6949 unsigned long flags;
6951 spin_lock_irqsave(&ctx->timeout_lock, flags);
6952 prev = req->timeout.head;
6953 req->timeout.head = NULL;
6956 * We don't expect the list to be empty, that will only happen if we
6957 * race with the completion of the linked work.
6960 io_remove_next_linked(prev);
6961 if (!req_ref_inc_not_zero(prev))
6964 list_del(&req->timeout.list);
6965 req->timeout.prev = prev;
6966 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6968 req->io_task_work.func = io_req_task_link_timeout;
6969 io_req_task_work_add(req);
6970 return HRTIMER_NORESTART;
6973 static void io_queue_linked_timeout(struct io_kiocb *req)
6975 struct io_ring_ctx *ctx = req->ctx;
6977 spin_lock_irq(&ctx->timeout_lock);
6979 * If the back reference is NULL, then our linked request finished
6980 * before we got a chance to setup the timer
6982 if (req->timeout.head) {
6983 struct io_timeout_data *data = req->async_data;
6985 data->timer.function = io_link_timeout_fn;
6986 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6988 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6990 spin_unlock_irq(&ctx->timeout_lock);
6991 /* drop submission reference */
6995 static void __io_queue_sqe(struct io_kiocb *req)
6996 __must_hold(&req->ctx->uring_lock)
6998 struct io_kiocb *linked_timeout;
7002 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
7005 * We async punt it if the file wasn't marked NOWAIT, or if the file
7006 * doesn't support non-blocking read/write attempts
7009 if (req->flags & REQ_F_COMPLETE_INLINE) {
7010 struct io_ring_ctx *ctx = req->ctx;
7011 struct io_submit_state *state = &ctx->submit_state;
7013 state->compl_reqs[state->compl_nr++] = req;
7014 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
7015 io_submit_flush_completions(ctx);
7019 linked_timeout = io_prep_linked_timeout(req);
7021 io_queue_linked_timeout(linked_timeout);
7022 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
7023 linked_timeout = io_prep_linked_timeout(req);
7025 switch (io_arm_poll_handler(req)) {
7026 case IO_APOLL_READY:
7028 io_queue_linked_timeout(linked_timeout);
7030 case IO_APOLL_ABORTED:
7032 * Queued up for async execution, worker will release
7033 * submit reference when the iocb is actually submitted.
7035 io_queue_async_work(req, NULL);
7040 io_queue_linked_timeout(linked_timeout);
7042 io_req_complete_failed(req, ret);
7046 static inline void io_queue_sqe(struct io_kiocb *req)
7047 __must_hold(&req->ctx->uring_lock)
7049 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7052 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7053 __io_queue_sqe(req);
7054 } else if (req->flags & REQ_F_FAIL) {
7055 io_req_complete_fail_submit(req);
7057 int ret = io_req_prep_async(req);
7060 io_req_complete_failed(req, ret);
7062 io_queue_async_work(req, NULL);
7067 * Check SQE restrictions (opcode and flags).
7069 * Returns 'true' if SQE is allowed, 'false' otherwise.
7071 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7072 struct io_kiocb *req,
7073 unsigned int sqe_flags)
7075 if (likely(!ctx->restricted))
7078 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7081 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7082 ctx->restrictions.sqe_flags_required)
7085 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7086 ctx->restrictions.sqe_flags_required))
7092 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7093 const struct io_uring_sqe *sqe)
7094 __must_hold(&ctx->uring_lock)
7096 struct io_submit_state *state;
7097 unsigned int sqe_flags;
7098 int personality, ret = 0;
7100 /* req is partially pre-initialised, see io_preinit_req() */
7101 req->opcode = READ_ONCE(sqe->opcode);
7102 /* same numerical values with corresponding REQ_F_*, safe to copy */
7103 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7104 req->user_data = READ_ONCE(sqe->user_data);
7106 req->fixed_rsrc_refs = NULL;
7107 req->task = current;
7109 /* enforce forwards compatibility on users */
7110 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7112 if (unlikely(req->opcode >= IORING_OP_LAST))
7114 if (!io_check_restriction(ctx, req, sqe_flags))
7117 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7118 !io_op_defs[req->opcode].buffer_select)
7120 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7121 ctx->drain_active = true;
7123 personality = READ_ONCE(sqe->personality);
7125 req->creds = xa_load(&ctx->personalities, personality);
7128 get_cred(req->creds);
7129 req->flags |= REQ_F_CREDS;
7131 state = &ctx->submit_state;
7134 * Plug now if we have more than 1 IO left after this, and the target
7135 * is potentially a read/write to block based storage.
7137 if (!state->plug_started && state->ios_left > 1 &&
7138 io_op_defs[req->opcode].plug) {
7139 blk_start_plug(&state->plug);
7140 state->plug_started = true;
7143 if (io_op_defs[req->opcode].needs_file) {
7144 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7145 (sqe_flags & IOSQE_FIXED_FILE));
7146 if (unlikely(!req->file))
7154 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7155 const struct io_uring_sqe *sqe)
7156 __must_hold(&ctx->uring_lock)
7158 struct io_submit_link *link = &ctx->submit_state.link;
7161 ret = io_init_req(ctx, req, sqe);
7162 if (unlikely(ret)) {
7164 /* fail even hard links since we don't submit */
7167 * we can judge a link req is failed or cancelled by if
7168 * REQ_F_FAIL is set, but the head is an exception since
7169 * it may be set REQ_F_FAIL because of other req's failure
7170 * so let's leverage req->result to distinguish if a head
7171 * is set REQ_F_FAIL because of its failure or other req's
7172 * failure so that we can set the correct ret code for it.
7173 * init result here to avoid affecting the normal path.
7175 if (!(link->head->flags & REQ_F_FAIL))
7176 req_fail_link_node(link->head, -ECANCELED);
7177 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7179 * the current req is a normal req, we should return
7180 * error and thus break the submittion loop.
7182 io_req_complete_failed(req, ret);
7185 req_fail_link_node(req, ret);
7187 ret = io_req_prep(req, sqe);
7192 /* don't need @sqe from now on */
7193 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7195 ctx->flags & IORING_SETUP_SQPOLL);
7198 * If we already have a head request, queue this one for async
7199 * submittal once the head completes. If we don't have a head but
7200 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7201 * submitted sync once the chain is complete. If none of those
7202 * conditions are true (normal request), then just queue it.
7205 struct io_kiocb *head = link->head;
7207 if (!(req->flags & REQ_F_FAIL)) {
7208 ret = io_req_prep_async(req);
7209 if (unlikely(ret)) {
7210 req_fail_link_node(req, ret);
7211 if (!(head->flags & REQ_F_FAIL))
7212 req_fail_link_node(head, -ECANCELED);
7215 trace_io_uring_link(ctx, req, head);
7216 link->last->link = req;
7219 /* last request of a link, enqueue the link */
7220 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7225 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7237 * Batched submission is done, ensure local IO is flushed out.
7239 static void io_submit_state_end(struct io_submit_state *state,
7240 struct io_ring_ctx *ctx)
7242 if (state->link.head)
7243 io_queue_sqe(state->link.head);
7244 if (state->compl_nr)
7245 io_submit_flush_completions(ctx);
7246 if (state->plug_started)
7247 blk_finish_plug(&state->plug);
7251 * Start submission side cache.
7253 static void io_submit_state_start(struct io_submit_state *state,
7254 unsigned int max_ios)
7256 state->plug_started = false;
7257 state->ios_left = max_ios;
7258 /* set only head, no need to init link_last in advance */
7259 state->link.head = NULL;
7262 static void io_commit_sqring(struct io_ring_ctx *ctx)
7264 struct io_rings *rings = ctx->rings;
7267 * Ensure any loads from the SQEs are done at this point,
7268 * since once we write the new head, the application could
7269 * write new data to them.
7271 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7275 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7276 * that is mapped by userspace. This means that care needs to be taken to
7277 * ensure that reads are stable, as we cannot rely on userspace always
7278 * being a good citizen. If members of the sqe are validated and then later
7279 * used, it's important that those reads are done through READ_ONCE() to
7280 * prevent a re-load down the line.
7282 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7284 unsigned head, mask = ctx->sq_entries - 1;
7285 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7288 * The cached sq head (or cq tail) serves two purposes:
7290 * 1) allows us to batch the cost of updating the user visible
7292 * 2) allows the kernel side to track the head on its own, even
7293 * though the application is the one updating it.
7295 head = READ_ONCE(ctx->sq_array[sq_idx]);
7296 if (likely(head < ctx->sq_entries))
7297 return &ctx->sq_sqes[head];
7299 /* drop invalid entries */
7301 WRITE_ONCE(ctx->rings->sq_dropped,
7302 READ_ONCE(ctx->rings->sq_dropped) + 1);
7306 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7307 __must_hold(&ctx->uring_lock)
7311 /* make sure SQ entry isn't read before tail */
7312 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7313 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7315 io_get_task_refs(nr);
7317 io_submit_state_start(&ctx->submit_state, nr);
7318 while (submitted < nr) {
7319 const struct io_uring_sqe *sqe;
7320 struct io_kiocb *req;
7322 req = io_alloc_req(ctx);
7323 if (unlikely(!req)) {
7325 submitted = -EAGAIN;
7328 sqe = io_get_sqe(ctx);
7329 if (unlikely(!sqe)) {
7330 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7333 /* will complete beyond this point, count as submitted */
7335 if (io_submit_sqe(ctx, req, sqe))
7339 if (unlikely(submitted != nr)) {
7340 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7341 int unused = nr - ref_used;
7343 current->io_uring->cached_refs += unused;
7344 percpu_ref_put_many(&ctx->refs, unused);
7347 io_submit_state_end(&ctx->submit_state, ctx);
7348 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7349 io_commit_sqring(ctx);
7354 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7356 return READ_ONCE(sqd->state);
7359 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7361 /* Tell userspace we may need a wakeup call */
7362 spin_lock(&ctx->completion_lock);
7363 WRITE_ONCE(ctx->rings->sq_flags,
7364 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7365 spin_unlock(&ctx->completion_lock);
7368 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7370 spin_lock(&ctx->completion_lock);
7371 WRITE_ONCE(ctx->rings->sq_flags,
7372 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7373 spin_unlock(&ctx->completion_lock);
7376 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7378 unsigned int to_submit;
7381 to_submit = io_sqring_entries(ctx);
7382 /* if we're handling multiple rings, cap submit size for fairness */
7383 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7384 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7386 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7387 unsigned nr_events = 0;
7388 const struct cred *creds = NULL;
7390 if (ctx->sq_creds != current_cred())
7391 creds = override_creds(ctx->sq_creds);
7393 mutex_lock(&ctx->uring_lock);
7394 if (!list_empty(&ctx->iopoll_list))
7395 io_do_iopoll(ctx, &nr_events, 0);
7398 * Don't submit if refs are dying, good for io_uring_register(),
7399 * but also it is relied upon by io_ring_exit_work()
7401 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7402 !(ctx->flags & IORING_SETUP_R_DISABLED))
7403 ret = io_submit_sqes(ctx, to_submit);
7404 mutex_unlock(&ctx->uring_lock);
7406 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7407 wake_up(&ctx->sqo_sq_wait);
7409 revert_creds(creds);
7415 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7417 struct io_ring_ctx *ctx;
7418 unsigned sq_thread_idle = 0;
7420 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7421 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7422 sqd->sq_thread_idle = sq_thread_idle;
7425 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7427 bool did_sig = false;
7428 struct ksignal ksig;
7430 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7431 signal_pending(current)) {
7432 mutex_unlock(&sqd->lock);
7433 if (signal_pending(current))
7434 did_sig = get_signal(&ksig);
7436 mutex_lock(&sqd->lock);
7438 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7441 static int io_sq_thread(void *data)
7443 struct io_sq_data *sqd = data;
7444 struct io_ring_ctx *ctx;
7445 unsigned long timeout = 0;
7446 char buf[TASK_COMM_LEN];
7449 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7450 set_task_comm(current, buf);
7452 if (sqd->sq_cpu != -1)
7453 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7455 set_cpus_allowed_ptr(current, cpu_online_mask);
7456 current->flags |= PF_NO_SETAFFINITY;
7458 mutex_lock(&sqd->lock);
7460 bool cap_entries, sqt_spin = false;
7462 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7463 if (io_sqd_handle_event(sqd))
7465 timeout = jiffies + sqd->sq_thread_idle;
7468 cap_entries = !list_is_singular(&sqd->ctx_list);
7469 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7470 int ret = __io_sq_thread(ctx, cap_entries);
7472 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7475 if (io_run_task_work())
7478 if (sqt_spin || !time_after(jiffies, timeout)) {
7481 timeout = jiffies + sqd->sq_thread_idle;
7485 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7486 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7487 bool needs_sched = true;
7489 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7490 io_ring_set_wakeup_flag(ctx);
7492 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7493 !list_empty_careful(&ctx->iopoll_list)) {
7494 needs_sched = false;
7497 if (io_sqring_entries(ctx)) {
7498 needs_sched = false;
7504 mutex_unlock(&sqd->lock);
7506 mutex_lock(&sqd->lock);
7508 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7509 io_ring_clear_wakeup_flag(ctx);
7512 finish_wait(&sqd->wait, &wait);
7513 timeout = jiffies + sqd->sq_thread_idle;
7516 io_uring_cancel_generic(true, sqd);
7518 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7519 io_ring_set_wakeup_flag(ctx);
7521 mutex_unlock(&sqd->lock);
7523 complete(&sqd->exited);
7527 struct io_wait_queue {
7528 struct wait_queue_entry wq;
7529 struct io_ring_ctx *ctx;
7531 unsigned nr_timeouts;
7534 static inline bool io_should_wake(struct io_wait_queue *iowq)
7536 struct io_ring_ctx *ctx = iowq->ctx;
7537 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7540 * Wake up if we have enough events, or if a timeout occurred since we
7541 * started waiting. For timeouts, we always want to return to userspace,
7542 * regardless of event count.
7544 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7547 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7548 int wake_flags, void *key)
7550 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7554 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7555 * the task, and the next invocation will do it.
7557 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7558 return autoremove_wake_function(curr, mode, wake_flags, key);
7562 static int io_run_task_work_sig(void)
7564 if (io_run_task_work())
7566 if (!signal_pending(current))
7568 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7569 return -ERESTARTSYS;
7573 /* when returns >0, the caller should retry */
7574 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7575 struct io_wait_queue *iowq,
7576 signed long *timeout)
7580 /* make sure we run task_work before checking for signals */
7581 ret = io_run_task_work_sig();
7582 if (ret || io_should_wake(iowq))
7584 /* let the caller flush overflows, retry */
7585 if (test_bit(0, &ctx->check_cq_overflow))
7588 *timeout = schedule_timeout(*timeout);
7589 return !*timeout ? -ETIME : 1;
7593 * Wait until events become available, if we don't already have some. The
7594 * application must reap them itself, as they reside on the shared cq ring.
7596 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7597 const sigset_t __user *sig, size_t sigsz,
7598 struct __kernel_timespec __user *uts)
7600 struct io_wait_queue iowq;
7601 struct io_rings *rings = ctx->rings;
7602 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7606 io_cqring_overflow_flush(ctx);
7607 if (io_cqring_events(ctx) >= min_events)
7609 if (!io_run_task_work())
7614 struct timespec64 ts;
7616 if (get_timespec64(&ts, uts))
7618 timeout = timespec64_to_jiffies(&ts);
7622 #ifdef CONFIG_COMPAT
7623 if (in_compat_syscall())
7624 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7628 ret = set_user_sigmask(sig, sigsz);
7634 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7635 iowq.wq.private = current;
7636 INIT_LIST_HEAD(&iowq.wq.entry);
7638 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7639 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7641 trace_io_uring_cqring_wait(ctx, min_events);
7643 /* if we can't even flush overflow, don't wait for more */
7644 if (!io_cqring_overflow_flush(ctx)) {
7648 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7649 TASK_INTERRUPTIBLE);
7650 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7651 finish_wait(&ctx->cq_wait, &iowq.wq);
7655 restore_saved_sigmask_unless(ret == -EINTR);
7657 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7660 static void io_free_page_table(void **table, size_t size)
7662 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7664 for (i = 0; i < nr_tables; i++)
7669 static void **io_alloc_page_table(size_t size)
7671 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7672 size_t init_size = size;
7675 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7679 for (i = 0; i < nr_tables; i++) {
7680 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7682 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7684 io_free_page_table(table, init_size);
7692 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7694 percpu_ref_exit(&ref_node->refs);
7698 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7700 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7701 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7702 unsigned long flags;
7703 bool first_add = false;
7705 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7708 while (!list_empty(&ctx->rsrc_ref_list)) {
7709 node = list_first_entry(&ctx->rsrc_ref_list,
7710 struct io_rsrc_node, node);
7711 /* recycle ref nodes in order */
7714 list_del(&node->node);
7715 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7717 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7720 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7723 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7725 struct io_rsrc_node *ref_node;
7727 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7731 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7736 INIT_LIST_HEAD(&ref_node->node);
7737 INIT_LIST_HEAD(&ref_node->rsrc_list);
7738 ref_node->done = false;
7742 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7743 struct io_rsrc_data *data_to_kill)
7745 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7746 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7749 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7751 rsrc_node->rsrc_data = data_to_kill;
7752 spin_lock_irq(&ctx->rsrc_ref_lock);
7753 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7754 spin_unlock_irq(&ctx->rsrc_ref_lock);
7756 atomic_inc(&data_to_kill->refs);
7757 percpu_ref_kill(&rsrc_node->refs);
7758 ctx->rsrc_node = NULL;
7761 if (!ctx->rsrc_node) {
7762 ctx->rsrc_node = ctx->rsrc_backup_node;
7763 ctx->rsrc_backup_node = NULL;
7767 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7769 if (ctx->rsrc_backup_node)
7771 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7772 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7775 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7779 /* As we may drop ->uring_lock, other task may have started quiesce */
7783 data->quiesce = true;
7785 ret = io_rsrc_node_switch_start(ctx);
7788 io_rsrc_node_switch(ctx, data);
7790 /* kill initial ref, already quiesced if zero */
7791 if (atomic_dec_and_test(&data->refs))
7793 mutex_unlock(&ctx->uring_lock);
7794 flush_delayed_work(&ctx->rsrc_put_work);
7795 ret = wait_for_completion_interruptible(&data->done);
7797 mutex_lock(&ctx->uring_lock);
7801 atomic_inc(&data->refs);
7802 /* wait for all works potentially completing data->done */
7803 flush_delayed_work(&ctx->rsrc_put_work);
7804 reinit_completion(&data->done);
7806 ret = io_run_task_work_sig();
7807 mutex_lock(&ctx->uring_lock);
7809 data->quiesce = false;
7814 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7816 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7817 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7819 return &data->tags[table_idx][off];
7822 static void io_rsrc_data_free(struct io_rsrc_data *data)
7824 size_t size = data->nr * sizeof(data->tags[0][0]);
7827 io_free_page_table((void **)data->tags, size);
7831 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7832 u64 __user *utags, unsigned nr,
7833 struct io_rsrc_data **pdata)
7835 struct io_rsrc_data *data;
7839 data = kzalloc(sizeof(*data), GFP_KERNEL);
7842 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7850 data->do_put = do_put;
7853 for (i = 0; i < nr; i++) {
7854 u64 *tag_slot = io_get_tag_slot(data, i);
7856 if (copy_from_user(tag_slot, &utags[i],
7862 atomic_set(&data->refs, 1);
7863 init_completion(&data->done);
7867 io_rsrc_data_free(data);
7871 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7873 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7874 GFP_KERNEL_ACCOUNT);
7875 return !!table->files;
7878 static void io_free_file_tables(struct io_file_table *table)
7880 kvfree(table->files);
7881 table->files = NULL;
7884 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7886 #if defined(CONFIG_UNIX)
7887 if (ctx->ring_sock) {
7888 struct sock *sock = ctx->ring_sock->sk;
7889 struct sk_buff *skb;
7891 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7897 for (i = 0; i < ctx->nr_user_files; i++) {
7900 file = io_file_from_index(ctx, i);
7905 io_free_file_tables(&ctx->file_table);
7906 io_rsrc_data_free(ctx->file_data);
7907 ctx->file_data = NULL;
7908 ctx->nr_user_files = 0;
7911 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7915 if (!ctx->file_data)
7917 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7919 __io_sqe_files_unregister(ctx);
7923 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7924 __releases(&sqd->lock)
7926 WARN_ON_ONCE(sqd->thread == current);
7929 * Do the dance but not conditional clear_bit() because it'd race with
7930 * other threads incrementing park_pending and setting the bit.
7932 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7933 if (atomic_dec_return(&sqd->park_pending))
7934 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7935 mutex_unlock(&sqd->lock);
7938 static void io_sq_thread_park(struct io_sq_data *sqd)
7939 __acquires(&sqd->lock)
7941 WARN_ON_ONCE(sqd->thread == current);
7943 atomic_inc(&sqd->park_pending);
7944 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7945 mutex_lock(&sqd->lock);
7947 wake_up_process(sqd->thread);
7950 static void io_sq_thread_stop(struct io_sq_data *sqd)
7952 WARN_ON_ONCE(sqd->thread == current);
7953 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7955 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7956 mutex_lock(&sqd->lock);
7958 wake_up_process(sqd->thread);
7959 mutex_unlock(&sqd->lock);
7960 wait_for_completion(&sqd->exited);
7963 static void io_put_sq_data(struct io_sq_data *sqd)
7965 if (refcount_dec_and_test(&sqd->refs)) {
7966 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7968 io_sq_thread_stop(sqd);
7973 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7975 struct io_sq_data *sqd = ctx->sq_data;
7978 io_sq_thread_park(sqd);
7979 list_del_init(&ctx->sqd_list);
7980 io_sqd_update_thread_idle(sqd);
7981 io_sq_thread_unpark(sqd);
7983 io_put_sq_data(sqd);
7984 ctx->sq_data = NULL;
7988 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7990 struct io_ring_ctx *ctx_attach;
7991 struct io_sq_data *sqd;
7994 f = fdget(p->wq_fd);
7996 return ERR_PTR(-ENXIO);
7997 if (f.file->f_op != &io_uring_fops) {
7999 return ERR_PTR(-EINVAL);
8002 ctx_attach = f.file->private_data;
8003 sqd = ctx_attach->sq_data;
8006 return ERR_PTR(-EINVAL);
8008 if (sqd->task_tgid != current->tgid) {
8010 return ERR_PTR(-EPERM);
8013 refcount_inc(&sqd->refs);
8018 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
8021 struct io_sq_data *sqd;
8024 if (p->flags & IORING_SETUP_ATTACH_WQ) {
8025 sqd = io_attach_sq_data(p);
8030 /* fall through for EPERM case, setup new sqd/task */
8031 if (PTR_ERR(sqd) != -EPERM)
8035 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8037 return ERR_PTR(-ENOMEM);
8039 atomic_set(&sqd->park_pending, 0);
8040 refcount_set(&sqd->refs, 1);
8041 INIT_LIST_HEAD(&sqd->ctx_list);
8042 mutex_init(&sqd->lock);
8043 init_waitqueue_head(&sqd->wait);
8044 init_completion(&sqd->exited);
8048 #if defined(CONFIG_UNIX)
8050 * Ensure the UNIX gc is aware of our file set, so we are certain that
8051 * the io_uring can be safely unregistered on process exit, even if we have
8052 * loops in the file referencing.
8054 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8056 struct sock *sk = ctx->ring_sock->sk;
8057 struct scm_fp_list *fpl;
8058 struct sk_buff *skb;
8061 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8065 skb = alloc_skb(0, GFP_KERNEL);
8074 fpl->user = get_uid(current_user());
8075 for (i = 0; i < nr; i++) {
8076 struct file *file = io_file_from_index(ctx, i + offset);
8080 fpl->fp[nr_files] = get_file(file);
8081 unix_inflight(fpl->user, fpl->fp[nr_files]);
8086 fpl->max = SCM_MAX_FD;
8087 fpl->count = nr_files;
8088 UNIXCB(skb).fp = fpl;
8089 skb->destructor = unix_destruct_scm;
8090 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8091 skb_queue_head(&sk->sk_receive_queue, skb);
8093 for (i = 0; i < nr_files; i++)
8104 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8105 * causes regular reference counting to break down. We rely on the UNIX
8106 * garbage collection to take care of this problem for us.
8108 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8110 unsigned left, total;
8114 left = ctx->nr_user_files;
8116 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8118 ret = __io_sqe_files_scm(ctx, this_files, total);
8122 total += this_files;
8128 while (total < ctx->nr_user_files) {
8129 struct file *file = io_file_from_index(ctx, total);
8139 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8145 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8147 struct file *file = prsrc->file;
8148 #if defined(CONFIG_UNIX)
8149 struct sock *sock = ctx->ring_sock->sk;
8150 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8151 struct sk_buff *skb;
8154 __skb_queue_head_init(&list);
8157 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8158 * remove this entry and rearrange the file array.
8160 skb = skb_dequeue(head);
8162 struct scm_fp_list *fp;
8164 fp = UNIXCB(skb).fp;
8165 for (i = 0; i < fp->count; i++) {
8168 if (fp->fp[i] != file)
8171 unix_notinflight(fp->user, fp->fp[i]);
8172 left = fp->count - 1 - i;
8174 memmove(&fp->fp[i], &fp->fp[i + 1],
8175 left * sizeof(struct file *));
8182 __skb_queue_tail(&list, skb);
8192 __skb_queue_tail(&list, skb);
8194 skb = skb_dequeue(head);
8197 if (skb_peek(&list)) {
8198 spin_lock_irq(&head->lock);
8199 while ((skb = __skb_dequeue(&list)) != NULL)
8200 __skb_queue_tail(head, skb);
8201 spin_unlock_irq(&head->lock);
8208 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8210 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8211 struct io_ring_ctx *ctx = rsrc_data->ctx;
8212 struct io_rsrc_put *prsrc, *tmp;
8214 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8215 list_del(&prsrc->list);
8218 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8220 io_ring_submit_lock(ctx, lock_ring);
8221 spin_lock(&ctx->completion_lock);
8222 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8224 io_commit_cqring(ctx);
8225 spin_unlock(&ctx->completion_lock);
8226 io_cqring_ev_posted(ctx);
8227 io_ring_submit_unlock(ctx, lock_ring);
8230 rsrc_data->do_put(ctx, prsrc);
8234 io_rsrc_node_destroy(ref_node);
8235 if (atomic_dec_and_test(&rsrc_data->refs))
8236 complete(&rsrc_data->done);
8239 static void io_rsrc_put_work(struct work_struct *work)
8241 struct io_ring_ctx *ctx;
8242 struct llist_node *node;
8244 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8245 node = llist_del_all(&ctx->rsrc_put_llist);
8248 struct io_rsrc_node *ref_node;
8249 struct llist_node *next = node->next;
8251 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8252 __io_rsrc_put_work(ref_node);
8257 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8258 unsigned nr_args, u64 __user *tags)
8260 __s32 __user *fds = (__s32 __user *) arg;
8269 if (nr_args > IORING_MAX_FIXED_FILES)
8271 if (nr_args > rlimit(RLIMIT_NOFILE))
8273 ret = io_rsrc_node_switch_start(ctx);
8276 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8282 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8285 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8286 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8290 /* allow sparse sets */
8293 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8300 if (unlikely(!file))
8304 * Don't allow io_uring instances to be registered. If UNIX
8305 * isn't enabled, then this causes a reference cycle and this
8306 * instance can never get freed. If UNIX is enabled we'll
8307 * handle it just fine, but there's still no point in allowing
8308 * a ring fd as it doesn't support regular read/write anyway.
8310 if (file->f_op == &io_uring_fops) {
8314 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8317 ret = io_sqe_files_scm(ctx);
8319 __io_sqe_files_unregister(ctx);
8323 io_rsrc_node_switch(ctx, NULL);
8326 for (i = 0; i < ctx->nr_user_files; i++) {
8327 file = io_file_from_index(ctx, i);
8331 io_free_file_tables(&ctx->file_table);
8332 ctx->nr_user_files = 0;
8334 io_rsrc_data_free(ctx->file_data);
8335 ctx->file_data = NULL;
8339 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8342 #if defined(CONFIG_UNIX)
8343 struct sock *sock = ctx->ring_sock->sk;
8344 struct sk_buff_head *head = &sock->sk_receive_queue;
8345 struct sk_buff *skb;
8348 * See if we can merge this file into an existing skb SCM_RIGHTS
8349 * file set. If there's no room, fall back to allocating a new skb
8350 * and filling it in.
8352 spin_lock_irq(&head->lock);
8353 skb = skb_peek(head);
8355 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8357 if (fpl->count < SCM_MAX_FD) {
8358 __skb_unlink(skb, head);
8359 spin_unlock_irq(&head->lock);
8360 fpl->fp[fpl->count] = get_file(file);
8361 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8363 spin_lock_irq(&head->lock);
8364 __skb_queue_head(head, skb);
8369 spin_unlock_irq(&head->lock);
8376 return __io_sqe_files_scm(ctx, 1, index);
8382 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8383 struct io_rsrc_node *node, void *rsrc)
8385 struct io_rsrc_put *prsrc;
8387 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8391 prsrc->tag = *io_get_tag_slot(data, idx);
8393 list_add(&prsrc->list, &node->rsrc_list);
8397 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8398 unsigned int issue_flags, u32 slot_index)
8400 struct io_ring_ctx *ctx = req->ctx;
8401 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8402 bool needs_switch = false;
8403 struct io_fixed_file *file_slot;
8406 io_ring_submit_lock(ctx, !force_nonblock);
8407 if (file->f_op == &io_uring_fops)
8410 if (!ctx->file_data)
8413 if (slot_index >= ctx->nr_user_files)
8416 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8417 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8419 if (file_slot->file_ptr) {
8420 struct file *old_file;
8422 ret = io_rsrc_node_switch_start(ctx);
8426 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8427 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8428 ctx->rsrc_node, old_file);
8431 file_slot->file_ptr = 0;
8432 needs_switch = true;
8435 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8436 io_fixed_file_set(file_slot, file);
8437 ret = io_sqe_file_register(ctx, file, slot_index);
8439 file_slot->file_ptr = 0;
8446 io_rsrc_node_switch(ctx, ctx->file_data);
8447 io_ring_submit_unlock(ctx, !force_nonblock);
8453 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8455 unsigned int offset = req->close.file_slot - 1;
8456 struct io_ring_ctx *ctx = req->ctx;
8457 struct io_fixed_file *file_slot;
8461 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8463 if (unlikely(!ctx->file_data))
8466 if (offset >= ctx->nr_user_files)
8468 ret = io_rsrc_node_switch_start(ctx);
8472 i = array_index_nospec(offset, ctx->nr_user_files);
8473 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8475 if (!file_slot->file_ptr)
8478 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8479 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8483 file_slot->file_ptr = 0;
8484 io_rsrc_node_switch(ctx, ctx->file_data);
8487 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8491 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8492 struct io_uring_rsrc_update2 *up,
8495 u64 __user *tags = u64_to_user_ptr(up->tags);
8496 __s32 __user *fds = u64_to_user_ptr(up->data);
8497 struct io_rsrc_data *data = ctx->file_data;
8498 struct io_fixed_file *file_slot;
8502 bool needs_switch = false;
8504 if (!ctx->file_data)
8506 if (up->offset + nr_args > ctx->nr_user_files)
8509 for (done = 0; done < nr_args; done++) {
8512 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8513 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8517 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8521 if (fd == IORING_REGISTER_FILES_SKIP)
8524 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8525 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8527 if (file_slot->file_ptr) {
8528 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8529 err = io_queue_rsrc_removal(data, up->offset + done,
8530 ctx->rsrc_node, file);
8533 file_slot->file_ptr = 0;
8534 needs_switch = true;
8543 * Don't allow io_uring instances to be registered. If
8544 * UNIX isn't enabled, then this causes a reference
8545 * cycle and this instance can never get freed. If UNIX
8546 * is enabled we'll handle it just fine, but there's
8547 * still no point in allowing a ring fd as it doesn't
8548 * support regular read/write anyway.
8550 if (file->f_op == &io_uring_fops) {
8555 *io_get_tag_slot(data, up->offset + done) = tag;
8556 io_fixed_file_set(file_slot, file);
8557 err = io_sqe_file_register(ctx, file, i);
8559 file_slot->file_ptr = 0;
8567 io_rsrc_node_switch(ctx, data);
8568 return done ? done : err;
8571 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8572 struct task_struct *task)
8574 struct io_wq_hash *hash;
8575 struct io_wq_data data;
8576 unsigned int concurrency;
8578 mutex_lock(&ctx->uring_lock);
8579 hash = ctx->hash_map;
8581 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8583 mutex_unlock(&ctx->uring_lock);
8584 return ERR_PTR(-ENOMEM);
8586 refcount_set(&hash->refs, 1);
8587 init_waitqueue_head(&hash->wait);
8588 ctx->hash_map = hash;
8590 mutex_unlock(&ctx->uring_lock);
8594 data.free_work = io_wq_free_work;
8595 data.do_work = io_wq_submit_work;
8597 /* Do QD, or 4 * CPUS, whatever is smallest */
8598 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8600 return io_wq_create(concurrency, &data);
8603 static int io_uring_alloc_task_context(struct task_struct *task,
8604 struct io_ring_ctx *ctx)
8606 struct io_uring_task *tctx;
8609 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8610 if (unlikely(!tctx))
8613 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8614 if (unlikely(ret)) {
8619 tctx->io_wq = io_init_wq_offload(ctx, task);
8620 if (IS_ERR(tctx->io_wq)) {
8621 ret = PTR_ERR(tctx->io_wq);
8622 percpu_counter_destroy(&tctx->inflight);
8628 init_waitqueue_head(&tctx->wait);
8629 atomic_set(&tctx->in_idle, 0);
8630 atomic_set(&tctx->inflight_tracked, 0);
8631 task->io_uring = tctx;
8632 spin_lock_init(&tctx->task_lock);
8633 INIT_WQ_LIST(&tctx->task_list);
8634 init_task_work(&tctx->task_work, tctx_task_work);
8638 void __io_uring_free(struct task_struct *tsk)
8640 struct io_uring_task *tctx = tsk->io_uring;
8642 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8643 WARN_ON_ONCE(tctx->io_wq);
8644 WARN_ON_ONCE(tctx->cached_refs);
8646 percpu_counter_destroy(&tctx->inflight);
8648 tsk->io_uring = NULL;
8651 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8652 struct io_uring_params *p)
8656 /* Retain compatibility with failing for an invalid attach attempt */
8657 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8658 IORING_SETUP_ATTACH_WQ) {
8661 f = fdget(p->wq_fd);
8664 if (f.file->f_op != &io_uring_fops) {
8670 if (ctx->flags & IORING_SETUP_SQPOLL) {
8671 struct task_struct *tsk;
8672 struct io_sq_data *sqd;
8675 sqd = io_get_sq_data(p, &attached);
8681 ctx->sq_creds = get_current_cred();
8683 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8684 if (!ctx->sq_thread_idle)
8685 ctx->sq_thread_idle = HZ;
8687 io_sq_thread_park(sqd);
8688 list_add(&ctx->sqd_list, &sqd->ctx_list);
8689 io_sqd_update_thread_idle(sqd);
8690 /* don't attach to a dying SQPOLL thread, would be racy */
8691 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8692 io_sq_thread_unpark(sqd);
8699 if (p->flags & IORING_SETUP_SQ_AFF) {
8700 int cpu = p->sq_thread_cpu;
8703 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8710 sqd->task_pid = current->pid;
8711 sqd->task_tgid = current->tgid;
8712 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8719 ret = io_uring_alloc_task_context(tsk, ctx);
8720 wake_up_new_task(tsk);
8723 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8724 /* Can't have SQ_AFF without SQPOLL */
8731 complete(&ctx->sq_data->exited);
8733 io_sq_thread_finish(ctx);
8737 static inline void __io_unaccount_mem(struct user_struct *user,
8738 unsigned long nr_pages)
8740 atomic_long_sub(nr_pages, &user->locked_vm);
8743 static inline int __io_account_mem(struct user_struct *user,
8744 unsigned long nr_pages)
8746 unsigned long page_limit, cur_pages, new_pages;
8748 /* Don't allow more pages than we can safely lock */
8749 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8752 cur_pages = atomic_long_read(&user->locked_vm);
8753 new_pages = cur_pages + nr_pages;
8754 if (new_pages > page_limit)
8756 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8757 new_pages) != cur_pages);
8762 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8765 __io_unaccount_mem(ctx->user, nr_pages);
8767 if (ctx->mm_account)
8768 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8771 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8776 ret = __io_account_mem(ctx->user, nr_pages);
8781 if (ctx->mm_account)
8782 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8787 static void io_mem_free(void *ptr)
8794 page = virt_to_head_page(ptr);
8795 if (put_page_testzero(page))
8796 free_compound_page(page);
8799 static void *io_mem_alloc(size_t size)
8801 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8802 __GFP_NORETRY | __GFP_ACCOUNT;
8804 return (void *) __get_free_pages(gfp_flags, get_order(size));
8807 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8810 struct io_rings *rings;
8811 size_t off, sq_array_size;
8813 off = struct_size(rings, cqes, cq_entries);
8814 if (off == SIZE_MAX)
8818 off = ALIGN(off, SMP_CACHE_BYTES);
8826 sq_array_size = array_size(sizeof(u32), sq_entries);
8827 if (sq_array_size == SIZE_MAX)
8830 if (check_add_overflow(off, sq_array_size, &off))
8836 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8838 struct io_mapped_ubuf *imu = *slot;
8841 if (imu != ctx->dummy_ubuf) {
8842 for (i = 0; i < imu->nr_bvecs; i++)
8843 unpin_user_page(imu->bvec[i].bv_page);
8844 if (imu->acct_pages)
8845 io_unaccount_mem(ctx, imu->acct_pages);
8851 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8853 io_buffer_unmap(ctx, &prsrc->buf);
8857 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8861 for (i = 0; i < ctx->nr_user_bufs; i++)
8862 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8863 kfree(ctx->user_bufs);
8864 io_rsrc_data_free(ctx->buf_data);
8865 ctx->user_bufs = NULL;
8866 ctx->buf_data = NULL;
8867 ctx->nr_user_bufs = 0;
8870 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8877 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8879 __io_sqe_buffers_unregister(ctx);
8883 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8884 void __user *arg, unsigned index)
8886 struct iovec __user *src;
8888 #ifdef CONFIG_COMPAT
8890 struct compat_iovec __user *ciovs;
8891 struct compat_iovec ciov;
8893 ciovs = (struct compat_iovec __user *) arg;
8894 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8897 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8898 dst->iov_len = ciov.iov_len;
8902 src = (struct iovec __user *) arg;
8903 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8909 * Not super efficient, but this is just a registration time. And we do cache
8910 * the last compound head, so generally we'll only do a full search if we don't
8913 * We check if the given compound head page has already been accounted, to
8914 * avoid double accounting it. This allows us to account the full size of the
8915 * page, not just the constituent pages of a huge page.
8917 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8918 int nr_pages, struct page *hpage)
8922 /* check current page array */
8923 for (i = 0; i < nr_pages; i++) {
8924 if (!PageCompound(pages[i]))
8926 if (compound_head(pages[i]) == hpage)
8930 /* check previously registered pages */
8931 for (i = 0; i < ctx->nr_user_bufs; i++) {
8932 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8934 for (j = 0; j < imu->nr_bvecs; j++) {
8935 if (!PageCompound(imu->bvec[j].bv_page))
8937 if (compound_head(imu->bvec[j].bv_page) == hpage)
8945 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8946 int nr_pages, struct io_mapped_ubuf *imu,
8947 struct page **last_hpage)
8951 imu->acct_pages = 0;
8952 for (i = 0; i < nr_pages; i++) {
8953 if (!PageCompound(pages[i])) {
8958 hpage = compound_head(pages[i]);
8959 if (hpage == *last_hpage)
8961 *last_hpage = hpage;
8962 if (headpage_already_acct(ctx, pages, i, hpage))
8964 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8968 if (!imu->acct_pages)
8971 ret = io_account_mem(ctx, imu->acct_pages);
8973 imu->acct_pages = 0;
8977 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8978 struct io_mapped_ubuf **pimu,
8979 struct page **last_hpage)
8981 struct io_mapped_ubuf *imu = NULL;
8982 struct vm_area_struct **vmas = NULL;
8983 struct page **pages = NULL;
8984 unsigned long off, start, end, ubuf;
8986 int ret, pret, nr_pages, i;
8988 if (!iov->iov_base) {
8989 *pimu = ctx->dummy_ubuf;
8993 ubuf = (unsigned long) iov->iov_base;
8994 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8995 start = ubuf >> PAGE_SHIFT;
8996 nr_pages = end - start;
9001 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9005 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9010 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9015 mmap_read_lock(current->mm);
9016 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9018 if (pret == nr_pages) {
9019 /* don't support file backed memory */
9020 for (i = 0; i < nr_pages; i++) {
9021 struct vm_area_struct *vma = vmas[i];
9023 if (vma_is_shmem(vma))
9026 !is_file_hugepages(vma->vm_file)) {
9032 ret = pret < 0 ? pret : -EFAULT;
9034 mmap_read_unlock(current->mm);
9037 * if we did partial map, or found file backed vmas,
9038 * release any pages we did get
9041 unpin_user_pages(pages, pret);
9045 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9047 unpin_user_pages(pages, pret);
9051 off = ubuf & ~PAGE_MASK;
9052 size = iov->iov_len;
9053 for (i = 0; i < nr_pages; i++) {
9056 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9057 imu->bvec[i].bv_page = pages[i];
9058 imu->bvec[i].bv_len = vec_len;
9059 imu->bvec[i].bv_offset = off;
9063 /* store original address for later verification */
9065 imu->ubuf_end = ubuf + iov->iov_len;
9066 imu->nr_bvecs = nr_pages;
9077 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9079 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9080 return ctx->user_bufs ? 0 : -ENOMEM;
9083 static int io_buffer_validate(struct iovec *iov)
9085 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9088 * Don't impose further limits on the size and buffer
9089 * constraints here, we'll -EINVAL later when IO is
9090 * submitted if they are wrong.
9093 return iov->iov_len ? -EFAULT : 0;
9097 /* arbitrary limit, but we need something */
9098 if (iov->iov_len > SZ_1G)
9101 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9107 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9108 unsigned int nr_args, u64 __user *tags)
9110 struct page *last_hpage = NULL;
9111 struct io_rsrc_data *data;
9117 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9119 ret = io_rsrc_node_switch_start(ctx);
9122 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9125 ret = io_buffers_map_alloc(ctx, nr_args);
9127 io_rsrc_data_free(data);
9131 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9132 ret = io_copy_iov(ctx, &iov, arg, i);
9135 ret = io_buffer_validate(&iov);
9138 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9143 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9149 WARN_ON_ONCE(ctx->buf_data);
9151 ctx->buf_data = data;
9153 __io_sqe_buffers_unregister(ctx);
9155 io_rsrc_node_switch(ctx, NULL);
9159 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9160 struct io_uring_rsrc_update2 *up,
9161 unsigned int nr_args)
9163 u64 __user *tags = u64_to_user_ptr(up->tags);
9164 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9165 struct page *last_hpage = NULL;
9166 bool needs_switch = false;
9172 if (up->offset + nr_args > ctx->nr_user_bufs)
9175 for (done = 0; done < nr_args; done++) {
9176 struct io_mapped_ubuf *imu;
9177 int offset = up->offset + done;
9180 err = io_copy_iov(ctx, &iov, iovs, done);
9183 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9187 err = io_buffer_validate(&iov);
9190 if (!iov.iov_base && tag) {
9194 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9198 i = array_index_nospec(offset, ctx->nr_user_bufs);
9199 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9200 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9201 ctx->rsrc_node, ctx->user_bufs[i]);
9202 if (unlikely(err)) {
9203 io_buffer_unmap(ctx, &imu);
9206 ctx->user_bufs[i] = NULL;
9207 needs_switch = true;
9210 ctx->user_bufs[i] = imu;
9211 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9215 io_rsrc_node_switch(ctx, ctx->buf_data);
9216 return done ? done : err;
9219 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9221 __s32 __user *fds = arg;
9227 if (copy_from_user(&fd, fds, sizeof(*fds)))
9230 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9231 if (IS_ERR(ctx->cq_ev_fd)) {
9232 int ret = PTR_ERR(ctx->cq_ev_fd);
9234 ctx->cq_ev_fd = NULL;
9241 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9243 if (ctx->cq_ev_fd) {
9244 eventfd_ctx_put(ctx->cq_ev_fd);
9245 ctx->cq_ev_fd = NULL;
9252 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9254 struct io_buffer *buf;
9255 unsigned long index;
9257 xa_for_each(&ctx->io_buffers, index, buf)
9258 __io_remove_buffers(ctx, buf, index, -1U);
9261 static void io_req_cache_free(struct list_head *list)
9263 struct io_kiocb *req, *nxt;
9265 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9266 list_del(&req->inflight_entry);
9267 kmem_cache_free(req_cachep, req);
9271 static void io_req_caches_free(struct io_ring_ctx *ctx)
9273 struct io_submit_state *state = &ctx->submit_state;
9275 mutex_lock(&ctx->uring_lock);
9277 if (state->free_reqs) {
9278 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9279 state->free_reqs = 0;
9282 io_flush_cached_locked_reqs(ctx, state);
9283 io_req_cache_free(&state->free_list);
9284 mutex_unlock(&ctx->uring_lock);
9287 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9289 if (data && !atomic_dec_and_test(&data->refs))
9290 wait_for_completion(&data->done);
9293 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9295 io_sq_thread_finish(ctx);
9297 if (ctx->mm_account) {
9298 mmdrop(ctx->mm_account);
9299 ctx->mm_account = NULL;
9302 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9303 io_wait_rsrc_data(ctx->buf_data);
9304 io_wait_rsrc_data(ctx->file_data);
9306 mutex_lock(&ctx->uring_lock);
9308 __io_sqe_buffers_unregister(ctx);
9310 __io_sqe_files_unregister(ctx);
9312 __io_cqring_overflow_flush(ctx, true);
9313 mutex_unlock(&ctx->uring_lock);
9314 io_eventfd_unregister(ctx);
9315 io_destroy_buffers(ctx);
9317 put_cred(ctx->sq_creds);
9319 /* there are no registered resources left, nobody uses it */
9321 io_rsrc_node_destroy(ctx->rsrc_node);
9322 if (ctx->rsrc_backup_node)
9323 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9324 flush_delayed_work(&ctx->rsrc_put_work);
9326 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9327 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9329 #if defined(CONFIG_UNIX)
9330 if (ctx->ring_sock) {
9331 ctx->ring_sock->file = NULL; /* so that iput() is called */
9332 sock_release(ctx->ring_sock);
9335 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9337 io_mem_free(ctx->rings);
9338 io_mem_free(ctx->sq_sqes);
9340 percpu_ref_exit(&ctx->refs);
9341 free_uid(ctx->user);
9342 io_req_caches_free(ctx);
9344 io_wq_put_hash(ctx->hash_map);
9345 kfree(ctx->cancel_hash);
9346 kfree(ctx->dummy_ubuf);
9350 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9352 struct io_ring_ctx *ctx = file->private_data;
9355 poll_wait(file, &ctx->poll_wait, wait);
9357 * synchronizes with barrier from wq_has_sleeper call in
9361 if (!io_sqring_full(ctx))
9362 mask |= EPOLLOUT | EPOLLWRNORM;
9365 * Don't flush cqring overflow list here, just do a simple check.
9366 * Otherwise there could possible be ABBA deadlock:
9369 * lock(&ctx->uring_lock);
9371 * lock(&ctx->uring_lock);
9374 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9375 * pushs them to do the flush.
9377 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9378 mask |= EPOLLIN | EPOLLRDNORM;
9383 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9385 const struct cred *creds;
9387 creds = xa_erase(&ctx->personalities, id);
9396 struct io_tctx_exit {
9397 struct callback_head task_work;
9398 struct completion completion;
9399 struct io_ring_ctx *ctx;
9402 static void io_tctx_exit_cb(struct callback_head *cb)
9404 struct io_uring_task *tctx = current->io_uring;
9405 struct io_tctx_exit *work;
9407 work = container_of(cb, struct io_tctx_exit, task_work);
9409 * When @in_idle, we're in cancellation and it's racy to remove the
9410 * node. It'll be removed by the end of cancellation, just ignore it.
9412 if (!atomic_read(&tctx->in_idle))
9413 io_uring_del_tctx_node((unsigned long)work->ctx);
9414 complete(&work->completion);
9417 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9419 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9421 return req->ctx == data;
9424 static void io_ring_exit_work(struct work_struct *work)
9426 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9427 unsigned long timeout = jiffies + HZ * 60 * 5;
9428 unsigned long interval = HZ / 20;
9429 struct io_tctx_exit exit;
9430 struct io_tctx_node *node;
9434 * If we're doing polled IO and end up having requests being
9435 * submitted async (out-of-line), then completions can come in while
9436 * we're waiting for refs to drop. We need to reap these manually,
9437 * as nobody else will be looking for them.
9440 io_uring_try_cancel_requests(ctx, NULL, true);
9442 struct io_sq_data *sqd = ctx->sq_data;
9443 struct task_struct *tsk;
9445 io_sq_thread_park(sqd);
9447 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9448 io_wq_cancel_cb(tsk->io_uring->io_wq,
9449 io_cancel_ctx_cb, ctx, true);
9450 io_sq_thread_unpark(sqd);
9453 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9454 /* there is little hope left, don't run it too often */
9457 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9459 init_completion(&exit.completion);
9460 init_task_work(&exit.task_work, io_tctx_exit_cb);
9463 * Some may use context even when all refs and requests have been put,
9464 * and they are free to do so while still holding uring_lock or
9465 * completion_lock, see io_req_task_submit(). Apart from other work,
9466 * this lock/unlock section also waits them to finish.
9468 mutex_lock(&ctx->uring_lock);
9469 while (!list_empty(&ctx->tctx_list)) {
9470 WARN_ON_ONCE(time_after(jiffies, timeout));
9472 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9474 /* don't spin on a single task if cancellation failed */
9475 list_rotate_left(&ctx->tctx_list);
9476 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9477 if (WARN_ON_ONCE(ret))
9479 wake_up_process(node->task);
9481 mutex_unlock(&ctx->uring_lock);
9482 wait_for_completion(&exit.completion);
9483 mutex_lock(&ctx->uring_lock);
9485 mutex_unlock(&ctx->uring_lock);
9486 spin_lock(&ctx->completion_lock);
9487 spin_unlock(&ctx->completion_lock);
9489 io_ring_ctx_free(ctx);
9492 /* Returns true if we found and killed one or more timeouts */
9493 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9496 struct io_kiocb *req, *tmp;
9499 spin_lock(&ctx->completion_lock);
9500 spin_lock_irq(&ctx->timeout_lock);
9501 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9502 if (io_match_task(req, tsk, cancel_all)) {
9503 io_kill_timeout(req, -ECANCELED);
9507 spin_unlock_irq(&ctx->timeout_lock);
9509 io_commit_cqring(ctx);
9510 spin_unlock(&ctx->completion_lock);
9512 io_cqring_ev_posted(ctx);
9513 return canceled != 0;
9516 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9518 unsigned long index;
9519 struct creds *creds;
9521 mutex_lock(&ctx->uring_lock);
9522 percpu_ref_kill(&ctx->refs);
9524 __io_cqring_overflow_flush(ctx, true);
9525 xa_for_each(&ctx->personalities, index, creds)
9526 io_unregister_personality(ctx, index);
9527 mutex_unlock(&ctx->uring_lock);
9529 io_kill_timeouts(ctx, NULL, true);
9530 io_poll_remove_all(ctx, NULL, true);
9532 /* if we failed setting up the ctx, we might not have any rings */
9533 io_iopoll_try_reap_events(ctx);
9535 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9537 * Use system_unbound_wq to avoid spawning tons of event kworkers
9538 * if we're exiting a ton of rings at the same time. It just adds
9539 * noise and overhead, there's no discernable change in runtime
9540 * over using system_wq.
9542 queue_work(system_unbound_wq, &ctx->exit_work);
9545 static int io_uring_release(struct inode *inode, struct file *file)
9547 struct io_ring_ctx *ctx = file->private_data;
9549 file->private_data = NULL;
9550 io_ring_ctx_wait_and_kill(ctx);
9554 struct io_task_cancel {
9555 struct task_struct *task;
9559 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9561 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9562 struct io_task_cancel *cancel = data;
9564 return io_match_task_safe(req, cancel->task, cancel->all);
9567 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9568 struct task_struct *task, bool cancel_all)
9570 struct io_defer_entry *de;
9573 spin_lock(&ctx->completion_lock);
9574 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9575 if (io_match_task_safe(de->req, task, cancel_all)) {
9576 list_cut_position(&list, &ctx->defer_list, &de->list);
9580 spin_unlock(&ctx->completion_lock);
9581 if (list_empty(&list))
9584 while (!list_empty(&list)) {
9585 de = list_first_entry(&list, struct io_defer_entry, list);
9586 list_del_init(&de->list);
9587 io_req_complete_failed(de->req, -ECANCELED);
9593 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9595 struct io_tctx_node *node;
9596 enum io_wq_cancel cret;
9599 mutex_lock(&ctx->uring_lock);
9600 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9601 struct io_uring_task *tctx = node->task->io_uring;
9604 * io_wq will stay alive while we hold uring_lock, because it's
9605 * killed after ctx nodes, which requires to take the lock.
9607 if (!tctx || !tctx->io_wq)
9609 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9610 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9612 mutex_unlock(&ctx->uring_lock);
9617 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9618 struct task_struct *task,
9621 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9622 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9625 enum io_wq_cancel cret;
9629 ret |= io_uring_try_cancel_iowq(ctx);
9630 } else if (tctx && tctx->io_wq) {
9632 * Cancels requests of all rings, not only @ctx, but
9633 * it's fine as the task is in exit/exec.
9635 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9637 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9640 /* SQPOLL thread does its own polling */
9641 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9642 (ctx->sq_data && ctx->sq_data->thread == current)) {
9643 while (!list_empty_careful(&ctx->iopoll_list)) {
9644 io_iopoll_try_reap_events(ctx);
9649 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9650 ret |= io_poll_remove_all(ctx, task, cancel_all);
9651 ret |= io_kill_timeouts(ctx, task, cancel_all);
9653 ret |= io_run_task_work();
9660 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9662 struct io_uring_task *tctx = current->io_uring;
9663 struct io_tctx_node *node;
9666 if (unlikely(!tctx)) {
9667 ret = io_uring_alloc_task_context(current, ctx);
9671 tctx = current->io_uring;
9672 if (ctx->iowq_limits_set) {
9673 unsigned int limits[2] = { ctx->iowq_limits[0],
9674 ctx->iowq_limits[1], };
9676 ret = io_wq_max_workers(tctx->io_wq, limits);
9681 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9682 node = kmalloc(sizeof(*node), GFP_KERNEL);
9686 node->task = current;
9688 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9695 mutex_lock(&ctx->uring_lock);
9696 list_add(&node->ctx_node, &ctx->tctx_list);
9697 mutex_unlock(&ctx->uring_lock);
9704 * Note that this task has used io_uring. We use it for cancelation purposes.
9706 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9708 struct io_uring_task *tctx = current->io_uring;
9710 if (likely(tctx && tctx->last == ctx))
9712 return __io_uring_add_tctx_node(ctx);
9716 * Remove this io_uring_file -> task mapping.
9718 static void io_uring_del_tctx_node(unsigned long index)
9720 struct io_uring_task *tctx = current->io_uring;
9721 struct io_tctx_node *node;
9725 node = xa_erase(&tctx->xa, index);
9729 WARN_ON_ONCE(current != node->task);
9730 WARN_ON_ONCE(list_empty(&node->ctx_node));
9732 mutex_lock(&node->ctx->uring_lock);
9733 list_del(&node->ctx_node);
9734 mutex_unlock(&node->ctx->uring_lock);
9736 if (tctx->last == node->ctx)
9741 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9743 struct io_wq *wq = tctx->io_wq;
9744 struct io_tctx_node *node;
9745 unsigned long index;
9747 xa_for_each(&tctx->xa, index, node) {
9748 io_uring_del_tctx_node(index);
9753 * Must be after io_uring_del_task_file() (removes nodes under
9754 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9756 io_wq_put_and_exit(wq);
9761 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9764 return atomic_read(&tctx->inflight_tracked);
9765 return percpu_counter_sum(&tctx->inflight);
9768 static void io_uring_drop_tctx_refs(struct task_struct *task)
9770 struct io_uring_task *tctx = task->io_uring;
9771 unsigned int refs = tctx->cached_refs;
9774 tctx->cached_refs = 0;
9775 percpu_counter_sub(&tctx->inflight, refs);
9776 put_task_struct_many(task, refs);
9781 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9782 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9784 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9786 struct io_uring_task *tctx = current->io_uring;
9787 struct io_ring_ctx *ctx;
9791 WARN_ON_ONCE(sqd && sqd->thread != current);
9793 if (!current->io_uring)
9796 io_wq_exit_start(tctx->io_wq);
9798 atomic_inc(&tctx->in_idle);
9800 io_uring_drop_tctx_refs(current);
9801 /* read completions before cancelations */
9802 inflight = tctx_inflight(tctx, !cancel_all);
9807 struct io_tctx_node *node;
9808 unsigned long index;
9810 xa_for_each(&tctx->xa, index, node) {
9811 /* sqpoll task will cancel all its requests */
9812 if (node->ctx->sq_data)
9814 io_uring_try_cancel_requests(node->ctx, current,
9818 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9819 io_uring_try_cancel_requests(ctx, current,
9823 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9825 io_uring_drop_tctx_refs(current);
9828 * If we've seen completions, retry without waiting. This
9829 * avoids a race where a completion comes in before we did
9830 * prepare_to_wait().
9832 if (inflight == tctx_inflight(tctx, !cancel_all))
9834 finish_wait(&tctx->wait, &wait);
9836 atomic_dec(&tctx->in_idle);
9838 io_uring_clean_tctx(tctx);
9840 /* for exec all current's requests should be gone, kill tctx */
9841 __io_uring_free(current);
9845 void __io_uring_cancel(bool cancel_all)
9847 io_uring_cancel_generic(cancel_all, NULL);
9850 static void *io_uring_validate_mmap_request(struct file *file,
9851 loff_t pgoff, size_t sz)
9853 struct io_ring_ctx *ctx = file->private_data;
9854 loff_t offset = pgoff << PAGE_SHIFT;
9859 case IORING_OFF_SQ_RING:
9860 case IORING_OFF_CQ_RING:
9863 case IORING_OFF_SQES:
9867 return ERR_PTR(-EINVAL);
9870 page = virt_to_head_page(ptr);
9871 if (sz > page_size(page))
9872 return ERR_PTR(-EINVAL);
9879 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9881 size_t sz = vma->vm_end - vma->vm_start;
9885 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9887 return PTR_ERR(ptr);
9889 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9890 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9893 #else /* !CONFIG_MMU */
9895 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9897 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9900 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9902 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9905 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9906 unsigned long addr, unsigned long len,
9907 unsigned long pgoff, unsigned long flags)
9911 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9913 return PTR_ERR(ptr);
9915 return (unsigned long) ptr;
9918 #endif /* !CONFIG_MMU */
9920 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9925 if (!io_sqring_full(ctx))
9927 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9929 if (!io_sqring_full(ctx))
9932 } while (!signal_pending(current));
9934 finish_wait(&ctx->sqo_sq_wait, &wait);
9938 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9939 struct __kernel_timespec __user **ts,
9940 const sigset_t __user **sig)
9942 struct io_uring_getevents_arg arg;
9945 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9946 * is just a pointer to the sigset_t.
9948 if (!(flags & IORING_ENTER_EXT_ARG)) {
9949 *sig = (const sigset_t __user *) argp;
9955 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9956 * timespec and sigset_t pointers if good.
9958 if (*argsz != sizeof(arg))
9960 if (copy_from_user(&arg, argp, sizeof(arg)))
9962 *sig = u64_to_user_ptr(arg.sigmask);
9963 *argsz = arg.sigmask_sz;
9964 *ts = u64_to_user_ptr(arg.ts);
9968 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9969 u32, min_complete, u32, flags, const void __user *, argp,
9972 struct io_ring_ctx *ctx;
9979 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9980 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9984 if (unlikely(!f.file))
9988 if (unlikely(f.file->f_op != &io_uring_fops))
9992 ctx = f.file->private_data;
9993 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9997 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10001 * For SQ polling, the thread will do all submissions and completions.
10002 * Just return the requested submit count, and wake the thread if
10003 * we were asked to.
10006 if (ctx->flags & IORING_SETUP_SQPOLL) {
10007 io_cqring_overflow_flush(ctx);
10009 if (unlikely(ctx->sq_data->thread == NULL)) {
10013 if (flags & IORING_ENTER_SQ_WAKEUP)
10014 wake_up(&ctx->sq_data->wait);
10015 if (flags & IORING_ENTER_SQ_WAIT) {
10016 ret = io_sqpoll_wait_sq(ctx);
10020 submitted = to_submit;
10021 } else if (to_submit) {
10022 ret = io_uring_add_tctx_node(ctx);
10025 mutex_lock(&ctx->uring_lock);
10026 submitted = io_submit_sqes(ctx, to_submit);
10027 mutex_unlock(&ctx->uring_lock);
10029 if (submitted != to_submit)
10032 if (flags & IORING_ENTER_GETEVENTS) {
10033 const sigset_t __user *sig;
10034 struct __kernel_timespec __user *ts;
10036 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10040 min_complete = min(min_complete, ctx->cq_entries);
10043 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10044 * space applications don't need to do io completion events
10045 * polling again, they can rely on io_sq_thread to do polling
10046 * work, which can reduce cpu usage and uring_lock contention.
10048 if (ctx->flags & IORING_SETUP_IOPOLL &&
10049 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10050 ret = io_iopoll_check(ctx, min_complete);
10052 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10057 percpu_ref_put(&ctx->refs);
10060 return submitted ? submitted : ret;
10063 #ifdef CONFIG_PROC_FS
10064 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10065 const struct cred *cred)
10067 struct user_namespace *uns = seq_user_ns(m);
10068 struct group_info *gi;
10073 seq_printf(m, "%5d\n", id);
10074 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10075 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10076 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10077 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10078 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10079 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10080 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10081 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10082 seq_puts(m, "\n\tGroups:\t");
10083 gi = cred->group_info;
10084 for (g = 0; g < gi->ngroups; g++) {
10085 seq_put_decimal_ull(m, g ? " " : "",
10086 from_kgid_munged(uns, gi->gid[g]));
10088 seq_puts(m, "\n\tCapEff:\t");
10089 cap = cred->cap_effective;
10090 CAP_FOR_EACH_U32(__capi)
10091 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10096 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10098 struct io_sq_data *sq = NULL;
10103 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10104 * since fdinfo case grabs it in the opposite direction of normal use
10105 * cases. If we fail to get the lock, we just don't iterate any
10106 * structures that could be going away outside the io_uring mutex.
10108 has_lock = mutex_trylock(&ctx->uring_lock);
10110 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10116 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10117 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10118 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10119 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10120 struct file *f = io_file_from_index(ctx, i);
10123 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10125 seq_printf(m, "%5u: <none>\n", i);
10127 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10128 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10129 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10130 unsigned int len = buf->ubuf_end - buf->ubuf;
10132 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10134 if (has_lock && !xa_empty(&ctx->personalities)) {
10135 unsigned long index;
10136 const struct cred *cred;
10138 seq_printf(m, "Personalities:\n");
10139 xa_for_each(&ctx->personalities, index, cred)
10140 io_uring_show_cred(m, index, cred);
10142 seq_printf(m, "PollList:\n");
10143 spin_lock(&ctx->completion_lock);
10144 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10145 struct hlist_head *list = &ctx->cancel_hash[i];
10146 struct io_kiocb *req;
10148 hlist_for_each_entry(req, list, hash_node)
10149 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10150 req->task->task_works != NULL);
10152 spin_unlock(&ctx->completion_lock);
10154 mutex_unlock(&ctx->uring_lock);
10157 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10159 struct io_ring_ctx *ctx = f->private_data;
10161 if (percpu_ref_tryget(&ctx->refs)) {
10162 __io_uring_show_fdinfo(ctx, m);
10163 percpu_ref_put(&ctx->refs);
10168 static const struct file_operations io_uring_fops = {
10169 .release = io_uring_release,
10170 .mmap = io_uring_mmap,
10172 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10173 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10175 .poll = io_uring_poll,
10176 #ifdef CONFIG_PROC_FS
10177 .show_fdinfo = io_uring_show_fdinfo,
10181 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10182 struct io_uring_params *p)
10184 struct io_rings *rings;
10185 size_t size, sq_array_offset;
10187 /* make sure these are sane, as we already accounted them */
10188 ctx->sq_entries = p->sq_entries;
10189 ctx->cq_entries = p->cq_entries;
10191 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10192 if (size == SIZE_MAX)
10195 rings = io_mem_alloc(size);
10199 ctx->rings = rings;
10200 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10201 rings->sq_ring_mask = p->sq_entries - 1;
10202 rings->cq_ring_mask = p->cq_entries - 1;
10203 rings->sq_ring_entries = p->sq_entries;
10204 rings->cq_ring_entries = p->cq_entries;
10206 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10207 if (size == SIZE_MAX) {
10208 io_mem_free(ctx->rings);
10213 ctx->sq_sqes = io_mem_alloc(size);
10214 if (!ctx->sq_sqes) {
10215 io_mem_free(ctx->rings);
10223 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10227 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10231 ret = io_uring_add_tctx_node(ctx);
10236 fd_install(fd, file);
10241 * Allocate an anonymous fd, this is what constitutes the application
10242 * visible backing of an io_uring instance. The application mmaps this
10243 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10244 * we have to tie this fd to a socket for file garbage collection purposes.
10246 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10249 #if defined(CONFIG_UNIX)
10252 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10255 return ERR_PTR(ret);
10258 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10259 O_RDWR | O_CLOEXEC);
10260 #if defined(CONFIG_UNIX)
10261 if (IS_ERR(file)) {
10262 sock_release(ctx->ring_sock);
10263 ctx->ring_sock = NULL;
10265 ctx->ring_sock->file = file;
10271 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10272 struct io_uring_params __user *params)
10274 struct io_ring_ctx *ctx;
10280 if (entries > IORING_MAX_ENTRIES) {
10281 if (!(p->flags & IORING_SETUP_CLAMP))
10283 entries = IORING_MAX_ENTRIES;
10287 * Use twice as many entries for the CQ ring. It's possible for the
10288 * application to drive a higher depth than the size of the SQ ring,
10289 * since the sqes are only used at submission time. This allows for
10290 * some flexibility in overcommitting a bit. If the application has
10291 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10292 * of CQ ring entries manually.
10294 p->sq_entries = roundup_pow_of_two(entries);
10295 if (p->flags & IORING_SETUP_CQSIZE) {
10297 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10298 * to a power-of-two, if it isn't already. We do NOT impose
10299 * any cq vs sq ring sizing.
10301 if (!p->cq_entries)
10303 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10304 if (!(p->flags & IORING_SETUP_CLAMP))
10306 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10308 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10309 if (p->cq_entries < p->sq_entries)
10312 p->cq_entries = 2 * p->sq_entries;
10315 ctx = io_ring_ctx_alloc(p);
10318 ctx->compat = in_compat_syscall();
10319 if (!capable(CAP_IPC_LOCK))
10320 ctx->user = get_uid(current_user());
10323 * This is just grabbed for accounting purposes. When a process exits,
10324 * the mm is exited and dropped before the files, hence we need to hang
10325 * on to this mm purely for the purposes of being able to unaccount
10326 * memory (locked/pinned vm). It's not used for anything else.
10328 mmgrab(current->mm);
10329 ctx->mm_account = current->mm;
10331 ret = io_allocate_scq_urings(ctx, p);
10335 ret = io_sq_offload_create(ctx, p);
10338 /* always set a rsrc node */
10339 ret = io_rsrc_node_switch_start(ctx);
10342 io_rsrc_node_switch(ctx, NULL);
10344 memset(&p->sq_off, 0, sizeof(p->sq_off));
10345 p->sq_off.head = offsetof(struct io_rings, sq.head);
10346 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10347 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10348 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10349 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10350 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10351 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10353 memset(&p->cq_off, 0, sizeof(p->cq_off));
10354 p->cq_off.head = offsetof(struct io_rings, cq.head);
10355 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10356 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10357 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10358 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10359 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10360 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10362 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10363 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10364 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10365 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10366 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10367 IORING_FEAT_RSRC_TAGS;
10369 if (copy_to_user(params, p, sizeof(*p))) {
10374 file = io_uring_get_file(ctx);
10375 if (IS_ERR(file)) {
10376 ret = PTR_ERR(file);
10381 * Install ring fd as the very last thing, so we don't risk someone
10382 * having closed it before we finish setup
10384 ret = io_uring_install_fd(ctx, file);
10386 /* fput will clean it up */
10391 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10394 io_ring_ctx_wait_and_kill(ctx);
10399 * Sets up an aio uring context, and returns the fd. Applications asks for a
10400 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10401 * params structure passed in.
10403 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10405 struct io_uring_params p;
10408 if (copy_from_user(&p, params, sizeof(p)))
10410 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10415 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10416 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10417 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10418 IORING_SETUP_R_DISABLED))
10421 return io_uring_create(entries, &p, params);
10424 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10425 struct io_uring_params __user *, params)
10427 return io_uring_setup(entries, params);
10430 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10432 struct io_uring_probe *p;
10436 size = struct_size(p, ops, nr_args);
10437 if (size == SIZE_MAX)
10439 p = kzalloc(size, GFP_KERNEL);
10444 if (copy_from_user(p, arg, size))
10447 if (memchr_inv(p, 0, size))
10450 p->last_op = IORING_OP_LAST - 1;
10451 if (nr_args > IORING_OP_LAST)
10452 nr_args = IORING_OP_LAST;
10454 for (i = 0; i < nr_args; i++) {
10456 if (!io_op_defs[i].not_supported)
10457 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10462 if (copy_to_user(arg, p, size))
10469 static int io_register_personality(struct io_ring_ctx *ctx)
10471 const struct cred *creds;
10475 creds = get_current_cred();
10477 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10478 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10486 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10487 unsigned int nr_args)
10489 struct io_uring_restriction *res;
10493 /* Restrictions allowed only if rings started disabled */
10494 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10497 /* We allow only a single restrictions registration */
10498 if (ctx->restrictions.registered)
10501 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10504 size = array_size(nr_args, sizeof(*res));
10505 if (size == SIZE_MAX)
10508 res = memdup_user(arg, size);
10510 return PTR_ERR(res);
10514 for (i = 0; i < nr_args; i++) {
10515 switch (res[i].opcode) {
10516 case IORING_RESTRICTION_REGISTER_OP:
10517 if (res[i].register_op >= IORING_REGISTER_LAST) {
10522 __set_bit(res[i].register_op,
10523 ctx->restrictions.register_op);
10525 case IORING_RESTRICTION_SQE_OP:
10526 if (res[i].sqe_op >= IORING_OP_LAST) {
10531 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10533 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10534 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10536 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10537 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10546 /* Reset all restrictions if an error happened */
10548 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10550 ctx->restrictions.registered = true;
10556 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10558 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10561 if (ctx->restrictions.registered)
10562 ctx->restricted = 1;
10564 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10565 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10566 wake_up(&ctx->sq_data->wait);
10570 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10571 struct io_uring_rsrc_update2 *up,
10579 if (check_add_overflow(up->offset, nr_args, &tmp))
10581 err = io_rsrc_node_switch_start(ctx);
10586 case IORING_RSRC_FILE:
10587 return __io_sqe_files_update(ctx, up, nr_args);
10588 case IORING_RSRC_BUFFER:
10589 return __io_sqe_buffers_update(ctx, up, nr_args);
10594 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10597 struct io_uring_rsrc_update2 up;
10601 memset(&up, 0, sizeof(up));
10602 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10604 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10607 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10608 unsigned size, unsigned type)
10610 struct io_uring_rsrc_update2 up;
10612 if (size != sizeof(up))
10614 if (copy_from_user(&up, arg, sizeof(up)))
10616 if (!up.nr || up.resv)
10618 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10621 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10622 unsigned int size, unsigned int type)
10624 struct io_uring_rsrc_register rr;
10626 /* keep it extendible */
10627 if (size != sizeof(rr))
10630 memset(&rr, 0, sizeof(rr));
10631 if (copy_from_user(&rr, arg, size))
10633 if (!rr.nr || rr.resv || rr.resv2)
10637 case IORING_RSRC_FILE:
10638 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10639 rr.nr, u64_to_user_ptr(rr.tags));
10640 case IORING_RSRC_BUFFER:
10641 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10642 rr.nr, u64_to_user_ptr(rr.tags));
10647 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10650 struct io_uring_task *tctx = current->io_uring;
10651 cpumask_var_t new_mask;
10654 if (!tctx || !tctx->io_wq)
10657 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10660 cpumask_clear(new_mask);
10661 if (len > cpumask_size())
10662 len = cpumask_size();
10664 if (copy_from_user(new_mask, arg, len)) {
10665 free_cpumask_var(new_mask);
10669 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10670 free_cpumask_var(new_mask);
10674 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10676 struct io_uring_task *tctx = current->io_uring;
10678 if (!tctx || !tctx->io_wq)
10681 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10684 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10686 __must_hold(&ctx->uring_lock)
10688 struct io_tctx_node *node;
10689 struct io_uring_task *tctx = NULL;
10690 struct io_sq_data *sqd = NULL;
10691 __u32 new_count[2];
10694 if (copy_from_user(new_count, arg, sizeof(new_count)))
10696 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10697 if (new_count[i] > INT_MAX)
10700 if (ctx->flags & IORING_SETUP_SQPOLL) {
10701 sqd = ctx->sq_data;
10704 * Observe the correct sqd->lock -> ctx->uring_lock
10705 * ordering. Fine to drop uring_lock here, we hold
10706 * a ref to the ctx.
10708 refcount_inc(&sqd->refs);
10709 mutex_unlock(&ctx->uring_lock);
10710 mutex_lock(&sqd->lock);
10711 mutex_lock(&ctx->uring_lock);
10713 tctx = sqd->thread->io_uring;
10716 tctx = current->io_uring;
10719 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10721 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10723 ctx->iowq_limits[i] = new_count[i];
10724 ctx->iowq_limits_set = true;
10727 if (tctx && tctx->io_wq) {
10728 ret = io_wq_max_workers(tctx->io_wq, new_count);
10732 memset(new_count, 0, sizeof(new_count));
10736 mutex_unlock(&sqd->lock);
10737 io_put_sq_data(sqd);
10740 if (copy_to_user(arg, new_count, sizeof(new_count)))
10743 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10747 /* now propagate the restriction to all registered users */
10748 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10749 struct io_uring_task *tctx = node->task->io_uring;
10751 if (WARN_ON_ONCE(!tctx->io_wq))
10754 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10755 new_count[i] = ctx->iowq_limits[i];
10756 /* ignore errors, it always returns zero anyway */
10757 (void)io_wq_max_workers(tctx->io_wq, new_count);
10762 mutex_unlock(&sqd->lock);
10763 io_put_sq_data(sqd);
10768 static bool io_register_op_must_quiesce(int op)
10771 case IORING_REGISTER_BUFFERS:
10772 case IORING_UNREGISTER_BUFFERS:
10773 case IORING_REGISTER_FILES:
10774 case IORING_UNREGISTER_FILES:
10775 case IORING_REGISTER_FILES_UPDATE:
10776 case IORING_REGISTER_PROBE:
10777 case IORING_REGISTER_PERSONALITY:
10778 case IORING_UNREGISTER_PERSONALITY:
10779 case IORING_REGISTER_FILES2:
10780 case IORING_REGISTER_FILES_UPDATE2:
10781 case IORING_REGISTER_BUFFERS2:
10782 case IORING_REGISTER_BUFFERS_UPDATE:
10783 case IORING_REGISTER_IOWQ_AFF:
10784 case IORING_UNREGISTER_IOWQ_AFF:
10785 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10792 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10796 percpu_ref_kill(&ctx->refs);
10799 * Drop uring mutex before waiting for references to exit. If another
10800 * thread is currently inside io_uring_enter() it might need to grab the
10801 * uring_lock to make progress. If we hold it here across the drain
10802 * wait, then we can deadlock. It's safe to drop the mutex here, since
10803 * no new references will come in after we've killed the percpu ref.
10805 mutex_unlock(&ctx->uring_lock);
10807 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10810 ret = io_run_task_work_sig();
10811 } while (ret >= 0);
10812 mutex_lock(&ctx->uring_lock);
10815 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10819 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10820 void __user *arg, unsigned nr_args)
10821 __releases(ctx->uring_lock)
10822 __acquires(ctx->uring_lock)
10827 * We're inside the ring mutex, if the ref is already dying, then
10828 * someone else killed the ctx or is already going through
10829 * io_uring_register().
10831 if (percpu_ref_is_dying(&ctx->refs))
10834 if (ctx->restricted) {
10835 if (opcode >= IORING_REGISTER_LAST)
10837 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10838 if (!test_bit(opcode, ctx->restrictions.register_op))
10842 if (io_register_op_must_quiesce(opcode)) {
10843 ret = io_ctx_quiesce(ctx);
10849 case IORING_REGISTER_BUFFERS:
10850 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10852 case IORING_UNREGISTER_BUFFERS:
10854 if (arg || nr_args)
10856 ret = io_sqe_buffers_unregister(ctx);
10858 case IORING_REGISTER_FILES:
10859 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10861 case IORING_UNREGISTER_FILES:
10863 if (arg || nr_args)
10865 ret = io_sqe_files_unregister(ctx);
10867 case IORING_REGISTER_FILES_UPDATE:
10868 ret = io_register_files_update(ctx, arg, nr_args);
10870 case IORING_REGISTER_EVENTFD:
10871 case IORING_REGISTER_EVENTFD_ASYNC:
10875 ret = io_eventfd_register(ctx, arg);
10878 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10879 ctx->eventfd_async = 1;
10881 ctx->eventfd_async = 0;
10883 case IORING_UNREGISTER_EVENTFD:
10885 if (arg || nr_args)
10887 ret = io_eventfd_unregister(ctx);
10889 case IORING_REGISTER_PROBE:
10891 if (!arg || nr_args > 256)
10893 ret = io_probe(ctx, arg, nr_args);
10895 case IORING_REGISTER_PERSONALITY:
10897 if (arg || nr_args)
10899 ret = io_register_personality(ctx);
10901 case IORING_UNREGISTER_PERSONALITY:
10905 ret = io_unregister_personality(ctx, nr_args);
10907 case IORING_REGISTER_ENABLE_RINGS:
10909 if (arg || nr_args)
10911 ret = io_register_enable_rings(ctx);
10913 case IORING_REGISTER_RESTRICTIONS:
10914 ret = io_register_restrictions(ctx, arg, nr_args);
10916 case IORING_REGISTER_FILES2:
10917 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10919 case IORING_REGISTER_FILES_UPDATE2:
10920 ret = io_register_rsrc_update(ctx, arg, nr_args,
10923 case IORING_REGISTER_BUFFERS2:
10924 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10926 case IORING_REGISTER_BUFFERS_UPDATE:
10927 ret = io_register_rsrc_update(ctx, arg, nr_args,
10928 IORING_RSRC_BUFFER);
10930 case IORING_REGISTER_IOWQ_AFF:
10932 if (!arg || !nr_args)
10934 ret = io_register_iowq_aff(ctx, arg, nr_args);
10936 case IORING_UNREGISTER_IOWQ_AFF:
10938 if (arg || nr_args)
10940 ret = io_unregister_iowq_aff(ctx);
10942 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10944 if (!arg || nr_args != 2)
10946 ret = io_register_iowq_max_workers(ctx, arg);
10953 if (io_register_op_must_quiesce(opcode)) {
10954 /* bring the ctx back to life */
10955 percpu_ref_reinit(&ctx->refs);
10956 reinit_completion(&ctx->ref_comp);
10961 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10962 void __user *, arg, unsigned int, nr_args)
10964 struct io_ring_ctx *ctx;
10973 if (f.file->f_op != &io_uring_fops)
10976 ctx = f.file->private_data;
10978 io_run_task_work();
10980 mutex_lock(&ctx->uring_lock);
10981 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10982 mutex_unlock(&ctx->uring_lock);
10983 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10984 ctx->cq_ev_fd != NULL, ret);
10990 static int __init io_uring_init(void)
10992 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10993 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10994 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10997 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10998 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10999 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11000 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11001 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11002 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11003 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11004 BUILD_BUG_SQE_ELEM(8, __u64, off);
11005 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11006 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11007 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11008 BUILD_BUG_SQE_ELEM(24, __u32, len);
11009 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11010 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11011 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11012 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11013 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11014 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11015 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11016 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11017 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11018 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11019 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11020 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11021 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11022 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11023 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11024 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11025 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11026 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11027 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11028 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11029 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11031 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11032 sizeof(struct io_uring_rsrc_update));
11033 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11034 sizeof(struct io_uring_rsrc_update2));
11036 /* ->buf_index is u16 */
11037 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11039 /* should fit into one byte */
11040 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11042 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11043 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11045 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11049 __initcall(io_uring_init);