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 __cold void io_uring_drop_tctx_refs(struct task_struct *task)
1765 struct io_uring_task *tctx = task->io_uring;
1766 unsigned int refs = tctx->cached_refs;
1769 tctx->cached_refs = 0;
1770 percpu_counter_sub(&tctx->inflight, refs);
1771 put_task_struct_many(task, refs);
1775 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1776 long res, unsigned int cflags)
1778 struct io_overflow_cqe *ocqe;
1780 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1783 * If we're in ring overflow flush mode, or in task cancel mode,
1784 * or cannot allocate an overflow entry, then we need to drop it
1787 io_account_cq_overflow(ctx);
1790 if (list_empty(&ctx->cq_overflow_list)) {
1791 set_bit(0, &ctx->check_cq_overflow);
1792 WRITE_ONCE(ctx->rings->sq_flags,
1793 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1796 ocqe->cqe.user_data = user_data;
1797 ocqe->cqe.res = res;
1798 ocqe->cqe.flags = cflags;
1799 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1803 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1804 long res, unsigned int cflags)
1806 struct io_uring_cqe *cqe;
1808 trace_io_uring_complete(ctx, user_data, res, cflags);
1811 * If we can't get a cq entry, userspace overflowed the
1812 * submission (by quite a lot). Increment the overflow count in
1815 cqe = io_get_cqe(ctx);
1817 WRITE_ONCE(cqe->user_data, user_data);
1818 WRITE_ONCE(cqe->res, res);
1819 WRITE_ONCE(cqe->flags, cflags);
1822 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1825 /* not as hot to bloat with inlining */
1826 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1827 long res, unsigned int cflags)
1829 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1832 static void io_req_complete_post(struct io_kiocb *req, long res,
1833 unsigned int cflags)
1835 struct io_ring_ctx *ctx = req->ctx;
1837 spin_lock(&ctx->completion_lock);
1838 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1840 * If we're the last reference to this request, add to our locked
1843 if (req_ref_put_and_test(req)) {
1844 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1845 if (req->flags & IO_DISARM_MASK)
1846 io_disarm_next(req);
1848 io_req_task_queue(req->link);
1852 io_dismantle_req(req);
1853 io_put_task(req->task, 1);
1854 list_add(&req->inflight_entry, &ctx->locked_free_list);
1855 ctx->locked_free_nr++;
1857 if (!percpu_ref_tryget(&ctx->refs))
1860 io_commit_cqring(ctx);
1861 spin_unlock(&ctx->completion_lock);
1864 io_cqring_ev_posted(ctx);
1865 percpu_ref_put(&ctx->refs);
1869 static inline bool io_req_needs_clean(struct io_kiocb *req)
1871 return req->flags & IO_REQ_CLEAN_FLAGS;
1874 static void io_req_complete_state(struct io_kiocb *req, long res,
1875 unsigned int cflags)
1877 if (io_req_needs_clean(req))
1880 req->compl.cflags = cflags;
1881 req->flags |= REQ_F_COMPLETE_INLINE;
1884 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1885 long res, unsigned cflags)
1887 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1888 io_req_complete_state(req, res, cflags);
1890 io_req_complete_post(req, res, cflags);
1893 static inline void io_req_complete(struct io_kiocb *req, long res)
1895 __io_req_complete(req, 0, res, 0);
1898 static void io_req_complete_failed(struct io_kiocb *req, long res)
1901 io_req_complete_post(req, res, 0);
1904 static void io_req_complete_fail_submit(struct io_kiocb *req)
1907 * We don't submit, fail them all, for that replace hardlinks with
1908 * normal links. Extra REQ_F_LINK is tolerated.
1910 req->flags &= ~REQ_F_HARDLINK;
1911 req->flags |= REQ_F_LINK;
1912 io_req_complete_failed(req, req->result);
1916 * Don't initialise the fields below on every allocation, but do that in
1917 * advance and keep them valid across allocations.
1919 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1923 req->async_data = NULL;
1924 /* not necessary, but safer to zero */
1928 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1929 struct io_submit_state *state)
1931 spin_lock(&ctx->completion_lock);
1932 list_splice_init(&ctx->locked_free_list, &state->free_list);
1933 ctx->locked_free_nr = 0;
1934 spin_unlock(&ctx->completion_lock);
1937 /* Returns true IFF there are requests in the cache */
1938 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1940 struct io_submit_state *state = &ctx->submit_state;
1944 * If we have more than a batch's worth of requests in our IRQ side
1945 * locked cache, grab the lock and move them over to our submission
1948 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1949 io_flush_cached_locked_reqs(ctx, state);
1951 nr = state->free_reqs;
1952 while (!list_empty(&state->free_list)) {
1953 struct io_kiocb *req = list_first_entry(&state->free_list,
1954 struct io_kiocb, inflight_entry);
1956 list_del(&req->inflight_entry);
1957 state->reqs[nr++] = req;
1958 if (nr == ARRAY_SIZE(state->reqs))
1962 state->free_reqs = nr;
1967 * A request might get retired back into the request caches even before opcode
1968 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1969 * Because of that, io_alloc_req() should be called only under ->uring_lock
1970 * and with extra caution to not get a request that is still worked on.
1972 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1973 __must_hold(&ctx->uring_lock)
1975 struct io_submit_state *state = &ctx->submit_state;
1976 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1979 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1981 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1984 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1988 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1989 * retry single alloc to be on the safe side.
1991 if (unlikely(ret <= 0)) {
1992 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1993 if (!state->reqs[0])
1998 for (i = 0; i < ret; i++)
1999 io_preinit_req(state->reqs[i], ctx);
2000 state->free_reqs = ret;
2003 return state->reqs[state->free_reqs];
2006 static inline void io_put_file(struct file *file)
2012 static void io_dismantle_req(struct io_kiocb *req)
2014 unsigned int flags = req->flags;
2016 if (io_req_needs_clean(req))
2018 if (!(flags & REQ_F_FIXED_FILE))
2019 io_put_file(req->file);
2020 if (req->fixed_rsrc_refs)
2021 percpu_ref_put(req->fixed_rsrc_refs);
2022 if (req->async_data) {
2023 kfree(req->async_data);
2024 req->async_data = NULL;
2028 static void __io_free_req(struct io_kiocb *req)
2030 struct io_ring_ctx *ctx = req->ctx;
2032 io_dismantle_req(req);
2033 io_put_task(req->task, 1);
2035 spin_lock(&ctx->completion_lock);
2036 list_add(&req->inflight_entry, &ctx->locked_free_list);
2037 ctx->locked_free_nr++;
2038 spin_unlock(&ctx->completion_lock);
2040 percpu_ref_put(&ctx->refs);
2043 static inline void io_remove_next_linked(struct io_kiocb *req)
2045 struct io_kiocb *nxt = req->link;
2047 req->link = nxt->link;
2051 static bool io_kill_linked_timeout(struct io_kiocb *req)
2052 __must_hold(&req->ctx->completion_lock)
2053 __must_hold(&req->ctx->timeout_lock)
2055 struct io_kiocb *link = req->link;
2057 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2058 struct io_timeout_data *io = link->async_data;
2060 io_remove_next_linked(req);
2061 link->timeout.head = NULL;
2062 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2063 list_del(&link->timeout.list);
2064 io_cqring_fill_event(link->ctx, link->user_data,
2066 io_put_req_deferred(link);
2073 static void io_fail_links(struct io_kiocb *req)
2074 __must_hold(&req->ctx->completion_lock)
2076 struct io_kiocb *nxt, *link = req->link;
2080 long res = -ECANCELED;
2082 if (link->flags & REQ_F_FAIL)
2088 trace_io_uring_fail_link(req, link);
2089 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2090 io_put_req_deferred(link);
2095 static bool io_disarm_next(struct io_kiocb *req)
2096 __must_hold(&req->ctx->completion_lock)
2098 bool posted = false;
2100 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2101 struct io_kiocb *link = req->link;
2103 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2104 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2105 io_remove_next_linked(req);
2106 io_cqring_fill_event(link->ctx, link->user_data,
2108 io_put_req_deferred(link);
2111 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2112 struct io_ring_ctx *ctx = req->ctx;
2114 spin_lock_irq(&ctx->timeout_lock);
2115 posted = io_kill_linked_timeout(req);
2116 spin_unlock_irq(&ctx->timeout_lock);
2118 if (unlikely((req->flags & REQ_F_FAIL) &&
2119 !(req->flags & REQ_F_HARDLINK))) {
2120 posted |= (req->link != NULL);
2126 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2128 struct io_kiocb *nxt;
2131 * If LINK is set, we have dependent requests in this chain. If we
2132 * didn't fail this request, queue the first one up, moving any other
2133 * dependencies to the next request. In case of failure, fail the rest
2136 if (req->flags & IO_DISARM_MASK) {
2137 struct io_ring_ctx *ctx = req->ctx;
2140 spin_lock(&ctx->completion_lock);
2141 posted = io_disarm_next(req);
2143 io_commit_cqring(req->ctx);
2144 spin_unlock(&ctx->completion_lock);
2146 io_cqring_ev_posted(ctx);
2153 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2155 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2157 return __io_req_find_next(req);
2160 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2165 if (ctx->submit_state.compl_nr)
2166 io_submit_flush_completions(ctx);
2167 mutex_unlock(&ctx->uring_lock);
2170 percpu_ref_put(&ctx->refs);
2173 static void tctx_task_work(struct callback_head *cb)
2175 bool locked = false;
2176 struct io_ring_ctx *ctx = NULL;
2177 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2181 struct io_wq_work_node *node;
2183 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2184 io_submit_flush_completions(ctx);
2186 spin_lock_irq(&tctx->task_lock);
2187 node = tctx->task_list.first;
2188 INIT_WQ_LIST(&tctx->task_list);
2190 tctx->task_running = false;
2191 spin_unlock_irq(&tctx->task_lock);
2196 struct io_wq_work_node *next = node->next;
2197 struct io_kiocb *req = container_of(node, struct io_kiocb,
2200 if (req->ctx != ctx) {
2201 ctx_flush_and_put(ctx, &locked);
2203 /* if not contended, grab and improve batching */
2204 locked = mutex_trylock(&ctx->uring_lock);
2205 percpu_ref_get(&ctx->refs);
2207 req->io_task_work.func(req, &locked);
2214 ctx_flush_and_put(ctx, &locked);
2216 /* relaxed read is enough as only the task itself sets ->in_idle */
2217 if (unlikely(atomic_read(&tctx->in_idle)))
2218 io_uring_drop_tctx_refs(current);
2221 static void io_req_task_work_add(struct io_kiocb *req)
2223 struct task_struct *tsk = req->task;
2224 struct io_uring_task *tctx = tsk->io_uring;
2225 enum task_work_notify_mode notify;
2226 struct io_wq_work_node *node;
2227 unsigned long flags;
2230 WARN_ON_ONCE(!tctx);
2232 spin_lock_irqsave(&tctx->task_lock, flags);
2233 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2234 running = tctx->task_running;
2236 tctx->task_running = true;
2237 spin_unlock_irqrestore(&tctx->task_lock, flags);
2239 /* task_work already pending, we're done */
2244 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2245 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2246 * processing task_work. There's no reliable way to tell if TWA_RESUME
2249 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2250 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2251 wake_up_process(tsk);
2255 spin_lock_irqsave(&tctx->task_lock, flags);
2256 tctx->task_running = false;
2257 node = tctx->task_list.first;
2258 INIT_WQ_LIST(&tctx->task_list);
2259 spin_unlock_irqrestore(&tctx->task_lock, flags);
2262 req = container_of(node, struct io_kiocb, io_task_work.node);
2264 if (llist_add(&req->io_task_work.fallback_node,
2265 &req->ctx->fallback_llist))
2266 schedule_delayed_work(&req->ctx->fallback_work, 1);
2270 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2272 struct io_ring_ctx *ctx = req->ctx;
2274 /* not needed for normal modes, but SQPOLL depends on it */
2275 io_tw_lock(ctx, locked);
2276 io_req_complete_failed(req, req->result);
2279 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2281 struct io_ring_ctx *ctx = req->ctx;
2283 io_tw_lock(ctx, locked);
2284 /* req->task == current here, checking PF_EXITING is safe */
2285 if (likely(!(req->task->flags & PF_EXITING)))
2286 __io_queue_sqe(req);
2288 io_req_complete_failed(req, -EFAULT);
2291 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2294 req->io_task_work.func = io_req_task_cancel;
2295 io_req_task_work_add(req);
2298 static void io_req_task_queue(struct io_kiocb *req)
2300 req->io_task_work.func = io_req_task_submit;
2301 io_req_task_work_add(req);
2304 static void io_req_task_queue_reissue(struct io_kiocb *req)
2306 req->io_task_work.func = io_queue_async_work;
2307 io_req_task_work_add(req);
2310 static inline void io_queue_next(struct io_kiocb *req)
2312 struct io_kiocb *nxt = io_req_find_next(req);
2315 io_req_task_queue(nxt);
2318 static void io_free_req(struct io_kiocb *req)
2324 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2330 struct task_struct *task;
2335 static inline void io_init_req_batch(struct req_batch *rb)
2342 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2343 struct req_batch *rb)
2346 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2348 io_put_task(rb->task, rb->task_refs);
2351 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2352 struct io_submit_state *state)
2355 io_dismantle_req(req);
2357 if (req->task != rb->task) {
2359 io_put_task(rb->task, rb->task_refs);
2360 rb->task = req->task;
2366 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2367 state->reqs[state->free_reqs++] = req;
2369 list_add(&req->inflight_entry, &state->free_list);
2372 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2373 __must_hold(&ctx->uring_lock)
2375 struct io_submit_state *state = &ctx->submit_state;
2376 int i, nr = state->compl_nr;
2377 struct req_batch rb;
2379 spin_lock(&ctx->completion_lock);
2380 for (i = 0; i < nr; i++) {
2381 struct io_kiocb *req = state->compl_reqs[i];
2383 __io_cqring_fill_event(ctx, req->user_data, req->result,
2386 io_commit_cqring(ctx);
2387 spin_unlock(&ctx->completion_lock);
2388 io_cqring_ev_posted(ctx);
2390 io_init_req_batch(&rb);
2391 for (i = 0; i < nr; i++) {
2392 struct io_kiocb *req = state->compl_reqs[i];
2394 if (req_ref_put_and_test(req))
2395 io_req_free_batch(&rb, req, &ctx->submit_state);
2398 io_req_free_batch_finish(ctx, &rb);
2399 state->compl_nr = 0;
2403 * Drop reference to request, return next in chain (if there is one) if this
2404 * was the last reference to this request.
2406 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2408 struct io_kiocb *nxt = NULL;
2410 if (req_ref_put_and_test(req)) {
2411 nxt = io_req_find_next(req);
2417 static inline void io_put_req(struct io_kiocb *req)
2419 if (req_ref_put_and_test(req))
2423 static inline void io_put_req_deferred(struct io_kiocb *req)
2425 if (req_ref_put_and_test(req)) {
2426 req->io_task_work.func = io_free_req_work;
2427 io_req_task_work_add(req);
2431 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2433 /* See comment at the top of this file */
2435 return __io_cqring_events(ctx);
2438 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2440 struct io_rings *rings = ctx->rings;
2442 /* make sure SQ entry isn't read before tail */
2443 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2446 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2448 unsigned int cflags;
2450 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2451 cflags |= IORING_CQE_F_BUFFER;
2452 req->flags &= ~REQ_F_BUFFER_SELECTED;
2457 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2459 struct io_buffer *kbuf;
2461 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2463 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2464 return io_put_kbuf(req, kbuf);
2467 static inline bool io_run_task_work(void)
2469 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2470 __set_current_state(TASK_RUNNING);
2471 tracehook_notify_signal();
2479 * Find and free completed poll iocbs
2481 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2482 struct list_head *done)
2484 struct req_batch rb;
2485 struct io_kiocb *req;
2487 /* order with ->result store in io_complete_rw_iopoll() */
2490 io_init_req_batch(&rb);
2491 while (!list_empty(done)) {
2492 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2493 list_del(&req->inflight_entry);
2495 __io_cqring_fill_event(ctx, req->user_data, req->result,
2496 io_put_rw_kbuf(req));
2499 if (req_ref_put_and_test(req))
2500 io_req_free_batch(&rb, req, &ctx->submit_state);
2503 io_commit_cqring(ctx);
2504 io_cqring_ev_posted_iopoll(ctx);
2505 io_req_free_batch_finish(ctx, &rb);
2508 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2511 struct io_kiocb *req, *tmp;
2516 * Only spin for completions if we don't have multiple devices hanging
2517 * off our complete list, and we're under the requested amount.
2519 spin = !ctx->poll_multi_queue && *nr_events < min;
2521 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2522 struct kiocb *kiocb = &req->rw.kiocb;
2526 * Move completed and retryable entries to our local lists.
2527 * If we find a request that requires polling, break out
2528 * and complete those lists first, if we have entries there.
2530 if (READ_ONCE(req->iopoll_completed)) {
2531 list_move_tail(&req->inflight_entry, &done);
2534 if (!list_empty(&done))
2537 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2538 if (unlikely(ret < 0))
2543 /* iopoll may have completed current req */
2544 if (READ_ONCE(req->iopoll_completed))
2545 list_move_tail(&req->inflight_entry, &done);
2548 if (!list_empty(&done))
2549 io_iopoll_complete(ctx, nr_events, &done);
2555 * We can't just wait for polled events to come to us, we have to actively
2556 * find and complete them.
2558 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2560 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2563 mutex_lock(&ctx->uring_lock);
2564 while (!list_empty(&ctx->iopoll_list)) {
2565 unsigned int nr_events = 0;
2567 io_do_iopoll(ctx, &nr_events, 0);
2569 /* let it sleep and repeat later if can't complete a request */
2573 * Ensure we allow local-to-the-cpu processing to take place,
2574 * in this case we need to ensure that we reap all events.
2575 * Also let task_work, etc. to progress by releasing the mutex
2577 if (need_resched()) {
2578 mutex_unlock(&ctx->uring_lock);
2580 mutex_lock(&ctx->uring_lock);
2583 mutex_unlock(&ctx->uring_lock);
2586 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2588 unsigned int nr_events = 0;
2592 * We disallow the app entering submit/complete with polling, but we
2593 * still need to lock the ring to prevent racing with polled issue
2594 * that got punted to a workqueue.
2596 mutex_lock(&ctx->uring_lock);
2598 * Don't enter poll loop if we already have events pending.
2599 * If we do, we can potentially be spinning for commands that
2600 * already triggered a CQE (eg in error).
2602 if (test_bit(0, &ctx->check_cq_overflow))
2603 __io_cqring_overflow_flush(ctx, false);
2604 if (io_cqring_events(ctx))
2608 * If a submit got punted to a workqueue, we can have the
2609 * application entering polling for a command before it gets
2610 * issued. That app will hold the uring_lock for the duration
2611 * of the poll right here, so we need to take a breather every
2612 * now and then to ensure that the issue has a chance to add
2613 * the poll to the issued list. Otherwise we can spin here
2614 * forever, while the workqueue is stuck trying to acquire the
2617 if (list_empty(&ctx->iopoll_list)) {
2618 u32 tail = ctx->cached_cq_tail;
2620 mutex_unlock(&ctx->uring_lock);
2622 mutex_lock(&ctx->uring_lock);
2624 /* some requests don't go through iopoll_list */
2625 if (tail != ctx->cached_cq_tail ||
2626 list_empty(&ctx->iopoll_list))
2629 ret = io_do_iopoll(ctx, &nr_events, min);
2630 } while (!ret && nr_events < min && !need_resched());
2632 mutex_unlock(&ctx->uring_lock);
2636 static void kiocb_end_write(struct io_kiocb *req)
2639 * Tell lockdep we inherited freeze protection from submission
2642 if (req->flags & REQ_F_ISREG) {
2643 struct super_block *sb = file_inode(req->file)->i_sb;
2645 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2651 static bool io_resubmit_prep(struct io_kiocb *req)
2653 struct io_async_rw *rw = req->async_data;
2656 return !io_req_prep_async(req);
2657 iov_iter_restore(&rw->iter, &rw->iter_state);
2661 static bool io_rw_should_reissue(struct io_kiocb *req)
2663 umode_t mode = file_inode(req->file)->i_mode;
2664 struct io_ring_ctx *ctx = req->ctx;
2666 if (!S_ISBLK(mode) && !S_ISREG(mode))
2668 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2669 !(ctx->flags & IORING_SETUP_IOPOLL)))
2672 * If ref is dying, we might be running poll reap from the exit work.
2673 * Don't attempt to reissue from that path, just let it fail with
2676 if (percpu_ref_is_dying(&ctx->refs))
2679 * Play it safe and assume not safe to re-import and reissue if we're
2680 * not in the original thread group (or in task context).
2682 if (!same_thread_group(req->task, current) || !in_task())
2687 static bool io_resubmit_prep(struct io_kiocb *req)
2691 static bool io_rw_should_reissue(struct io_kiocb *req)
2697 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2699 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2700 kiocb_end_write(req);
2701 if (res != req->result) {
2702 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2703 io_rw_should_reissue(req)) {
2704 req->flags |= REQ_F_REISSUE;
2713 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2715 unsigned int cflags = io_put_rw_kbuf(req);
2716 long res = req->result;
2719 struct io_ring_ctx *ctx = req->ctx;
2720 struct io_submit_state *state = &ctx->submit_state;
2722 io_req_complete_state(req, res, cflags);
2723 state->compl_reqs[state->compl_nr++] = req;
2724 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2725 io_submit_flush_completions(ctx);
2727 io_req_complete_post(req, res, cflags);
2731 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2732 unsigned int issue_flags)
2734 if (__io_complete_rw_common(req, res))
2736 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2739 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2741 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2743 if (__io_complete_rw_common(req, res))
2746 req->io_task_work.func = io_req_task_complete;
2747 io_req_task_work_add(req);
2750 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2752 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2754 if (kiocb->ki_flags & IOCB_WRITE)
2755 kiocb_end_write(req);
2756 if (unlikely(res != req->result)) {
2757 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2758 req->flags |= REQ_F_REISSUE;
2763 WRITE_ONCE(req->result, res);
2764 /* order with io_iopoll_complete() checking ->result */
2766 WRITE_ONCE(req->iopoll_completed, 1);
2770 * After the iocb has been issued, it's safe to be found on the poll list.
2771 * Adding the kiocb to the list AFTER submission ensures that we don't
2772 * find it from a io_do_iopoll() thread before the issuer is done
2773 * accessing the kiocb cookie.
2775 static void io_iopoll_req_issued(struct io_kiocb *req)
2777 struct io_ring_ctx *ctx = req->ctx;
2778 const bool in_async = io_wq_current_is_worker();
2780 /* workqueue context doesn't hold uring_lock, grab it now */
2781 if (unlikely(in_async))
2782 mutex_lock(&ctx->uring_lock);
2785 * Track whether we have multiple files in our lists. This will impact
2786 * how we do polling eventually, not spinning if we're on potentially
2787 * different devices.
2789 if (list_empty(&ctx->iopoll_list)) {
2790 ctx->poll_multi_queue = false;
2791 } else if (!ctx->poll_multi_queue) {
2792 struct io_kiocb *list_req;
2793 unsigned int queue_num0, queue_num1;
2795 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2798 if (list_req->file != req->file) {
2799 ctx->poll_multi_queue = true;
2801 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2802 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2803 if (queue_num0 != queue_num1)
2804 ctx->poll_multi_queue = true;
2809 * For fast devices, IO may have already completed. If it has, add
2810 * it to the front so we find it first.
2812 if (READ_ONCE(req->iopoll_completed))
2813 list_add(&req->inflight_entry, &ctx->iopoll_list);
2815 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2817 if (unlikely(in_async)) {
2819 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2820 * in sq thread task context or in io worker task context. If
2821 * current task context is sq thread, we don't need to check
2822 * whether should wake up sq thread.
2824 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2825 wq_has_sleeper(&ctx->sq_data->wait))
2826 wake_up(&ctx->sq_data->wait);
2828 mutex_unlock(&ctx->uring_lock);
2832 static bool io_bdev_nowait(struct block_device *bdev)
2834 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2838 * If we tracked the file through the SCM inflight mechanism, we could support
2839 * any file. For now, just ensure that anything potentially problematic is done
2842 static bool __io_file_supports_nowait(struct file *file, int rw)
2844 umode_t mode = file_inode(file)->i_mode;
2846 if (S_ISBLK(mode)) {
2847 if (IS_ENABLED(CONFIG_BLOCK) &&
2848 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2854 if (S_ISREG(mode)) {
2855 if (IS_ENABLED(CONFIG_BLOCK) &&
2856 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2857 file->f_op != &io_uring_fops)
2862 /* any ->read/write should understand O_NONBLOCK */
2863 if (file->f_flags & O_NONBLOCK)
2866 if (!(file->f_mode & FMODE_NOWAIT))
2870 return file->f_op->read_iter != NULL;
2872 return file->f_op->write_iter != NULL;
2875 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2877 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2879 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2882 return __io_file_supports_nowait(req->file, rw);
2885 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2888 struct io_ring_ctx *ctx = req->ctx;
2889 struct kiocb *kiocb = &req->rw.kiocb;
2890 struct file *file = req->file;
2894 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2895 req->flags |= REQ_F_ISREG;
2897 kiocb->ki_pos = READ_ONCE(sqe->off);
2898 if (kiocb->ki_pos == -1) {
2899 if (!(file->f_mode & FMODE_STREAM)) {
2900 req->flags |= REQ_F_CUR_POS;
2901 kiocb->ki_pos = file->f_pos;
2906 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2907 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2908 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2913 * If the file is marked O_NONBLOCK, still allow retry for it if it
2914 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2915 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2917 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2918 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2919 req->flags |= REQ_F_NOWAIT;
2921 ioprio = READ_ONCE(sqe->ioprio);
2923 ret = ioprio_check_cap(ioprio);
2927 kiocb->ki_ioprio = ioprio;
2929 kiocb->ki_ioprio = get_current_ioprio();
2931 if (ctx->flags & IORING_SETUP_IOPOLL) {
2932 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2933 !kiocb->ki_filp->f_op->iopoll)
2936 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2937 kiocb->ki_complete = io_complete_rw_iopoll;
2938 req->iopoll_completed = 0;
2940 if (kiocb->ki_flags & IOCB_HIPRI)
2942 kiocb->ki_complete = io_complete_rw;
2945 if (req->opcode == IORING_OP_READ_FIXED ||
2946 req->opcode == IORING_OP_WRITE_FIXED) {
2948 io_req_set_rsrc_node(req);
2951 req->rw.addr = READ_ONCE(sqe->addr);
2952 req->rw.len = READ_ONCE(sqe->len);
2953 req->buf_index = READ_ONCE(sqe->buf_index);
2957 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2963 case -ERESTARTNOINTR:
2964 case -ERESTARTNOHAND:
2965 case -ERESTART_RESTARTBLOCK:
2967 * We can't just restart the syscall, since previously
2968 * submitted sqes may already be in progress. Just fail this
2974 kiocb->ki_complete(kiocb, ret, 0);
2978 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2979 unsigned int issue_flags)
2981 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2982 struct io_async_rw *io = req->async_data;
2984 /* add previously done IO, if any */
2985 if (io && io->bytes_done > 0) {
2987 ret = io->bytes_done;
2989 ret += io->bytes_done;
2992 if (req->flags & REQ_F_CUR_POS)
2993 req->file->f_pos = kiocb->ki_pos;
2994 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2995 __io_complete_rw(req, ret, 0, issue_flags);
2997 io_rw_done(kiocb, ret);
2999 if (req->flags & REQ_F_REISSUE) {
3000 req->flags &= ~REQ_F_REISSUE;
3001 if (io_resubmit_prep(req)) {
3002 io_req_task_queue_reissue(req);
3004 unsigned int cflags = io_put_rw_kbuf(req);
3005 struct io_ring_ctx *ctx = req->ctx;
3008 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
3009 mutex_lock(&ctx->uring_lock);
3010 __io_req_complete(req, issue_flags, ret, cflags);
3011 mutex_unlock(&ctx->uring_lock);
3013 __io_req_complete(req, issue_flags, ret, cflags);
3019 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3020 struct io_mapped_ubuf *imu)
3022 size_t len = req->rw.len;
3023 u64 buf_end, buf_addr = req->rw.addr;
3026 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
3028 /* not inside the mapped region */
3029 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
3033 * May not be a start of buffer, set size appropriately
3034 * and advance us to the beginning.
3036 offset = buf_addr - imu->ubuf;
3037 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3041 * Don't use iov_iter_advance() here, as it's really slow for
3042 * using the latter parts of a big fixed buffer - it iterates
3043 * over each segment manually. We can cheat a bit here, because
3046 * 1) it's a BVEC iter, we set it up
3047 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3048 * first and last bvec
3050 * So just find our index, and adjust the iterator afterwards.
3051 * If the offset is within the first bvec (or the whole first
3052 * bvec, just use iov_iter_advance(). This makes it easier
3053 * since we can just skip the first segment, which may not
3054 * be PAGE_SIZE aligned.
3056 const struct bio_vec *bvec = imu->bvec;
3058 if (offset <= bvec->bv_len) {
3059 iov_iter_advance(iter, offset);
3061 unsigned long seg_skip;
3063 /* skip first vec */
3064 offset -= bvec->bv_len;
3065 seg_skip = 1 + (offset >> PAGE_SHIFT);
3067 iter->bvec = bvec + seg_skip;
3068 iter->nr_segs -= seg_skip;
3069 iter->count -= bvec->bv_len + offset;
3070 iter->iov_offset = offset & ~PAGE_MASK;
3077 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3079 struct io_ring_ctx *ctx = req->ctx;
3080 struct io_mapped_ubuf *imu = req->imu;
3081 u16 index, buf_index = req->buf_index;
3084 if (unlikely(buf_index >= ctx->nr_user_bufs))
3086 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3087 imu = READ_ONCE(ctx->user_bufs[index]);
3090 return __io_import_fixed(req, rw, iter, imu);
3093 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3096 mutex_unlock(&ctx->uring_lock);
3099 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3102 * "Normal" inline submissions always hold the uring_lock, since we
3103 * grab it from the system call. Same is true for the SQPOLL offload.
3104 * The only exception is when we've detached the request and issue it
3105 * from an async worker thread, grab the lock for that case.
3108 mutex_lock(&ctx->uring_lock);
3111 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3112 int bgid, struct io_buffer *kbuf,
3115 struct io_buffer *head;
3117 if (req->flags & REQ_F_BUFFER_SELECTED)
3120 io_ring_submit_lock(req->ctx, needs_lock);
3122 lockdep_assert_held(&req->ctx->uring_lock);
3124 head = xa_load(&req->ctx->io_buffers, bgid);
3126 if (!list_empty(&head->list)) {
3127 kbuf = list_last_entry(&head->list, struct io_buffer,
3129 list_del(&kbuf->list);
3132 xa_erase(&req->ctx->io_buffers, bgid);
3134 if (*len > kbuf->len)
3137 kbuf = ERR_PTR(-ENOBUFS);
3140 io_ring_submit_unlock(req->ctx, needs_lock);
3145 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3148 struct io_buffer *kbuf;
3151 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3152 bgid = req->buf_index;
3153 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3156 req->rw.addr = (u64) (unsigned long) kbuf;
3157 req->flags |= REQ_F_BUFFER_SELECTED;
3158 return u64_to_user_ptr(kbuf->addr);
3161 #ifdef CONFIG_COMPAT
3162 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3165 struct compat_iovec __user *uiov;
3166 compat_ssize_t clen;
3170 uiov = u64_to_user_ptr(req->rw.addr);
3171 if (!access_ok(uiov, sizeof(*uiov)))
3173 if (__get_user(clen, &uiov->iov_len))
3179 buf = io_rw_buffer_select(req, &len, needs_lock);
3181 return PTR_ERR(buf);
3182 iov[0].iov_base = buf;
3183 iov[0].iov_len = (compat_size_t) len;
3188 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3191 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3195 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3198 len = iov[0].iov_len;
3201 buf = io_rw_buffer_select(req, &len, needs_lock);
3203 return PTR_ERR(buf);
3204 iov[0].iov_base = buf;
3205 iov[0].iov_len = len;
3209 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3212 if (req->flags & REQ_F_BUFFER_SELECTED) {
3213 struct io_buffer *kbuf;
3215 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3216 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3217 iov[0].iov_len = kbuf->len;
3220 if (req->rw.len != 1)
3223 #ifdef CONFIG_COMPAT
3224 if (req->ctx->compat)
3225 return io_compat_import(req, iov, needs_lock);
3228 return __io_iov_buffer_select(req, iov, needs_lock);
3231 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3232 struct iov_iter *iter, bool needs_lock)
3234 void __user *buf = u64_to_user_ptr(req->rw.addr);
3235 size_t sqe_len = req->rw.len;
3236 u8 opcode = req->opcode;
3239 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3241 return io_import_fixed(req, rw, iter);
3244 /* buffer index only valid with fixed read/write, or buffer select */
3245 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3248 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3249 if (req->flags & REQ_F_BUFFER_SELECT) {
3250 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3252 return PTR_ERR(buf);
3253 req->rw.len = sqe_len;
3256 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3261 if (req->flags & REQ_F_BUFFER_SELECT) {
3262 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3264 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3269 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3273 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3275 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3279 * For files that don't have ->read_iter() and ->write_iter(), handle them
3280 * by looping over ->read() or ->write() manually.
3282 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3284 struct kiocb *kiocb = &req->rw.kiocb;
3285 struct file *file = req->file;
3289 * Don't support polled IO through this interface, and we can't
3290 * support non-blocking either. For the latter, this just causes
3291 * the kiocb to be handled from an async context.
3293 if (kiocb->ki_flags & IOCB_HIPRI)
3295 if (kiocb->ki_flags & IOCB_NOWAIT)
3298 while (iov_iter_count(iter)) {
3302 if (!iov_iter_is_bvec(iter)) {
3303 iovec = iov_iter_iovec(iter);
3305 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3306 iovec.iov_len = req->rw.len;
3310 nr = file->f_op->read(file, iovec.iov_base,
3311 iovec.iov_len, io_kiocb_ppos(kiocb));
3313 nr = file->f_op->write(file, iovec.iov_base,
3314 iovec.iov_len, io_kiocb_ppos(kiocb));
3323 if (!iov_iter_is_bvec(iter)) {
3324 iov_iter_advance(iter, nr);
3331 if (nr != iovec.iov_len)
3338 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3339 const struct iovec *fast_iov, struct iov_iter *iter)
3341 struct io_async_rw *rw = req->async_data;
3343 memcpy(&rw->iter, iter, sizeof(*iter));
3344 rw->free_iovec = iovec;
3346 /* can only be fixed buffers, no need to do anything */
3347 if (iov_iter_is_bvec(iter))
3350 unsigned iov_off = 0;
3352 rw->iter.iov = rw->fast_iov;
3353 if (iter->iov != fast_iov) {
3354 iov_off = iter->iov - fast_iov;
3355 rw->iter.iov += iov_off;
3357 if (rw->fast_iov != fast_iov)
3358 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3359 sizeof(struct iovec) * iter->nr_segs);
3361 req->flags |= REQ_F_NEED_CLEANUP;
3365 static inline int io_alloc_async_data(struct io_kiocb *req)
3367 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3368 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3369 return req->async_data == NULL;
3372 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3373 const struct iovec *fast_iov,
3374 struct iov_iter *iter, bool force)
3376 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3378 if (!req->async_data) {
3379 struct io_async_rw *iorw;
3381 if (io_alloc_async_data(req)) {
3386 io_req_map_rw(req, iovec, fast_iov, iter);
3387 iorw = req->async_data;
3388 /* we've copied and mapped the iter, ensure state is saved */
3389 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3394 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3396 struct io_async_rw *iorw = req->async_data;
3397 struct iovec *iov = iorw->fast_iov;
3400 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3401 if (unlikely(ret < 0))
3404 iorw->bytes_done = 0;
3405 iorw->free_iovec = iov;
3407 req->flags |= REQ_F_NEED_CLEANUP;
3408 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3412 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3414 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3416 return io_prep_rw(req, sqe, READ);
3420 * This is our waitqueue callback handler, registered through lock_page_async()
3421 * when we initially tried to do the IO with the iocb armed our waitqueue.
3422 * This gets called when the page is unlocked, and we generally expect that to
3423 * happen when the page IO is completed and the page is now uptodate. This will
3424 * queue a task_work based retry of the operation, attempting to copy the data
3425 * again. If the latter fails because the page was NOT uptodate, then we will
3426 * do a thread based blocking retry of the operation. That's the unexpected
3429 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3430 int sync, void *arg)
3432 struct wait_page_queue *wpq;
3433 struct io_kiocb *req = wait->private;
3434 struct wait_page_key *key = arg;
3436 wpq = container_of(wait, struct wait_page_queue, wait);
3438 if (!wake_page_match(wpq, key))
3441 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3442 list_del_init(&wait->entry);
3443 io_req_task_queue(req);
3448 * This controls whether a given IO request should be armed for async page
3449 * based retry. If we return false here, the request is handed to the async
3450 * worker threads for retry. If we're doing buffered reads on a regular file,
3451 * we prepare a private wait_page_queue entry and retry the operation. This
3452 * will either succeed because the page is now uptodate and unlocked, or it
3453 * will register a callback when the page is unlocked at IO completion. Through
3454 * that callback, io_uring uses task_work to setup a retry of the operation.
3455 * That retry will attempt the buffered read again. The retry will generally
3456 * succeed, or in rare cases where it fails, we then fall back to using the
3457 * async worker threads for a blocking retry.
3459 static bool io_rw_should_retry(struct io_kiocb *req)
3461 struct io_async_rw *rw = req->async_data;
3462 struct wait_page_queue *wait = &rw->wpq;
3463 struct kiocb *kiocb = &req->rw.kiocb;
3465 /* never retry for NOWAIT, we just complete with -EAGAIN */
3466 if (req->flags & REQ_F_NOWAIT)
3469 /* Only for buffered IO */
3470 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3474 * just use poll if we can, and don't attempt if the fs doesn't
3475 * support callback based unlocks
3477 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3480 wait->wait.func = io_async_buf_func;
3481 wait->wait.private = req;
3482 wait->wait.flags = 0;
3483 INIT_LIST_HEAD(&wait->wait.entry);
3484 kiocb->ki_flags |= IOCB_WAITQ;
3485 kiocb->ki_flags &= ~IOCB_NOWAIT;
3486 kiocb->ki_waitq = wait;
3490 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3492 if (req->file->f_op->read_iter)
3493 return call_read_iter(req->file, &req->rw.kiocb, iter);
3494 else if (req->file->f_op->read)
3495 return loop_rw_iter(READ, req, iter);
3500 static bool need_read_all(struct io_kiocb *req)
3502 return req->flags & REQ_F_ISREG ||
3503 S_ISBLK(file_inode(req->file)->i_mode);
3506 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3508 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3509 struct kiocb *kiocb = &req->rw.kiocb;
3510 struct iov_iter __iter, *iter = &__iter;
3511 struct io_async_rw *rw = req->async_data;
3512 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3513 struct iov_iter_state __state, *state;
3518 state = &rw->iter_state;
3520 * We come here from an earlier attempt, restore our state to
3521 * match in case it doesn't. It's cheap enough that we don't
3522 * need to make this conditional.
3524 iov_iter_restore(iter, state);
3527 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3531 iov_iter_save_state(iter, state);
3533 req->result = iov_iter_count(iter);
3535 /* Ensure we clear previously set non-block flag */
3536 if (!force_nonblock)
3537 kiocb->ki_flags &= ~IOCB_NOWAIT;
3539 kiocb->ki_flags |= IOCB_NOWAIT;
3541 /* If the file doesn't support async, just async punt */
3542 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3543 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3544 return ret ?: -EAGAIN;
3547 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3548 if (unlikely(ret)) {
3553 ret = io_iter_do_read(req, iter);
3555 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3556 req->flags &= ~REQ_F_REISSUE;
3557 /* IOPOLL retry should happen for io-wq threads */
3558 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3560 /* no retry on NONBLOCK nor RWF_NOWAIT */
3561 if (req->flags & REQ_F_NOWAIT)
3564 } else if (ret == -EIOCBQUEUED) {
3566 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3567 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3568 /* read all, failed, already did sync or don't want to retry */
3573 * Don't depend on the iter state matching what was consumed, or being
3574 * untouched in case of error. Restore it and we'll advance it
3575 * manually if we need to.
3577 iov_iter_restore(iter, state);
3579 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3584 rw = req->async_data;
3586 * Now use our persistent iterator and state, if we aren't already.
3587 * We've restored and mapped the iter to match.
3589 if (iter != &rw->iter) {
3591 state = &rw->iter_state;
3596 * We end up here because of a partial read, either from
3597 * above or inside this loop. Advance the iter by the bytes
3598 * that were consumed.
3600 iov_iter_advance(iter, ret);
3601 if (!iov_iter_count(iter))
3603 rw->bytes_done += ret;
3604 iov_iter_save_state(iter, state);
3606 /* if we can retry, do so with the callbacks armed */
3607 if (!io_rw_should_retry(req)) {
3608 kiocb->ki_flags &= ~IOCB_WAITQ;
3613 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3614 * we get -EIOCBQUEUED, then we'll get a notification when the
3615 * desired page gets unlocked. We can also get a partial read
3616 * here, and if we do, then just retry at the new offset.
3618 ret = io_iter_do_read(req, iter);
3619 if (ret == -EIOCBQUEUED)
3621 /* we got some bytes, but not all. retry. */
3622 kiocb->ki_flags &= ~IOCB_WAITQ;
3623 iov_iter_restore(iter, state);
3626 kiocb_done(kiocb, ret, issue_flags);
3628 /* it's faster to check here then delegate to kfree */
3634 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3636 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3638 return io_prep_rw(req, sqe, WRITE);
3641 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3643 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3644 struct kiocb *kiocb = &req->rw.kiocb;
3645 struct iov_iter __iter, *iter = &__iter;
3646 struct io_async_rw *rw = req->async_data;
3647 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3648 struct iov_iter_state __state, *state;
3653 state = &rw->iter_state;
3654 iov_iter_restore(iter, state);
3657 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3661 iov_iter_save_state(iter, state);
3663 req->result = iov_iter_count(iter);
3665 /* Ensure we clear previously set non-block flag */
3666 if (!force_nonblock)
3667 kiocb->ki_flags &= ~IOCB_NOWAIT;
3669 kiocb->ki_flags |= IOCB_NOWAIT;
3671 /* If the file doesn't support async, just async punt */
3672 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3675 /* file path doesn't support NOWAIT for non-direct_IO */
3676 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3677 (req->flags & REQ_F_ISREG))
3680 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3685 * Open-code file_start_write here to grab freeze protection,
3686 * which will be released by another thread in
3687 * io_complete_rw(). Fool lockdep by telling it the lock got
3688 * released so that it doesn't complain about the held lock when
3689 * we return to userspace.
3691 if (req->flags & REQ_F_ISREG) {
3692 sb_start_write(file_inode(req->file)->i_sb);
3693 __sb_writers_release(file_inode(req->file)->i_sb,
3696 kiocb->ki_flags |= IOCB_WRITE;
3698 if (req->file->f_op->write_iter)
3699 ret2 = call_write_iter(req->file, kiocb, iter);
3700 else if (req->file->f_op->write)
3701 ret2 = loop_rw_iter(WRITE, req, iter);
3705 if (req->flags & REQ_F_REISSUE) {
3706 req->flags &= ~REQ_F_REISSUE;
3711 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3712 * retry them without IOCB_NOWAIT.
3714 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3716 /* no retry on NONBLOCK nor RWF_NOWAIT */
3717 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3719 if (!force_nonblock || ret2 != -EAGAIN) {
3720 /* IOPOLL retry should happen for io-wq threads */
3721 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3724 kiocb_done(kiocb, ret2, issue_flags);
3727 iov_iter_restore(iter, state);
3728 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3729 return ret ?: -EAGAIN;
3732 /* it's reportedly faster than delegating the null check to kfree() */
3738 static int io_renameat_prep(struct io_kiocb *req,
3739 const struct io_uring_sqe *sqe)
3741 struct io_rename *ren = &req->rename;
3742 const char __user *oldf, *newf;
3744 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3746 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3748 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3751 ren->old_dfd = READ_ONCE(sqe->fd);
3752 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3753 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3754 ren->new_dfd = READ_ONCE(sqe->len);
3755 ren->flags = READ_ONCE(sqe->rename_flags);
3757 ren->oldpath = getname(oldf);
3758 if (IS_ERR(ren->oldpath))
3759 return PTR_ERR(ren->oldpath);
3761 ren->newpath = getname(newf);
3762 if (IS_ERR(ren->newpath)) {
3763 putname(ren->oldpath);
3764 return PTR_ERR(ren->newpath);
3767 req->flags |= REQ_F_NEED_CLEANUP;
3771 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3773 struct io_rename *ren = &req->rename;
3776 if (issue_flags & IO_URING_F_NONBLOCK)
3779 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3780 ren->newpath, ren->flags);
3782 req->flags &= ~REQ_F_NEED_CLEANUP;
3785 io_req_complete(req, ret);
3789 static int io_unlinkat_prep(struct io_kiocb *req,
3790 const struct io_uring_sqe *sqe)
3792 struct io_unlink *un = &req->unlink;
3793 const char __user *fname;
3795 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3797 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3800 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3803 un->dfd = READ_ONCE(sqe->fd);
3805 un->flags = READ_ONCE(sqe->unlink_flags);
3806 if (un->flags & ~AT_REMOVEDIR)
3809 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3810 un->filename = getname(fname);
3811 if (IS_ERR(un->filename))
3812 return PTR_ERR(un->filename);
3814 req->flags |= REQ_F_NEED_CLEANUP;
3818 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3820 struct io_unlink *un = &req->unlink;
3823 if (issue_flags & IO_URING_F_NONBLOCK)
3826 if (un->flags & AT_REMOVEDIR)
3827 ret = do_rmdir(un->dfd, un->filename);
3829 ret = do_unlinkat(un->dfd, un->filename);
3831 req->flags &= ~REQ_F_NEED_CLEANUP;
3834 io_req_complete(req, ret);
3838 static int io_mkdirat_prep(struct io_kiocb *req,
3839 const struct io_uring_sqe *sqe)
3841 struct io_mkdir *mkd = &req->mkdir;
3842 const char __user *fname;
3844 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3846 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3849 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3852 mkd->dfd = READ_ONCE(sqe->fd);
3853 mkd->mode = READ_ONCE(sqe->len);
3855 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3856 mkd->filename = getname(fname);
3857 if (IS_ERR(mkd->filename))
3858 return PTR_ERR(mkd->filename);
3860 req->flags |= REQ_F_NEED_CLEANUP;
3864 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3866 struct io_mkdir *mkd = &req->mkdir;
3869 if (issue_flags & IO_URING_F_NONBLOCK)
3872 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3874 req->flags &= ~REQ_F_NEED_CLEANUP;
3877 io_req_complete(req, ret);
3881 static int io_symlinkat_prep(struct io_kiocb *req,
3882 const struct io_uring_sqe *sqe)
3884 struct io_symlink *sl = &req->symlink;
3885 const char __user *oldpath, *newpath;
3887 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3889 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3892 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3895 sl->new_dfd = READ_ONCE(sqe->fd);
3896 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3897 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3899 sl->oldpath = getname(oldpath);
3900 if (IS_ERR(sl->oldpath))
3901 return PTR_ERR(sl->oldpath);
3903 sl->newpath = getname(newpath);
3904 if (IS_ERR(sl->newpath)) {
3905 putname(sl->oldpath);
3906 return PTR_ERR(sl->newpath);
3909 req->flags |= REQ_F_NEED_CLEANUP;
3913 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3915 struct io_symlink *sl = &req->symlink;
3918 if (issue_flags & IO_URING_F_NONBLOCK)
3921 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3923 req->flags &= ~REQ_F_NEED_CLEANUP;
3926 io_req_complete(req, ret);
3930 static int io_linkat_prep(struct io_kiocb *req,
3931 const struct io_uring_sqe *sqe)
3933 struct io_hardlink *lnk = &req->hardlink;
3934 const char __user *oldf, *newf;
3936 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3938 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3940 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3943 lnk->old_dfd = READ_ONCE(sqe->fd);
3944 lnk->new_dfd = READ_ONCE(sqe->len);
3945 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3946 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3947 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3949 lnk->oldpath = getname(oldf);
3950 if (IS_ERR(lnk->oldpath))
3951 return PTR_ERR(lnk->oldpath);
3953 lnk->newpath = getname(newf);
3954 if (IS_ERR(lnk->newpath)) {
3955 putname(lnk->oldpath);
3956 return PTR_ERR(lnk->newpath);
3959 req->flags |= REQ_F_NEED_CLEANUP;
3963 static int io_linkat(struct io_kiocb *req, int issue_flags)
3965 struct io_hardlink *lnk = &req->hardlink;
3968 if (issue_flags & IO_URING_F_NONBLOCK)
3971 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3972 lnk->newpath, lnk->flags);
3974 req->flags &= ~REQ_F_NEED_CLEANUP;
3977 io_req_complete(req, ret);
3981 static int io_shutdown_prep(struct io_kiocb *req,
3982 const struct io_uring_sqe *sqe)
3984 #if defined(CONFIG_NET)
3985 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3987 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3988 sqe->buf_index || sqe->splice_fd_in))
3991 req->shutdown.how = READ_ONCE(sqe->len);
3998 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
4000 #if defined(CONFIG_NET)
4001 struct socket *sock;
4004 if (issue_flags & IO_URING_F_NONBLOCK)
4007 sock = sock_from_file(req->file);
4008 if (unlikely(!sock))
4011 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4014 io_req_complete(req, ret);
4021 static int __io_splice_prep(struct io_kiocb *req,
4022 const struct io_uring_sqe *sqe)
4024 struct io_splice *sp = &req->splice;
4025 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4027 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4031 sp->len = READ_ONCE(sqe->len);
4032 sp->flags = READ_ONCE(sqe->splice_flags);
4034 if (unlikely(sp->flags & ~valid_flags))
4037 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
4038 (sp->flags & SPLICE_F_FD_IN_FIXED));
4041 req->flags |= REQ_F_NEED_CLEANUP;
4045 static int io_tee_prep(struct io_kiocb *req,
4046 const struct io_uring_sqe *sqe)
4048 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4050 return __io_splice_prep(req, sqe);
4053 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4055 struct io_splice *sp = &req->splice;
4056 struct file *in = sp->file_in;
4057 struct file *out = sp->file_out;
4058 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4061 if (issue_flags & IO_URING_F_NONBLOCK)
4064 ret = do_tee(in, out, sp->len, flags);
4066 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4068 req->flags &= ~REQ_F_NEED_CLEANUP;
4072 io_req_complete(req, ret);
4076 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4078 struct io_splice *sp = &req->splice;
4080 sp->off_in = READ_ONCE(sqe->splice_off_in);
4081 sp->off_out = READ_ONCE(sqe->off);
4082 return __io_splice_prep(req, sqe);
4085 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4087 struct io_splice *sp = &req->splice;
4088 struct file *in = sp->file_in;
4089 struct file *out = sp->file_out;
4090 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4091 loff_t *poff_in, *poff_out;
4094 if (issue_flags & IO_URING_F_NONBLOCK)
4097 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4098 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4101 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4103 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4105 req->flags &= ~REQ_F_NEED_CLEANUP;
4109 io_req_complete(req, ret);
4114 * IORING_OP_NOP just posts a completion event, nothing else.
4116 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4118 struct io_ring_ctx *ctx = req->ctx;
4120 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4123 __io_req_complete(req, issue_flags, 0, 0);
4127 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4129 struct io_ring_ctx *ctx = req->ctx;
4134 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4136 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4140 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4141 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4144 req->sync.off = READ_ONCE(sqe->off);
4145 req->sync.len = READ_ONCE(sqe->len);
4149 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4151 loff_t end = req->sync.off + req->sync.len;
4154 /* fsync always requires a blocking context */
4155 if (issue_flags & IO_URING_F_NONBLOCK)
4158 ret = vfs_fsync_range(req->file, req->sync.off,
4159 end > 0 ? end : LLONG_MAX,
4160 req->sync.flags & IORING_FSYNC_DATASYNC);
4163 io_req_complete(req, ret);
4167 static int io_fallocate_prep(struct io_kiocb *req,
4168 const struct io_uring_sqe *sqe)
4170 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4173 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4176 req->sync.off = READ_ONCE(sqe->off);
4177 req->sync.len = READ_ONCE(sqe->addr);
4178 req->sync.mode = READ_ONCE(sqe->len);
4182 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4186 /* fallocate always requiring blocking context */
4187 if (issue_flags & IO_URING_F_NONBLOCK)
4189 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4193 io_req_complete(req, ret);
4197 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4199 const char __user *fname;
4202 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4204 if (unlikely(sqe->ioprio || sqe->buf_index))
4206 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4209 /* open.how should be already initialised */
4210 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4211 req->open.how.flags |= O_LARGEFILE;
4213 req->open.dfd = READ_ONCE(sqe->fd);
4214 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4215 req->open.filename = getname(fname);
4216 if (IS_ERR(req->open.filename)) {
4217 ret = PTR_ERR(req->open.filename);
4218 req->open.filename = NULL;
4222 req->open.file_slot = READ_ONCE(sqe->file_index);
4223 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4226 req->open.nofile = rlimit(RLIMIT_NOFILE);
4227 req->flags |= REQ_F_NEED_CLEANUP;
4231 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4233 u64 mode = READ_ONCE(sqe->len);
4234 u64 flags = READ_ONCE(sqe->open_flags);
4236 req->open.how = build_open_how(flags, mode);
4237 return __io_openat_prep(req, sqe);
4240 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4242 struct open_how __user *how;
4246 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4247 len = READ_ONCE(sqe->len);
4248 if (len < OPEN_HOW_SIZE_VER0)
4251 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4256 return __io_openat_prep(req, sqe);
4259 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4261 struct open_flags op;
4263 bool resolve_nonblock, nonblock_set;
4264 bool fixed = !!req->open.file_slot;
4267 ret = build_open_flags(&req->open.how, &op);
4270 nonblock_set = op.open_flag & O_NONBLOCK;
4271 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4272 if (issue_flags & IO_URING_F_NONBLOCK) {
4274 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4275 * it'll always -EAGAIN
4277 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4279 op.lookup_flags |= LOOKUP_CACHED;
4280 op.open_flag |= O_NONBLOCK;
4284 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4289 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4292 * We could hang on to this 'fd' on retrying, but seems like
4293 * marginal gain for something that is now known to be a slower
4294 * path. So just put it, and we'll get a new one when we retry.
4299 ret = PTR_ERR(file);
4300 /* only retry if RESOLVE_CACHED wasn't already set by application */
4301 if (ret == -EAGAIN &&
4302 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4307 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4308 file->f_flags &= ~O_NONBLOCK;
4309 fsnotify_open(file);
4312 fd_install(ret, file);
4314 ret = io_install_fixed_file(req, file, issue_flags,
4315 req->open.file_slot - 1);
4317 putname(req->open.filename);
4318 req->flags &= ~REQ_F_NEED_CLEANUP;
4321 __io_req_complete(req, issue_flags, ret, 0);
4325 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4327 return io_openat2(req, issue_flags);
4330 static int io_remove_buffers_prep(struct io_kiocb *req,
4331 const struct io_uring_sqe *sqe)
4333 struct io_provide_buf *p = &req->pbuf;
4336 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4340 tmp = READ_ONCE(sqe->fd);
4341 if (!tmp || tmp > USHRT_MAX)
4344 memset(p, 0, sizeof(*p));
4346 p->bgid = READ_ONCE(sqe->buf_group);
4350 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4351 int bgid, unsigned nbufs)
4355 /* shouldn't happen */
4359 /* the head kbuf is the list itself */
4360 while (!list_empty(&buf->list)) {
4361 struct io_buffer *nxt;
4363 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4364 list_del(&nxt->list);
4372 xa_erase(&ctx->io_buffers, bgid);
4377 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4379 struct io_provide_buf *p = &req->pbuf;
4380 struct io_ring_ctx *ctx = req->ctx;
4381 struct io_buffer *head;
4383 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4385 io_ring_submit_lock(ctx, !force_nonblock);
4387 lockdep_assert_held(&ctx->uring_lock);
4390 head = xa_load(&ctx->io_buffers, p->bgid);
4392 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4396 /* complete before unlock, IOPOLL may need the lock */
4397 __io_req_complete(req, issue_flags, ret, 0);
4398 io_ring_submit_unlock(ctx, !force_nonblock);
4402 static int io_provide_buffers_prep(struct io_kiocb *req,
4403 const struct io_uring_sqe *sqe)
4405 unsigned long size, tmp_check;
4406 struct io_provide_buf *p = &req->pbuf;
4409 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4412 tmp = READ_ONCE(sqe->fd);
4413 if (!tmp || tmp > USHRT_MAX)
4416 p->addr = READ_ONCE(sqe->addr);
4417 p->len = READ_ONCE(sqe->len);
4419 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4422 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4425 size = (unsigned long)p->len * p->nbufs;
4426 if (!access_ok(u64_to_user_ptr(p->addr), size))
4429 p->bgid = READ_ONCE(sqe->buf_group);
4430 tmp = READ_ONCE(sqe->off);
4431 if (tmp > USHRT_MAX)
4437 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4439 struct io_buffer *buf;
4440 u64 addr = pbuf->addr;
4441 int i, bid = pbuf->bid;
4443 for (i = 0; i < pbuf->nbufs; i++) {
4444 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4449 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4454 INIT_LIST_HEAD(&buf->list);
4457 list_add_tail(&buf->list, &(*head)->list);
4462 return i ? i : -ENOMEM;
4465 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4467 struct io_provide_buf *p = &req->pbuf;
4468 struct io_ring_ctx *ctx = req->ctx;
4469 struct io_buffer *head, *list;
4471 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4473 io_ring_submit_lock(ctx, !force_nonblock);
4475 lockdep_assert_held(&ctx->uring_lock);
4477 list = head = xa_load(&ctx->io_buffers, p->bgid);
4479 ret = io_add_buffers(p, &head);
4480 if (ret >= 0 && !list) {
4481 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4483 __io_remove_buffers(ctx, head, p->bgid, -1U);
4487 /* complete before unlock, IOPOLL may need the lock */
4488 __io_req_complete(req, issue_flags, ret, 0);
4489 io_ring_submit_unlock(ctx, !force_nonblock);
4493 static int io_epoll_ctl_prep(struct io_kiocb *req,
4494 const struct io_uring_sqe *sqe)
4496 #if defined(CONFIG_EPOLL)
4497 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4499 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4502 req->epoll.epfd = READ_ONCE(sqe->fd);
4503 req->epoll.op = READ_ONCE(sqe->len);
4504 req->epoll.fd = READ_ONCE(sqe->off);
4506 if (ep_op_has_event(req->epoll.op)) {
4507 struct epoll_event __user *ev;
4509 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4510 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4520 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4522 #if defined(CONFIG_EPOLL)
4523 struct io_epoll *ie = &req->epoll;
4525 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4527 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4528 if (force_nonblock && ret == -EAGAIN)
4533 __io_req_complete(req, issue_flags, ret, 0);
4540 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4542 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4543 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4548 req->madvise.addr = READ_ONCE(sqe->addr);
4549 req->madvise.len = READ_ONCE(sqe->len);
4550 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4557 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4559 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4560 struct io_madvise *ma = &req->madvise;
4563 if (issue_flags & IO_URING_F_NONBLOCK)
4566 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4569 io_req_complete(req, ret);
4576 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4578 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4583 req->fadvise.offset = READ_ONCE(sqe->off);
4584 req->fadvise.len = READ_ONCE(sqe->len);
4585 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4589 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4591 struct io_fadvise *fa = &req->fadvise;
4594 if (issue_flags & IO_URING_F_NONBLOCK) {
4595 switch (fa->advice) {
4596 case POSIX_FADV_NORMAL:
4597 case POSIX_FADV_RANDOM:
4598 case POSIX_FADV_SEQUENTIAL:
4605 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4608 __io_req_complete(req, issue_flags, ret, 0);
4612 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4614 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4616 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4618 if (req->flags & REQ_F_FIXED_FILE)
4621 req->statx.dfd = READ_ONCE(sqe->fd);
4622 req->statx.mask = READ_ONCE(sqe->len);
4623 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4624 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4625 req->statx.flags = READ_ONCE(sqe->statx_flags);
4630 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4632 struct io_statx *ctx = &req->statx;
4635 if (issue_flags & IO_URING_F_NONBLOCK)
4638 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4643 io_req_complete(req, ret);
4647 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4649 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4651 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4652 sqe->rw_flags || sqe->buf_index)
4654 if (req->flags & REQ_F_FIXED_FILE)
4657 req->close.fd = READ_ONCE(sqe->fd);
4658 req->close.file_slot = READ_ONCE(sqe->file_index);
4659 if (req->close.file_slot && req->close.fd)
4665 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4667 struct files_struct *files = current->files;
4668 struct io_close *close = &req->close;
4669 struct fdtable *fdt;
4670 struct file *file = NULL;
4673 if (req->close.file_slot) {
4674 ret = io_close_fixed(req, issue_flags);
4678 spin_lock(&files->file_lock);
4679 fdt = files_fdtable(files);
4680 if (close->fd >= fdt->max_fds) {
4681 spin_unlock(&files->file_lock);
4684 file = fdt->fd[close->fd];
4685 if (!file || file->f_op == &io_uring_fops) {
4686 spin_unlock(&files->file_lock);
4691 /* if the file has a flush method, be safe and punt to async */
4692 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4693 spin_unlock(&files->file_lock);
4697 ret = __close_fd_get_file(close->fd, &file);
4698 spin_unlock(&files->file_lock);
4705 /* No ->flush() or already async, safely close from here */
4706 ret = filp_close(file, current->files);
4712 __io_req_complete(req, issue_flags, ret, 0);
4716 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718 struct io_ring_ctx *ctx = req->ctx;
4720 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4722 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4726 req->sync.off = READ_ONCE(sqe->off);
4727 req->sync.len = READ_ONCE(sqe->len);
4728 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4732 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4736 /* sync_file_range always requires a blocking context */
4737 if (issue_flags & IO_URING_F_NONBLOCK)
4740 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4744 io_req_complete(req, ret);
4748 #if defined(CONFIG_NET)
4749 static int io_setup_async_msg(struct io_kiocb *req,
4750 struct io_async_msghdr *kmsg)
4752 struct io_async_msghdr *async_msg = req->async_data;
4756 if (io_alloc_async_data(req)) {
4757 kfree(kmsg->free_iov);
4760 async_msg = req->async_data;
4761 req->flags |= REQ_F_NEED_CLEANUP;
4762 memcpy(async_msg, kmsg, sizeof(*kmsg));
4763 async_msg->msg.msg_name = &async_msg->addr;
4764 /* if were using fast_iov, set it to the new one */
4765 if (!async_msg->free_iov)
4766 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4771 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4772 struct io_async_msghdr *iomsg)
4774 iomsg->msg.msg_name = &iomsg->addr;
4775 iomsg->free_iov = iomsg->fast_iov;
4776 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4777 req->sr_msg.msg_flags, &iomsg->free_iov);
4780 static int io_sendmsg_prep_async(struct io_kiocb *req)
4784 ret = io_sendmsg_copy_hdr(req, req->async_data);
4786 req->flags |= REQ_F_NEED_CLEANUP;
4790 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4792 struct io_sr_msg *sr = &req->sr_msg;
4794 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4797 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4798 sr->len = READ_ONCE(sqe->len);
4799 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4800 if (sr->msg_flags & MSG_DONTWAIT)
4801 req->flags |= REQ_F_NOWAIT;
4803 #ifdef CONFIG_COMPAT
4804 if (req->ctx->compat)
4805 sr->msg_flags |= MSG_CMSG_COMPAT;
4810 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4812 struct io_async_msghdr iomsg, *kmsg;
4813 struct socket *sock;
4818 sock = sock_from_file(req->file);
4819 if (unlikely(!sock))
4822 kmsg = req->async_data;
4824 ret = io_sendmsg_copy_hdr(req, &iomsg);
4830 flags = req->sr_msg.msg_flags;
4831 if (issue_flags & IO_URING_F_NONBLOCK)
4832 flags |= MSG_DONTWAIT;
4833 if (flags & MSG_WAITALL)
4834 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4836 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4837 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4838 return io_setup_async_msg(req, kmsg);
4839 if (ret == -ERESTARTSYS)
4842 /* fast path, check for non-NULL to avoid function call */
4844 kfree(kmsg->free_iov);
4845 req->flags &= ~REQ_F_NEED_CLEANUP;
4848 __io_req_complete(req, issue_flags, ret, 0);
4852 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4854 struct io_sr_msg *sr = &req->sr_msg;
4857 struct socket *sock;
4862 sock = sock_from_file(req->file);
4863 if (unlikely(!sock))
4866 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4870 msg.msg_name = NULL;
4871 msg.msg_control = NULL;
4872 msg.msg_controllen = 0;
4873 msg.msg_namelen = 0;
4875 flags = req->sr_msg.msg_flags;
4876 if (issue_flags & IO_URING_F_NONBLOCK)
4877 flags |= MSG_DONTWAIT;
4878 if (flags & MSG_WAITALL)
4879 min_ret = iov_iter_count(&msg.msg_iter);
4881 msg.msg_flags = flags;
4882 ret = sock_sendmsg(sock, &msg);
4883 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4885 if (ret == -ERESTARTSYS)
4890 __io_req_complete(req, issue_flags, ret, 0);
4894 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4895 struct io_async_msghdr *iomsg)
4897 struct io_sr_msg *sr = &req->sr_msg;
4898 struct iovec __user *uiov;
4902 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4903 &iomsg->uaddr, &uiov, &iov_len);
4907 if (req->flags & REQ_F_BUFFER_SELECT) {
4910 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4912 sr->len = iomsg->fast_iov[0].iov_len;
4913 iomsg->free_iov = NULL;
4915 iomsg->free_iov = iomsg->fast_iov;
4916 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4917 &iomsg->free_iov, &iomsg->msg.msg_iter,
4926 #ifdef CONFIG_COMPAT
4927 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4928 struct io_async_msghdr *iomsg)
4930 struct io_sr_msg *sr = &req->sr_msg;
4931 struct compat_iovec __user *uiov;
4936 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4941 uiov = compat_ptr(ptr);
4942 if (req->flags & REQ_F_BUFFER_SELECT) {
4943 compat_ssize_t clen;
4947 if (!access_ok(uiov, sizeof(*uiov)))
4949 if (__get_user(clen, &uiov->iov_len))
4954 iomsg->free_iov = NULL;
4956 iomsg->free_iov = iomsg->fast_iov;
4957 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4958 UIO_FASTIOV, &iomsg->free_iov,
4959 &iomsg->msg.msg_iter, true);
4968 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4969 struct io_async_msghdr *iomsg)
4971 iomsg->msg.msg_name = &iomsg->addr;
4973 #ifdef CONFIG_COMPAT
4974 if (req->ctx->compat)
4975 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4978 return __io_recvmsg_copy_hdr(req, iomsg);
4981 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4984 struct io_sr_msg *sr = &req->sr_msg;
4985 struct io_buffer *kbuf;
4987 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4992 req->flags |= REQ_F_BUFFER_SELECTED;
4996 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4998 return io_put_kbuf(req, req->sr_msg.kbuf);
5001 static int io_recvmsg_prep_async(struct io_kiocb *req)
5005 ret = io_recvmsg_copy_hdr(req, req->async_data);
5007 req->flags |= REQ_F_NEED_CLEANUP;
5011 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5013 struct io_sr_msg *sr = &req->sr_msg;
5015 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5018 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5019 sr->len = READ_ONCE(sqe->len);
5020 sr->bgid = READ_ONCE(sqe->buf_group);
5021 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5022 if (sr->msg_flags & MSG_DONTWAIT)
5023 req->flags |= REQ_F_NOWAIT;
5025 #ifdef CONFIG_COMPAT
5026 if (req->ctx->compat)
5027 sr->msg_flags |= MSG_CMSG_COMPAT;
5032 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5034 struct io_async_msghdr iomsg, *kmsg;
5035 struct socket *sock;
5036 struct io_buffer *kbuf;
5039 int ret, cflags = 0;
5040 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5042 sock = sock_from_file(req->file);
5043 if (unlikely(!sock))
5046 kmsg = req->async_data;
5048 ret = io_recvmsg_copy_hdr(req, &iomsg);
5054 if (req->flags & REQ_F_BUFFER_SELECT) {
5055 kbuf = io_recv_buffer_select(req, !force_nonblock);
5057 return PTR_ERR(kbuf);
5058 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5059 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5060 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5061 1, req->sr_msg.len);
5064 flags = req->sr_msg.msg_flags;
5066 flags |= MSG_DONTWAIT;
5067 if (flags & MSG_WAITALL)
5068 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5070 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5071 kmsg->uaddr, flags);
5072 if (force_nonblock && ret == -EAGAIN)
5073 return io_setup_async_msg(req, kmsg);
5074 if (ret == -ERESTARTSYS)
5077 if (req->flags & REQ_F_BUFFER_SELECTED)
5078 cflags = io_put_recv_kbuf(req);
5079 /* fast path, check for non-NULL to avoid function call */
5081 kfree(kmsg->free_iov);
5082 req->flags &= ~REQ_F_NEED_CLEANUP;
5083 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5085 __io_req_complete(req, issue_flags, ret, cflags);
5089 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5091 struct io_buffer *kbuf;
5092 struct io_sr_msg *sr = &req->sr_msg;
5094 void __user *buf = sr->buf;
5095 struct socket *sock;
5099 int ret, cflags = 0;
5100 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5102 sock = sock_from_file(req->file);
5103 if (unlikely(!sock))
5106 if (req->flags & REQ_F_BUFFER_SELECT) {
5107 kbuf = io_recv_buffer_select(req, !force_nonblock);
5109 return PTR_ERR(kbuf);
5110 buf = u64_to_user_ptr(kbuf->addr);
5113 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5117 msg.msg_name = NULL;
5118 msg.msg_control = NULL;
5119 msg.msg_controllen = 0;
5120 msg.msg_namelen = 0;
5121 msg.msg_iocb = NULL;
5124 flags = req->sr_msg.msg_flags;
5126 flags |= MSG_DONTWAIT;
5127 if (flags & MSG_WAITALL)
5128 min_ret = iov_iter_count(&msg.msg_iter);
5130 ret = sock_recvmsg(sock, &msg, flags);
5131 if (force_nonblock && ret == -EAGAIN)
5133 if (ret == -ERESTARTSYS)
5136 if (req->flags & REQ_F_BUFFER_SELECTED)
5137 cflags = io_put_recv_kbuf(req);
5138 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5140 __io_req_complete(req, issue_flags, ret, cflags);
5144 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5146 struct io_accept *accept = &req->accept;
5148 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5150 if (sqe->ioprio || sqe->len || sqe->buf_index)
5153 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5154 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5155 accept->flags = READ_ONCE(sqe->accept_flags);
5156 accept->nofile = rlimit(RLIMIT_NOFILE);
5158 accept->file_slot = READ_ONCE(sqe->file_index);
5159 if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))
5161 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5163 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5164 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5168 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5170 struct io_accept *accept = &req->accept;
5171 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5172 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5173 bool fixed = !!accept->file_slot;
5177 if (req->file->f_flags & O_NONBLOCK)
5178 req->flags |= REQ_F_NOWAIT;
5181 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5182 if (unlikely(fd < 0))
5185 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5190 ret = PTR_ERR(file);
5191 if (ret == -EAGAIN && force_nonblock)
5193 if (ret == -ERESTARTSYS)
5196 } else if (!fixed) {
5197 fd_install(fd, file);
5200 ret = io_install_fixed_file(req, file, issue_flags,
5201 accept->file_slot - 1);
5203 __io_req_complete(req, issue_flags, ret, 0);
5207 static int io_connect_prep_async(struct io_kiocb *req)
5209 struct io_async_connect *io = req->async_data;
5210 struct io_connect *conn = &req->connect;
5212 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5215 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5217 struct io_connect *conn = &req->connect;
5219 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5221 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5225 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5226 conn->addr_len = READ_ONCE(sqe->addr2);
5230 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5232 struct io_async_connect __io, *io;
5233 unsigned file_flags;
5235 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5237 if (req->async_data) {
5238 io = req->async_data;
5240 ret = move_addr_to_kernel(req->connect.addr,
5241 req->connect.addr_len,
5248 file_flags = force_nonblock ? O_NONBLOCK : 0;
5250 ret = __sys_connect_file(req->file, &io->address,
5251 req->connect.addr_len, file_flags);
5252 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5253 if (req->async_data)
5255 if (io_alloc_async_data(req)) {
5259 memcpy(req->async_data, &__io, sizeof(__io));
5262 if (ret == -ERESTARTSYS)
5267 __io_req_complete(req, issue_flags, ret, 0);
5270 #else /* !CONFIG_NET */
5271 #define IO_NETOP_FN(op) \
5272 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5274 return -EOPNOTSUPP; \
5277 #define IO_NETOP_PREP(op) \
5279 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5281 return -EOPNOTSUPP; \
5284 #define IO_NETOP_PREP_ASYNC(op) \
5286 static int io_##op##_prep_async(struct io_kiocb *req) \
5288 return -EOPNOTSUPP; \
5291 IO_NETOP_PREP_ASYNC(sendmsg);
5292 IO_NETOP_PREP_ASYNC(recvmsg);
5293 IO_NETOP_PREP_ASYNC(connect);
5294 IO_NETOP_PREP(accept);
5297 #endif /* CONFIG_NET */
5299 struct io_poll_table {
5300 struct poll_table_struct pt;
5301 struct io_kiocb *req;
5306 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5307 __poll_t mask, io_req_tw_func_t func)
5309 /* for instances that support it check for an event match first: */
5310 if (mask && !(mask & poll->events))
5313 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5315 list_del_init(&poll->wait.entry);
5318 req->io_task_work.func = func;
5321 * If this fails, then the task is exiting. When a task exits, the
5322 * work gets canceled, so just cancel this request as well instead
5323 * of executing it. We can't safely execute it anyway, as we may not
5324 * have the needed state needed for it anyway.
5326 io_req_task_work_add(req);
5330 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5331 __acquires(&req->ctx->completion_lock)
5333 struct io_ring_ctx *ctx = req->ctx;
5335 /* req->task == current here, checking PF_EXITING is safe */
5336 if (unlikely(req->task->flags & PF_EXITING))
5337 WRITE_ONCE(poll->canceled, true);
5339 if (!req->result && !READ_ONCE(poll->canceled)) {
5340 struct poll_table_struct pt = { ._key = poll->events };
5342 req->result = vfs_poll(req->file, &pt) & poll->events;
5345 spin_lock(&ctx->completion_lock);
5346 if (!req->result && !READ_ONCE(poll->canceled)) {
5347 add_wait_queue(poll->head, &poll->wait);
5354 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5356 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5357 if (req->opcode == IORING_OP_POLL_ADD)
5358 return req->async_data;
5359 return req->apoll->double_poll;
5362 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5364 if (req->opcode == IORING_OP_POLL_ADD)
5366 return &req->apoll->poll;
5369 static void io_poll_remove_double(struct io_kiocb *req)
5370 __must_hold(&req->ctx->completion_lock)
5372 struct io_poll_iocb *poll = io_poll_get_double(req);
5374 lockdep_assert_held(&req->ctx->completion_lock);
5376 if (poll && poll->head) {
5377 struct wait_queue_head *head = poll->head;
5379 spin_lock_irq(&head->lock);
5380 list_del_init(&poll->wait.entry);
5381 if (poll->wait.private)
5384 spin_unlock_irq(&head->lock);
5388 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5389 __must_hold(&req->ctx->completion_lock)
5391 struct io_ring_ctx *ctx = req->ctx;
5392 unsigned flags = IORING_CQE_F_MORE;
5395 if (READ_ONCE(req->poll.canceled)) {
5397 req->poll.events |= EPOLLONESHOT;
5399 error = mangle_poll(mask);
5401 if (req->poll.events & EPOLLONESHOT)
5403 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5404 req->poll.events |= EPOLLONESHOT;
5407 if (flags & IORING_CQE_F_MORE)
5410 return !(flags & IORING_CQE_F_MORE);
5413 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5414 __must_hold(&req->ctx->completion_lock)
5418 done = __io_poll_complete(req, mask);
5419 io_commit_cqring(req->ctx);
5423 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5425 struct io_ring_ctx *ctx = req->ctx;
5426 struct io_kiocb *nxt;
5428 if (io_poll_rewait(req, &req->poll)) {
5429 spin_unlock(&ctx->completion_lock);
5433 if (req->poll.done) {
5434 spin_unlock(&ctx->completion_lock);
5437 done = __io_poll_complete(req, req->result);
5439 io_poll_remove_double(req);
5440 hash_del(&req->hash_node);
5441 req->poll.done = true;
5444 add_wait_queue(req->poll.head, &req->poll.wait);
5446 io_commit_cqring(ctx);
5447 spin_unlock(&ctx->completion_lock);
5448 io_cqring_ev_posted(ctx);
5451 nxt = io_put_req_find_next(req);
5453 io_req_task_submit(nxt, locked);
5458 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5459 int sync, void *key)
5461 struct io_kiocb *req = wait->private;
5462 struct io_poll_iocb *poll = io_poll_get_single(req);
5463 __poll_t mask = key_to_poll(key);
5464 unsigned long flags;
5466 /* for instances that support it check for an event match first: */
5467 if (mask && !(mask & poll->events))
5469 if (!(poll->events & EPOLLONESHOT))
5470 return poll->wait.func(&poll->wait, mode, sync, key);
5472 list_del_init(&wait->entry);
5477 spin_lock_irqsave(&poll->head->lock, flags);
5478 done = list_empty(&poll->wait.entry);
5480 list_del_init(&poll->wait.entry);
5481 /* make sure double remove sees this as being gone */
5482 wait->private = NULL;
5483 spin_unlock_irqrestore(&poll->head->lock, flags);
5485 /* use wait func handler, so it matches the rq type */
5486 poll->wait.func(&poll->wait, mode, sync, key);
5493 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5494 wait_queue_func_t wake_func)
5498 poll->canceled = false;
5499 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5500 /* mask in events that we always want/need */
5501 poll->events = events | IO_POLL_UNMASK;
5502 INIT_LIST_HEAD(&poll->wait.entry);
5503 init_waitqueue_func_entry(&poll->wait, wake_func);
5506 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5507 struct wait_queue_head *head,
5508 struct io_poll_iocb **poll_ptr)
5510 struct io_kiocb *req = pt->req;
5513 * The file being polled uses multiple waitqueues for poll handling
5514 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5517 if (unlikely(pt->nr_entries)) {
5518 struct io_poll_iocb *poll_one = poll;
5520 /* double add on the same waitqueue head, ignore */
5521 if (poll_one->head == head)
5523 /* already have a 2nd entry, fail a third attempt */
5525 if ((*poll_ptr)->head == head)
5527 pt->error = -EINVAL;
5531 * Can't handle multishot for double wait for now, turn it
5532 * into one-shot mode.
5534 if (!(poll_one->events & EPOLLONESHOT))
5535 poll_one->events |= EPOLLONESHOT;
5536 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5538 pt->error = -ENOMEM;
5541 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5543 poll->wait.private = req;
5550 if (poll->events & EPOLLEXCLUSIVE)
5551 add_wait_queue_exclusive(head, &poll->wait);
5553 add_wait_queue(head, &poll->wait);
5556 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5557 struct poll_table_struct *p)
5559 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5560 struct async_poll *apoll = pt->req->apoll;
5562 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5565 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5567 struct async_poll *apoll = req->apoll;
5568 struct io_ring_ctx *ctx = req->ctx;
5570 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5572 if (io_poll_rewait(req, &apoll->poll)) {
5573 spin_unlock(&ctx->completion_lock);
5577 hash_del(&req->hash_node);
5578 io_poll_remove_double(req);
5579 apoll->poll.done = true;
5580 spin_unlock(&ctx->completion_lock);
5582 if (!READ_ONCE(apoll->poll.canceled))
5583 io_req_task_submit(req, locked);
5585 io_req_complete_failed(req, -ECANCELED);
5588 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5591 struct io_kiocb *req = wait->private;
5592 struct io_poll_iocb *poll = &req->apoll->poll;
5594 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5597 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5600 static void io_poll_req_insert(struct io_kiocb *req)
5602 struct io_ring_ctx *ctx = req->ctx;
5603 struct hlist_head *list;
5605 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5606 hlist_add_head(&req->hash_node, list);
5609 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5610 struct io_poll_iocb *poll,
5611 struct io_poll_table *ipt, __poll_t mask,
5612 wait_queue_func_t wake_func)
5613 __acquires(&ctx->completion_lock)
5615 struct io_ring_ctx *ctx = req->ctx;
5616 bool cancel = false;
5618 INIT_HLIST_NODE(&req->hash_node);
5619 io_init_poll_iocb(poll, mask, wake_func);
5620 poll->file = req->file;
5621 poll->wait.private = req;
5623 ipt->pt._key = mask;
5626 ipt->nr_entries = 0;
5628 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5629 if (unlikely(!ipt->nr_entries) && !ipt->error)
5630 ipt->error = -EINVAL;
5632 spin_lock(&ctx->completion_lock);
5633 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5634 io_poll_remove_double(req);
5635 if (likely(poll->head)) {
5636 spin_lock_irq(&poll->head->lock);
5637 if (unlikely(list_empty(&poll->wait.entry))) {
5643 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5644 list_del_init(&poll->wait.entry);
5646 WRITE_ONCE(poll->canceled, true);
5647 else if (!poll->done) /* actually waiting for an event */
5648 io_poll_req_insert(req);
5649 spin_unlock_irq(&poll->head->lock);
5661 static int io_arm_poll_handler(struct io_kiocb *req)
5663 const struct io_op_def *def = &io_op_defs[req->opcode];
5664 struct io_ring_ctx *ctx = req->ctx;
5665 struct async_poll *apoll;
5666 struct io_poll_table ipt;
5667 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5670 if (!req->file || !file_can_poll(req->file))
5671 return IO_APOLL_ABORTED;
5672 if (req->flags & REQ_F_POLLED)
5673 return IO_APOLL_ABORTED;
5674 if (!def->pollin && !def->pollout)
5675 return IO_APOLL_ABORTED;
5679 mask |= POLLIN | POLLRDNORM;
5681 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5682 if ((req->opcode == IORING_OP_RECVMSG) &&
5683 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5687 mask |= POLLOUT | POLLWRNORM;
5690 /* if we can't nonblock try, then no point in arming a poll handler */
5691 if (!io_file_supports_nowait(req, rw))
5692 return IO_APOLL_ABORTED;
5694 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5695 if (unlikely(!apoll))
5696 return IO_APOLL_ABORTED;
5697 apoll->double_poll = NULL;
5699 req->flags |= REQ_F_POLLED;
5700 ipt.pt._qproc = io_async_queue_proc;
5701 io_req_set_refcount(req);
5703 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5705 spin_unlock(&ctx->completion_lock);
5706 if (ret || ipt.error)
5707 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5709 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5710 mask, apoll->poll.events);
5714 static bool __io_poll_remove_one(struct io_kiocb *req,
5715 struct io_poll_iocb *poll, bool do_cancel)
5716 __must_hold(&req->ctx->completion_lock)
5718 bool do_complete = false;
5722 spin_lock_irq(&poll->head->lock);
5724 WRITE_ONCE(poll->canceled, true);
5725 if (!list_empty(&poll->wait.entry)) {
5726 list_del_init(&poll->wait.entry);
5729 spin_unlock_irq(&poll->head->lock);
5730 hash_del(&req->hash_node);
5734 static bool io_poll_remove_one(struct io_kiocb *req)
5735 __must_hold(&req->ctx->completion_lock)
5739 io_poll_remove_double(req);
5740 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5743 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5744 io_commit_cqring(req->ctx);
5746 io_put_req_deferred(req);
5752 * Returns true if we found and killed one or more poll requests
5754 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5757 struct hlist_node *tmp;
5758 struct io_kiocb *req;
5761 spin_lock(&ctx->completion_lock);
5762 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5763 struct hlist_head *list;
5765 list = &ctx->cancel_hash[i];
5766 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5767 if (io_match_task_safe(req, tsk, cancel_all))
5768 posted += io_poll_remove_one(req);
5771 spin_unlock(&ctx->completion_lock);
5774 io_cqring_ev_posted(ctx);
5779 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5781 __must_hold(&ctx->completion_lock)
5783 struct hlist_head *list;
5784 struct io_kiocb *req;
5786 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5787 hlist_for_each_entry(req, list, hash_node) {
5788 if (sqe_addr != req->user_data)
5790 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5797 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5799 __must_hold(&ctx->completion_lock)
5801 struct io_kiocb *req;
5803 req = io_poll_find(ctx, sqe_addr, poll_only);
5806 if (io_poll_remove_one(req))
5812 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5817 events = READ_ONCE(sqe->poll32_events);
5819 events = swahw32(events);
5821 if (!(flags & IORING_POLL_ADD_MULTI))
5822 events |= EPOLLONESHOT;
5823 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5826 static int io_poll_update_prep(struct io_kiocb *req,
5827 const struct io_uring_sqe *sqe)
5829 struct io_poll_update *upd = &req->poll_update;
5832 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5834 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5836 flags = READ_ONCE(sqe->len);
5837 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5838 IORING_POLL_ADD_MULTI))
5840 /* meaningless without update */
5841 if (flags == IORING_POLL_ADD_MULTI)
5844 upd->old_user_data = READ_ONCE(sqe->addr);
5845 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5846 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5848 upd->new_user_data = READ_ONCE(sqe->off);
5849 if (!upd->update_user_data && upd->new_user_data)
5851 if (upd->update_events)
5852 upd->events = io_poll_parse_events(sqe, flags);
5853 else if (sqe->poll32_events)
5859 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5862 struct io_kiocb *req = wait->private;
5863 struct io_poll_iocb *poll = &req->poll;
5865 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5868 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5869 struct poll_table_struct *p)
5871 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5873 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5876 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5878 struct io_poll_iocb *poll = &req->poll;
5881 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5883 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5885 flags = READ_ONCE(sqe->len);
5886 if (flags & ~IORING_POLL_ADD_MULTI)
5889 io_req_set_refcount(req);
5890 poll->events = io_poll_parse_events(sqe, flags);
5894 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5896 struct io_poll_iocb *poll = &req->poll;
5897 struct io_ring_ctx *ctx = req->ctx;
5898 struct io_poll_table ipt;
5902 ipt.pt._qproc = io_poll_queue_proc;
5904 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5907 if (mask) { /* no async, we'd stolen it */
5909 done = io_poll_complete(req, mask);
5911 spin_unlock(&ctx->completion_lock);
5914 io_cqring_ev_posted(ctx);
5921 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5923 struct io_ring_ctx *ctx = req->ctx;
5924 struct io_kiocb *preq;
5928 spin_lock(&ctx->completion_lock);
5929 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5935 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5937 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5942 * Don't allow racy completion with singleshot, as we cannot safely
5943 * update those. For multishot, if we're racing with completion, just
5944 * let completion re-add it.
5946 io_poll_remove_double(preq);
5947 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5948 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5952 /* we now have a detached poll request. reissue. */
5956 spin_unlock(&ctx->completion_lock);
5958 io_req_complete(req, ret);
5961 /* only mask one event flags, keep behavior flags */
5962 if (req->poll_update.update_events) {
5963 preq->poll.events &= ~0xffff;
5964 preq->poll.events |= req->poll_update.events & 0xffff;
5965 preq->poll.events |= IO_POLL_UNMASK;
5967 if (req->poll_update.update_user_data)
5968 preq->user_data = req->poll_update.new_user_data;
5969 spin_unlock(&ctx->completion_lock);
5971 /* complete update request, we're done with it */
5972 io_req_complete(req, ret);
5975 ret = io_poll_add(preq, issue_flags);
5978 io_req_complete(preq, ret);
5984 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5987 io_req_complete_post(req, -ETIME, 0);
5990 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5992 struct io_timeout_data *data = container_of(timer,
5993 struct io_timeout_data, timer);
5994 struct io_kiocb *req = data->req;
5995 struct io_ring_ctx *ctx = req->ctx;
5996 unsigned long flags;
5998 spin_lock_irqsave(&ctx->timeout_lock, flags);
5999 list_del_init(&req->timeout.list);
6000 atomic_set(&req->ctx->cq_timeouts,
6001 atomic_read(&req->ctx->cq_timeouts) + 1);
6002 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6004 req->io_task_work.func = io_req_task_timeout;
6005 io_req_task_work_add(req);
6006 return HRTIMER_NORESTART;
6009 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
6011 __must_hold(&ctx->timeout_lock)
6013 struct io_timeout_data *io;
6014 struct io_kiocb *req;
6017 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
6018 found = user_data == req->user_data;
6023 return ERR_PTR(-ENOENT);
6025 io = req->async_data;
6026 if (hrtimer_try_to_cancel(&io->timer) == -1)
6027 return ERR_PTR(-EALREADY);
6028 list_del_init(&req->timeout.list);
6032 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
6033 __must_hold(&ctx->completion_lock)
6034 __must_hold(&ctx->timeout_lock)
6036 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6039 return PTR_ERR(req);
6042 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
6043 io_put_req_deferred(req);
6047 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6049 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6050 case IORING_TIMEOUT_BOOTTIME:
6051 return CLOCK_BOOTTIME;
6052 case IORING_TIMEOUT_REALTIME:
6053 return CLOCK_REALTIME;
6055 /* can't happen, vetted at prep time */
6059 return CLOCK_MONOTONIC;
6063 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6064 struct timespec64 *ts, enum hrtimer_mode mode)
6065 __must_hold(&ctx->timeout_lock)
6067 struct io_timeout_data *io;
6068 struct io_kiocb *req;
6071 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6072 found = user_data == req->user_data;
6079 io = req->async_data;
6080 if (hrtimer_try_to_cancel(&io->timer) == -1)
6082 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6083 io->timer.function = io_link_timeout_fn;
6084 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6088 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6089 struct timespec64 *ts, enum hrtimer_mode mode)
6090 __must_hold(&ctx->timeout_lock)
6092 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6093 struct io_timeout_data *data;
6096 return PTR_ERR(req);
6098 req->timeout.off = 0; /* noseq */
6099 data = req->async_data;
6100 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6101 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6102 data->timer.function = io_timeout_fn;
6103 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6107 static int io_timeout_remove_prep(struct io_kiocb *req,
6108 const struct io_uring_sqe *sqe)
6110 struct io_timeout_rem *tr = &req->timeout_rem;
6112 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6114 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6116 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6119 tr->ltimeout = false;
6120 tr->addr = READ_ONCE(sqe->addr);
6121 tr->flags = READ_ONCE(sqe->timeout_flags);
6122 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6123 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6125 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6126 tr->ltimeout = true;
6127 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6129 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6131 } else if (tr->flags) {
6132 /* timeout removal doesn't support flags */
6139 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6141 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6146 * Remove or update an existing timeout command
6148 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6150 struct io_timeout_rem *tr = &req->timeout_rem;
6151 struct io_ring_ctx *ctx = req->ctx;
6154 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6155 spin_lock(&ctx->completion_lock);
6156 spin_lock_irq(&ctx->timeout_lock);
6157 ret = io_timeout_cancel(ctx, tr->addr);
6158 spin_unlock_irq(&ctx->timeout_lock);
6159 spin_unlock(&ctx->completion_lock);
6161 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6163 spin_lock_irq(&ctx->timeout_lock);
6165 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6167 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6168 spin_unlock_irq(&ctx->timeout_lock);
6173 io_req_complete_post(req, ret, 0);
6177 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6178 bool is_timeout_link)
6180 struct io_timeout_data *data;
6182 u32 off = READ_ONCE(sqe->off);
6184 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6186 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6189 if (off && is_timeout_link)
6191 flags = READ_ONCE(sqe->timeout_flags);
6192 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6194 /* more than one clock specified is invalid, obviously */
6195 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6198 INIT_LIST_HEAD(&req->timeout.list);
6199 req->timeout.off = off;
6200 if (unlikely(off && !req->ctx->off_timeout_used))
6201 req->ctx->off_timeout_used = true;
6203 if (!req->async_data && io_alloc_async_data(req))
6206 data = req->async_data;
6208 data->flags = flags;
6210 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6213 data->mode = io_translate_timeout_mode(flags);
6214 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6216 if (is_timeout_link) {
6217 struct io_submit_link *link = &req->ctx->submit_state.link;
6221 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6223 req->timeout.head = link->last;
6224 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6229 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6231 struct io_ring_ctx *ctx = req->ctx;
6232 struct io_timeout_data *data = req->async_data;
6233 struct list_head *entry;
6234 u32 tail, off = req->timeout.off;
6236 spin_lock_irq(&ctx->timeout_lock);
6239 * sqe->off holds how many events that need to occur for this
6240 * timeout event to be satisfied. If it isn't set, then this is
6241 * a pure timeout request, sequence isn't used.
6243 if (io_is_timeout_noseq(req)) {
6244 entry = ctx->timeout_list.prev;
6248 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6249 req->timeout.target_seq = tail + off;
6251 /* Update the last seq here in case io_flush_timeouts() hasn't.
6252 * This is safe because ->completion_lock is held, and submissions
6253 * and completions are never mixed in the same ->completion_lock section.
6255 ctx->cq_last_tm_flush = tail;
6258 * Insertion sort, ensuring the first entry in the list is always
6259 * the one we need first.
6261 list_for_each_prev(entry, &ctx->timeout_list) {
6262 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6265 if (io_is_timeout_noseq(nxt))
6267 /* nxt.seq is behind @tail, otherwise would've been completed */
6268 if (off >= nxt->timeout.target_seq - tail)
6272 list_add(&req->timeout.list, entry);
6273 data->timer.function = io_timeout_fn;
6274 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6275 spin_unlock_irq(&ctx->timeout_lock);
6279 struct io_cancel_data {
6280 struct io_ring_ctx *ctx;
6284 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6286 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6287 struct io_cancel_data *cd = data;
6289 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6292 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6293 struct io_ring_ctx *ctx)
6295 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6296 enum io_wq_cancel cancel_ret;
6299 if (!tctx || !tctx->io_wq)
6302 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6303 switch (cancel_ret) {
6304 case IO_WQ_CANCEL_OK:
6307 case IO_WQ_CANCEL_RUNNING:
6310 case IO_WQ_CANCEL_NOTFOUND:
6318 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6320 struct io_ring_ctx *ctx = req->ctx;
6323 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6325 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6329 spin_lock(&ctx->completion_lock);
6330 spin_lock_irq(&ctx->timeout_lock);
6331 ret = io_timeout_cancel(ctx, sqe_addr);
6332 spin_unlock_irq(&ctx->timeout_lock);
6335 ret = io_poll_cancel(ctx, sqe_addr, false);
6337 spin_unlock(&ctx->completion_lock);
6341 static int io_async_cancel_prep(struct io_kiocb *req,
6342 const struct io_uring_sqe *sqe)
6344 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6346 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6348 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6352 req->cancel.addr = READ_ONCE(sqe->addr);
6356 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6358 struct io_ring_ctx *ctx = req->ctx;
6359 u64 sqe_addr = req->cancel.addr;
6360 struct io_tctx_node *node;
6363 ret = io_try_cancel_userdata(req, sqe_addr);
6367 /* slow path, try all io-wq's */
6368 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6370 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6371 struct io_uring_task *tctx = node->task->io_uring;
6373 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6377 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6381 io_req_complete_post(req, ret, 0);
6385 static int io_rsrc_update_prep(struct io_kiocb *req,
6386 const struct io_uring_sqe *sqe)
6388 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6390 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6393 req->rsrc_update.offset = READ_ONCE(sqe->off);
6394 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6395 if (!req->rsrc_update.nr_args)
6397 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6401 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6403 struct io_ring_ctx *ctx = req->ctx;
6404 struct io_uring_rsrc_update2 up;
6407 up.offset = req->rsrc_update.offset;
6408 up.data = req->rsrc_update.arg;
6413 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6414 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6415 &up, req->rsrc_update.nr_args);
6416 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6420 __io_req_complete(req, issue_flags, ret, 0);
6424 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6426 switch (req->opcode) {
6429 case IORING_OP_READV:
6430 case IORING_OP_READ_FIXED:
6431 case IORING_OP_READ:
6432 return io_read_prep(req, sqe);
6433 case IORING_OP_WRITEV:
6434 case IORING_OP_WRITE_FIXED:
6435 case IORING_OP_WRITE:
6436 return io_write_prep(req, sqe);
6437 case IORING_OP_POLL_ADD:
6438 return io_poll_add_prep(req, sqe);
6439 case IORING_OP_POLL_REMOVE:
6440 return io_poll_update_prep(req, sqe);
6441 case IORING_OP_FSYNC:
6442 return io_fsync_prep(req, sqe);
6443 case IORING_OP_SYNC_FILE_RANGE:
6444 return io_sfr_prep(req, sqe);
6445 case IORING_OP_SENDMSG:
6446 case IORING_OP_SEND:
6447 return io_sendmsg_prep(req, sqe);
6448 case IORING_OP_RECVMSG:
6449 case IORING_OP_RECV:
6450 return io_recvmsg_prep(req, sqe);
6451 case IORING_OP_CONNECT:
6452 return io_connect_prep(req, sqe);
6453 case IORING_OP_TIMEOUT:
6454 return io_timeout_prep(req, sqe, false);
6455 case IORING_OP_TIMEOUT_REMOVE:
6456 return io_timeout_remove_prep(req, sqe);
6457 case IORING_OP_ASYNC_CANCEL:
6458 return io_async_cancel_prep(req, sqe);
6459 case IORING_OP_LINK_TIMEOUT:
6460 return io_timeout_prep(req, sqe, true);
6461 case IORING_OP_ACCEPT:
6462 return io_accept_prep(req, sqe);
6463 case IORING_OP_FALLOCATE:
6464 return io_fallocate_prep(req, sqe);
6465 case IORING_OP_OPENAT:
6466 return io_openat_prep(req, sqe);
6467 case IORING_OP_CLOSE:
6468 return io_close_prep(req, sqe);
6469 case IORING_OP_FILES_UPDATE:
6470 return io_rsrc_update_prep(req, sqe);
6471 case IORING_OP_STATX:
6472 return io_statx_prep(req, sqe);
6473 case IORING_OP_FADVISE:
6474 return io_fadvise_prep(req, sqe);
6475 case IORING_OP_MADVISE:
6476 return io_madvise_prep(req, sqe);
6477 case IORING_OP_OPENAT2:
6478 return io_openat2_prep(req, sqe);
6479 case IORING_OP_EPOLL_CTL:
6480 return io_epoll_ctl_prep(req, sqe);
6481 case IORING_OP_SPLICE:
6482 return io_splice_prep(req, sqe);
6483 case IORING_OP_PROVIDE_BUFFERS:
6484 return io_provide_buffers_prep(req, sqe);
6485 case IORING_OP_REMOVE_BUFFERS:
6486 return io_remove_buffers_prep(req, sqe);
6488 return io_tee_prep(req, sqe);
6489 case IORING_OP_SHUTDOWN:
6490 return io_shutdown_prep(req, sqe);
6491 case IORING_OP_RENAMEAT:
6492 return io_renameat_prep(req, sqe);
6493 case IORING_OP_UNLINKAT:
6494 return io_unlinkat_prep(req, sqe);
6495 case IORING_OP_MKDIRAT:
6496 return io_mkdirat_prep(req, sqe);
6497 case IORING_OP_SYMLINKAT:
6498 return io_symlinkat_prep(req, sqe);
6499 case IORING_OP_LINKAT:
6500 return io_linkat_prep(req, sqe);
6503 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6508 static int io_req_prep_async(struct io_kiocb *req)
6510 if (!io_op_defs[req->opcode].needs_async_setup)
6512 if (WARN_ON_ONCE(req->async_data))
6514 if (io_alloc_async_data(req))
6517 switch (req->opcode) {
6518 case IORING_OP_READV:
6519 return io_rw_prep_async(req, READ);
6520 case IORING_OP_WRITEV:
6521 return io_rw_prep_async(req, WRITE);
6522 case IORING_OP_SENDMSG:
6523 return io_sendmsg_prep_async(req);
6524 case IORING_OP_RECVMSG:
6525 return io_recvmsg_prep_async(req);
6526 case IORING_OP_CONNECT:
6527 return io_connect_prep_async(req);
6529 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6534 static u32 io_get_sequence(struct io_kiocb *req)
6536 u32 seq = req->ctx->cached_sq_head;
6538 /* need original cached_sq_head, but it was increased for each req */
6539 io_for_each_link(req, req)
6544 static bool io_drain_req(struct io_kiocb *req)
6546 struct io_kiocb *pos;
6547 struct io_ring_ctx *ctx = req->ctx;
6548 struct io_defer_entry *de;
6552 if (req->flags & REQ_F_FAIL) {
6553 io_req_complete_fail_submit(req);
6558 * If we need to drain a request in the middle of a link, drain the
6559 * head request and the next request/link after the current link.
6560 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6561 * maintained for every request of our link.
6563 if (ctx->drain_next) {
6564 req->flags |= REQ_F_IO_DRAIN;
6565 ctx->drain_next = false;
6567 /* not interested in head, start from the first linked */
6568 io_for_each_link(pos, req->link) {
6569 if (pos->flags & REQ_F_IO_DRAIN) {
6570 ctx->drain_next = true;
6571 req->flags |= REQ_F_IO_DRAIN;
6576 /* Still need defer if there is pending req in defer list. */
6577 spin_lock(&ctx->completion_lock);
6578 if (likely(list_empty_careful(&ctx->defer_list) &&
6579 !(req->flags & REQ_F_IO_DRAIN))) {
6580 spin_unlock(&ctx->completion_lock);
6581 ctx->drain_active = false;
6584 spin_unlock(&ctx->completion_lock);
6586 seq = io_get_sequence(req);
6587 /* Still a chance to pass the sequence check */
6588 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6591 ret = io_req_prep_async(req);
6594 io_prep_async_link(req);
6595 de = kmalloc(sizeof(*de), GFP_KERNEL);
6599 io_req_complete_failed(req, ret);
6603 spin_lock(&ctx->completion_lock);
6604 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6605 spin_unlock(&ctx->completion_lock);
6607 io_queue_async_work(req, NULL);
6611 trace_io_uring_defer(ctx, req, req->user_data);
6614 list_add_tail(&de->list, &ctx->defer_list);
6615 spin_unlock(&ctx->completion_lock);
6619 static void io_clean_op(struct io_kiocb *req)
6621 if (req->flags & REQ_F_BUFFER_SELECTED) {
6622 switch (req->opcode) {
6623 case IORING_OP_READV:
6624 case IORING_OP_READ_FIXED:
6625 case IORING_OP_READ:
6626 kfree((void *)(unsigned long)req->rw.addr);
6628 case IORING_OP_RECVMSG:
6629 case IORING_OP_RECV:
6630 kfree(req->sr_msg.kbuf);
6635 if (req->flags & REQ_F_NEED_CLEANUP) {
6636 switch (req->opcode) {
6637 case IORING_OP_READV:
6638 case IORING_OP_READ_FIXED:
6639 case IORING_OP_READ:
6640 case IORING_OP_WRITEV:
6641 case IORING_OP_WRITE_FIXED:
6642 case IORING_OP_WRITE: {
6643 struct io_async_rw *io = req->async_data;
6645 kfree(io->free_iovec);
6648 case IORING_OP_RECVMSG:
6649 case IORING_OP_SENDMSG: {
6650 struct io_async_msghdr *io = req->async_data;
6652 kfree(io->free_iov);
6655 case IORING_OP_SPLICE:
6657 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6658 io_put_file(req->splice.file_in);
6660 case IORING_OP_OPENAT:
6661 case IORING_OP_OPENAT2:
6662 if (req->open.filename)
6663 putname(req->open.filename);
6665 case IORING_OP_RENAMEAT:
6666 putname(req->rename.oldpath);
6667 putname(req->rename.newpath);
6669 case IORING_OP_UNLINKAT:
6670 putname(req->unlink.filename);
6672 case IORING_OP_MKDIRAT:
6673 putname(req->mkdir.filename);
6675 case IORING_OP_SYMLINKAT:
6676 putname(req->symlink.oldpath);
6677 putname(req->symlink.newpath);
6679 case IORING_OP_LINKAT:
6680 putname(req->hardlink.oldpath);
6681 putname(req->hardlink.newpath);
6685 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6686 kfree(req->apoll->double_poll);
6690 if (req->flags & REQ_F_INFLIGHT) {
6691 struct io_uring_task *tctx = req->task->io_uring;
6693 atomic_dec(&tctx->inflight_tracked);
6695 if (req->flags & REQ_F_CREDS)
6696 put_cred(req->creds);
6698 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6701 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6703 struct io_ring_ctx *ctx = req->ctx;
6704 const struct cred *creds = NULL;
6707 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6708 creds = override_creds(req->creds);
6710 switch (req->opcode) {
6712 ret = io_nop(req, issue_flags);
6714 case IORING_OP_READV:
6715 case IORING_OP_READ_FIXED:
6716 case IORING_OP_READ:
6717 ret = io_read(req, issue_flags);
6719 case IORING_OP_WRITEV:
6720 case IORING_OP_WRITE_FIXED:
6721 case IORING_OP_WRITE:
6722 ret = io_write(req, issue_flags);
6724 case IORING_OP_FSYNC:
6725 ret = io_fsync(req, issue_flags);
6727 case IORING_OP_POLL_ADD:
6728 ret = io_poll_add(req, issue_flags);
6730 case IORING_OP_POLL_REMOVE:
6731 ret = io_poll_update(req, issue_flags);
6733 case IORING_OP_SYNC_FILE_RANGE:
6734 ret = io_sync_file_range(req, issue_flags);
6736 case IORING_OP_SENDMSG:
6737 ret = io_sendmsg(req, issue_flags);
6739 case IORING_OP_SEND:
6740 ret = io_send(req, issue_flags);
6742 case IORING_OP_RECVMSG:
6743 ret = io_recvmsg(req, issue_flags);
6745 case IORING_OP_RECV:
6746 ret = io_recv(req, issue_flags);
6748 case IORING_OP_TIMEOUT:
6749 ret = io_timeout(req, issue_flags);
6751 case IORING_OP_TIMEOUT_REMOVE:
6752 ret = io_timeout_remove(req, issue_flags);
6754 case IORING_OP_ACCEPT:
6755 ret = io_accept(req, issue_flags);
6757 case IORING_OP_CONNECT:
6758 ret = io_connect(req, issue_flags);
6760 case IORING_OP_ASYNC_CANCEL:
6761 ret = io_async_cancel(req, issue_flags);
6763 case IORING_OP_FALLOCATE:
6764 ret = io_fallocate(req, issue_flags);
6766 case IORING_OP_OPENAT:
6767 ret = io_openat(req, issue_flags);
6769 case IORING_OP_CLOSE:
6770 ret = io_close(req, issue_flags);
6772 case IORING_OP_FILES_UPDATE:
6773 ret = io_files_update(req, issue_flags);
6775 case IORING_OP_STATX:
6776 ret = io_statx(req, issue_flags);
6778 case IORING_OP_FADVISE:
6779 ret = io_fadvise(req, issue_flags);
6781 case IORING_OP_MADVISE:
6782 ret = io_madvise(req, issue_flags);
6784 case IORING_OP_OPENAT2:
6785 ret = io_openat2(req, issue_flags);
6787 case IORING_OP_EPOLL_CTL:
6788 ret = io_epoll_ctl(req, issue_flags);
6790 case IORING_OP_SPLICE:
6791 ret = io_splice(req, issue_flags);
6793 case IORING_OP_PROVIDE_BUFFERS:
6794 ret = io_provide_buffers(req, issue_flags);
6796 case IORING_OP_REMOVE_BUFFERS:
6797 ret = io_remove_buffers(req, issue_flags);
6800 ret = io_tee(req, issue_flags);
6802 case IORING_OP_SHUTDOWN:
6803 ret = io_shutdown(req, issue_flags);
6805 case IORING_OP_RENAMEAT:
6806 ret = io_renameat(req, issue_flags);
6808 case IORING_OP_UNLINKAT:
6809 ret = io_unlinkat(req, issue_flags);
6811 case IORING_OP_MKDIRAT:
6812 ret = io_mkdirat(req, issue_flags);
6814 case IORING_OP_SYMLINKAT:
6815 ret = io_symlinkat(req, issue_flags);
6817 case IORING_OP_LINKAT:
6818 ret = io_linkat(req, issue_flags);
6826 revert_creds(creds);
6829 /* If the op doesn't have a file, we're not polling for it */
6830 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6831 io_iopoll_req_issued(req);
6836 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6838 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6840 req = io_put_req_find_next(req);
6841 return req ? &req->work : NULL;
6844 static void io_wq_submit_work(struct io_wq_work *work)
6846 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6847 struct io_kiocb *timeout;
6850 /* one will be dropped by ->io_free_work() after returning to io-wq */
6851 if (!(req->flags & REQ_F_REFCOUNT))
6852 __io_req_set_refcount(req, 2);
6856 timeout = io_prep_linked_timeout(req);
6858 io_queue_linked_timeout(timeout);
6860 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6861 if (work->flags & IO_WQ_WORK_CANCEL)
6866 ret = io_issue_sqe(req, 0);
6868 * We can get EAGAIN for polled IO even though we're
6869 * forcing a sync submission from here, since we can't
6870 * wait for request slots on the block side.
6878 /* avoid locking problems by failing it from a clean context */
6880 io_req_task_queue_fail(req, ret);
6883 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6886 return &table->files[i];
6889 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6892 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6894 return (struct file *) (slot->file_ptr & FFS_MASK);
6897 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6899 unsigned long file_ptr = (unsigned long) file;
6901 if (__io_file_supports_nowait(file, READ))
6902 file_ptr |= FFS_ASYNC_READ;
6903 if (__io_file_supports_nowait(file, WRITE))
6904 file_ptr |= FFS_ASYNC_WRITE;
6905 if (S_ISREG(file_inode(file)->i_mode))
6906 file_ptr |= FFS_ISREG;
6907 file_slot->file_ptr = file_ptr;
6910 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6911 struct io_kiocb *req, int fd)
6914 unsigned long file_ptr;
6916 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6918 fd = array_index_nospec(fd, ctx->nr_user_files);
6919 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6920 file = (struct file *) (file_ptr & FFS_MASK);
6921 file_ptr &= ~FFS_MASK;
6922 /* mask in overlapping REQ_F and FFS bits */
6923 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6924 io_req_set_rsrc_node(req);
6928 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6929 struct io_kiocb *req, int fd)
6931 struct file *file = fget(fd);
6933 trace_io_uring_file_get(ctx, fd);
6935 /* we don't allow fixed io_uring files */
6936 if (file && unlikely(file->f_op == &io_uring_fops))
6937 io_req_track_inflight(req);
6941 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6942 struct io_kiocb *req, int fd, bool fixed)
6945 return io_file_get_fixed(ctx, req, fd);
6947 return io_file_get_normal(ctx, req, fd);
6950 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6952 struct io_kiocb *prev = req->timeout.prev;
6956 if (!(req->task->flags & PF_EXITING))
6957 ret = io_try_cancel_userdata(req, prev->user_data);
6958 io_req_complete_post(req, ret ?: -ETIME, 0);
6961 io_req_complete_post(req, -ETIME, 0);
6965 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6967 struct io_timeout_data *data = container_of(timer,
6968 struct io_timeout_data, timer);
6969 struct io_kiocb *prev, *req = data->req;
6970 struct io_ring_ctx *ctx = req->ctx;
6971 unsigned long flags;
6973 spin_lock_irqsave(&ctx->timeout_lock, flags);
6974 prev = req->timeout.head;
6975 req->timeout.head = NULL;
6978 * We don't expect the list to be empty, that will only happen if we
6979 * race with the completion of the linked work.
6982 io_remove_next_linked(prev);
6983 if (!req_ref_inc_not_zero(prev))
6986 list_del(&req->timeout.list);
6987 req->timeout.prev = prev;
6988 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6990 req->io_task_work.func = io_req_task_link_timeout;
6991 io_req_task_work_add(req);
6992 return HRTIMER_NORESTART;
6995 static void io_queue_linked_timeout(struct io_kiocb *req)
6997 struct io_ring_ctx *ctx = req->ctx;
6999 spin_lock_irq(&ctx->timeout_lock);
7001 * If the back reference is NULL, then our linked request finished
7002 * before we got a chance to setup the timer
7004 if (req->timeout.head) {
7005 struct io_timeout_data *data = req->async_data;
7007 data->timer.function = io_link_timeout_fn;
7008 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
7010 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
7012 spin_unlock_irq(&ctx->timeout_lock);
7013 /* drop submission reference */
7017 static void __io_queue_sqe(struct io_kiocb *req)
7018 __must_hold(&req->ctx->uring_lock)
7020 struct io_kiocb *linked_timeout;
7024 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
7027 * We async punt it if the file wasn't marked NOWAIT, or if the file
7028 * doesn't support non-blocking read/write attempts
7031 if (req->flags & REQ_F_COMPLETE_INLINE) {
7032 struct io_ring_ctx *ctx = req->ctx;
7033 struct io_submit_state *state = &ctx->submit_state;
7035 state->compl_reqs[state->compl_nr++] = req;
7036 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
7037 io_submit_flush_completions(ctx);
7041 linked_timeout = io_prep_linked_timeout(req);
7043 io_queue_linked_timeout(linked_timeout);
7044 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
7045 linked_timeout = io_prep_linked_timeout(req);
7047 switch (io_arm_poll_handler(req)) {
7048 case IO_APOLL_READY:
7050 io_queue_linked_timeout(linked_timeout);
7052 case IO_APOLL_ABORTED:
7054 * Queued up for async execution, worker will release
7055 * submit reference when the iocb is actually submitted.
7057 io_queue_async_work(req, NULL);
7062 io_queue_linked_timeout(linked_timeout);
7064 io_req_complete_failed(req, ret);
7068 static inline void io_queue_sqe(struct io_kiocb *req)
7069 __must_hold(&req->ctx->uring_lock)
7071 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7074 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7075 __io_queue_sqe(req);
7076 } else if (req->flags & REQ_F_FAIL) {
7077 io_req_complete_fail_submit(req);
7079 int ret = io_req_prep_async(req);
7082 io_req_complete_failed(req, ret);
7084 io_queue_async_work(req, NULL);
7089 * Check SQE restrictions (opcode and flags).
7091 * Returns 'true' if SQE is allowed, 'false' otherwise.
7093 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7094 struct io_kiocb *req,
7095 unsigned int sqe_flags)
7097 if (likely(!ctx->restricted))
7100 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7103 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7104 ctx->restrictions.sqe_flags_required)
7107 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7108 ctx->restrictions.sqe_flags_required))
7114 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7115 const struct io_uring_sqe *sqe)
7116 __must_hold(&ctx->uring_lock)
7118 struct io_submit_state *state;
7119 unsigned int sqe_flags;
7120 int personality, ret = 0;
7122 /* req is partially pre-initialised, see io_preinit_req() */
7123 req->opcode = READ_ONCE(sqe->opcode);
7124 /* same numerical values with corresponding REQ_F_*, safe to copy */
7125 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7126 req->user_data = READ_ONCE(sqe->user_data);
7128 req->fixed_rsrc_refs = NULL;
7129 req->task = current;
7131 /* enforce forwards compatibility on users */
7132 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7134 if (unlikely(req->opcode >= IORING_OP_LAST))
7136 if (!io_check_restriction(ctx, req, sqe_flags))
7139 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7140 !io_op_defs[req->opcode].buffer_select)
7142 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7143 ctx->drain_active = true;
7145 personality = READ_ONCE(sqe->personality);
7147 req->creds = xa_load(&ctx->personalities, personality);
7150 get_cred(req->creds);
7151 req->flags |= REQ_F_CREDS;
7153 state = &ctx->submit_state;
7156 * Plug now if we have more than 1 IO left after this, and the target
7157 * is potentially a read/write to block based storage.
7159 if (!state->plug_started && state->ios_left > 1 &&
7160 io_op_defs[req->opcode].plug) {
7161 blk_start_plug(&state->plug);
7162 state->plug_started = true;
7165 if (io_op_defs[req->opcode].needs_file) {
7166 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7167 (sqe_flags & IOSQE_FIXED_FILE));
7168 if (unlikely(!req->file))
7176 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7177 const struct io_uring_sqe *sqe)
7178 __must_hold(&ctx->uring_lock)
7180 struct io_submit_link *link = &ctx->submit_state.link;
7183 ret = io_init_req(ctx, req, sqe);
7184 if (unlikely(ret)) {
7186 /* fail even hard links since we don't submit */
7189 * we can judge a link req is failed or cancelled by if
7190 * REQ_F_FAIL is set, but the head is an exception since
7191 * it may be set REQ_F_FAIL because of other req's failure
7192 * so let's leverage req->result to distinguish if a head
7193 * is set REQ_F_FAIL because of its failure or other req's
7194 * failure so that we can set the correct ret code for it.
7195 * init result here to avoid affecting the normal path.
7197 if (!(link->head->flags & REQ_F_FAIL))
7198 req_fail_link_node(link->head, -ECANCELED);
7199 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7201 * the current req is a normal req, we should return
7202 * error and thus break the submittion loop.
7204 io_req_complete_failed(req, ret);
7207 req_fail_link_node(req, ret);
7209 ret = io_req_prep(req, sqe);
7214 /* don't need @sqe from now on */
7215 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7217 ctx->flags & IORING_SETUP_SQPOLL);
7220 * If we already have a head request, queue this one for async
7221 * submittal once the head completes. If we don't have a head but
7222 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7223 * submitted sync once the chain is complete. If none of those
7224 * conditions are true (normal request), then just queue it.
7227 struct io_kiocb *head = link->head;
7229 if (!(req->flags & REQ_F_FAIL)) {
7230 ret = io_req_prep_async(req);
7231 if (unlikely(ret)) {
7232 req_fail_link_node(req, ret);
7233 if (!(head->flags & REQ_F_FAIL))
7234 req_fail_link_node(head, -ECANCELED);
7237 trace_io_uring_link(ctx, req, head);
7238 link->last->link = req;
7241 /* last request of a link, enqueue the link */
7242 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7247 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7259 * Batched submission is done, ensure local IO is flushed out.
7261 static void io_submit_state_end(struct io_submit_state *state,
7262 struct io_ring_ctx *ctx)
7264 if (state->link.head)
7265 io_queue_sqe(state->link.head);
7266 if (state->compl_nr)
7267 io_submit_flush_completions(ctx);
7268 if (state->plug_started)
7269 blk_finish_plug(&state->plug);
7273 * Start submission side cache.
7275 static void io_submit_state_start(struct io_submit_state *state,
7276 unsigned int max_ios)
7278 state->plug_started = false;
7279 state->ios_left = max_ios;
7280 /* set only head, no need to init link_last in advance */
7281 state->link.head = NULL;
7284 static void io_commit_sqring(struct io_ring_ctx *ctx)
7286 struct io_rings *rings = ctx->rings;
7289 * Ensure any loads from the SQEs are done at this point,
7290 * since once we write the new head, the application could
7291 * write new data to them.
7293 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7297 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7298 * that is mapped by userspace. This means that care needs to be taken to
7299 * ensure that reads are stable, as we cannot rely on userspace always
7300 * being a good citizen. If members of the sqe are validated and then later
7301 * used, it's important that those reads are done through READ_ONCE() to
7302 * prevent a re-load down the line.
7304 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7306 unsigned head, mask = ctx->sq_entries - 1;
7307 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7310 * The cached sq head (or cq tail) serves two purposes:
7312 * 1) allows us to batch the cost of updating the user visible
7314 * 2) allows the kernel side to track the head on its own, even
7315 * though the application is the one updating it.
7317 head = READ_ONCE(ctx->sq_array[sq_idx]);
7318 if (likely(head < ctx->sq_entries))
7319 return &ctx->sq_sqes[head];
7321 /* drop invalid entries */
7323 WRITE_ONCE(ctx->rings->sq_dropped,
7324 READ_ONCE(ctx->rings->sq_dropped) + 1);
7328 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7329 __must_hold(&ctx->uring_lock)
7333 /* make sure SQ entry isn't read before tail */
7334 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7335 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7337 io_get_task_refs(nr);
7339 io_submit_state_start(&ctx->submit_state, nr);
7340 while (submitted < nr) {
7341 const struct io_uring_sqe *sqe;
7342 struct io_kiocb *req;
7344 req = io_alloc_req(ctx);
7345 if (unlikely(!req)) {
7347 submitted = -EAGAIN;
7350 sqe = io_get_sqe(ctx);
7351 if (unlikely(!sqe)) {
7352 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7355 /* will complete beyond this point, count as submitted */
7357 if (io_submit_sqe(ctx, req, sqe))
7361 if (unlikely(submitted != nr)) {
7362 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7363 int unused = nr - ref_used;
7365 current->io_uring->cached_refs += unused;
7366 percpu_ref_put_many(&ctx->refs, unused);
7369 io_submit_state_end(&ctx->submit_state, ctx);
7370 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7371 io_commit_sqring(ctx);
7376 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7378 return READ_ONCE(sqd->state);
7381 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7383 /* Tell userspace we may need a wakeup call */
7384 spin_lock(&ctx->completion_lock);
7385 WRITE_ONCE(ctx->rings->sq_flags,
7386 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7387 spin_unlock(&ctx->completion_lock);
7390 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7392 spin_lock(&ctx->completion_lock);
7393 WRITE_ONCE(ctx->rings->sq_flags,
7394 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7395 spin_unlock(&ctx->completion_lock);
7398 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7400 unsigned int to_submit;
7403 to_submit = io_sqring_entries(ctx);
7404 /* if we're handling multiple rings, cap submit size for fairness */
7405 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7406 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7408 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7409 unsigned nr_events = 0;
7410 const struct cred *creds = NULL;
7412 if (ctx->sq_creds != current_cred())
7413 creds = override_creds(ctx->sq_creds);
7415 mutex_lock(&ctx->uring_lock);
7416 if (!list_empty(&ctx->iopoll_list))
7417 io_do_iopoll(ctx, &nr_events, 0);
7420 * Don't submit if refs are dying, good for io_uring_register(),
7421 * but also it is relied upon by io_ring_exit_work()
7423 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7424 !(ctx->flags & IORING_SETUP_R_DISABLED))
7425 ret = io_submit_sqes(ctx, to_submit);
7426 mutex_unlock(&ctx->uring_lock);
7428 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7429 wake_up(&ctx->sqo_sq_wait);
7431 revert_creds(creds);
7437 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7439 struct io_ring_ctx *ctx;
7440 unsigned sq_thread_idle = 0;
7442 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7443 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7444 sqd->sq_thread_idle = sq_thread_idle;
7447 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7449 bool did_sig = false;
7450 struct ksignal ksig;
7452 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7453 signal_pending(current)) {
7454 mutex_unlock(&sqd->lock);
7455 if (signal_pending(current))
7456 did_sig = get_signal(&ksig);
7458 mutex_lock(&sqd->lock);
7460 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7463 static int io_sq_thread(void *data)
7465 struct io_sq_data *sqd = data;
7466 struct io_ring_ctx *ctx;
7467 unsigned long timeout = 0;
7468 char buf[TASK_COMM_LEN];
7471 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7472 set_task_comm(current, buf);
7474 if (sqd->sq_cpu != -1)
7475 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7477 set_cpus_allowed_ptr(current, cpu_online_mask);
7478 current->flags |= PF_NO_SETAFFINITY;
7480 mutex_lock(&sqd->lock);
7482 bool cap_entries, sqt_spin = false;
7484 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7485 if (io_sqd_handle_event(sqd))
7487 timeout = jiffies + sqd->sq_thread_idle;
7490 cap_entries = !list_is_singular(&sqd->ctx_list);
7491 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7492 int ret = __io_sq_thread(ctx, cap_entries);
7494 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7497 if (io_run_task_work())
7500 if (sqt_spin || !time_after(jiffies, timeout)) {
7503 timeout = jiffies + sqd->sq_thread_idle;
7507 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7508 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7509 bool needs_sched = true;
7511 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7512 io_ring_set_wakeup_flag(ctx);
7514 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7515 !list_empty_careful(&ctx->iopoll_list)) {
7516 needs_sched = false;
7519 if (io_sqring_entries(ctx)) {
7520 needs_sched = false;
7526 mutex_unlock(&sqd->lock);
7528 mutex_lock(&sqd->lock);
7530 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7531 io_ring_clear_wakeup_flag(ctx);
7534 finish_wait(&sqd->wait, &wait);
7535 timeout = jiffies + sqd->sq_thread_idle;
7538 io_uring_cancel_generic(true, sqd);
7540 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7541 io_ring_set_wakeup_flag(ctx);
7543 mutex_unlock(&sqd->lock);
7545 complete(&sqd->exited);
7549 struct io_wait_queue {
7550 struct wait_queue_entry wq;
7551 struct io_ring_ctx *ctx;
7553 unsigned nr_timeouts;
7556 static inline bool io_should_wake(struct io_wait_queue *iowq)
7558 struct io_ring_ctx *ctx = iowq->ctx;
7559 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7562 * Wake up if we have enough events, or if a timeout occurred since we
7563 * started waiting. For timeouts, we always want to return to userspace,
7564 * regardless of event count.
7566 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7569 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7570 int wake_flags, void *key)
7572 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7576 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7577 * the task, and the next invocation will do it.
7579 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7580 return autoremove_wake_function(curr, mode, wake_flags, key);
7584 static int io_run_task_work_sig(void)
7586 if (io_run_task_work())
7588 if (!signal_pending(current))
7590 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7591 return -ERESTARTSYS;
7595 /* when returns >0, the caller should retry */
7596 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7597 struct io_wait_queue *iowq,
7602 /* make sure we run task_work before checking for signals */
7603 ret = io_run_task_work_sig();
7604 if (ret || io_should_wake(iowq))
7606 /* let the caller flush overflows, retry */
7607 if (test_bit(0, &ctx->check_cq_overflow))
7610 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
7616 * Wait until events become available, if we don't already have some. The
7617 * application must reap them itself, as they reside on the shared cq ring.
7619 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7620 const sigset_t __user *sig, size_t sigsz,
7621 struct __kernel_timespec __user *uts)
7623 struct io_wait_queue iowq;
7624 struct io_rings *rings = ctx->rings;
7625 ktime_t timeout = KTIME_MAX;
7629 io_cqring_overflow_flush(ctx);
7630 if (io_cqring_events(ctx) >= min_events)
7632 if (!io_run_task_work())
7637 struct timespec64 ts;
7639 if (get_timespec64(&ts, uts))
7641 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
7645 #ifdef CONFIG_COMPAT
7646 if (in_compat_syscall())
7647 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7651 ret = set_user_sigmask(sig, sigsz);
7657 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7658 iowq.wq.private = current;
7659 INIT_LIST_HEAD(&iowq.wq.entry);
7661 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7662 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7664 trace_io_uring_cqring_wait(ctx, min_events);
7666 /* if we can't even flush overflow, don't wait for more */
7667 if (!io_cqring_overflow_flush(ctx)) {
7671 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7672 TASK_INTERRUPTIBLE);
7673 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
7674 finish_wait(&ctx->cq_wait, &iowq.wq);
7678 restore_saved_sigmask_unless(ret == -EINTR);
7680 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7683 static void io_free_page_table(void **table, size_t size)
7685 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7687 for (i = 0; i < nr_tables; i++)
7692 static void **io_alloc_page_table(size_t size)
7694 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7695 size_t init_size = size;
7698 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7702 for (i = 0; i < nr_tables; i++) {
7703 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7705 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7707 io_free_page_table(table, init_size);
7715 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7717 percpu_ref_exit(&ref_node->refs);
7721 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7723 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7724 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7725 unsigned long flags;
7726 bool first_add = false;
7727 unsigned long delay = HZ;
7729 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7732 /* if we are mid-quiesce then do not delay */
7733 if (node->rsrc_data->quiesce)
7736 while (!list_empty(&ctx->rsrc_ref_list)) {
7737 node = list_first_entry(&ctx->rsrc_ref_list,
7738 struct io_rsrc_node, node);
7739 /* recycle ref nodes in order */
7742 list_del(&node->node);
7743 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7745 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7748 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7751 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7753 struct io_rsrc_node *ref_node;
7755 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7759 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7764 INIT_LIST_HEAD(&ref_node->node);
7765 INIT_LIST_HEAD(&ref_node->rsrc_list);
7766 ref_node->done = false;
7770 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7771 struct io_rsrc_data *data_to_kill)
7773 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7774 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7777 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7779 rsrc_node->rsrc_data = data_to_kill;
7780 spin_lock_irq(&ctx->rsrc_ref_lock);
7781 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7782 spin_unlock_irq(&ctx->rsrc_ref_lock);
7784 atomic_inc(&data_to_kill->refs);
7785 percpu_ref_kill(&rsrc_node->refs);
7786 ctx->rsrc_node = NULL;
7789 if (!ctx->rsrc_node) {
7790 ctx->rsrc_node = ctx->rsrc_backup_node;
7791 ctx->rsrc_backup_node = NULL;
7795 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7797 if (ctx->rsrc_backup_node)
7799 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7800 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7803 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7807 /* As we may drop ->uring_lock, other task may have started quiesce */
7811 data->quiesce = true;
7813 ret = io_rsrc_node_switch_start(ctx);
7816 io_rsrc_node_switch(ctx, data);
7818 /* kill initial ref, already quiesced if zero */
7819 if (atomic_dec_and_test(&data->refs))
7821 mutex_unlock(&ctx->uring_lock);
7822 flush_delayed_work(&ctx->rsrc_put_work);
7823 ret = wait_for_completion_interruptible(&data->done);
7825 mutex_lock(&ctx->uring_lock);
7826 if (atomic_read(&data->refs) > 0) {
7828 * it has been revived by another thread while
7831 mutex_unlock(&ctx->uring_lock);
7837 atomic_inc(&data->refs);
7838 /* wait for all works potentially completing data->done */
7839 flush_delayed_work(&ctx->rsrc_put_work);
7840 reinit_completion(&data->done);
7842 ret = io_run_task_work_sig();
7843 mutex_lock(&ctx->uring_lock);
7845 data->quiesce = false;
7850 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7852 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7853 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7855 return &data->tags[table_idx][off];
7858 static void io_rsrc_data_free(struct io_rsrc_data *data)
7860 size_t size = data->nr * sizeof(data->tags[0][0]);
7863 io_free_page_table((void **)data->tags, size);
7867 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7868 u64 __user *utags, unsigned nr,
7869 struct io_rsrc_data **pdata)
7871 struct io_rsrc_data *data;
7875 data = kzalloc(sizeof(*data), GFP_KERNEL);
7878 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7886 data->do_put = do_put;
7889 for (i = 0; i < nr; i++) {
7890 u64 *tag_slot = io_get_tag_slot(data, i);
7892 if (copy_from_user(tag_slot, &utags[i],
7898 atomic_set(&data->refs, 1);
7899 init_completion(&data->done);
7903 io_rsrc_data_free(data);
7907 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7909 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7910 GFP_KERNEL_ACCOUNT);
7911 return !!table->files;
7914 static void io_free_file_tables(struct io_file_table *table)
7916 kvfree(table->files);
7917 table->files = NULL;
7920 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7922 #if defined(CONFIG_UNIX)
7923 if (ctx->ring_sock) {
7924 struct sock *sock = ctx->ring_sock->sk;
7925 struct sk_buff *skb;
7927 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7933 for (i = 0; i < ctx->nr_user_files; i++) {
7936 file = io_file_from_index(ctx, i);
7941 io_free_file_tables(&ctx->file_table);
7942 io_rsrc_data_free(ctx->file_data);
7943 ctx->file_data = NULL;
7944 ctx->nr_user_files = 0;
7947 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7951 if (!ctx->file_data)
7953 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7955 __io_sqe_files_unregister(ctx);
7959 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7960 __releases(&sqd->lock)
7962 WARN_ON_ONCE(sqd->thread == current);
7965 * Do the dance but not conditional clear_bit() because it'd race with
7966 * other threads incrementing park_pending and setting the bit.
7968 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7969 if (atomic_dec_return(&sqd->park_pending))
7970 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7971 mutex_unlock(&sqd->lock);
7974 static void io_sq_thread_park(struct io_sq_data *sqd)
7975 __acquires(&sqd->lock)
7977 WARN_ON_ONCE(sqd->thread == current);
7979 atomic_inc(&sqd->park_pending);
7980 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7981 mutex_lock(&sqd->lock);
7983 wake_up_process(sqd->thread);
7986 static void io_sq_thread_stop(struct io_sq_data *sqd)
7988 WARN_ON_ONCE(sqd->thread == current);
7989 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7991 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7992 mutex_lock(&sqd->lock);
7994 wake_up_process(sqd->thread);
7995 mutex_unlock(&sqd->lock);
7996 wait_for_completion(&sqd->exited);
7999 static void io_put_sq_data(struct io_sq_data *sqd)
8001 if (refcount_dec_and_test(&sqd->refs)) {
8002 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
8004 io_sq_thread_stop(sqd);
8009 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
8011 struct io_sq_data *sqd = ctx->sq_data;
8014 io_sq_thread_park(sqd);
8015 list_del_init(&ctx->sqd_list);
8016 io_sqd_update_thread_idle(sqd);
8017 io_sq_thread_unpark(sqd);
8019 io_put_sq_data(sqd);
8020 ctx->sq_data = NULL;
8024 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
8026 struct io_ring_ctx *ctx_attach;
8027 struct io_sq_data *sqd;
8030 f = fdget(p->wq_fd);
8032 return ERR_PTR(-ENXIO);
8033 if (f.file->f_op != &io_uring_fops) {
8035 return ERR_PTR(-EINVAL);
8038 ctx_attach = f.file->private_data;
8039 sqd = ctx_attach->sq_data;
8042 return ERR_PTR(-EINVAL);
8044 if (sqd->task_tgid != current->tgid) {
8046 return ERR_PTR(-EPERM);
8049 refcount_inc(&sqd->refs);
8054 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
8057 struct io_sq_data *sqd;
8060 if (p->flags & IORING_SETUP_ATTACH_WQ) {
8061 sqd = io_attach_sq_data(p);
8066 /* fall through for EPERM case, setup new sqd/task */
8067 if (PTR_ERR(sqd) != -EPERM)
8071 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8073 return ERR_PTR(-ENOMEM);
8075 atomic_set(&sqd->park_pending, 0);
8076 refcount_set(&sqd->refs, 1);
8077 INIT_LIST_HEAD(&sqd->ctx_list);
8078 mutex_init(&sqd->lock);
8079 init_waitqueue_head(&sqd->wait);
8080 init_completion(&sqd->exited);
8084 #if defined(CONFIG_UNIX)
8086 * Ensure the UNIX gc is aware of our file set, so we are certain that
8087 * the io_uring can be safely unregistered on process exit, even if we have
8088 * loops in the file referencing.
8090 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8092 struct sock *sk = ctx->ring_sock->sk;
8093 struct scm_fp_list *fpl;
8094 struct sk_buff *skb;
8097 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8101 skb = alloc_skb(0, GFP_KERNEL);
8110 fpl->user = get_uid(current_user());
8111 for (i = 0; i < nr; i++) {
8112 struct file *file = io_file_from_index(ctx, i + offset);
8116 fpl->fp[nr_files] = get_file(file);
8117 unix_inflight(fpl->user, fpl->fp[nr_files]);
8122 fpl->max = SCM_MAX_FD;
8123 fpl->count = nr_files;
8124 UNIXCB(skb).fp = fpl;
8125 skb->destructor = unix_destruct_scm;
8126 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8127 skb_queue_head(&sk->sk_receive_queue, skb);
8129 for (i = 0; i < nr_files; i++)
8140 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8141 * causes regular reference counting to break down. We rely on the UNIX
8142 * garbage collection to take care of this problem for us.
8144 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8146 unsigned left, total;
8150 left = ctx->nr_user_files;
8152 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8154 ret = __io_sqe_files_scm(ctx, this_files, total);
8158 total += this_files;
8164 while (total < ctx->nr_user_files) {
8165 struct file *file = io_file_from_index(ctx, total);
8175 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8181 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8183 struct file *file = prsrc->file;
8184 #if defined(CONFIG_UNIX)
8185 struct sock *sock = ctx->ring_sock->sk;
8186 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8187 struct sk_buff *skb;
8190 __skb_queue_head_init(&list);
8193 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8194 * remove this entry and rearrange the file array.
8196 skb = skb_dequeue(head);
8198 struct scm_fp_list *fp;
8200 fp = UNIXCB(skb).fp;
8201 for (i = 0; i < fp->count; i++) {
8204 if (fp->fp[i] != file)
8207 unix_notinflight(fp->user, fp->fp[i]);
8208 left = fp->count - 1 - i;
8210 memmove(&fp->fp[i], &fp->fp[i + 1],
8211 left * sizeof(struct file *));
8218 __skb_queue_tail(&list, skb);
8228 __skb_queue_tail(&list, skb);
8230 skb = skb_dequeue(head);
8233 if (skb_peek(&list)) {
8234 spin_lock_irq(&head->lock);
8235 while ((skb = __skb_dequeue(&list)) != NULL)
8236 __skb_queue_tail(head, skb);
8237 spin_unlock_irq(&head->lock);
8244 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8246 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8247 struct io_ring_ctx *ctx = rsrc_data->ctx;
8248 struct io_rsrc_put *prsrc, *tmp;
8250 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8251 list_del(&prsrc->list);
8254 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8256 io_ring_submit_lock(ctx, lock_ring);
8257 spin_lock(&ctx->completion_lock);
8258 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8260 io_commit_cqring(ctx);
8261 spin_unlock(&ctx->completion_lock);
8262 io_cqring_ev_posted(ctx);
8263 io_ring_submit_unlock(ctx, lock_ring);
8266 rsrc_data->do_put(ctx, prsrc);
8270 io_rsrc_node_destroy(ref_node);
8271 if (atomic_dec_and_test(&rsrc_data->refs))
8272 complete(&rsrc_data->done);
8275 static void io_rsrc_put_work(struct work_struct *work)
8277 struct io_ring_ctx *ctx;
8278 struct llist_node *node;
8280 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8281 node = llist_del_all(&ctx->rsrc_put_llist);
8284 struct io_rsrc_node *ref_node;
8285 struct llist_node *next = node->next;
8287 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8288 __io_rsrc_put_work(ref_node);
8293 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8294 unsigned nr_args, u64 __user *tags)
8296 __s32 __user *fds = (__s32 __user *) arg;
8305 if (nr_args > IORING_MAX_FIXED_FILES)
8307 if (nr_args > rlimit(RLIMIT_NOFILE))
8309 ret = io_rsrc_node_switch_start(ctx);
8312 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8318 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8321 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8322 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8326 /* allow sparse sets */
8329 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8336 if (unlikely(!file))
8340 * Don't allow io_uring instances to be registered. If UNIX
8341 * isn't enabled, then this causes a reference cycle and this
8342 * instance can never get freed. If UNIX is enabled we'll
8343 * handle it just fine, but there's still no point in allowing
8344 * a ring fd as it doesn't support regular read/write anyway.
8346 if (file->f_op == &io_uring_fops) {
8350 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8353 ret = io_sqe_files_scm(ctx);
8355 __io_sqe_files_unregister(ctx);
8359 io_rsrc_node_switch(ctx, NULL);
8362 for (i = 0; i < ctx->nr_user_files; i++) {
8363 file = io_file_from_index(ctx, i);
8367 io_free_file_tables(&ctx->file_table);
8368 ctx->nr_user_files = 0;
8370 io_rsrc_data_free(ctx->file_data);
8371 ctx->file_data = NULL;
8375 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8378 #if defined(CONFIG_UNIX)
8379 struct sock *sock = ctx->ring_sock->sk;
8380 struct sk_buff_head *head = &sock->sk_receive_queue;
8381 struct sk_buff *skb;
8384 * See if we can merge this file into an existing skb SCM_RIGHTS
8385 * file set. If there's no room, fall back to allocating a new skb
8386 * and filling it in.
8388 spin_lock_irq(&head->lock);
8389 skb = skb_peek(head);
8391 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8393 if (fpl->count < SCM_MAX_FD) {
8394 __skb_unlink(skb, head);
8395 spin_unlock_irq(&head->lock);
8396 fpl->fp[fpl->count] = get_file(file);
8397 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8399 spin_lock_irq(&head->lock);
8400 __skb_queue_head(head, skb);
8405 spin_unlock_irq(&head->lock);
8412 return __io_sqe_files_scm(ctx, 1, index);
8418 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8419 struct io_rsrc_node *node, void *rsrc)
8421 struct io_rsrc_put *prsrc;
8423 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8427 prsrc->tag = *io_get_tag_slot(data, idx);
8429 list_add(&prsrc->list, &node->rsrc_list);
8433 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8434 unsigned int issue_flags, u32 slot_index)
8436 struct io_ring_ctx *ctx = req->ctx;
8437 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8438 bool needs_switch = false;
8439 struct io_fixed_file *file_slot;
8442 io_ring_submit_lock(ctx, !force_nonblock);
8443 if (file->f_op == &io_uring_fops)
8446 if (!ctx->file_data)
8449 if (slot_index >= ctx->nr_user_files)
8452 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8453 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8455 if (file_slot->file_ptr) {
8456 struct file *old_file;
8458 ret = io_rsrc_node_switch_start(ctx);
8462 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8463 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8464 ctx->rsrc_node, old_file);
8467 file_slot->file_ptr = 0;
8468 needs_switch = true;
8471 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8472 io_fixed_file_set(file_slot, file);
8473 ret = io_sqe_file_register(ctx, file, slot_index);
8475 file_slot->file_ptr = 0;
8482 io_rsrc_node_switch(ctx, ctx->file_data);
8483 io_ring_submit_unlock(ctx, !force_nonblock);
8489 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8491 unsigned int offset = req->close.file_slot - 1;
8492 struct io_ring_ctx *ctx = req->ctx;
8493 struct io_fixed_file *file_slot;
8497 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8499 if (unlikely(!ctx->file_data))
8502 if (offset >= ctx->nr_user_files)
8504 ret = io_rsrc_node_switch_start(ctx);
8508 i = array_index_nospec(offset, ctx->nr_user_files);
8509 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8511 if (!file_slot->file_ptr)
8514 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8515 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8519 file_slot->file_ptr = 0;
8520 io_rsrc_node_switch(ctx, ctx->file_data);
8523 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8527 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8528 struct io_uring_rsrc_update2 *up,
8531 u64 __user *tags = u64_to_user_ptr(up->tags);
8532 __s32 __user *fds = u64_to_user_ptr(up->data);
8533 struct io_rsrc_data *data = ctx->file_data;
8534 struct io_fixed_file *file_slot;
8538 bool needs_switch = false;
8540 if (!ctx->file_data)
8542 if (up->offset + nr_args > ctx->nr_user_files)
8545 for (done = 0; done < nr_args; done++) {
8548 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8549 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8553 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8557 if (fd == IORING_REGISTER_FILES_SKIP)
8560 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8561 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8563 if (file_slot->file_ptr) {
8564 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8565 err = io_queue_rsrc_removal(data, up->offset + done,
8566 ctx->rsrc_node, file);
8569 file_slot->file_ptr = 0;
8570 needs_switch = true;
8579 * Don't allow io_uring instances to be registered. If
8580 * UNIX isn't enabled, then this causes a reference
8581 * cycle and this instance can never get freed. If UNIX
8582 * is enabled we'll handle it just fine, but there's
8583 * still no point in allowing a ring fd as it doesn't
8584 * support regular read/write anyway.
8586 if (file->f_op == &io_uring_fops) {
8591 *io_get_tag_slot(data, up->offset + done) = tag;
8592 io_fixed_file_set(file_slot, file);
8593 err = io_sqe_file_register(ctx, file, i);
8595 file_slot->file_ptr = 0;
8603 io_rsrc_node_switch(ctx, data);
8604 return done ? done : err;
8607 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8608 struct task_struct *task)
8610 struct io_wq_hash *hash;
8611 struct io_wq_data data;
8612 unsigned int concurrency;
8614 mutex_lock(&ctx->uring_lock);
8615 hash = ctx->hash_map;
8617 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8619 mutex_unlock(&ctx->uring_lock);
8620 return ERR_PTR(-ENOMEM);
8622 refcount_set(&hash->refs, 1);
8623 init_waitqueue_head(&hash->wait);
8624 ctx->hash_map = hash;
8626 mutex_unlock(&ctx->uring_lock);
8630 data.free_work = io_wq_free_work;
8631 data.do_work = io_wq_submit_work;
8633 /* Do QD, or 4 * CPUS, whatever is smallest */
8634 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8636 return io_wq_create(concurrency, &data);
8639 static int io_uring_alloc_task_context(struct task_struct *task,
8640 struct io_ring_ctx *ctx)
8642 struct io_uring_task *tctx;
8645 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8646 if (unlikely(!tctx))
8649 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8650 if (unlikely(ret)) {
8655 tctx->io_wq = io_init_wq_offload(ctx, task);
8656 if (IS_ERR(tctx->io_wq)) {
8657 ret = PTR_ERR(tctx->io_wq);
8658 percpu_counter_destroy(&tctx->inflight);
8664 init_waitqueue_head(&tctx->wait);
8665 atomic_set(&tctx->in_idle, 0);
8666 atomic_set(&tctx->inflight_tracked, 0);
8667 task->io_uring = tctx;
8668 spin_lock_init(&tctx->task_lock);
8669 INIT_WQ_LIST(&tctx->task_list);
8670 init_task_work(&tctx->task_work, tctx_task_work);
8674 void __io_uring_free(struct task_struct *tsk)
8676 struct io_uring_task *tctx = tsk->io_uring;
8678 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8679 WARN_ON_ONCE(tctx->io_wq);
8680 WARN_ON_ONCE(tctx->cached_refs);
8682 percpu_counter_destroy(&tctx->inflight);
8684 tsk->io_uring = NULL;
8687 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8688 struct io_uring_params *p)
8692 /* Retain compatibility with failing for an invalid attach attempt */
8693 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8694 IORING_SETUP_ATTACH_WQ) {
8697 f = fdget(p->wq_fd);
8700 if (f.file->f_op != &io_uring_fops) {
8706 if (ctx->flags & IORING_SETUP_SQPOLL) {
8707 struct task_struct *tsk;
8708 struct io_sq_data *sqd;
8711 sqd = io_get_sq_data(p, &attached);
8717 ctx->sq_creds = get_current_cred();
8719 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8720 if (!ctx->sq_thread_idle)
8721 ctx->sq_thread_idle = HZ;
8723 io_sq_thread_park(sqd);
8724 list_add(&ctx->sqd_list, &sqd->ctx_list);
8725 io_sqd_update_thread_idle(sqd);
8726 /* don't attach to a dying SQPOLL thread, would be racy */
8727 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8728 io_sq_thread_unpark(sqd);
8735 if (p->flags & IORING_SETUP_SQ_AFF) {
8736 int cpu = p->sq_thread_cpu;
8739 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8746 sqd->task_pid = current->pid;
8747 sqd->task_tgid = current->tgid;
8748 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8755 ret = io_uring_alloc_task_context(tsk, ctx);
8756 wake_up_new_task(tsk);
8759 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8760 /* Can't have SQ_AFF without SQPOLL */
8767 complete(&ctx->sq_data->exited);
8769 io_sq_thread_finish(ctx);
8773 static inline void __io_unaccount_mem(struct user_struct *user,
8774 unsigned long nr_pages)
8776 atomic_long_sub(nr_pages, &user->locked_vm);
8779 static inline int __io_account_mem(struct user_struct *user,
8780 unsigned long nr_pages)
8782 unsigned long page_limit, cur_pages, new_pages;
8784 /* Don't allow more pages than we can safely lock */
8785 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8788 cur_pages = atomic_long_read(&user->locked_vm);
8789 new_pages = cur_pages + nr_pages;
8790 if (new_pages > page_limit)
8792 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8793 new_pages) != cur_pages);
8798 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8801 __io_unaccount_mem(ctx->user, nr_pages);
8803 if (ctx->mm_account)
8804 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8807 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8812 ret = __io_account_mem(ctx->user, nr_pages);
8817 if (ctx->mm_account)
8818 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8823 static void io_mem_free(void *ptr)
8830 page = virt_to_head_page(ptr);
8831 if (put_page_testzero(page))
8832 free_compound_page(page);
8835 static void *io_mem_alloc(size_t size)
8837 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
8839 return (void *) __get_free_pages(gfp, get_order(size));
8842 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8845 struct io_rings *rings;
8846 size_t off, sq_array_size;
8848 off = struct_size(rings, cqes, cq_entries);
8849 if (off == SIZE_MAX)
8853 off = ALIGN(off, SMP_CACHE_BYTES);
8861 sq_array_size = array_size(sizeof(u32), sq_entries);
8862 if (sq_array_size == SIZE_MAX)
8865 if (check_add_overflow(off, sq_array_size, &off))
8871 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8873 struct io_mapped_ubuf *imu = *slot;
8876 if (imu != ctx->dummy_ubuf) {
8877 for (i = 0; i < imu->nr_bvecs; i++)
8878 unpin_user_page(imu->bvec[i].bv_page);
8879 if (imu->acct_pages)
8880 io_unaccount_mem(ctx, imu->acct_pages);
8886 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8888 io_buffer_unmap(ctx, &prsrc->buf);
8892 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8896 for (i = 0; i < ctx->nr_user_bufs; i++)
8897 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8898 kfree(ctx->user_bufs);
8899 io_rsrc_data_free(ctx->buf_data);
8900 ctx->user_bufs = NULL;
8901 ctx->buf_data = NULL;
8902 ctx->nr_user_bufs = 0;
8905 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8912 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8914 __io_sqe_buffers_unregister(ctx);
8918 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8919 void __user *arg, unsigned index)
8921 struct iovec __user *src;
8923 #ifdef CONFIG_COMPAT
8925 struct compat_iovec __user *ciovs;
8926 struct compat_iovec ciov;
8928 ciovs = (struct compat_iovec __user *) arg;
8929 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8932 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8933 dst->iov_len = ciov.iov_len;
8937 src = (struct iovec __user *) arg;
8938 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8944 * Not super efficient, but this is just a registration time. And we do cache
8945 * the last compound head, so generally we'll only do a full search if we don't
8948 * We check if the given compound head page has already been accounted, to
8949 * avoid double accounting it. This allows us to account the full size of the
8950 * page, not just the constituent pages of a huge page.
8952 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8953 int nr_pages, struct page *hpage)
8957 /* check current page array */
8958 for (i = 0; i < nr_pages; i++) {
8959 if (!PageCompound(pages[i]))
8961 if (compound_head(pages[i]) == hpage)
8965 /* check previously registered pages */
8966 for (i = 0; i < ctx->nr_user_bufs; i++) {
8967 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8969 for (j = 0; j < imu->nr_bvecs; j++) {
8970 if (!PageCompound(imu->bvec[j].bv_page))
8972 if (compound_head(imu->bvec[j].bv_page) == hpage)
8980 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8981 int nr_pages, struct io_mapped_ubuf *imu,
8982 struct page **last_hpage)
8986 imu->acct_pages = 0;
8987 for (i = 0; i < nr_pages; i++) {
8988 if (!PageCompound(pages[i])) {
8993 hpage = compound_head(pages[i]);
8994 if (hpage == *last_hpage)
8996 *last_hpage = hpage;
8997 if (headpage_already_acct(ctx, pages, i, hpage))
8999 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
9003 if (!imu->acct_pages)
9006 ret = io_account_mem(ctx, imu->acct_pages);
9008 imu->acct_pages = 0;
9012 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
9013 struct io_mapped_ubuf **pimu,
9014 struct page **last_hpage)
9016 struct io_mapped_ubuf *imu = NULL;
9017 struct vm_area_struct **vmas = NULL;
9018 struct page **pages = NULL;
9019 unsigned long off, start, end, ubuf;
9021 int ret, pret, nr_pages, i;
9023 if (!iov->iov_base) {
9024 *pimu = ctx->dummy_ubuf;
9028 ubuf = (unsigned long) iov->iov_base;
9029 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9030 start = ubuf >> PAGE_SHIFT;
9031 nr_pages = end - start;
9036 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9040 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9045 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9050 mmap_read_lock(current->mm);
9051 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9053 if (pret == nr_pages) {
9054 /* don't support file backed memory */
9055 for (i = 0; i < nr_pages; i++) {
9056 struct vm_area_struct *vma = vmas[i];
9058 if (vma_is_shmem(vma))
9061 !is_file_hugepages(vma->vm_file)) {
9067 ret = pret < 0 ? pret : -EFAULT;
9069 mmap_read_unlock(current->mm);
9072 * if we did partial map, or found file backed vmas,
9073 * release any pages we did get
9076 unpin_user_pages(pages, pret);
9080 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9082 unpin_user_pages(pages, pret);
9086 off = ubuf & ~PAGE_MASK;
9087 size = iov->iov_len;
9088 for (i = 0; i < nr_pages; i++) {
9091 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9092 imu->bvec[i].bv_page = pages[i];
9093 imu->bvec[i].bv_len = vec_len;
9094 imu->bvec[i].bv_offset = off;
9098 /* store original address for later verification */
9100 imu->ubuf_end = ubuf + iov->iov_len;
9101 imu->nr_bvecs = nr_pages;
9112 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9114 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9115 return ctx->user_bufs ? 0 : -ENOMEM;
9118 static int io_buffer_validate(struct iovec *iov)
9120 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9123 * Don't impose further limits on the size and buffer
9124 * constraints here, we'll -EINVAL later when IO is
9125 * submitted if they are wrong.
9128 return iov->iov_len ? -EFAULT : 0;
9132 /* arbitrary limit, but we need something */
9133 if (iov->iov_len > SZ_1G)
9136 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9142 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9143 unsigned int nr_args, u64 __user *tags)
9145 struct page *last_hpage = NULL;
9146 struct io_rsrc_data *data;
9152 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9154 ret = io_rsrc_node_switch_start(ctx);
9157 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9160 ret = io_buffers_map_alloc(ctx, nr_args);
9162 io_rsrc_data_free(data);
9166 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9167 ret = io_copy_iov(ctx, &iov, arg, i);
9170 ret = io_buffer_validate(&iov);
9173 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9178 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9184 WARN_ON_ONCE(ctx->buf_data);
9186 ctx->buf_data = data;
9188 __io_sqe_buffers_unregister(ctx);
9190 io_rsrc_node_switch(ctx, NULL);
9194 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9195 struct io_uring_rsrc_update2 *up,
9196 unsigned int nr_args)
9198 u64 __user *tags = u64_to_user_ptr(up->tags);
9199 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9200 struct page *last_hpage = NULL;
9201 bool needs_switch = false;
9207 if (up->offset + nr_args > ctx->nr_user_bufs)
9210 for (done = 0; done < nr_args; done++) {
9211 struct io_mapped_ubuf *imu;
9212 int offset = up->offset + done;
9215 err = io_copy_iov(ctx, &iov, iovs, done);
9218 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9222 err = io_buffer_validate(&iov);
9225 if (!iov.iov_base && tag) {
9229 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9233 i = array_index_nospec(offset, ctx->nr_user_bufs);
9234 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9235 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9236 ctx->rsrc_node, ctx->user_bufs[i]);
9237 if (unlikely(err)) {
9238 io_buffer_unmap(ctx, &imu);
9241 ctx->user_bufs[i] = NULL;
9242 needs_switch = true;
9245 ctx->user_bufs[i] = imu;
9246 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9250 io_rsrc_node_switch(ctx, ctx->buf_data);
9251 return done ? done : err;
9254 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9256 __s32 __user *fds = arg;
9262 if (copy_from_user(&fd, fds, sizeof(*fds)))
9265 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9266 if (IS_ERR(ctx->cq_ev_fd)) {
9267 int ret = PTR_ERR(ctx->cq_ev_fd);
9269 ctx->cq_ev_fd = NULL;
9276 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9278 if (ctx->cq_ev_fd) {
9279 eventfd_ctx_put(ctx->cq_ev_fd);
9280 ctx->cq_ev_fd = NULL;
9287 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9289 struct io_buffer *buf;
9290 unsigned long index;
9292 xa_for_each(&ctx->io_buffers, index, buf)
9293 __io_remove_buffers(ctx, buf, index, -1U);
9296 static void io_req_cache_free(struct list_head *list)
9298 struct io_kiocb *req, *nxt;
9300 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9301 list_del(&req->inflight_entry);
9302 kmem_cache_free(req_cachep, req);
9306 static void io_req_caches_free(struct io_ring_ctx *ctx)
9308 struct io_submit_state *state = &ctx->submit_state;
9310 mutex_lock(&ctx->uring_lock);
9312 if (state->free_reqs) {
9313 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9314 state->free_reqs = 0;
9317 io_flush_cached_locked_reqs(ctx, state);
9318 io_req_cache_free(&state->free_list);
9319 mutex_unlock(&ctx->uring_lock);
9322 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9324 if (data && !atomic_dec_and_test(&data->refs))
9325 wait_for_completion(&data->done);
9328 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9330 io_sq_thread_finish(ctx);
9332 if (ctx->mm_account) {
9333 mmdrop(ctx->mm_account);
9334 ctx->mm_account = NULL;
9337 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9338 io_wait_rsrc_data(ctx->buf_data);
9339 io_wait_rsrc_data(ctx->file_data);
9341 mutex_lock(&ctx->uring_lock);
9343 __io_sqe_buffers_unregister(ctx);
9345 __io_sqe_files_unregister(ctx);
9347 __io_cqring_overflow_flush(ctx, true);
9348 mutex_unlock(&ctx->uring_lock);
9349 io_eventfd_unregister(ctx);
9350 io_destroy_buffers(ctx);
9352 put_cred(ctx->sq_creds);
9354 /* there are no registered resources left, nobody uses it */
9356 io_rsrc_node_destroy(ctx->rsrc_node);
9357 if (ctx->rsrc_backup_node)
9358 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9359 flush_delayed_work(&ctx->rsrc_put_work);
9361 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9362 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9364 #if defined(CONFIG_UNIX)
9365 if (ctx->ring_sock) {
9366 ctx->ring_sock->file = NULL; /* so that iput() is called */
9367 sock_release(ctx->ring_sock);
9370 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9372 io_mem_free(ctx->rings);
9373 io_mem_free(ctx->sq_sqes);
9375 percpu_ref_exit(&ctx->refs);
9376 free_uid(ctx->user);
9377 io_req_caches_free(ctx);
9379 io_wq_put_hash(ctx->hash_map);
9380 kfree(ctx->cancel_hash);
9381 kfree(ctx->dummy_ubuf);
9385 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9387 struct io_ring_ctx *ctx = file->private_data;
9390 poll_wait(file, &ctx->poll_wait, wait);
9392 * synchronizes with barrier from wq_has_sleeper call in
9396 if (!io_sqring_full(ctx))
9397 mask |= EPOLLOUT | EPOLLWRNORM;
9400 * Don't flush cqring overflow list here, just do a simple check.
9401 * Otherwise there could possible be ABBA deadlock:
9404 * lock(&ctx->uring_lock);
9406 * lock(&ctx->uring_lock);
9409 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9410 * pushs them to do the flush.
9412 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9413 mask |= EPOLLIN | EPOLLRDNORM;
9418 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9420 const struct cred *creds;
9422 creds = xa_erase(&ctx->personalities, id);
9431 struct io_tctx_exit {
9432 struct callback_head task_work;
9433 struct completion completion;
9434 struct io_ring_ctx *ctx;
9437 static void io_tctx_exit_cb(struct callback_head *cb)
9439 struct io_uring_task *tctx = current->io_uring;
9440 struct io_tctx_exit *work;
9442 work = container_of(cb, struct io_tctx_exit, task_work);
9444 * When @in_idle, we're in cancellation and it's racy to remove the
9445 * node. It'll be removed by the end of cancellation, just ignore it.
9447 if (!atomic_read(&tctx->in_idle))
9448 io_uring_del_tctx_node((unsigned long)work->ctx);
9449 complete(&work->completion);
9452 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9454 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9456 return req->ctx == data;
9459 static void io_ring_exit_work(struct work_struct *work)
9461 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9462 unsigned long timeout = jiffies + HZ * 60 * 5;
9463 unsigned long interval = HZ / 20;
9464 struct io_tctx_exit exit;
9465 struct io_tctx_node *node;
9469 * If we're doing polled IO and end up having requests being
9470 * submitted async (out-of-line), then completions can come in while
9471 * we're waiting for refs to drop. We need to reap these manually,
9472 * as nobody else will be looking for them.
9475 io_uring_try_cancel_requests(ctx, NULL, true);
9477 struct io_sq_data *sqd = ctx->sq_data;
9478 struct task_struct *tsk;
9480 io_sq_thread_park(sqd);
9482 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9483 io_wq_cancel_cb(tsk->io_uring->io_wq,
9484 io_cancel_ctx_cb, ctx, true);
9485 io_sq_thread_unpark(sqd);
9488 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9489 /* there is little hope left, don't run it too often */
9492 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9494 init_completion(&exit.completion);
9495 init_task_work(&exit.task_work, io_tctx_exit_cb);
9498 * Some may use context even when all refs and requests have been put,
9499 * and they are free to do so while still holding uring_lock or
9500 * completion_lock, see io_req_task_submit(). Apart from other work,
9501 * this lock/unlock section also waits them to finish.
9503 mutex_lock(&ctx->uring_lock);
9504 while (!list_empty(&ctx->tctx_list)) {
9505 WARN_ON_ONCE(time_after(jiffies, timeout));
9507 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9509 /* don't spin on a single task if cancellation failed */
9510 list_rotate_left(&ctx->tctx_list);
9511 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9512 if (WARN_ON_ONCE(ret))
9514 wake_up_process(node->task);
9516 mutex_unlock(&ctx->uring_lock);
9517 wait_for_completion(&exit.completion);
9518 mutex_lock(&ctx->uring_lock);
9520 mutex_unlock(&ctx->uring_lock);
9521 spin_lock(&ctx->completion_lock);
9522 spin_unlock(&ctx->completion_lock);
9524 io_ring_ctx_free(ctx);
9527 /* Returns true if we found and killed one or more timeouts */
9528 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9531 struct io_kiocb *req, *tmp;
9534 spin_lock(&ctx->completion_lock);
9535 spin_lock_irq(&ctx->timeout_lock);
9536 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9537 if (io_match_task(req, tsk, cancel_all)) {
9538 io_kill_timeout(req, -ECANCELED);
9542 spin_unlock_irq(&ctx->timeout_lock);
9544 io_commit_cqring(ctx);
9545 spin_unlock(&ctx->completion_lock);
9547 io_cqring_ev_posted(ctx);
9548 return canceled != 0;
9551 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9553 unsigned long index;
9554 struct creds *creds;
9556 mutex_lock(&ctx->uring_lock);
9557 percpu_ref_kill(&ctx->refs);
9559 __io_cqring_overflow_flush(ctx, true);
9560 xa_for_each(&ctx->personalities, index, creds)
9561 io_unregister_personality(ctx, index);
9562 mutex_unlock(&ctx->uring_lock);
9564 io_kill_timeouts(ctx, NULL, true);
9565 io_poll_remove_all(ctx, NULL, true);
9567 /* if we failed setting up the ctx, we might not have any rings */
9568 io_iopoll_try_reap_events(ctx);
9570 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9572 * Use system_unbound_wq to avoid spawning tons of event kworkers
9573 * if we're exiting a ton of rings at the same time. It just adds
9574 * noise and overhead, there's no discernable change in runtime
9575 * over using system_wq.
9577 queue_work(system_unbound_wq, &ctx->exit_work);
9580 static int io_uring_release(struct inode *inode, struct file *file)
9582 struct io_ring_ctx *ctx = file->private_data;
9584 file->private_data = NULL;
9585 io_ring_ctx_wait_and_kill(ctx);
9589 struct io_task_cancel {
9590 struct task_struct *task;
9594 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9596 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9597 struct io_task_cancel *cancel = data;
9599 return io_match_task_safe(req, cancel->task, cancel->all);
9602 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9603 struct task_struct *task, bool cancel_all)
9605 struct io_defer_entry *de;
9608 spin_lock(&ctx->completion_lock);
9609 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9610 if (io_match_task_safe(de->req, task, cancel_all)) {
9611 list_cut_position(&list, &ctx->defer_list, &de->list);
9615 spin_unlock(&ctx->completion_lock);
9616 if (list_empty(&list))
9619 while (!list_empty(&list)) {
9620 de = list_first_entry(&list, struct io_defer_entry, list);
9621 list_del_init(&de->list);
9622 io_req_complete_failed(de->req, -ECANCELED);
9628 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9630 struct io_tctx_node *node;
9631 enum io_wq_cancel cret;
9634 mutex_lock(&ctx->uring_lock);
9635 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9636 struct io_uring_task *tctx = node->task->io_uring;
9639 * io_wq will stay alive while we hold uring_lock, because it's
9640 * killed after ctx nodes, which requires to take the lock.
9642 if (!tctx || !tctx->io_wq)
9644 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9645 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9647 mutex_unlock(&ctx->uring_lock);
9652 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9653 struct task_struct *task,
9656 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9657 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9660 enum io_wq_cancel cret;
9664 ret |= io_uring_try_cancel_iowq(ctx);
9665 } else if (tctx && tctx->io_wq) {
9667 * Cancels requests of all rings, not only @ctx, but
9668 * it's fine as the task is in exit/exec.
9670 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9672 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9675 /* SQPOLL thread does its own polling */
9676 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9677 (ctx->sq_data && ctx->sq_data->thread == current)) {
9678 while (!list_empty_careful(&ctx->iopoll_list)) {
9679 io_iopoll_try_reap_events(ctx);
9684 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9685 ret |= io_poll_remove_all(ctx, task, cancel_all);
9686 ret |= io_kill_timeouts(ctx, task, cancel_all);
9688 ret |= io_run_task_work();
9695 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9697 struct io_uring_task *tctx = current->io_uring;
9698 struct io_tctx_node *node;
9701 if (unlikely(!tctx)) {
9702 ret = io_uring_alloc_task_context(current, ctx);
9706 tctx = current->io_uring;
9707 if (ctx->iowq_limits_set) {
9708 unsigned int limits[2] = { ctx->iowq_limits[0],
9709 ctx->iowq_limits[1], };
9711 ret = io_wq_max_workers(tctx->io_wq, limits);
9716 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9717 node = kmalloc(sizeof(*node), GFP_KERNEL);
9721 node->task = current;
9723 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9730 mutex_lock(&ctx->uring_lock);
9731 list_add(&node->ctx_node, &ctx->tctx_list);
9732 mutex_unlock(&ctx->uring_lock);
9739 * Note that this task has used io_uring. We use it for cancelation purposes.
9741 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9743 struct io_uring_task *tctx = current->io_uring;
9745 if (likely(tctx && tctx->last == ctx))
9747 return __io_uring_add_tctx_node(ctx);
9751 * Remove this io_uring_file -> task mapping.
9753 static void io_uring_del_tctx_node(unsigned long index)
9755 struct io_uring_task *tctx = current->io_uring;
9756 struct io_tctx_node *node;
9760 node = xa_erase(&tctx->xa, index);
9764 WARN_ON_ONCE(current != node->task);
9765 WARN_ON_ONCE(list_empty(&node->ctx_node));
9767 mutex_lock(&node->ctx->uring_lock);
9768 list_del(&node->ctx_node);
9769 mutex_unlock(&node->ctx->uring_lock);
9771 if (tctx->last == node->ctx)
9776 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9778 struct io_wq *wq = tctx->io_wq;
9779 struct io_tctx_node *node;
9780 unsigned long index;
9782 xa_for_each(&tctx->xa, index, node) {
9783 io_uring_del_tctx_node(index);
9788 * Must be after io_uring_del_task_file() (removes nodes under
9789 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9791 io_wq_put_and_exit(wq);
9796 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9799 return atomic_read(&tctx->inflight_tracked);
9800 return percpu_counter_sum(&tctx->inflight);
9804 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9805 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9807 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9809 struct io_uring_task *tctx = current->io_uring;
9810 struct io_ring_ctx *ctx;
9814 WARN_ON_ONCE(sqd && sqd->thread != current);
9816 if (!current->io_uring)
9819 io_wq_exit_start(tctx->io_wq);
9821 atomic_inc(&tctx->in_idle);
9823 io_uring_drop_tctx_refs(current);
9824 /* read completions before cancelations */
9825 inflight = tctx_inflight(tctx, !cancel_all);
9830 struct io_tctx_node *node;
9831 unsigned long index;
9833 xa_for_each(&tctx->xa, index, node) {
9834 /* sqpoll task will cancel all its requests */
9835 if (node->ctx->sq_data)
9837 io_uring_try_cancel_requests(node->ctx, current,
9841 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9842 io_uring_try_cancel_requests(ctx, current,
9846 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9848 io_uring_drop_tctx_refs(current);
9851 * If we've seen completions, retry without waiting. This
9852 * avoids a race where a completion comes in before we did
9853 * prepare_to_wait().
9855 if (inflight == tctx_inflight(tctx, !cancel_all))
9857 finish_wait(&tctx->wait, &wait);
9860 io_uring_clean_tctx(tctx);
9863 * We shouldn't run task_works after cancel, so just leave
9864 * ->in_idle set for normal exit.
9866 atomic_dec(&tctx->in_idle);
9867 /* for exec all current's requests should be gone, kill tctx */
9868 __io_uring_free(current);
9872 void __io_uring_cancel(bool cancel_all)
9874 io_uring_cancel_generic(cancel_all, NULL);
9877 static void *io_uring_validate_mmap_request(struct file *file,
9878 loff_t pgoff, size_t sz)
9880 struct io_ring_ctx *ctx = file->private_data;
9881 loff_t offset = pgoff << PAGE_SHIFT;
9886 case IORING_OFF_SQ_RING:
9887 case IORING_OFF_CQ_RING:
9890 case IORING_OFF_SQES:
9894 return ERR_PTR(-EINVAL);
9897 page = virt_to_head_page(ptr);
9898 if (sz > page_size(page))
9899 return ERR_PTR(-EINVAL);
9906 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9908 size_t sz = vma->vm_end - vma->vm_start;
9912 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9914 return PTR_ERR(ptr);
9916 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9917 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9920 #else /* !CONFIG_MMU */
9922 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9924 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9927 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9929 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9932 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9933 unsigned long addr, unsigned long len,
9934 unsigned long pgoff, unsigned long flags)
9938 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9940 return PTR_ERR(ptr);
9942 return (unsigned long) ptr;
9945 #endif /* !CONFIG_MMU */
9947 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9952 if (!io_sqring_full(ctx))
9954 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9956 if (!io_sqring_full(ctx))
9959 } while (!signal_pending(current));
9961 finish_wait(&ctx->sqo_sq_wait, &wait);
9965 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9966 struct __kernel_timespec __user **ts,
9967 const sigset_t __user **sig)
9969 struct io_uring_getevents_arg arg;
9972 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9973 * is just a pointer to the sigset_t.
9975 if (!(flags & IORING_ENTER_EXT_ARG)) {
9976 *sig = (const sigset_t __user *) argp;
9982 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9983 * timespec and sigset_t pointers if good.
9985 if (*argsz != sizeof(arg))
9987 if (copy_from_user(&arg, argp, sizeof(arg)))
9989 *sig = u64_to_user_ptr(arg.sigmask);
9990 *argsz = arg.sigmask_sz;
9991 *ts = u64_to_user_ptr(arg.ts);
9995 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9996 u32, min_complete, u32, flags, const void __user *, argp,
9999 struct io_ring_ctx *ctx;
10004 io_run_task_work();
10006 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
10007 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
10011 if (unlikely(!f.file))
10015 if (unlikely(f.file->f_op != &io_uring_fops))
10019 ctx = f.file->private_data;
10020 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
10024 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10028 * For SQ polling, the thread will do all submissions and completions.
10029 * Just return the requested submit count, and wake the thread if
10030 * we were asked to.
10033 if (ctx->flags & IORING_SETUP_SQPOLL) {
10034 io_cqring_overflow_flush(ctx);
10036 if (unlikely(ctx->sq_data->thread == NULL)) {
10040 if (flags & IORING_ENTER_SQ_WAKEUP)
10041 wake_up(&ctx->sq_data->wait);
10042 if (flags & IORING_ENTER_SQ_WAIT) {
10043 ret = io_sqpoll_wait_sq(ctx);
10047 submitted = to_submit;
10048 } else if (to_submit) {
10049 ret = io_uring_add_tctx_node(ctx);
10052 mutex_lock(&ctx->uring_lock);
10053 submitted = io_submit_sqes(ctx, to_submit);
10054 mutex_unlock(&ctx->uring_lock);
10056 if (submitted != to_submit)
10059 if (flags & IORING_ENTER_GETEVENTS) {
10060 const sigset_t __user *sig;
10061 struct __kernel_timespec __user *ts;
10063 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10067 min_complete = min(min_complete, ctx->cq_entries);
10070 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10071 * space applications don't need to do io completion events
10072 * polling again, they can rely on io_sq_thread to do polling
10073 * work, which can reduce cpu usage and uring_lock contention.
10075 if (ctx->flags & IORING_SETUP_IOPOLL &&
10076 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10077 ret = io_iopoll_check(ctx, min_complete);
10079 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10084 percpu_ref_put(&ctx->refs);
10087 return submitted ? submitted : ret;
10090 #ifdef CONFIG_PROC_FS
10091 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10092 const struct cred *cred)
10094 struct user_namespace *uns = seq_user_ns(m);
10095 struct group_info *gi;
10100 seq_printf(m, "%5d\n", id);
10101 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10102 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10103 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10104 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10105 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10106 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10107 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10108 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10109 seq_puts(m, "\n\tGroups:\t");
10110 gi = cred->group_info;
10111 for (g = 0; g < gi->ngroups; g++) {
10112 seq_put_decimal_ull(m, g ? " " : "",
10113 from_kgid_munged(uns, gi->gid[g]));
10115 seq_puts(m, "\n\tCapEff:\t");
10116 cap = cred->cap_effective;
10117 CAP_FOR_EACH_U32(__capi)
10118 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10123 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10125 struct io_sq_data *sq = NULL;
10130 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10131 * since fdinfo case grabs it in the opposite direction of normal use
10132 * cases. If we fail to get the lock, we just don't iterate any
10133 * structures that could be going away outside the io_uring mutex.
10135 has_lock = mutex_trylock(&ctx->uring_lock);
10137 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10143 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10144 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10145 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10146 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10147 struct file *f = io_file_from_index(ctx, i);
10150 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10152 seq_printf(m, "%5u: <none>\n", i);
10154 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10155 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10156 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10157 unsigned int len = buf->ubuf_end - buf->ubuf;
10159 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10161 if (has_lock && !xa_empty(&ctx->personalities)) {
10162 unsigned long index;
10163 const struct cred *cred;
10165 seq_printf(m, "Personalities:\n");
10166 xa_for_each(&ctx->personalities, index, cred)
10167 io_uring_show_cred(m, index, cred);
10169 seq_printf(m, "PollList:\n");
10170 spin_lock(&ctx->completion_lock);
10171 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10172 struct hlist_head *list = &ctx->cancel_hash[i];
10173 struct io_kiocb *req;
10175 hlist_for_each_entry(req, list, hash_node)
10176 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10177 req->task->task_works != NULL);
10179 spin_unlock(&ctx->completion_lock);
10181 mutex_unlock(&ctx->uring_lock);
10184 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10186 struct io_ring_ctx *ctx = f->private_data;
10188 if (percpu_ref_tryget(&ctx->refs)) {
10189 __io_uring_show_fdinfo(ctx, m);
10190 percpu_ref_put(&ctx->refs);
10195 static const struct file_operations io_uring_fops = {
10196 .release = io_uring_release,
10197 .mmap = io_uring_mmap,
10199 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10200 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10202 .poll = io_uring_poll,
10203 #ifdef CONFIG_PROC_FS
10204 .show_fdinfo = io_uring_show_fdinfo,
10208 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10209 struct io_uring_params *p)
10211 struct io_rings *rings;
10212 size_t size, sq_array_offset;
10214 /* make sure these are sane, as we already accounted them */
10215 ctx->sq_entries = p->sq_entries;
10216 ctx->cq_entries = p->cq_entries;
10218 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10219 if (size == SIZE_MAX)
10222 rings = io_mem_alloc(size);
10226 ctx->rings = rings;
10227 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10228 rings->sq_ring_mask = p->sq_entries - 1;
10229 rings->cq_ring_mask = p->cq_entries - 1;
10230 rings->sq_ring_entries = p->sq_entries;
10231 rings->cq_ring_entries = p->cq_entries;
10233 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10234 if (size == SIZE_MAX) {
10235 io_mem_free(ctx->rings);
10240 ctx->sq_sqes = io_mem_alloc(size);
10241 if (!ctx->sq_sqes) {
10242 io_mem_free(ctx->rings);
10250 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10254 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10258 ret = io_uring_add_tctx_node(ctx);
10263 fd_install(fd, file);
10268 * Allocate an anonymous fd, this is what constitutes the application
10269 * visible backing of an io_uring instance. The application mmaps this
10270 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10271 * we have to tie this fd to a socket for file garbage collection purposes.
10273 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10276 #if defined(CONFIG_UNIX)
10279 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10282 return ERR_PTR(ret);
10285 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10286 O_RDWR | O_CLOEXEC);
10287 #if defined(CONFIG_UNIX)
10288 if (IS_ERR(file)) {
10289 sock_release(ctx->ring_sock);
10290 ctx->ring_sock = NULL;
10292 ctx->ring_sock->file = file;
10298 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10299 struct io_uring_params __user *params)
10301 struct io_ring_ctx *ctx;
10307 if (entries > IORING_MAX_ENTRIES) {
10308 if (!(p->flags & IORING_SETUP_CLAMP))
10310 entries = IORING_MAX_ENTRIES;
10314 * Use twice as many entries for the CQ ring. It's possible for the
10315 * application to drive a higher depth than the size of the SQ ring,
10316 * since the sqes are only used at submission time. This allows for
10317 * some flexibility in overcommitting a bit. If the application has
10318 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10319 * of CQ ring entries manually.
10321 p->sq_entries = roundup_pow_of_two(entries);
10322 if (p->flags & IORING_SETUP_CQSIZE) {
10324 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10325 * to a power-of-two, if it isn't already. We do NOT impose
10326 * any cq vs sq ring sizing.
10328 if (!p->cq_entries)
10330 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10331 if (!(p->flags & IORING_SETUP_CLAMP))
10333 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10335 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10336 if (p->cq_entries < p->sq_entries)
10339 p->cq_entries = 2 * p->sq_entries;
10342 ctx = io_ring_ctx_alloc(p);
10345 ctx->compat = in_compat_syscall();
10346 if (!capable(CAP_IPC_LOCK))
10347 ctx->user = get_uid(current_user());
10350 * This is just grabbed for accounting purposes. When a process exits,
10351 * the mm is exited and dropped before the files, hence we need to hang
10352 * on to this mm purely for the purposes of being able to unaccount
10353 * memory (locked/pinned vm). It's not used for anything else.
10355 mmgrab(current->mm);
10356 ctx->mm_account = current->mm;
10358 ret = io_allocate_scq_urings(ctx, p);
10362 ret = io_sq_offload_create(ctx, p);
10365 /* always set a rsrc node */
10366 ret = io_rsrc_node_switch_start(ctx);
10369 io_rsrc_node_switch(ctx, NULL);
10371 memset(&p->sq_off, 0, sizeof(p->sq_off));
10372 p->sq_off.head = offsetof(struct io_rings, sq.head);
10373 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10374 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10375 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10376 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10377 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10378 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10380 memset(&p->cq_off, 0, sizeof(p->cq_off));
10381 p->cq_off.head = offsetof(struct io_rings, cq.head);
10382 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10383 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10384 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10385 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10386 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10387 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10389 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10390 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10391 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10392 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10393 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10394 IORING_FEAT_RSRC_TAGS;
10396 if (copy_to_user(params, p, sizeof(*p))) {
10401 file = io_uring_get_file(ctx);
10402 if (IS_ERR(file)) {
10403 ret = PTR_ERR(file);
10408 * Install ring fd as the very last thing, so we don't risk someone
10409 * having closed it before we finish setup
10411 ret = io_uring_install_fd(ctx, file);
10413 /* fput will clean it up */
10418 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10421 io_ring_ctx_wait_and_kill(ctx);
10426 * Sets up an aio uring context, and returns the fd. Applications asks for a
10427 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10428 * params structure passed in.
10430 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10432 struct io_uring_params p;
10435 if (copy_from_user(&p, params, sizeof(p)))
10437 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10442 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10443 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10444 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10445 IORING_SETUP_R_DISABLED))
10448 return io_uring_create(entries, &p, params);
10451 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10452 struct io_uring_params __user *, params)
10454 return io_uring_setup(entries, params);
10457 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10459 struct io_uring_probe *p;
10463 size = struct_size(p, ops, nr_args);
10464 if (size == SIZE_MAX)
10466 p = kzalloc(size, GFP_KERNEL);
10471 if (copy_from_user(p, arg, size))
10474 if (memchr_inv(p, 0, size))
10477 p->last_op = IORING_OP_LAST - 1;
10478 if (nr_args > IORING_OP_LAST)
10479 nr_args = IORING_OP_LAST;
10481 for (i = 0; i < nr_args; i++) {
10483 if (!io_op_defs[i].not_supported)
10484 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10489 if (copy_to_user(arg, p, size))
10496 static int io_register_personality(struct io_ring_ctx *ctx)
10498 const struct cred *creds;
10502 creds = get_current_cred();
10504 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10505 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10513 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10514 unsigned int nr_args)
10516 struct io_uring_restriction *res;
10520 /* Restrictions allowed only if rings started disabled */
10521 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10524 /* We allow only a single restrictions registration */
10525 if (ctx->restrictions.registered)
10528 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10531 size = array_size(nr_args, sizeof(*res));
10532 if (size == SIZE_MAX)
10535 res = memdup_user(arg, size);
10537 return PTR_ERR(res);
10541 for (i = 0; i < nr_args; i++) {
10542 switch (res[i].opcode) {
10543 case IORING_RESTRICTION_REGISTER_OP:
10544 if (res[i].register_op >= IORING_REGISTER_LAST) {
10549 __set_bit(res[i].register_op,
10550 ctx->restrictions.register_op);
10552 case IORING_RESTRICTION_SQE_OP:
10553 if (res[i].sqe_op >= IORING_OP_LAST) {
10558 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10560 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10561 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10563 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10564 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10573 /* Reset all restrictions if an error happened */
10575 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10577 ctx->restrictions.registered = true;
10583 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10585 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10588 if (ctx->restrictions.registered)
10589 ctx->restricted = 1;
10591 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10592 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10593 wake_up(&ctx->sq_data->wait);
10597 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10598 struct io_uring_rsrc_update2 *up,
10606 if (check_add_overflow(up->offset, nr_args, &tmp))
10608 err = io_rsrc_node_switch_start(ctx);
10613 case IORING_RSRC_FILE:
10614 return __io_sqe_files_update(ctx, up, nr_args);
10615 case IORING_RSRC_BUFFER:
10616 return __io_sqe_buffers_update(ctx, up, nr_args);
10621 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10624 struct io_uring_rsrc_update2 up;
10628 memset(&up, 0, sizeof(up));
10629 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10631 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10634 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10635 unsigned size, unsigned type)
10637 struct io_uring_rsrc_update2 up;
10639 if (size != sizeof(up))
10641 if (copy_from_user(&up, arg, sizeof(up)))
10643 if (!up.nr || up.resv)
10645 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10648 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10649 unsigned int size, unsigned int type)
10651 struct io_uring_rsrc_register rr;
10653 /* keep it extendible */
10654 if (size != sizeof(rr))
10657 memset(&rr, 0, sizeof(rr));
10658 if (copy_from_user(&rr, arg, size))
10660 if (!rr.nr || rr.resv || rr.resv2)
10664 case IORING_RSRC_FILE:
10665 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10666 rr.nr, u64_to_user_ptr(rr.tags));
10667 case IORING_RSRC_BUFFER:
10668 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10669 rr.nr, u64_to_user_ptr(rr.tags));
10674 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10677 struct io_uring_task *tctx = current->io_uring;
10678 cpumask_var_t new_mask;
10681 if (!tctx || !tctx->io_wq)
10684 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10687 cpumask_clear(new_mask);
10688 if (len > cpumask_size())
10689 len = cpumask_size();
10691 if (copy_from_user(new_mask, arg, len)) {
10692 free_cpumask_var(new_mask);
10696 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10697 free_cpumask_var(new_mask);
10701 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10703 struct io_uring_task *tctx = current->io_uring;
10705 if (!tctx || !tctx->io_wq)
10708 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10711 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10713 __must_hold(&ctx->uring_lock)
10715 struct io_tctx_node *node;
10716 struct io_uring_task *tctx = NULL;
10717 struct io_sq_data *sqd = NULL;
10718 __u32 new_count[2];
10721 if (copy_from_user(new_count, arg, sizeof(new_count)))
10723 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10724 if (new_count[i] > INT_MAX)
10727 if (ctx->flags & IORING_SETUP_SQPOLL) {
10728 sqd = ctx->sq_data;
10731 * Observe the correct sqd->lock -> ctx->uring_lock
10732 * ordering. Fine to drop uring_lock here, we hold
10733 * a ref to the ctx.
10735 refcount_inc(&sqd->refs);
10736 mutex_unlock(&ctx->uring_lock);
10737 mutex_lock(&sqd->lock);
10738 mutex_lock(&ctx->uring_lock);
10740 tctx = sqd->thread->io_uring;
10743 tctx = current->io_uring;
10746 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10748 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10750 ctx->iowq_limits[i] = new_count[i];
10751 ctx->iowq_limits_set = true;
10754 if (tctx && tctx->io_wq) {
10755 ret = io_wq_max_workers(tctx->io_wq, new_count);
10759 memset(new_count, 0, sizeof(new_count));
10763 mutex_unlock(&sqd->lock);
10764 io_put_sq_data(sqd);
10767 if (copy_to_user(arg, new_count, sizeof(new_count)))
10770 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10774 /* now propagate the restriction to all registered users */
10775 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10776 struct io_uring_task *tctx = node->task->io_uring;
10778 if (WARN_ON_ONCE(!tctx->io_wq))
10781 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10782 new_count[i] = ctx->iowq_limits[i];
10783 /* ignore errors, it always returns zero anyway */
10784 (void)io_wq_max_workers(tctx->io_wq, new_count);
10789 mutex_unlock(&sqd->lock);
10790 io_put_sq_data(sqd);
10795 static bool io_register_op_must_quiesce(int op)
10798 case IORING_REGISTER_BUFFERS:
10799 case IORING_UNREGISTER_BUFFERS:
10800 case IORING_REGISTER_FILES:
10801 case IORING_UNREGISTER_FILES:
10802 case IORING_REGISTER_FILES_UPDATE:
10803 case IORING_REGISTER_PROBE:
10804 case IORING_REGISTER_PERSONALITY:
10805 case IORING_UNREGISTER_PERSONALITY:
10806 case IORING_REGISTER_FILES2:
10807 case IORING_REGISTER_FILES_UPDATE2:
10808 case IORING_REGISTER_BUFFERS2:
10809 case IORING_REGISTER_BUFFERS_UPDATE:
10810 case IORING_REGISTER_IOWQ_AFF:
10811 case IORING_UNREGISTER_IOWQ_AFF:
10812 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10819 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10823 percpu_ref_kill(&ctx->refs);
10826 * Drop uring mutex before waiting for references to exit. If another
10827 * thread is currently inside io_uring_enter() it might need to grab the
10828 * uring_lock to make progress. If we hold it here across the drain
10829 * wait, then we can deadlock. It's safe to drop the mutex here, since
10830 * no new references will come in after we've killed the percpu ref.
10832 mutex_unlock(&ctx->uring_lock);
10834 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10837 ret = io_run_task_work_sig();
10838 } while (ret >= 0);
10839 mutex_lock(&ctx->uring_lock);
10842 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10846 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10847 void __user *arg, unsigned nr_args)
10848 __releases(ctx->uring_lock)
10849 __acquires(ctx->uring_lock)
10854 * We're inside the ring mutex, if the ref is already dying, then
10855 * someone else killed the ctx or is already going through
10856 * io_uring_register().
10858 if (percpu_ref_is_dying(&ctx->refs))
10861 if (ctx->restricted) {
10862 if (opcode >= IORING_REGISTER_LAST)
10864 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10865 if (!test_bit(opcode, ctx->restrictions.register_op))
10869 if (io_register_op_must_quiesce(opcode)) {
10870 ret = io_ctx_quiesce(ctx);
10876 case IORING_REGISTER_BUFFERS:
10877 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10879 case IORING_UNREGISTER_BUFFERS:
10881 if (arg || nr_args)
10883 ret = io_sqe_buffers_unregister(ctx);
10885 case IORING_REGISTER_FILES:
10886 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10888 case IORING_UNREGISTER_FILES:
10890 if (arg || nr_args)
10892 ret = io_sqe_files_unregister(ctx);
10894 case IORING_REGISTER_FILES_UPDATE:
10895 ret = io_register_files_update(ctx, arg, nr_args);
10897 case IORING_REGISTER_EVENTFD:
10898 case IORING_REGISTER_EVENTFD_ASYNC:
10902 ret = io_eventfd_register(ctx, arg);
10905 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10906 ctx->eventfd_async = 1;
10908 ctx->eventfd_async = 0;
10910 case IORING_UNREGISTER_EVENTFD:
10912 if (arg || nr_args)
10914 ret = io_eventfd_unregister(ctx);
10916 case IORING_REGISTER_PROBE:
10918 if (!arg || nr_args > 256)
10920 ret = io_probe(ctx, arg, nr_args);
10922 case IORING_REGISTER_PERSONALITY:
10924 if (arg || nr_args)
10926 ret = io_register_personality(ctx);
10928 case IORING_UNREGISTER_PERSONALITY:
10932 ret = io_unregister_personality(ctx, nr_args);
10934 case IORING_REGISTER_ENABLE_RINGS:
10936 if (arg || nr_args)
10938 ret = io_register_enable_rings(ctx);
10940 case IORING_REGISTER_RESTRICTIONS:
10941 ret = io_register_restrictions(ctx, arg, nr_args);
10943 case IORING_REGISTER_FILES2:
10944 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10946 case IORING_REGISTER_FILES_UPDATE2:
10947 ret = io_register_rsrc_update(ctx, arg, nr_args,
10950 case IORING_REGISTER_BUFFERS2:
10951 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10953 case IORING_REGISTER_BUFFERS_UPDATE:
10954 ret = io_register_rsrc_update(ctx, arg, nr_args,
10955 IORING_RSRC_BUFFER);
10957 case IORING_REGISTER_IOWQ_AFF:
10959 if (!arg || !nr_args)
10961 ret = io_register_iowq_aff(ctx, arg, nr_args);
10963 case IORING_UNREGISTER_IOWQ_AFF:
10965 if (arg || nr_args)
10967 ret = io_unregister_iowq_aff(ctx);
10969 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10971 if (!arg || nr_args != 2)
10973 ret = io_register_iowq_max_workers(ctx, arg);
10980 if (io_register_op_must_quiesce(opcode)) {
10981 /* bring the ctx back to life */
10982 percpu_ref_reinit(&ctx->refs);
10983 reinit_completion(&ctx->ref_comp);
10988 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10989 void __user *, arg, unsigned int, nr_args)
10991 struct io_ring_ctx *ctx;
11000 if (f.file->f_op != &io_uring_fops)
11003 ctx = f.file->private_data;
11005 io_run_task_work();
11007 mutex_lock(&ctx->uring_lock);
11008 ret = __io_uring_register(ctx, opcode, arg, nr_args);
11009 mutex_unlock(&ctx->uring_lock);
11010 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
11011 ctx->cq_ev_fd != NULL, ret);
11017 static int __init io_uring_init(void)
11019 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11020 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11021 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11024 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11025 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11026 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11027 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11028 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11029 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11030 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11031 BUILD_BUG_SQE_ELEM(8, __u64, off);
11032 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11033 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11034 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11035 BUILD_BUG_SQE_ELEM(24, __u32, len);
11036 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11037 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11038 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11039 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11040 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11041 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11042 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11043 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11044 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11045 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11046 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11047 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11048 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11049 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11050 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11051 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11052 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11053 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11054 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11055 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11056 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11058 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11059 sizeof(struct io_uring_rsrc_update));
11060 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11061 sizeof(struct io_uring_rsrc_update2));
11063 /* ->buf_index is u16 */
11064 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11066 /* should fit into one byte */
11067 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11069 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11070 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11072 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11076 __initcall(io_uring_init);