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));
3322 if (!iov_iter_is_bvec(iter)) {
3323 iov_iter_advance(iter, nr);
3329 if (nr != iovec.iov_len)
3336 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3337 const struct iovec *fast_iov, struct iov_iter *iter)
3339 struct io_async_rw *rw = req->async_data;
3341 memcpy(&rw->iter, iter, sizeof(*iter));
3342 rw->free_iovec = iovec;
3344 /* can only be fixed buffers, no need to do anything */
3345 if (iov_iter_is_bvec(iter))
3348 unsigned iov_off = 0;
3350 rw->iter.iov = rw->fast_iov;
3351 if (iter->iov != fast_iov) {
3352 iov_off = iter->iov - fast_iov;
3353 rw->iter.iov += iov_off;
3355 if (rw->fast_iov != fast_iov)
3356 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3357 sizeof(struct iovec) * iter->nr_segs);
3359 req->flags |= REQ_F_NEED_CLEANUP;
3363 static inline int io_alloc_async_data(struct io_kiocb *req)
3365 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3366 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3367 return req->async_data == NULL;
3370 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3371 const struct iovec *fast_iov,
3372 struct iov_iter *iter, bool force)
3374 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3376 if (!req->async_data) {
3377 struct io_async_rw *iorw;
3379 if (io_alloc_async_data(req)) {
3384 io_req_map_rw(req, iovec, fast_iov, iter);
3385 iorw = req->async_data;
3386 /* we've copied and mapped the iter, ensure state is saved */
3387 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3392 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3394 struct io_async_rw *iorw = req->async_data;
3395 struct iovec *iov = iorw->fast_iov;
3398 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3399 if (unlikely(ret < 0))
3402 iorw->bytes_done = 0;
3403 iorw->free_iovec = iov;
3405 req->flags |= REQ_F_NEED_CLEANUP;
3406 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3410 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3412 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3414 return io_prep_rw(req, sqe, READ);
3418 * This is our waitqueue callback handler, registered through lock_page_async()
3419 * when we initially tried to do the IO with the iocb armed our waitqueue.
3420 * This gets called when the page is unlocked, and we generally expect that to
3421 * happen when the page IO is completed and the page is now uptodate. This will
3422 * queue a task_work based retry of the operation, attempting to copy the data
3423 * again. If the latter fails because the page was NOT uptodate, then we will
3424 * do a thread based blocking retry of the operation. That's the unexpected
3427 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3428 int sync, void *arg)
3430 struct wait_page_queue *wpq;
3431 struct io_kiocb *req = wait->private;
3432 struct wait_page_key *key = arg;
3434 wpq = container_of(wait, struct wait_page_queue, wait);
3436 if (!wake_page_match(wpq, key))
3439 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3440 list_del_init(&wait->entry);
3441 io_req_task_queue(req);
3446 * This controls whether a given IO request should be armed for async page
3447 * based retry. If we return false here, the request is handed to the async
3448 * worker threads for retry. If we're doing buffered reads on a regular file,
3449 * we prepare a private wait_page_queue entry and retry the operation. This
3450 * will either succeed because the page is now uptodate and unlocked, or it
3451 * will register a callback when the page is unlocked at IO completion. Through
3452 * that callback, io_uring uses task_work to setup a retry of the operation.
3453 * That retry will attempt the buffered read again. The retry will generally
3454 * succeed, or in rare cases where it fails, we then fall back to using the
3455 * async worker threads for a blocking retry.
3457 static bool io_rw_should_retry(struct io_kiocb *req)
3459 struct io_async_rw *rw = req->async_data;
3460 struct wait_page_queue *wait = &rw->wpq;
3461 struct kiocb *kiocb = &req->rw.kiocb;
3463 /* never retry for NOWAIT, we just complete with -EAGAIN */
3464 if (req->flags & REQ_F_NOWAIT)
3467 /* Only for buffered IO */
3468 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3472 * just use poll if we can, and don't attempt if the fs doesn't
3473 * support callback based unlocks
3475 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3478 wait->wait.func = io_async_buf_func;
3479 wait->wait.private = req;
3480 wait->wait.flags = 0;
3481 INIT_LIST_HEAD(&wait->wait.entry);
3482 kiocb->ki_flags |= IOCB_WAITQ;
3483 kiocb->ki_flags &= ~IOCB_NOWAIT;
3484 kiocb->ki_waitq = wait;
3488 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3490 if (req->file->f_op->read_iter)
3491 return call_read_iter(req->file, &req->rw.kiocb, iter);
3492 else if (req->file->f_op->read)
3493 return loop_rw_iter(READ, req, iter);
3498 static bool need_read_all(struct io_kiocb *req)
3500 return req->flags & REQ_F_ISREG ||
3501 S_ISBLK(file_inode(req->file)->i_mode);
3504 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3506 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3507 struct kiocb *kiocb = &req->rw.kiocb;
3508 struct iov_iter __iter, *iter = &__iter;
3509 struct io_async_rw *rw = req->async_data;
3510 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3511 struct iov_iter_state __state, *state;
3516 state = &rw->iter_state;
3518 * We come here from an earlier attempt, restore our state to
3519 * match in case it doesn't. It's cheap enough that we don't
3520 * need to make this conditional.
3522 iov_iter_restore(iter, state);
3525 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3529 iov_iter_save_state(iter, state);
3531 req->result = iov_iter_count(iter);
3533 /* Ensure we clear previously set non-block flag */
3534 if (!force_nonblock)
3535 kiocb->ki_flags &= ~IOCB_NOWAIT;
3537 kiocb->ki_flags |= IOCB_NOWAIT;
3539 /* If the file doesn't support async, just async punt */
3540 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3541 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3542 return ret ?: -EAGAIN;
3545 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3546 if (unlikely(ret)) {
3551 ret = io_iter_do_read(req, iter);
3553 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3554 req->flags &= ~REQ_F_REISSUE;
3555 /* IOPOLL retry should happen for io-wq threads */
3556 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3558 /* no retry on NONBLOCK nor RWF_NOWAIT */
3559 if (req->flags & REQ_F_NOWAIT)
3562 } else if (ret == -EIOCBQUEUED) {
3564 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3565 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3566 /* read all, failed, already did sync or don't want to retry */
3571 * Don't depend on the iter state matching what was consumed, or being
3572 * untouched in case of error. Restore it and we'll advance it
3573 * manually if we need to.
3575 iov_iter_restore(iter, state);
3577 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3582 rw = req->async_data;
3584 * Now use our persistent iterator and state, if we aren't already.
3585 * We've restored and mapped the iter to match.
3587 if (iter != &rw->iter) {
3589 state = &rw->iter_state;
3594 * We end up here because of a partial read, either from
3595 * above or inside this loop. Advance the iter by the bytes
3596 * that were consumed.
3598 iov_iter_advance(iter, ret);
3599 if (!iov_iter_count(iter))
3601 rw->bytes_done += ret;
3602 iov_iter_save_state(iter, state);
3604 /* if we can retry, do so with the callbacks armed */
3605 if (!io_rw_should_retry(req)) {
3606 kiocb->ki_flags &= ~IOCB_WAITQ;
3611 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3612 * we get -EIOCBQUEUED, then we'll get a notification when the
3613 * desired page gets unlocked. We can also get a partial read
3614 * here, and if we do, then just retry at the new offset.
3616 ret = io_iter_do_read(req, iter);
3617 if (ret == -EIOCBQUEUED)
3619 /* we got some bytes, but not all. retry. */
3620 kiocb->ki_flags &= ~IOCB_WAITQ;
3621 iov_iter_restore(iter, state);
3624 kiocb_done(kiocb, ret, issue_flags);
3626 /* it's faster to check here then delegate to kfree */
3632 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3634 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3636 return io_prep_rw(req, sqe, WRITE);
3639 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3641 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3642 struct kiocb *kiocb = &req->rw.kiocb;
3643 struct iov_iter __iter, *iter = &__iter;
3644 struct io_async_rw *rw = req->async_data;
3645 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3646 struct iov_iter_state __state, *state;
3651 state = &rw->iter_state;
3652 iov_iter_restore(iter, state);
3655 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3659 iov_iter_save_state(iter, state);
3661 req->result = iov_iter_count(iter);
3663 /* Ensure we clear previously set non-block flag */
3664 if (!force_nonblock)
3665 kiocb->ki_flags &= ~IOCB_NOWAIT;
3667 kiocb->ki_flags |= IOCB_NOWAIT;
3669 /* If the file doesn't support async, just async punt */
3670 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3673 /* file path doesn't support NOWAIT for non-direct_IO */
3674 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3675 (req->flags & REQ_F_ISREG))
3678 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3683 * Open-code file_start_write here to grab freeze protection,
3684 * which will be released by another thread in
3685 * io_complete_rw(). Fool lockdep by telling it the lock got
3686 * released so that it doesn't complain about the held lock when
3687 * we return to userspace.
3689 if (req->flags & REQ_F_ISREG) {
3690 sb_start_write(file_inode(req->file)->i_sb);
3691 __sb_writers_release(file_inode(req->file)->i_sb,
3694 kiocb->ki_flags |= IOCB_WRITE;
3696 if (req->file->f_op->write_iter)
3697 ret2 = call_write_iter(req->file, kiocb, iter);
3698 else if (req->file->f_op->write)
3699 ret2 = loop_rw_iter(WRITE, req, iter);
3703 if (req->flags & REQ_F_REISSUE) {
3704 req->flags &= ~REQ_F_REISSUE;
3709 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3710 * retry them without IOCB_NOWAIT.
3712 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3714 /* no retry on NONBLOCK nor RWF_NOWAIT */
3715 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3717 if (!force_nonblock || ret2 != -EAGAIN) {
3718 /* IOPOLL retry should happen for io-wq threads */
3719 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3722 kiocb_done(kiocb, ret2, issue_flags);
3725 iov_iter_restore(iter, state);
3726 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3727 return ret ?: -EAGAIN;
3730 /* it's reportedly faster than delegating the null check to kfree() */
3736 static int io_renameat_prep(struct io_kiocb *req,
3737 const struct io_uring_sqe *sqe)
3739 struct io_rename *ren = &req->rename;
3740 const char __user *oldf, *newf;
3742 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3744 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3746 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3749 ren->old_dfd = READ_ONCE(sqe->fd);
3750 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3751 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3752 ren->new_dfd = READ_ONCE(sqe->len);
3753 ren->flags = READ_ONCE(sqe->rename_flags);
3755 ren->oldpath = getname(oldf);
3756 if (IS_ERR(ren->oldpath))
3757 return PTR_ERR(ren->oldpath);
3759 ren->newpath = getname(newf);
3760 if (IS_ERR(ren->newpath)) {
3761 putname(ren->oldpath);
3762 return PTR_ERR(ren->newpath);
3765 req->flags |= REQ_F_NEED_CLEANUP;
3769 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3771 struct io_rename *ren = &req->rename;
3774 if (issue_flags & IO_URING_F_NONBLOCK)
3777 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3778 ren->newpath, ren->flags);
3780 req->flags &= ~REQ_F_NEED_CLEANUP;
3783 io_req_complete(req, ret);
3787 static int io_unlinkat_prep(struct io_kiocb *req,
3788 const struct io_uring_sqe *sqe)
3790 struct io_unlink *un = &req->unlink;
3791 const char __user *fname;
3793 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3795 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3798 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3801 un->dfd = READ_ONCE(sqe->fd);
3803 un->flags = READ_ONCE(sqe->unlink_flags);
3804 if (un->flags & ~AT_REMOVEDIR)
3807 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3808 un->filename = getname(fname);
3809 if (IS_ERR(un->filename))
3810 return PTR_ERR(un->filename);
3812 req->flags |= REQ_F_NEED_CLEANUP;
3816 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3818 struct io_unlink *un = &req->unlink;
3821 if (issue_flags & IO_URING_F_NONBLOCK)
3824 if (un->flags & AT_REMOVEDIR)
3825 ret = do_rmdir(un->dfd, un->filename);
3827 ret = do_unlinkat(un->dfd, un->filename);
3829 req->flags &= ~REQ_F_NEED_CLEANUP;
3832 io_req_complete(req, ret);
3836 static int io_mkdirat_prep(struct io_kiocb *req,
3837 const struct io_uring_sqe *sqe)
3839 struct io_mkdir *mkd = &req->mkdir;
3840 const char __user *fname;
3842 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3844 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3847 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3850 mkd->dfd = READ_ONCE(sqe->fd);
3851 mkd->mode = READ_ONCE(sqe->len);
3853 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3854 mkd->filename = getname(fname);
3855 if (IS_ERR(mkd->filename))
3856 return PTR_ERR(mkd->filename);
3858 req->flags |= REQ_F_NEED_CLEANUP;
3862 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3864 struct io_mkdir *mkd = &req->mkdir;
3867 if (issue_flags & IO_URING_F_NONBLOCK)
3870 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3872 req->flags &= ~REQ_F_NEED_CLEANUP;
3875 io_req_complete(req, ret);
3879 static int io_symlinkat_prep(struct io_kiocb *req,
3880 const struct io_uring_sqe *sqe)
3882 struct io_symlink *sl = &req->symlink;
3883 const char __user *oldpath, *newpath;
3885 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3887 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3890 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3893 sl->new_dfd = READ_ONCE(sqe->fd);
3894 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3895 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3897 sl->oldpath = getname(oldpath);
3898 if (IS_ERR(sl->oldpath))
3899 return PTR_ERR(sl->oldpath);
3901 sl->newpath = getname(newpath);
3902 if (IS_ERR(sl->newpath)) {
3903 putname(sl->oldpath);
3904 return PTR_ERR(sl->newpath);
3907 req->flags |= REQ_F_NEED_CLEANUP;
3911 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3913 struct io_symlink *sl = &req->symlink;
3916 if (issue_flags & IO_URING_F_NONBLOCK)
3919 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3921 req->flags &= ~REQ_F_NEED_CLEANUP;
3924 io_req_complete(req, ret);
3928 static int io_linkat_prep(struct io_kiocb *req,
3929 const struct io_uring_sqe *sqe)
3931 struct io_hardlink *lnk = &req->hardlink;
3932 const char __user *oldf, *newf;
3934 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3936 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3938 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3941 lnk->old_dfd = READ_ONCE(sqe->fd);
3942 lnk->new_dfd = READ_ONCE(sqe->len);
3943 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3944 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3945 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3947 lnk->oldpath = getname(oldf);
3948 if (IS_ERR(lnk->oldpath))
3949 return PTR_ERR(lnk->oldpath);
3951 lnk->newpath = getname(newf);
3952 if (IS_ERR(lnk->newpath)) {
3953 putname(lnk->oldpath);
3954 return PTR_ERR(lnk->newpath);
3957 req->flags |= REQ_F_NEED_CLEANUP;
3961 static int io_linkat(struct io_kiocb *req, int issue_flags)
3963 struct io_hardlink *lnk = &req->hardlink;
3966 if (issue_flags & IO_URING_F_NONBLOCK)
3969 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3970 lnk->newpath, lnk->flags);
3972 req->flags &= ~REQ_F_NEED_CLEANUP;
3975 io_req_complete(req, ret);
3979 static int io_shutdown_prep(struct io_kiocb *req,
3980 const struct io_uring_sqe *sqe)
3982 #if defined(CONFIG_NET)
3983 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3985 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3986 sqe->buf_index || sqe->splice_fd_in))
3989 req->shutdown.how = READ_ONCE(sqe->len);
3996 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3998 #if defined(CONFIG_NET)
3999 struct socket *sock;
4002 if (issue_flags & IO_URING_F_NONBLOCK)
4005 sock = sock_from_file(req->file);
4006 if (unlikely(!sock))
4009 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4012 io_req_complete(req, ret);
4019 static int __io_splice_prep(struct io_kiocb *req,
4020 const struct io_uring_sqe *sqe)
4022 struct io_splice *sp = &req->splice;
4023 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4025 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4029 sp->len = READ_ONCE(sqe->len);
4030 sp->flags = READ_ONCE(sqe->splice_flags);
4032 if (unlikely(sp->flags & ~valid_flags))
4035 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
4036 (sp->flags & SPLICE_F_FD_IN_FIXED));
4039 req->flags |= REQ_F_NEED_CLEANUP;
4043 static int io_tee_prep(struct io_kiocb *req,
4044 const struct io_uring_sqe *sqe)
4046 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4048 return __io_splice_prep(req, sqe);
4051 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4053 struct io_splice *sp = &req->splice;
4054 struct file *in = sp->file_in;
4055 struct file *out = sp->file_out;
4056 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4059 if (issue_flags & IO_URING_F_NONBLOCK)
4062 ret = do_tee(in, out, sp->len, flags);
4064 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4066 req->flags &= ~REQ_F_NEED_CLEANUP;
4070 io_req_complete(req, ret);
4074 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4076 struct io_splice *sp = &req->splice;
4078 sp->off_in = READ_ONCE(sqe->splice_off_in);
4079 sp->off_out = READ_ONCE(sqe->off);
4080 return __io_splice_prep(req, sqe);
4083 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4085 struct io_splice *sp = &req->splice;
4086 struct file *in = sp->file_in;
4087 struct file *out = sp->file_out;
4088 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4089 loff_t *poff_in, *poff_out;
4092 if (issue_flags & IO_URING_F_NONBLOCK)
4095 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4096 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4099 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4101 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4103 req->flags &= ~REQ_F_NEED_CLEANUP;
4107 io_req_complete(req, ret);
4112 * IORING_OP_NOP just posts a completion event, nothing else.
4114 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4116 struct io_ring_ctx *ctx = req->ctx;
4118 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4121 __io_req_complete(req, issue_flags, 0, 0);
4125 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4127 struct io_ring_ctx *ctx = req->ctx;
4132 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4134 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4138 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4139 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4142 req->sync.off = READ_ONCE(sqe->off);
4143 req->sync.len = READ_ONCE(sqe->len);
4147 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4149 loff_t end = req->sync.off + req->sync.len;
4152 /* fsync always requires a blocking context */
4153 if (issue_flags & IO_URING_F_NONBLOCK)
4156 ret = vfs_fsync_range(req->file, req->sync.off,
4157 end > 0 ? end : LLONG_MAX,
4158 req->sync.flags & IORING_FSYNC_DATASYNC);
4161 io_req_complete(req, ret);
4165 static int io_fallocate_prep(struct io_kiocb *req,
4166 const struct io_uring_sqe *sqe)
4168 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4171 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4174 req->sync.off = READ_ONCE(sqe->off);
4175 req->sync.len = READ_ONCE(sqe->addr);
4176 req->sync.mode = READ_ONCE(sqe->len);
4180 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4184 /* fallocate always requiring blocking context */
4185 if (issue_flags & IO_URING_F_NONBLOCK)
4187 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4191 io_req_complete(req, ret);
4195 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4197 const char __user *fname;
4200 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4202 if (unlikely(sqe->ioprio || sqe->buf_index))
4204 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4207 /* open.how should be already initialised */
4208 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4209 req->open.how.flags |= O_LARGEFILE;
4211 req->open.dfd = READ_ONCE(sqe->fd);
4212 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4213 req->open.filename = getname(fname);
4214 if (IS_ERR(req->open.filename)) {
4215 ret = PTR_ERR(req->open.filename);
4216 req->open.filename = NULL;
4220 req->open.file_slot = READ_ONCE(sqe->file_index);
4221 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4224 req->open.nofile = rlimit(RLIMIT_NOFILE);
4225 req->flags |= REQ_F_NEED_CLEANUP;
4229 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4231 u64 mode = READ_ONCE(sqe->len);
4232 u64 flags = READ_ONCE(sqe->open_flags);
4234 req->open.how = build_open_how(flags, mode);
4235 return __io_openat_prep(req, sqe);
4238 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4240 struct open_how __user *how;
4244 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4245 len = READ_ONCE(sqe->len);
4246 if (len < OPEN_HOW_SIZE_VER0)
4249 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4254 return __io_openat_prep(req, sqe);
4257 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4259 struct open_flags op;
4261 bool resolve_nonblock, nonblock_set;
4262 bool fixed = !!req->open.file_slot;
4265 ret = build_open_flags(&req->open.how, &op);
4268 nonblock_set = op.open_flag & O_NONBLOCK;
4269 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4270 if (issue_flags & IO_URING_F_NONBLOCK) {
4272 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4273 * it'll always -EAGAIN
4275 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4277 op.lookup_flags |= LOOKUP_CACHED;
4278 op.open_flag |= O_NONBLOCK;
4282 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4287 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4290 * We could hang on to this 'fd' on retrying, but seems like
4291 * marginal gain for something that is now known to be a slower
4292 * path. So just put it, and we'll get a new one when we retry.
4297 ret = PTR_ERR(file);
4298 /* only retry if RESOLVE_CACHED wasn't already set by application */
4299 if (ret == -EAGAIN &&
4300 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4305 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4306 file->f_flags &= ~O_NONBLOCK;
4307 fsnotify_open(file);
4310 fd_install(ret, file);
4312 ret = io_install_fixed_file(req, file, issue_flags,
4313 req->open.file_slot - 1);
4315 putname(req->open.filename);
4316 req->flags &= ~REQ_F_NEED_CLEANUP;
4319 __io_req_complete(req, issue_flags, ret, 0);
4323 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4325 return io_openat2(req, issue_flags);
4328 static int io_remove_buffers_prep(struct io_kiocb *req,
4329 const struct io_uring_sqe *sqe)
4331 struct io_provide_buf *p = &req->pbuf;
4334 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4338 tmp = READ_ONCE(sqe->fd);
4339 if (!tmp || tmp > USHRT_MAX)
4342 memset(p, 0, sizeof(*p));
4344 p->bgid = READ_ONCE(sqe->buf_group);
4348 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4349 int bgid, unsigned nbufs)
4353 /* shouldn't happen */
4357 /* the head kbuf is the list itself */
4358 while (!list_empty(&buf->list)) {
4359 struct io_buffer *nxt;
4361 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4362 list_del(&nxt->list);
4370 xa_erase(&ctx->io_buffers, bgid);
4375 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4377 struct io_provide_buf *p = &req->pbuf;
4378 struct io_ring_ctx *ctx = req->ctx;
4379 struct io_buffer *head;
4381 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4383 io_ring_submit_lock(ctx, !force_nonblock);
4385 lockdep_assert_held(&ctx->uring_lock);
4388 head = xa_load(&ctx->io_buffers, p->bgid);
4390 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4394 /* complete before unlock, IOPOLL may need the lock */
4395 __io_req_complete(req, issue_flags, ret, 0);
4396 io_ring_submit_unlock(ctx, !force_nonblock);
4400 static int io_provide_buffers_prep(struct io_kiocb *req,
4401 const struct io_uring_sqe *sqe)
4403 unsigned long size, tmp_check;
4404 struct io_provide_buf *p = &req->pbuf;
4407 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4410 tmp = READ_ONCE(sqe->fd);
4411 if (!tmp || tmp > USHRT_MAX)
4414 p->addr = READ_ONCE(sqe->addr);
4415 p->len = READ_ONCE(sqe->len);
4417 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4420 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4423 size = (unsigned long)p->len * p->nbufs;
4424 if (!access_ok(u64_to_user_ptr(p->addr), size))
4427 p->bgid = READ_ONCE(sqe->buf_group);
4428 tmp = READ_ONCE(sqe->off);
4429 if (tmp > USHRT_MAX)
4435 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4437 struct io_buffer *buf;
4438 u64 addr = pbuf->addr;
4439 int i, bid = pbuf->bid;
4441 for (i = 0; i < pbuf->nbufs; i++) {
4442 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4447 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4452 INIT_LIST_HEAD(&buf->list);
4455 list_add_tail(&buf->list, &(*head)->list);
4460 return i ? i : -ENOMEM;
4463 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4465 struct io_provide_buf *p = &req->pbuf;
4466 struct io_ring_ctx *ctx = req->ctx;
4467 struct io_buffer *head, *list;
4469 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4471 io_ring_submit_lock(ctx, !force_nonblock);
4473 lockdep_assert_held(&ctx->uring_lock);
4475 list = head = xa_load(&ctx->io_buffers, p->bgid);
4477 ret = io_add_buffers(p, &head);
4478 if (ret >= 0 && !list) {
4479 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4481 __io_remove_buffers(ctx, head, p->bgid, -1U);
4485 /* complete before unlock, IOPOLL may need the lock */
4486 __io_req_complete(req, issue_flags, ret, 0);
4487 io_ring_submit_unlock(ctx, !force_nonblock);
4491 static int io_epoll_ctl_prep(struct io_kiocb *req,
4492 const struct io_uring_sqe *sqe)
4494 #if defined(CONFIG_EPOLL)
4495 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4497 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4500 req->epoll.epfd = READ_ONCE(sqe->fd);
4501 req->epoll.op = READ_ONCE(sqe->len);
4502 req->epoll.fd = READ_ONCE(sqe->off);
4504 if (ep_op_has_event(req->epoll.op)) {
4505 struct epoll_event __user *ev;
4507 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4508 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4518 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4520 #if defined(CONFIG_EPOLL)
4521 struct io_epoll *ie = &req->epoll;
4523 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4525 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4526 if (force_nonblock && ret == -EAGAIN)
4531 __io_req_complete(req, issue_flags, ret, 0);
4538 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4540 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4541 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4546 req->madvise.addr = READ_ONCE(sqe->addr);
4547 req->madvise.len = READ_ONCE(sqe->len);
4548 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4555 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4557 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4558 struct io_madvise *ma = &req->madvise;
4561 if (issue_flags & IO_URING_F_NONBLOCK)
4564 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4567 io_req_complete(req, ret);
4574 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4576 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4578 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4581 req->fadvise.offset = READ_ONCE(sqe->off);
4582 req->fadvise.len = READ_ONCE(sqe->len);
4583 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4587 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4589 struct io_fadvise *fa = &req->fadvise;
4592 if (issue_flags & IO_URING_F_NONBLOCK) {
4593 switch (fa->advice) {
4594 case POSIX_FADV_NORMAL:
4595 case POSIX_FADV_RANDOM:
4596 case POSIX_FADV_SEQUENTIAL:
4603 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4606 __io_req_complete(req, issue_flags, ret, 0);
4610 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4612 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4614 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4616 if (req->flags & REQ_F_FIXED_FILE)
4619 req->statx.dfd = READ_ONCE(sqe->fd);
4620 req->statx.mask = READ_ONCE(sqe->len);
4621 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4622 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4623 req->statx.flags = READ_ONCE(sqe->statx_flags);
4628 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4630 struct io_statx *ctx = &req->statx;
4633 if (issue_flags & IO_URING_F_NONBLOCK)
4636 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4641 io_req_complete(req, ret);
4645 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4647 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4649 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4650 sqe->rw_flags || sqe->buf_index)
4652 if (req->flags & REQ_F_FIXED_FILE)
4655 req->close.fd = READ_ONCE(sqe->fd);
4656 req->close.file_slot = READ_ONCE(sqe->file_index);
4657 if (req->close.file_slot && req->close.fd)
4663 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4665 struct files_struct *files = current->files;
4666 struct io_close *close = &req->close;
4667 struct fdtable *fdt;
4668 struct file *file = NULL;
4671 if (req->close.file_slot) {
4672 ret = io_close_fixed(req, issue_flags);
4676 spin_lock(&files->file_lock);
4677 fdt = files_fdtable(files);
4678 if (close->fd >= fdt->max_fds) {
4679 spin_unlock(&files->file_lock);
4682 file = fdt->fd[close->fd];
4683 if (!file || file->f_op == &io_uring_fops) {
4684 spin_unlock(&files->file_lock);
4689 /* if the file has a flush method, be safe and punt to async */
4690 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4691 spin_unlock(&files->file_lock);
4695 ret = __close_fd_get_file(close->fd, &file);
4696 spin_unlock(&files->file_lock);
4703 /* No ->flush() or already async, safely close from here */
4704 ret = filp_close(file, current->files);
4710 __io_req_complete(req, issue_flags, ret, 0);
4714 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4716 struct io_ring_ctx *ctx = req->ctx;
4718 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4720 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4724 req->sync.off = READ_ONCE(sqe->off);
4725 req->sync.len = READ_ONCE(sqe->len);
4726 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4730 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4734 /* sync_file_range always requires a blocking context */
4735 if (issue_flags & IO_URING_F_NONBLOCK)
4738 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4742 io_req_complete(req, ret);
4746 #if defined(CONFIG_NET)
4747 static int io_setup_async_msg(struct io_kiocb *req,
4748 struct io_async_msghdr *kmsg)
4750 struct io_async_msghdr *async_msg = req->async_data;
4754 if (io_alloc_async_data(req)) {
4755 kfree(kmsg->free_iov);
4758 async_msg = req->async_data;
4759 req->flags |= REQ_F_NEED_CLEANUP;
4760 memcpy(async_msg, kmsg, sizeof(*kmsg));
4761 async_msg->msg.msg_name = &async_msg->addr;
4762 /* if were using fast_iov, set it to the new one */
4763 if (!async_msg->free_iov)
4764 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4769 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4770 struct io_async_msghdr *iomsg)
4772 iomsg->msg.msg_name = &iomsg->addr;
4773 iomsg->free_iov = iomsg->fast_iov;
4774 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4775 req->sr_msg.msg_flags, &iomsg->free_iov);
4778 static int io_sendmsg_prep_async(struct io_kiocb *req)
4782 ret = io_sendmsg_copy_hdr(req, req->async_data);
4784 req->flags |= REQ_F_NEED_CLEANUP;
4788 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4790 struct io_sr_msg *sr = &req->sr_msg;
4792 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4795 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4796 sr->len = READ_ONCE(sqe->len);
4797 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4798 if (sr->msg_flags & MSG_DONTWAIT)
4799 req->flags |= REQ_F_NOWAIT;
4801 #ifdef CONFIG_COMPAT
4802 if (req->ctx->compat)
4803 sr->msg_flags |= MSG_CMSG_COMPAT;
4808 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4810 struct io_async_msghdr iomsg, *kmsg;
4811 struct socket *sock;
4816 sock = sock_from_file(req->file);
4817 if (unlikely(!sock))
4820 kmsg = req->async_data;
4822 ret = io_sendmsg_copy_hdr(req, &iomsg);
4828 flags = req->sr_msg.msg_flags;
4829 if (issue_flags & IO_URING_F_NONBLOCK)
4830 flags |= MSG_DONTWAIT;
4831 if (flags & MSG_WAITALL)
4832 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4834 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4835 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4836 return io_setup_async_msg(req, kmsg);
4837 if (ret == -ERESTARTSYS)
4840 /* fast path, check for non-NULL to avoid function call */
4842 kfree(kmsg->free_iov);
4843 req->flags &= ~REQ_F_NEED_CLEANUP;
4846 __io_req_complete(req, issue_flags, ret, 0);
4850 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4852 struct io_sr_msg *sr = &req->sr_msg;
4855 struct socket *sock;
4860 sock = sock_from_file(req->file);
4861 if (unlikely(!sock))
4864 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4868 msg.msg_name = NULL;
4869 msg.msg_control = NULL;
4870 msg.msg_controllen = 0;
4871 msg.msg_namelen = 0;
4873 flags = req->sr_msg.msg_flags;
4874 if (issue_flags & IO_URING_F_NONBLOCK)
4875 flags |= MSG_DONTWAIT;
4876 if (flags & MSG_WAITALL)
4877 min_ret = iov_iter_count(&msg.msg_iter);
4879 msg.msg_flags = flags;
4880 ret = sock_sendmsg(sock, &msg);
4881 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4883 if (ret == -ERESTARTSYS)
4888 __io_req_complete(req, issue_flags, ret, 0);
4892 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4893 struct io_async_msghdr *iomsg)
4895 struct io_sr_msg *sr = &req->sr_msg;
4896 struct iovec __user *uiov;
4900 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4901 &iomsg->uaddr, &uiov, &iov_len);
4905 if (req->flags & REQ_F_BUFFER_SELECT) {
4908 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4910 sr->len = iomsg->fast_iov[0].iov_len;
4911 iomsg->free_iov = NULL;
4913 iomsg->free_iov = iomsg->fast_iov;
4914 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4915 &iomsg->free_iov, &iomsg->msg.msg_iter,
4924 #ifdef CONFIG_COMPAT
4925 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4926 struct io_async_msghdr *iomsg)
4928 struct io_sr_msg *sr = &req->sr_msg;
4929 struct compat_iovec __user *uiov;
4934 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4939 uiov = compat_ptr(ptr);
4940 if (req->flags & REQ_F_BUFFER_SELECT) {
4941 compat_ssize_t clen;
4945 if (!access_ok(uiov, sizeof(*uiov)))
4947 if (__get_user(clen, &uiov->iov_len))
4952 iomsg->free_iov = NULL;
4954 iomsg->free_iov = iomsg->fast_iov;
4955 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4956 UIO_FASTIOV, &iomsg->free_iov,
4957 &iomsg->msg.msg_iter, true);
4966 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4967 struct io_async_msghdr *iomsg)
4969 iomsg->msg.msg_name = &iomsg->addr;
4971 #ifdef CONFIG_COMPAT
4972 if (req->ctx->compat)
4973 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4976 return __io_recvmsg_copy_hdr(req, iomsg);
4979 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4982 struct io_sr_msg *sr = &req->sr_msg;
4983 struct io_buffer *kbuf;
4985 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4990 req->flags |= REQ_F_BUFFER_SELECTED;
4994 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4996 return io_put_kbuf(req, req->sr_msg.kbuf);
4999 static int io_recvmsg_prep_async(struct io_kiocb *req)
5003 ret = io_recvmsg_copy_hdr(req, req->async_data);
5005 req->flags |= REQ_F_NEED_CLEANUP;
5009 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5011 struct io_sr_msg *sr = &req->sr_msg;
5013 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5016 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5017 sr->len = READ_ONCE(sqe->len);
5018 sr->bgid = READ_ONCE(sqe->buf_group);
5019 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5020 if (sr->msg_flags & MSG_DONTWAIT)
5021 req->flags |= REQ_F_NOWAIT;
5023 #ifdef CONFIG_COMPAT
5024 if (req->ctx->compat)
5025 sr->msg_flags |= MSG_CMSG_COMPAT;
5030 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5032 struct io_async_msghdr iomsg, *kmsg;
5033 struct socket *sock;
5034 struct io_buffer *kbuf;
5037 int ret, cflags = 0;
5038 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5040 sock = sock_from_file(req->file);
5041 if (unlikely(!sock))
5044 kmsg = req->async_data;
5046 ret = io_recvmsg_copy_hdr(req, &iomsg);
5052 if (req->flags & REQ_F_BUFFER_SELECT) {
5053 kbuf = io_recv_buffer_select(req, !force_nonblock);
5055 return PTR_ERR(kbuf);
5056 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5057 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5058 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5059 1, req->sr_msg.len);
5062 flags = req->sr_msg.msg_flags;
5064 flags |= MSG_DONTWAIT;
5065 if (flags & MSG_WAITALL)
5066 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5068 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5069 kmsg->uaddr, flags);
5070 if (force_nonblock && ret == -EAGAIN)
5071 return io_setup_async_msg(req, kmsg);
5072 if (ret == -ERESTARTSYS)
5075 if (req->flags & REQ_F_BUFFER_SELECTED)
5076 cflags = io_put_recv_kbuf(req);
5077 /* fast path, check for non-NULL to avoid function call */
5079 kfree(kmsg->free_iov);
5080 req->flags &= ~REQ_F_NEED_CLEANUP;
5081 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5083 __io_req_complete(req, issue_flags, ret, cflags);
5087 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5089 struct io_buffer *kbuf;
5090 struct io_sr_msg *sr = &req->sr_msg;
5092 void __user *buf = sr->buf;
5093 struct socket *sock;
5097 int ret, cflags = 0;
5098 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5100 sock = sock_from_file(req->file);
5101 if (unlikely(!sock))
5104 if (req->flags & REQ_F_BUFFER_SELECT) {
5105 kbuf = io_recv_buffer_select(req, !force_nonblock);
5107 return PTR_ERR(kbuf);
5108 buf = u64_to_user_ptr(kbuf->addr);
5111 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5115 msg.msg_name = NULL;
5116 msg.msg_control = NULL;
5117 msg.msg_controllen = 0;
5118 msg.msg_namelen = 0;
5119 msg.msg_iocb = NULL;
5122 flags = req->sr_msg.msg_flags;
5124 flags |= MSG_DONTWAIT;
5125 if (flags & MSG_WAITALL)
5126 min_ret = iov_iter_count(&msg.msg_iter);
5128 ret = sock_recvmsg(sock, &msg, flags);
5129 if (force_nonblock && ret == -EAGAIN)
5131 if (ret == -ERESTARTSYS)
5134 if (req->flags & REQ_F_BUFFER_SELECTED)
5135 cflags = io_put_recv_kbuf(req);
5136 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5138 __io_req_complete(req, issue_flags, ret, cflags);
5142 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5144 struct io_accept *accept = &req->accept;
5146 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5148 if (sqe->ioprio || sqe->len || sqe->buf_index)
5151 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5152 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5153 accept->flags = READ_ONCE(sqe->accept_flags);
5154 accept->nofile = rlimit(RLIMIT_NOFILE);
5156 accept->file_slot = READ_ONCE(sqe->file_index);
5157 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5158 (accept->flags & SOCK_CLOEXEC)))
5160 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5162 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5163 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5167 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5169 struct io_accept *accept = &req->accept;
5170 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5171 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5172 bool fixed = !!accept->file_slot;
5176 if (req->file->f_flags & O_NONBLOCK)
5177 req->flags |= REQ_F_NOWAIT;
5180 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5181 if (unlikely(fd < 0))
5184 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5189 ret = PTR_ERR(file);
5190 if (ret == -EAGAIN && force_nonblock)
5192 if (ret == -ERESTARTSYS)
5195 } else if (!fixed) {
5196 fd_install(fd, file);
5199 ret = io_install_fixed_file(req, file, issue_flags,
5200 accept->file_slot - 1);
5202 __io_req_complete(req, issue_flags, ret, 0);
5206 static int io_connect_prep_async(struct io_kiocb *req)
5208 struct io_async_connect *io = req->async_data;
5209 struct io_connect *conn = &req->connect;
5211 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5214 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5216 struct io_connect *conn = &req->connect;
5218 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5220 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5224 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5225 conn->addr_len = READ_ONCE(sqe->addr2);
5229 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5231 struct io_async_connect __io, *io;
5232 unsigned file_flags;
5234 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5236 if (req->async_data) {
5237 io = req->async_data;
5239 ret = move_addr_to_kernel(req->connect.addr,
5240 req->connect.addr_len,
5247 file_flags = force_nonblock ? O_NONBLOCK : 0;
5249 ret = __sys_connect_file(req->file, &io->address,
5250 req->connect.addr_len, file_flags);
5251 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5252 if (req->async_data)
5254 if (io_alloc_async_data(req)) {
5258 memcpy(req->async_data, &__io, sizeof(__io));
5261 if (ret == -ERESTARTSYS)
5266 __io_req_complete(req, issue_flags, ret, 0);
5269 #else /* !CONFIG_NET */
5270 #define IO_NETOP_FN(op) \
5271 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5273 return -EOPNOTSUPP; \
5276 #define IO_NETOP_PREP(op) \
5278 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5280 return -EOPNOTSUPP; \
5283 #define IO_NETOP_PREP_ASYNC(op) \
5285 static int io_##op##_prep_async(struct io_kiocb *req) \
5287 return -EOPNOTSUPP; \
5290 IO_NETOP_PREP_ASYNC(sendmsg);
5291 IO_NETOP_PREP_ASYNC(recvmsg);
5292 IO_NETOP_PREP_ASYNC(connect);
5293 IO_NETOP_PREP(accept);
5296 #endif /* CONFIG_NET */
5298 struct io_poll_table {
5299 struct poll_table_struct pt;
5300 struct io_kiocb *req;
5305 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5306 __poll_t mask, io_req_tw_func_t func)
5308 /* for instances that support it check for an event match first: */
5309 if (mask && !(mask & poll->events))
5312 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5314 list_del_init(&poll->wait.entry);
5317 req->io_task_work.func = func;
5320 * If this fails, then the task is exiting. When a task exits, the
5321 * work gets canceled, so just cancel this request as well instead
5322 * of executing it. We can't safely execute it anyway, as we may not
5323 * have the needed state needed for it anyway.
5325 io_req_task_work_add(req);
5329 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5330 __acquires(&req->ctx->completion_lock)
5332 struct io_ring_ctx *ctx = req->ctx;
5334 /* req->task == current here, checking PF_EXITING is safe */
5335 if (unlikely(req->task->flags & PF_EXITING))
5336 WRITE_ONCE(poll->canceled, true);
5338 if (!req->result && !READ_ONCE(poll->canceled)) {
5339 struct poll_table_struct pt = { ._key = poll->events };
5341 req->result = vfs_poll(req->file, &pt) & poll->events;
5344 spin_lock(&ctx->completion_lock);
5345 if (!req->result && !READ_ONCE(poll->canceled)) {
5346 add_wait_queue(poll->head, &poll->wait);
5353 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5355 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5356 if (req->opcode == IORING_OP_POLL_ADD)
5357 return req->async_data;
5358 return req->apoll->double_poll;
5361 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5363 if (req->opcode == IORING_OP_POLL_ADD)
5365 return &req->apoll->poll;
5368 static void io_poll_remove_double(struct io_kiocb *req)
5369 __must_hold(&req->ctx->completion_lock)
5371 struct io_poll_iocb *poll = io_poll_get_double(req);
5373 lockdep_assert_held(&req->ctx->completion_lock);
5375 if (poll && poll->head) {
5376 struct wait_queue_head *head = poll->head;
5378 spin_lock_irq(&head->lock);
5379 list_del_init(&poll->wait.entry);
5380 if (poll->wait.private)
5383 spin_unlock_irq(&head->lock);
5387 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5388 __must_hold(&req->ctx->completion_lock)
5390 struct io_ring_ctx *ctx = req->ctx;
5391 unsigned flags = IORING_CQE_F_MORE;
5394 if (READ_ONCE(req->poll.canceled)) {
5396 req->poll.events |= EPOLLONESHOT;
5398 error = mangle_poll(mask);
5400 if (req->poll.events & EPOLLONESHOT)
5402 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5403 req->poll.events |= EPOLLONESHOT;
5406 if (flags & IORING_CQE_F_MORE)
5409 return !(flags & IORING_CQE_F_MORE);
5412 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5413 __must_hold(&req->ctx->completion_lock)
5417 done = __io_poll_complete(req, mask);
5418 io_commit_cqring(req->ctx);
5422 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5424 struct io_ring_ctx *ctx = req->ctx;
5425 struct io_kiocb *nxt;
5427 if (io_poll_rewait(req, &req->poll)) {
5428 spin_unlock(&ctx->completion_lock);
5432 if (req->poll.done) {
5433 spin_unlock(&ctx->completion_lock);
5436 done = __io_poll_complete(req, req->result);
5438 io_poll_remove_double(req);
5439 hash_del(&req->hash_node);
5440 req->poll.done = true;
5443 add_wait_queue(req->poll.head, &req->poll.wait);
5445 io_commit_cqring(ctx);
5446 spin_unlock(&ctx->completion_lock);
5447 io_cqring_ev_posted(ctx);
5450 nxt = io_put_req_find_next(req);
5452 io_req_task_submit(nxt, locked);
5457 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5458 int sync, void *key)
5460 struct io_kiocb *req = wait->private;
5461 struct io_poll_iocb *poll = io_poll_get_single(req);
5462 __poll_t mask = key_to_poll(key);
5463 unsigned long flags;
5465 /* for instances that support it check for an event match first: */
5466 if (mask && !(mask & poll->events))
5468 if (!(poll->events & EPOLLONESHOT))
5469 return poll->wait.func(&poll->wait, mode, sync, key);
5471 list_del_init(&wait->entry);
5476 spin_lock_irqsave(&poll->head->lock, flags);
5477 done = list_empty(&poll->wait.entry);
5479 list_del_init(&poll->wait.entry);
5480 /* make sure double remove sees this as being gone */
5481 wait->private = NULL;
5482 spin_unlock_irqrestore(&poll->head->lock, flags);
5484 /* use wait func handler, so it matches the rq type */
5485 poll->wait.func(&poll->wait, mode, sync, key);
5492 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5493 wait_queue_func_t wake_func)
5497 poll->canceled = false;
5498 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5499 /* mask in events that we always want/need */
5500 poll->events = events | IO_POLL_UNMASK;
5501 INIT_LIST_HEAD(&poll->wait.entry);
5502 init_waitqueue_func_entry(&poll->wait, wake_func);
5505 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5506 struct wait_queue_head *head,
5507 struct io_poll_iocb **poll_ptr)
5509 struct io_kiocb *req = pt->req;
5512 * The file being polled uses multiple waitqueues for poll handling
5513 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5516 if (unlikely(pt->nr_entries)) {
5517 struct io_poll_iocb *poll_one = poll;
5519 /* double add on the same waitqueue head, ignore */
5520 if (poll_one->head == head)
5522 /* already have a 2nd entry, fail a third attempt */
5524 if ((*poll_ptr)->head == head)
5526 pt->error = -EINVAL;
5530 * Can't handle multishot for double wait for now, turn it
5531 * into one-shot mode.
5533 if (!(poll_one->events & EPOLLONESHOT))
5534 poll_one->events |= EPOLLONESHOT;
5535 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5537 pt->error = -ENOMEM;
5540 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5542 poll->wait.private = req;
5549 if (poll->events & EPOLLEXCLUSIVE)
5550 add_wait_queue_exclusive(head, &poll->wait);
5552 add_wait_queue(head, &poll->wait);
5555 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5556 struct poll_table_struct *p)
5558 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5559 struct async_poll *apoll = pt->req->apoll;
5561 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5564 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5566 struct async_poll *apoll = req->apoll;
5567 struct io_ring_ctx *ctx = req->ctx;
5569 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5571 if (io_poll_rewait(req, &apoll->poll)) {
5572 spin_unlock(&ctx->completion_lock);
5576 hash_del(&req->hash_node);
5577 io_poll_remove_double(req);
5578 apoll->poll.done = true;
5579 spin_unlock(&ctx->completion_lock);
5581 if (!READ_ONCE(apoll->poll.canceled))
5582 io_req_task_submit(req, locked);
5584 io_req_complete_failed(req, -ECANCELED);
5587 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5590 struct io_kiocb *req = wait->private;
5591 struct io_poll_iocb *poll = &req->apoll->poll;
5593 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5596 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5599 static void io_poll_req_insert(struct io_kiocb *req)
5601 struct io_ring_ctx *ctx = req->ctx;
5602 struct hlist_head *list;
5604 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5605 hlist_add_head(&req->hash_node, list);
5608 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5609 struct io_poll_iocb *poll,
5610 struct io_poll_table *ipt, __poll_t mask,
5611 wait_queue_func_t wake_func)
5612 __acquires(&ctx->completion_lock)
5614 struct io_ring_ctx *ctx = req->ctx;
5615 bool cancel = false;
5617 INIT_HLIST_NODE(&req->hash_node);
5618 io_init_poll_iocb(poll, mask, wake_func);
5619 poll->file = req->file;
5620 poll->wait.private = req;
5622 ipt->pt._key = mask;
5625 ipt->nr_entries = 0;
5627 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5628 if (unlikely(!ipt->nr_entries) && !ipt->error)
5629 ipt->error = -EINVAL;
5631 spin_lock(&ctx->completion_lock);
5632 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5633 io_poll_remove_double(req);
5634 if (likely(poll->head)) {
5635 spin_lock_irq(&poll->head->lock);
5636 if (unlikely(list_empty(&poll->wait.entry))) {
5642 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5643 list_del_init(&poll->wait.entry);
5645 WRITE_ONCE(poll->canceled, true);
5646 else if (!poll->done) /* actually waiting for an event */
5647 io_poll_req_insert(req);
5648 spin_unlock_irq(&poll->head->lock);
5660 static int io_arm_poll_handler(struct io_kiocb *req)
5662 const struct io_op_def *def = &io_op_defs[req->opcode];
5663 struct io_ring_ctx *ctx = req->ctx;
5664 struct async_poll *apoll;
5665 struct io_poll_table ipt;
5666 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5669 if (!req->file || !file_can_poll(req->file))
5670 return IO_APOLL_ABORTED;
5671 if (req->flags & REQ_F_POLLED)
5672 return IO_APOLL_ABORTED;
5673 if (!def->pollin && !def->pollout)
5674 return IO_APOLL_ABORTED;
5678 mask |= POLLIN | POLLRDNORM;
5680 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5681 if ((req->opcode == IORING_OP_RECVMSG) &&
5682 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5686 mask |= POLLOUT | POLLWRNORM;
5689 /* if we can't nonblock try, then no point in arming a poll handler */
5690 if (!io_file_supports_nowait(req, rw))
5691 return IO_APOLL_ABORTED;
5693 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5694 if (unlikely(!apoll))
5695 return IO_APOLL_ABORTED;
5696 apoll->double_poll = NULL;
5698 req->flags |= REQ_F_POLLED;
5699 ipt.pt._qproc = io_async_queue_proc;
5700 io_req_set_refcount(req);
5702 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5704 spin_unlock(&ctx->completion_lock);
5705 if (ret || ipt.error)
5706 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5708 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5709 mask, apoll->poll.events);
5713 static bool __io_poll_remove_one(struct io_kiocb *req,
5714 struct io_poll_iocb *poll, bool do_cancel)
5715 __must_hold(&req->ctx->completion_lock)
5717 bool do_complete = false;
5721 spin_lock_irq(&poll->head->lock);
5723 WRITE_ONCE(poll->canceled, true);
5724 if (!list_empty(&poll->wait.entry)) {
5725 list_del_init(&poll->wait.entry);
5728 spin_unlock_irq(&poll->head->lock);
5729 hash_del(&req->hash_node);
5733 static bool io_poll_remove_one(struct io_kiocb *req)
5734 __must_hold(&req->ctx->completion_lock)
5738 io_poll_remove_double(req);
5739 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5742 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5743 io_commit_cqring(req->ctx);
5745 io_put_req_deferred(req);
5751 * Returns true if we found and killed one or more poll requests
5753 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5756 struct hlist_node *tmp;
5757 struct io_kiocb *req;
5760 spin_lock(&ctx->completion_lock);
5761 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5762 struct hlist_head *list;
5764 list = &ctx->cancel_hash[i];
5765 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5766 if (io_match_task_safe(req, tsk, cancel_all))
5767 posted += io_poll_remove_one(req);
5770 spin_unlock(&ctx->completion_lock);
5773 io_cqring_ev_posted(ctx);
5778 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5780 __must_hold(&ctx->completion_lock)
5782 struct hlist_head *list;
5783 struct io_kiocb *req;
5785 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5786 hlist_for_each_entry(req, list, hash_node) {
5787 if (sqe_addr != req->user_data)
5789 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5796 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5798 __must_hold(&ctx->completion_lock)
5800 struct io_kiocb *req;
5802 req = io_poll_find(ctx, sqe_addr, poll_only);
5805 if (io_poll_remove_one(req))
5811 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5816 events = READ_ONCE(sqe->poll32_events);
5818 events = swahw32(events);
5820 if (!(flags & IORING_POLL_ADD_MULTI))
5821 events |= EPOLLONESHOT;
5822 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5825 static int io_poll_update_prep(struct io_kiocb *req,
5826 const struct io_uring_sqe *sqe)
5828 struct io_poll_update *upd = &req->poll_update;
5831 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5833 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5835 flags = READ_ONCE(sqe->len);
5836 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5837 IORING_POLL_ADD_MULTI))
5839 /* meaningless without update */
5840 if (flags == IORING_POLL_ADD_MULTI)
5843 upd->old_user_data = READ_ONCE(sqe->addr);
5844 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5845 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5847 upd->new_user_data = READ_ONCE(sqe->off);
5848 if (!upd->update_user_data && upd->new_user_data)
5850 if (upd->update_events)
5851 upd->events = io_poll_parse_events(sqe, flags);
5852 else if (sqe->poll32_events)
5858 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5861 struct io_kiocb *req = wait->private;
5862 struct io_poll_iocb *poll = &req->poll;
5864 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5867 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5868 struct poll_table_struct *p)
5870 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5872 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5875 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5877 struct io_poll_iocb *poll = &req->poll;
5880 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5882 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5884 flags = READ_ONCE(sqe->len);
5885 if (flags & ~IORING_POLL_ADD_MULTI)
5888 io_req_set_refcount(req);
5889 poll->events = io_poll_parse_events(sqe, flags);
5893 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5895 struct io_poll_iocb *poll = &req->poll;
5896 struct io_ring_ctx *ctx = req->ctx;
5897 struct io_poll_table ipt;
5901 ipt.pt._qproc = io_poll_queue_proc;
5903 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5906 if (mask) { /* no async, we'd stolen it */
5908 done = io_poll_complete(req, mask);
5910 spin_unlock(&ctx->completion_lock);
5913 io_cqring_ev_posted(ctx);
5920 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5922 struct io_ring_ctx *ctx = req->ctx;
5923 struct io_kiocb *preq;
5927 spin_lock(&ctx->completion_lock);
5928 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5934 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5936 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5941 * Don't allow racy completion with singleshot, as we cannot safely
5942 * update those. For multishot, if we're racing with completion, just
5943 * let completion re-add it.
5945 io_poll_remove_double(preq);
5946 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5947 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5951 /* we now have a detached poll request. reissue. */
5955 spin_unlock(&ctx->completion_lock);
5957 io_req_complete(req, ret);
5960 /* only mask one event flags, keep behavior flags */
5961 if (req->poll_update.update_events) {
5962 preq->poll.events &= ~0xffff;
5963 preq->poll.events |= req->poll_update.events & 0xffff;
5964 preq->poll.events |= IO_POLL_UNMASK;
5966 if (req->poll_update.update_user_data)
5967 preq->user_data = req->poll_update.new_user_data;
5968 spin_unlock(&ctx->completion_lock);
5970 /* complete update request, we're done with it */
5971 io_req_complete(req, ret);
5974 ret = io_poll_add(preq, issue_flags);
5977 io_req_complete(preq, ret);
5983 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5986 io_req_complete_post(req, -ETIME, 0);
5989 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5991 struct io_timeout_data *data = container_of(timer,
5992 struct io_timeout_data, timer);
5993 struct io_kiocb *req = data->req;
5994 struct io_ring_ctx *ctx = req->ctx;
5995 unsigned long flags;
5997 spin_lock_irqsave(&ctx->timeout_lock, flags);
5998 list_del_init(&req->timeout.list);
5999 atomic_set(&req->ctx->cq_timeouts,
6000 atomic_read(&req->ctx->cq_timeouts) + 1);
6001 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6003 req->io_task_work.func = io_req_task_timeout;
6004 io_req_task_work_add(req);
6005 return HRTIMER_NORESTART;
6008 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
6010 __must_hold(&ctx->timeout_lock)
6012 struct io_timeout_data *io;
6013 struct io_kiocb *req;
6016 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
6017 found = user_data == req->user_data;
6022 return ERR_PTR(-ENOENT);
6024 io = req->async_data;
6025 if (hrtimer_try_to_cancel(&io->timer) == -1)
6026 return ERR_PTR(-EALREADY);
6027 list_del_init(&req->timeout.list);
6031 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
6032 __must_hold(&ctx->completion_lock)
6033 __must_hold(&ctx->timeout_lock)
6035 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6038 return PTR_ERR(req);
6041 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
6042 io_put_req_deferred(req);
6046 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6048 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6049 case IORING_TIMEOUT_BOOTTIME:
6050 return CLOCK_BOOTTIME;
6051 case IORING_TIMEOUT_REALTIME:
6052 return CLOCK_REALTIME;
6054 /* can't happen, vetted at prep time */
6058 return CLOCK_MONOTONIC;
6062 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6063 struct timespec64 *ts, enum hrtimer_mode mode)
6064 __must_hold(&ctx->timeout_lock)
6066 struct io_timeout_data *io;
6067 struct io_kiocb *req;
6070 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6071 found = user_data == req->user_data;
6078 io = req->async_data;
6079 if (hrtimer_try_to_cancel(&io->timer) == -1)
6081 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6082 io->timer.function = io_link_timeout_fn;
6083 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6087 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6088 struct timespec64 *ts, enum hrtimer_mode mode)
6089 __must_hold(&ctx->timeout_lock)
6091 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6092 struct io_timeout_data *data;
6095 return PTR_ERR(req);
6097 req->timeout.off = 0; /* noseq */
6098 data = req->async_data;
6099 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6100 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6101 data->timer.function = io_timeout_fn;
6102 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6106 static int io_timeout_remove_prep(struct io_kiocb *req,
6107 const struct io_uring_sqe *sqe)
6109 struct io_timeout_rem *tr = &req->timeout_rem;
6111 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6113 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6115 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6118 tr->ltimeout = false;
6119 tr->addr = READ_ONCE(sqe->addr);
6120 tr->flags = READ_ONCE(sqe->timeout_flags);
6121 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6122 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6124 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6125 tr->ltimeout = true;
6126 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6128 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6130 } else if (tr->flags) {
6131 /* timeout removal doesn't support flags */
6138 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6140 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6145 * Remove or update an existing timeout command
6147 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6149 struct io_timeout_rem *tr = &req->timeout_rem;
6150 struct io_ring_ctx *ctx = req->ctx;
6153 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6154 spin_lock(&ctx->completion_lock);
6155 spin_lock_irq(&ctx->timeout_lock);
6156 ret = io_timeout_cancel(ctx, tr->addr);
6157 spin_unlock_irq(&ctx->timeout_lock);
6158 spin_unlock(&ctx->completion_lock);
6160 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6162 spin_lock_irq(&ctx->timeout_lock);
6164 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6166 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6167 spin_unlock_irq(&ctx->timeout_lock);
6172 io_req_complete_post(req, ret, 0);
6176 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6177 bool is_timeout_link)
6179 struct io_timeout_data *data;
6181 u32 off = READ_ONCE(sqe->off);
6183 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6185 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6188 if (off && is_timeout_link)
6190 flags = READ_ONCE(sqe->timeout_flags);
6191 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6193 /* more than one clock specified is invalid, obviously */
6194 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6197 INIT_LIST_HEAD(&req->timeout.list);
6198 req->timeout.off = off;
6199 if (unlikely(off && !req->ctx->off_timeout_used))
6200 req->ctx->off_timeout_used = true;
6202 if (!req->async_data && io_alloc_async_data(req))
6205 data = req->async_data;
6207 data->flags = flags;
6209 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6212 data->mode = io_translate_timeout_mode(flags);
6213 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6215 if (is_timeout_link) {
6216 struct io_submit_link *link = &req->ctx->submit_state.link;
6220 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6222 req->timeout.head = link->last;
6223 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6228 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6230 struct io_ring_ctx *ctx = req->ctx;
6231 struct io_timeout_data *data = req->async_data;
6232 struct list_head *entry;
6233 u32 tail, off = req->timeout.off;
6235 spin_lock_irq(&ctx->timeout_lock);
6238 * sqe->off holds how many events that need to occur for this
6239 * timeout event to be satisfied. If it isn't set, then this is
6240 * a pure timeout request, sequence isn't used.
6242 if (io_is_timeout_noseq(req)) {
6243 entry = ctx->timeout_list.prev;
6247 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6248 req->timeout.target_seq = tail + off;
6250 /* Update the last seq here in case io_flush_timeouts() hasn't.
6251 * This is safe because ->completion_lock is held, and submissions
6252 * and completions are never mixed in the same ->completion_lock section.
6254 ctx->cq_last_tm_flush = tail;
6257 * Insertion sort, ensuring the first entry in the list is always
6258 * the one we need first.
6260 list_for_each_prev(entry, &ctx->timeout_list) {
6261 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6264 if (io_is_timeout_noseq(nxt))
6266 /* nxt.seq is behind @tail, otherwise would've been completed */
6267 if (off >= nxt->timeout.target_seq - tail)
6271 list_add(&req->timeout.list, entry);
6272 data->timer.function = io_timeout_fn;
6273 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6274 spin_unlock_irq(&ctx->timeout_lock);
6278 struct io_cancel_data {
6279 struct io_ring_ctx *ctx;
6283 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6285 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6286 struct io_cancel_data *cd = data;
6288 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6291 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6292 struct io_ring_ctx *ctx)
6294 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6295 enum io_wq_cancel cancel_ret;
6298 if (!tctx || !tctx->io_wq)
6301 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6302 switch (cancel_ret) {
6303 case IO_WQ_CANCEL_OK:
6306 case IO_WQ_CANCEL_RUNNING:
6309 case IO_WQ_CANCEL_NOTFOUND:
6317 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6319 struct io_ring_ctx *ctx = req->ctx;
6322 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6324 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6328 spin_lock(&ctx->completion_lock);
6329 spin_lock_irq(&ctx->timeout_lock);
6330 ret = io_timeout_cancel(ctx, sqe_addr);
6331 spin_unlock_irq(&ctx->timeout_lock);
6334 ret = io_poll_cancel(ctx, sqe_addr, false);
6336 spin_unlock(&ctx->completion_lock);
6340 static int io_async_cancel_prep(struct io_kiocb *req,
6341 const struct io_uring_sqe *sqe)
6343 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6345 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6347 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6351 req->cancel.addr = READ_ONCE(sqe->addr);
6355 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6357 struct io_ring_ctx *ctx = req->ctx;
6358 u64 sqe_addr = req->cancel.addr;
6359 struct io_tctx_node *node;
6362 ret = io_try_cancel_userdata(req, sqe_addr);
6366 /* slow path, try all io-wq's */
6367 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6369 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6370 struct io_uring_task *tctx = node->task->io_uring;
6372 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6376 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6380 io_req_complete_post(req, ret, 0);
6384 static int io_rsrc_update_prep(struct io_kiocb *req,
6385 const struct io_uring_sqe *sqe)
6387 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6389 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6392 req->rsrc_update.offset = READ_ONCE(sqe->off);
6393 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6394 if (!req->rsrc_update.nr_args)
6396 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6400 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6402 struct io_ring_ctx *ctx = req->ctx;
6403 struct io_uring_rsrc_update2 up;
6406 up.offset = req->rsrc_update.offset;
6407 up.data = req->rsrc_update.arg;
6412 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6413 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6414 &up, req->rsrc_update.nr_args);
6415 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6419 __io_req_complete(req, issue_flags, ret, 0);
6423 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6425 switch (req->opcode) {
6428 case IORING_OP_READV:
6429 case IORING_OP_READ_FIXED:
6430 case IORING_OP_READ:
6431 return io_read_prep(req, sqe);
6432 case IORING_OP_WRITEV:
6433 case IORING_OP_WRITE_FIXED:
6434 case IORING_OP_WRITE:
6435 return io_write_prep(req, sqe);
6436 case IORING_OP_POLL_ADD:
6437 return io_poll_add_prep(req, sqe);
6438 case IORING_OP_POLL_REMOVE:
6439 return io_poll_update_prep(req, sqe);
6440 case IORING_OP_FSYNC:
6441 return io_fsync_prep(req, sqe);
6442 case IORING_OP_SYNC_FILE_RANGE:
6443 return io_sfr_prep(req, sqe);
6444 case IORING_OP_SENDMSG:
6445 case IORING_OP_SEND:
6446 return io_sendmsg_prep(req, sqe);
6447 case IORING_OP_RECVMSG:
6448 case IORING_OP_RECV:
6449 return io_recvmsg_prep(req, sqe);
6450 case IORING_OP_CONNECT:
6451 return io_connect_prep(req, sqe);
6452 case IORING_OP_TIMEOUT:
6453 return io_timeout_prep(req, sqe, false);
6454 case IORING_OP_TIMEOUT_REMOVE:
6455 return io_timeout_remove_prep(req, sqe);
6456 case IORING_OP_ASYNC_CANCEL:
6457 return io_async_cancel_prep(req, sqe);
6458 case IORING_OP_LINK_TIMEOUT:
6459 return io_timeout_prep(req, sqe, true);
6460 case IORING_OP_ACCEPT:
6461 return io_accept_prep(req, sqe);
6462 case IORING_OP_FALLOCATE:
6463 return io_fallocate_prep(req, sqe);
6464 case IORING_OP_OPENAT:
6465 return io_openat_prep(req, sqe);
6466 case IORING_OP_CLOSE:
6467 return io_close_prep(req, sqe);
6468 case IORING_OP_FILES_UPDATE:
6469 return io_rsrc_update_prep(req, sqe);
6470 case IORING_OP_STATX:
6471 return io_statx_prep(req, sqe);
6472 case IORING_OP_FADVISE:
6473 return io_fadvise_prep(req, sqe);
6474 case IORING_OP_MADVISE:
6475 return io_madvise_prep(req, sqe);
6476 case IORING_OP_OPENAT2:
6477 return io_openat2_prep(req, sqe);
6478 case IORING_OP_EPOLL_CTL:
6479 return io_epoll_ctl_prep(req, sqe);
6480 case IORING_OP_SPLICE:
6481 return io_splice_prep(req, sqe);
6482 case IORING_OP_PROVIDE_BUFFERS:
6483 return io_provide_buffers_prep(req, sqe);
6484 case IORING_OP_REMOVE_BUFFERS:
6485 return io_remove_buffers_prep(req, sqe);
6487 return io_tee_prep(req, sqe);
6488 case IORING_OP_SHUTDOWN:
6489 return io_shutdown_prep(req, sqe);
6490 case IORING_OP_RENAMEAT:
6491 return io_renameat_prep(req, sqe);
6492 case IORING_OP_UNLINKAT:
6493 return io_unlinkat_prep(req, sqe);
6494 case IORING_OP_MKDIRAT:
6495 return io_mkdirat_prep(req, sqe);
6496 case IORING_OP_SYMLINKAT:
6497 return io_symlinkat_prep(req, sqe);
6498 case IORING_OP_LINKAT:
6499 return io_linkat_prep(req, sqe);
6502 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6507 static int io_req_prep_async(struct io_kiocb *req)
6509 if (!io_op_defs[req->opcode].needs_async_setup)
6511 if (WARN_ON_ONCE(req->async_data))
6513 if (io_alloc_async_data(req))
6516 switch (req->opcode) {
6517 case IORING_OP_READV:
6518 return io_rw_prep_async(req, READ);
6519 case IORING_OP_WRITEV:
6520 return io_rw_prep_async(req, WRITE);
6521 case IORING_OP_SENDMSG:
6522 return io_sendmsg_prep_async(req);
6523 case IORING_OP_RECVMSG:
6524 return io_recvmsg_prep_async(req);
6525 case IORING_OP_CONNECT:
6526 return io_connect_prep_async(req);
6528 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6533 static u32 io_get_sequence(struct io_kiocb *req)
6535 u32 seq = req->ctx->cached_sq_head;
6537 /* need original cached_sq_head, but it was increased for each req */
6538 io_for_each_link(req, req)
6543 static bool io_drain_req(struct io_kiocb *req)
6545 struct io_kiocb *pos;
6546 struct io_ring_ctx *ctx = req->ctx;
6547 struct io_defer_entry *de;
6551 if (req->flags & REQ_F_FAIL) {
6552 io_req_complete_fail_submit(req);
6557 * If we need to drain a request in the middle of a link, drain the
6558 * head request and the next request/link after the current link.
6559 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6560 * maintained for every request of our link.
6562 if (ctx->drain_next) {
6563 req->flags |= REQ_F_IO_DRAIN;
6564 ctx->drain_next = false;
6566 /* not interested in head, start from the first linked */
6567 io_for_each_link(pos, req->link) {
6568 if (pos->flags & REQ_F_IO_DRAIN) {
6569 ctx->drain_next = true;
6570 req->flags |= REQ_F_IO_DRAIN;
6575 /* Still need defer if there is pending req in defer list. */
6576 spin_lock(&ctx->completion_lock);
6577 if (likely(list_empty_careful(&ctx->defer_list) &&
6578 !(req->flags & REQ_F_IO_DRAIN))) {
6579 spin_unlock(&ctx->completion_lock);
6580 ctx->drain_active = false;
6583 spin_unlock(&ctx->completion_lock);
6585 seq = io_get_sequence(req);
6586 /* Still a chance to pass the sequence check */
6587 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6590 ret = io_req_prep_async(req);
6593 io_prep_async_link(req);
6594 de = kmalloc(sizeof(*de), GFP_KERNEL);
6598 io_req_complete_failed(req, ret);
6602 spin_lock(&ctx->completion_lock);
6603 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6604 spin_unlock(&ctx->completion_lock);
6606 io_queue_async_work(req, NULL);
6610 trace_io_uring_defer(ctx, req, req->user_data);
6613 list_add_tail(&de->list, &ctx->defer_list);
6614 spin_unlock(&ctx->completion_lock);
6618 static void io_clean_op(struct io_kiocb *req)
6620 if (req->flags & REQ_F_BUFFER_SELECTED) {
6621 switch (req->opcode) {
6622 case IORING_OP_READV:
6623 case IORING_OP_READ_FIXED:
6624 case IORING_OP_READ:
6625 kfree((void *)(unsigned long)req->rw.addr);
6627 case IORING_OP_RECVMSG:
6628 case IORING_OP_RECV:
6629 kfree(req->sr_msg.kbuf);
6634 if (req->flags & REQ_F_NEED_CLEANUP) {
6635 switch (req->opcode) {
6636 case IORING_OP_READV:
6637 case IORING_OP_READ_FIXED:
6638 case IORING_OP_READ:
6639 case IORING_OP_WRITEV:
6640 case IORING_OP_WRITE_FIXED:
6641 case IORING_OP_WRITE: {
6642 struct io_async_rw *io = req->async_data;
6644 kfree(io->free_iovec);
6647 case IORING_OP_RECVMSG:
6648 case IORING_OP_SENDMSG: {
6649 struct io_async_msghdr *io = req->async_data;
6651 kfree(io->free_iov);
6654 case IORING_OP_SPLICE:
6656 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6657 io_put_file(req->splice.file_in);
6659 case IORING_OP_OPENAT:
6660 case IORING_OP_OPENAT2:
6661 if (req->open.filename)
6662 putname(req->open.filename);
6664 case IORING_OP_RENAMEAT:
6665 putname(req->rename.oldpath);
6666 putname(req->rename.newpath);
6668 case IORING_OP_UNLINKAT:
6669 putname(req->unlink.filename);
6671 case IORING_OP_MKDIRAT:
6672 putname(req->mkdir.filename);
6674 case IORING_OP_SYMLINKAT:
6675 putname(req->symlink.oldpath);
6676 putname(req->symlink.newpath);
6678 case IORING_OP_LINKAT:
6679 putname(req->hardlink.oldpath);
6680 putname(req->hardlink.newpath);
6684 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6685 kfree(req->apoll->double_poll);
6689 if (req->flags & REQ_F_INFLIGHT) {
6690 struct io_uring_task *tctx = req->task->io_uring;
6692 atomic_dec(&tctx->inflight_tracked);
6694 if (req->flags & REQ_F_CREDS)
6695 put_cred(req->creds);
6697 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6700 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6702 struct io_ring_ctx *ctx = req->ctx;
6703 const struct cred *creds = NULL;
6706 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6707 creds = override_creds(req->creds);
6709 switch (req->opcode) {
6711 ret = io_nop(req, issue_flags);
6713 case IORING_OP_READV:
6714 case IORING_OP_READ_FIXED:
6715 case IORING_OP_READ:
6716 ret = io_read(req, issue_flags);
6718 case IORING_OP_WRITEV:
6719 case IORING_OP_WRITE_FIXED:
6720 case IORING_OP_WRITE:
6721 ret = io_write(req, issue_flags);
6723 case IORING_OP_FSYNC:
6724 ret = io_fsync(req, issue_flags);
6726 case IORING_OP_POLL_ADD:
6727 ret = io_poll_add(req, issue_flags);
6729 case IORING_OP_POLL_REMOVE:
6730 ret = io_poll_update(req, issue_flags);
6732 case IORING_OP_SYNC_FILE_RANGE:
6733 ret = io_sync_file_range(req, issue_flags);
6735 case IORING_OP_SENDMSG:
6736 ret = io_sendmsg(req, issue_flags);
6738 case IORING_OP_SEND:
6739 ret = io_send(req, issue_flags);
6741 case IORING_OP_RECVMSG:
6742 ret = io_recvmsg(req, issue_flags);
6744 case IORING_OP_RECV:
6745 ret = io_recv(req, issue_flags);
6747 case IORING_OP_TIMEOUT:
6748 ret = io_timeout(req, issue_flags);
6750 case IORING_OP_TIMEOUT_REMOVE:
6751 ret = io_timeout_remove(req, issue_flags);
6753 case IORING_OP_ACCEPT:
6754 ret = io_accept(req, issue_flags);
6756 case IORING_OP_CONNECT:
6757 ret = io_connect(req, issue_flags);
6759 case IORING_OP_ASYNC_CANCEL:
6760 ret = io_async_cancel(req, issue_flags);
6762 case IORING_OP_FALLOCATE:
6763 ret = io_fallocate(req, issue_flags);
6765 case IORING_OP_OPENAT:
6766 ret = io_openat(req, issue_flags);
6768 case IORING_OP_CLOSE:
6769 ret = io_close(req, issue_flags);
6771 case IORING_OP_FILES_UPDATE:
6772 ret = io_files_update(req, issue_flags);
6774 case IORING_OP_STATX:
6775 ret = io_statx(req, issue_flags);
6777 case IORING_OP_FADVISE:
6778 ret = io_fadvise(req, issue_flags);
6780 case IORING_OP_MADVISE:
6781 ret = io_madvise(req, issue_flags);
6783 case IORING_OP_OPENAT2:
6784 ret = io_openat2(req, issue_flags);
6786 case IORING_OP_EPOLL_CTL:
6787 ret = io_epoll_ctl(req, issue_flags);
6789 case IORING_OP_SPLICE:
6790 ret = io_splice(req, issue_flags);
6792 case IORING_OP_PROVIDE_BUFFERS:
6793 ret = io_provide_buffers(req, issue_flags);
6795 case IORING_OP_REMOVE_BUFFERS:
6796 ret = io_remove_buffers(req, issue_flags);
6799 ret = io_tee(req, issue_flags);
6801 case IORING_OP_SHUTDOWN:
6802 ret = io_shutdown(req, issue_flags);
6804 case IORING_OP_RENAMEAT:
6805 ret = io_renameat(req, issue_flags);
6807 case IORING_OP_UNLINKAT:
6808 ret = io_unlinkat(req, issue_flags);
6810 case IORING_OP_MKDIRAT:
6811 ret = io_mkdirat(req, issue_flags);
6813 case IORING_OP_SYMLINKAT:
6814 ret = io_symlinkat(req, issue_flags);
6816 case IORING_OP_LINKAT:
6817 ret = io_linkat(req, issue_flags);
6825 revert_creds(creds);
6828 /* If the op doesn't have a file, we're not polling for it */
6829 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6830 io_iopoll_req_issued(req);
6835 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6837 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6839 req = io_put_req_find_next(req);
6840 return req ? &req->work : NULL;
6843 static void io_wq_submit_work(struct io_wq_work *work)
6845 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6846 struct io_kiocb *timeout;
6849 /* one will be dropped by ->io_free_work() after returning to io-wq */
6850 if (!(req->flags & REQ_F_REFCOUNT))
6851 __io_req_set_refcount(req, 2);
6855 timeout = io_prep_linked_timeout(req);
6857 io_queue_linked_timeout(timeout);
6859 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6860 if (work->flags & IO_WQ_WORK_CANCEL)
6865 ret = io_issue_sqe(req, 0);
6867 * We can get EAGAIN for polled IO even though we're
6868 * forcing a sync submission from here, since we can't
6869 * wait for request slots on the block side.
6877 /* avoid locking problems by failing it from a clean context */
6879 io_req_task_queue_fail(req, ret);
6882 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6885 return &table->files[i];
6888 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6891 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6893 return (struct file *) (slot->file_ptr & FFS_MASK);
6896 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6898 unsigned long file_ptr = (unsigned long) file;
6900 if (__io_file_supports_nowait(file, READ))
6901 file_ptr |= FFS_ASYNC_READ;
6902 if (__io_file_supports_nowait(file, WRITE))
6903 file_ptr |= FFS_ASYNC_WRITE;
6904 if (S_ISREG(file_inode(file)->i_mode))
6905 file_ptr |= FFS_ISREG;
6906 file_slot->file_ptr = file_ptr;
6909 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6910 struct io_kiocb *req, int fd)
6913 unsigned long file_ptr;
6915 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6917 fd = array_index_nospec(fd, ctx->nr_user_files);
6918 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6919 file = (struct file *) (file_ptr & FFS_MASK);
6920 file_ptr &= ~FFS_MASK;
6921 /* mask in overlapping REQ_F and FFS bits */
6922 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6923 io_req_set_rsrc_node(req);
6927 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6928 struct io_kiocb *req, int fd)
6930 struct file *file = fget(fd);
6932 trace_io_uring_file_get(ctx, fd);
6934 /* we don't allow fixed io_uring files */
6935 if (file && unlikely(file->f_op == &io_uring_fops))
6936 io_req_track_inflight(req);
6940 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6941 struct io_kiocb *req, int fd, bool fixed)
6944 return io_file_get_fixed(ctx, req, fd);
6946 return io_file_get_normal(ctx, req, fd);
6949 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6951 struct io_kiocb *prev = req->timeout.prev;
6955 if (!(req->task->flags & PF_EXITING))
6956 ret = io_try_cancel_userdata(req, prev->user_data);
6957 io_req_complete_post(req, ret ?: -ETIME, 0);
6960 io_req_complete_post(req, -ETIME, 0);
6964 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6966 struct io_timeout_data *data = container_of(timer,
6967 struct io_timeout_data, timer);
6968 struct io_kiocb *prev, *req = data->req;
6969 struct io_ring_ctx *ctx = req->ctx;
6970 unsigned long flags;
6972 spin_lock_irqsave(&ctx->timeout_lock, flags);
6973 prev = req->timeout.head;
6974 req->timeout.head = NULL;
6977 * We don't expect the list to be empty, that will only happen if we
6978 * race with the completion of the linked work.
6981 io_remove_next_linked(prev);
6982 if (!req_ref_inc_not_zero(prev))
6985 list_del(&req->timeout.list);
6986 req->timeout.prev = prev;
6987 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6989 req->io_task_work.func = io_req_task_link_timeout;
6990 io_req_task_work_add(req);
6991 return HRTIMER_NORESTART;
6994 static void io_queue_linked_timeout(struct io_kiocb *req)
6996 struct io_ring_ctx *ctx = req->ctx;
6998 spin_lock_irq(&ctx->timeout_lock);
7000 * If the back reference is NULL, then our linked request finished
7001 * before we got a chance to setup the timer
7003 if (req->timeout.head) {
7004 struct io_timeout_data *data = req->async_data;
7006 data->timer.function = io_link_timeout_fn;
7007 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
7009 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
7011 spin_unlock_irq(&ctx->timeout_lock);
7012 /* drop submission reference */
7016 static void __io_queue_sqe(struct io_kiocb *req)
7017 __must_hold(&req->ctx->uring_lock)
7019 struct io_kiocb *linked_timeout;
7023 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
7026 * We async punt it if the file wasn't marked NOWAIT, or if the file
7027 * doesn't support non-blocking read/write attempts
7030 if (req->flags & REQ_F_COMPLETE_INLINE) {
7031 struct io_ring_ctx *ctx = req->ctx;
7032 struct io_submit_state *state = &ctx->submit_state;
7034 state->compl_reqs[state->compl_nr++] = req;
7035 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
7036 io_submit_flush_completions(ctx);
7040 linked_timeout = io_prep_linked_timeout(req);
7042 io_queue_linked_timeout(linked_timeout);
7043 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
7044 linked_timeout = io_prep_linked_timeout(req);
7046 switch (io_arm_poll_handler(req)) {
7047 case IO_APOLL_READY:
7049 io_queue_linked_timeout(linked_timeout);
7051 case IO_APOLL_ABORTED:
7053 * Queued up for async execution, worker will release
7054 * submit reference when the iocb is actually submitted.
7056 io_queue_async_work(req, NULL);
7061 io_queue_linked_timeout(linked_timeout);
7063 io_req_complete_failed(req, ret);
7067 static inline void io_queue_sqe(struct io_kiocb *req)
7068 __must_hold(&req->ctx->uring_lock)
7070 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7073 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7074 __io_queue_sqe(req);
7075 } else if (req->flags & REQ_F_FAIL) {
7076 io_req_complete_fail_submit(req);
7078 int ret = io_req_prep_async(req);
7081 io_req_complete_failed(req, ret);
7083 io_queue_async_work(req, NULL);
7088 * Check SQE restrictions (opcode and flags).
7090 * Returns 'true' if SQE is allowed, 'false' otherwise.
7092 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7093 struct io_kiocb *req,
7094 unsigned int sqe_flags)
7096 if (likely(!ctx->restricted))
7099 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7102 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7103 ctx->restrictions.sqe_flags_required)
7106 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7107 ctx->restrictions.sqe_flags_required))
7113 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7114 const struct io_uring_sqe *sqe)
7115 __must_hold(&ctx->uring_lock)
7117 struct io_submit_state *state;
7118 unsigned int sqe_flags;
7119 int personality, ret = 0;
7121 /* req is partially pre-initialised, see io_preinit_req() */
7122 req->opcode = READ_ONCE(sqe->opcode);
7123 /* same numerical values with corresponding REQ_F_*, safe to copy */
7124 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7125 req->user_data = READ_ONCE(sqe->user_data);
7127 req->fixed_rsrc_refs = NULL;
7128 req->task = current;
7130 /* enforce forwards compatibility on users */
7131 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7133 if (unlikely(req->opcode >= IORING_OP_LAST))
7135 if (!io_check_restriction(ctx, req, sqe_flags))
7138 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7139 !io_op_defs[req->opcode].buffer_select)
7141 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7142 ctx->drain_active = true;
7144 personality = READ_ONCE(sqe->personality);
7146 req->creds = xa_load(&ctx->personalities, personality);
7149 get_cred(req->creds);
7150 req->flags |= REQ_F_CREDS;
7152 state = &ctx->submit_state;
7155 * Plug now if we have more than 1 IO left after this, and the target
7156 * is potentially a read/write to block based storage.
7158 if (!state->plug_started && state->ios_left > 1 &&
7159 io_op_defs[req->opcode].plug) {
7160 blk_start_plug(&state->plug);
7161 state->plug_started = true;
7164 if (io_op_defs[req->opcode].needs_file) {
7165 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7166 (sqe_flags & IOSQE_FIXED_FILE));
7167 if (unlikely(!req->file))
7175 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7176 const struct io_uring_sqe *sqe)
7177 __must_hold(&ctx->uring_lock)
7179 struct io_submit_link *link = &ctx->submit_state.link;
7182 ret = io_init_req(ctx, req, sqe);
7183 if (unlikely(ret)) {
7185 /* fail even hard links since we don't submit */
7188 * we can judge a link req is failed or cancelled by if
7189 * REQ_F_FAIL is set, but the head is an exception since
7190 * it may be set REQ_F_FAIL because of other req's failure
7191 * so let's leverage req->result to distinguish if a head
7192 * is set REQ_F_FAIL because of its failure or other req's
7193 * failure so that we can set the correct ret code for it.
7194 * init result here to avoid affecting the normal path.
7196 if (!(link->head->flags & REQ_F_FAIL))
7197 req_fail_link_node(link->head, -ECANCELED);
7198 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7200 * the current req is a normal req, we should return
7201 * error and thus break the submittion loop.
7203 io_req_complete_failed(req, ret);
7206 req_fail_link_node(req, ret);
7208 ret = io_req_prep(req, sqe);
7213 /* don't need @sqe from now on */
7214 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7216 ctx->flags & IORING_SETUP_SQPOLL);
7219 * If we already have a head request, queue this one for async
7220 * submittal once the head completes. If we don't have a head but
7221 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7222 * submitted sync once the chain is complete. If none of those
7223 * conditions are true (normal request), then just queue it.
7226 struct io_kiocb *head = link->head;
7228 if (!(req->flags & REQ_F_FAIL)) {
7229 ret = io_req_prep_async(req);
7230 if (unlikely(ret)) {
7231 req_fail_link_node(req, ret);
7232 if (!(head->flags & REQ_F_FAIL))
7233 req_fail_link_node(head, -ECANCELED);
7236 trace_io_uring_link(ctx, req, head);
7237 link->last->link = req;
7240 /* last request of a link, enqueue the link */
7241 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7246 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7258 * Batched submission is done, ensure local IO is flushed out.
7260 static void io_submit_state_end(struct io_submit_state *state,
7261 struct io_ring_ctx *ctx)
7263 if (state->link.head)
7264 io_queue_sqe(state->link.head);
7265 if (state->compl_nr)
7266 io_submit_flush_completions(ctx);
7267 if (state->plug_started)
7268 blk_finish_plug(&state->plug);
7272 * Start submission side cache.
7274 static void io_submit_state_start(struct io_submit_state *state,
7275 unsigned int max_ios)
7277 state->plug_started = false;
7278 state->ios_left = max_ios;
7279 /* set only head, no need to init link_last in advance */
7280 state->link.head = NULL;
7283 static void io_commit_sqring(struct io_ring_ctx *ctx)
7285 struct io_rings *rings = ctx->rings;
7288 * Ensure any loads from the SQEs are done at this point,
7289 * since once we write the new head, the application could
7290 * write new data to them.
7292 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7296 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7297 * that is mapped by userspace. This means that care needs to be taken to
7298 * ensure that reads are stable, as we cannot rely on userspace always
7299 * being a good citizen. If members of the sqe are validated and then later
7300 * used, it's important that those reads are done through READ_ONCE() to
7301 * prevent a re-load down the line.
7303 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7305 unsigned head, mask = ctx->sq_entries - 1;
7306 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7309 * The cached sq head (or cq tail) serves two purposes:
7311 * 1) allows us to batch the cost of updating the user visible
7313 * 2) allows the kernel side to track the head on its own, even
7314 * though the application is the one updating it.
7316 head = READ_ONCE(ctx->sq_array[sq_idx]);
7317 if (likely(head < ctx->sq_entries))
7318 return &ctx->sq_sqes[head];
7320 /* drop invalid entries */
7322 WRITE_ONCE(ctx->rings->sq_dropped,
7323 READ_ONCE(ctx->rings->sq_dropped) + 1);
7327 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7328 __must_hold(&ctx->uring_lock)
7332 /* make sure SQ entry isn't read before tail */
7333 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7334 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7336 io_get_task_refs(nr);
7338 io_submit_state_start(&ctx->submit_state, nr);
7339 while (submitted < nr) {
7340 const struct io_uring_sqe *sqe;
7341 struct io_kiocb *req;
7343 req = io_alloc_req(ctx);
7344 if (unlikely(!req)) {
7346 submitted = -EAGAIN;
7349 sqe = io_get_sqe(ctx);
7350 if (unlikely(!sqe)) {
7351 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7354 /* will complete beyond this point, count as submitted */
7356 if (io_submit_sqe(ctx, req, sqe))
7360 if (unlikely(submitted != nr)) {
7361 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7362 int unused = nr - ref_used;
7364 current->io_uring->cached_refs += unused;
7365 percpu_ref_put_many(&ctx->refs, unused);
7368 io_submit_state_end(&ctx->submit_state, ctx);
7369 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7370 io_commit_sqring(ctx);
7375 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7377 return READ_ONCE(sqd->state);
7380 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7382 /* Tell userspace we may need a wakeup call */
7383 spin_lock(&ctx->completion_lock);
7384 WRITE_ONCE(ctx->rings->sq_flags,
7385 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7386 spin_unlock(&ctx->completion_lock);
7389 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7391 spin_lock(&ctx->completion_lock);
7392 WRITE_ONCE(ctx->rings->sq_flags,
7393 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7394 spin_unlock(&ctx->completion_lock);
7397 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7399 unsigned int to_submit;
7402 to_submit = io_sqring_entries(ctx);
7403 /* if we're handling multiple rings, cap submit size for fairness */
7404 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7405 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7407 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7408 unsigned nr_events = 0;
7409 const struct cred *creds = NULL;
7411 if (ctx->sq_creds != current_cred())
7412 creds = override_creds(ctx->sq_creds);
7414 mutex_lock(&ctx->uring_lock);
7415 if (!list_empty(&ctx->iopoll_list))
7416 io_do_iopoll(ctx, &nr_events, 0);
7419 * Don't submit if refs are dying, good for io_uring_register(),
7420 * but also it is relied upon by io_ring_exit_work()
7422 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7423 !(ctx->flags & IORING_SETUP_R_DISABLED))
7424 ret = io_submit_sqes(ctx, to_submit);
7425 mutex_unlock(&ctx->uring_lock);
7427 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7428 wake_up(&ctx->sqo_sq_wait);
7430 revert_creds(creds);
7436 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7438 struct io_ring_ctx *ctx;
7439 unsigned sq_thread_idle = 0;
7441 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7442 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7443 sqd->sq_thread_idle = sq_thread_idle;
7446 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7448 bool did_sig = false;
7449 struct ksignal ksig;
7451 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7452 signal_pending(current)) {
7453 mutex_unlock(&sqd->lock);
7454 if (signal_pending(current))
7455 did_sig = get_signal(&ksig);
7457 mutex_lock(&sqd->lock);
7459 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7462 static int io_sq_thread(void *data)
7464 struct io_sq_data *sqd = data;
7465 struct io_ring_ctx *ctx;
7466 unsigned long timeout = 0;
7467 char buf[TASK_COMM_LEN];
7470 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7471 set_task_comm(current, buf);
7473 if (sqd->sq_cpu != -1)
7474 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7476 set_cpus_allowed_ptr(current, cpu_online_mask);
7477 current->flags |= PF_NO_SETAFFINITY;
7479 mutex_lock(&sqd->lock);
7481 bool cap_entries, sqt_spin = false;
7483 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7484 if (io_sqd_handle_event(sqd))
7486 timeout = jiffies + sqd->sq_thread_idle;
7489 cap_entries = !list_is_singular(&sqd->ctx_list);
7490 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7491 int ret = __io_sq_thread(ctx, cap_entries);
7493 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7496 if (io_run_task_work())
7499 if (sqt_spin || !time_after(jiffies, timeout)) {
7502 timeout = jiffies + sqd->sq_thread_idle;
7506 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7507 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7508 bool needs_sched = true;
7510 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7511 io_ring_set_wakeup_flag(ctx);
7513 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7514 !list_empty_careful(&ctx->iopoll_list)) {
7515 needs_sched = false;
7518 if (io_sqring_entries(ctx)) {
7519 needs_sched = false;
7525 mutex_unlock(&sqd->lock);
7527 mutex_lock(&sqd->lock);
7529 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7530 io_ring_clear_wakeup_flag(ctx);
7533 finish_wait(&sqd->wait, &wait);
7534 timeout = jiffies + sqd->sq_thread_idle;
7537 io_uring_cancel_generic(true, sqd);
7539 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7540 io_ring_set_wakeup_flag(ctx);
7542 mutex_unlock(&sqd->lock);
7544 complete(&sqd->exited);
7548 struct io_wait_queue {
7549 struct wait_queue_entry wq;
7550 struct io_ring_ctx *ctx;
7552 unsigned nr_timeouts;
7555 static inline bool io_should_wake(struct io_wait_queue *iowq)
7557 struct io_ring_ctx *ctx = iowq->ctx;
7558 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7561 * Wake up if we have enough events, or if a timeout occurred since we
7562 * started waiting. For timeouts, we always want to return to userspace,
7563 * regardless of event count.
7565 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7568 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7569 int wake_flags, void *key)
7571 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7575 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7576 * the task, and the next invocation will do it.
7578 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7579 return autoremove_wake_function(curr, mode, wake_flags, key);
7583 static int io_run_task_work_sig(void)
7585 if (io_run_task_work())
7587 if (!signal_pending(current))
7589 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7590 return -ERESTARTSYS;
7594 /* when returns >0, the caller should retry */
7595 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7596 struct io_wait_queue *iowq,
7601 /* make sure we run task_work before checking for signals */
7602 ret = io_run_task_work_sig();
7603 if (ret || io_should_wake(iowq))
7605 /* let the caller flush overflows, retry */
7606 if (test_bit(0, &ctx->check_cq_overflow))
7609 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
7615 * Wait until events become available, if we don't already have some. The
7616 * application must reap them itself, as they reside on the shared cq ring.
7618 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7619 const sigset_t __user *sig, size_t sigsz,
7620 struct __kernel_timespec __user *uts)
7622 struct io_wait_queue iowq;
7623 struct io_rings *rings = ctx->rings;
7624 ktime_t timeout = KTIME_MAX;
7628 io_cqring_overflow_flush(ctx);
7629 if (io_cqring_events(ctx) >= min_events)
7631 if (!io_run_task_work())
7636 struct timespec64 ts;
7638 if (get_timespec64(&ts, uts))
7640 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
7644 #ifdef CONFIG_COMPAT
7645 if (in_compat_syscall())
7646 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7650 ret = set_user_sigmask(sig, sigsz);
7656 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7657 iowq.wq.private = current;
7658 INIT_LIST_HEAD(&iowq.wq.entry);
7660 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7661 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7663 trace_io_uring_cqring_wait(ctx, min_events);
7665 /* if we can't even flush overflow, don't wait for more */
7666 if (!io_cqring_overflow_flush(ctx)) {
7670 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7671 TASK_INTERRUPTIBLE);
7672 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
7673 finish_wait(&ctx->cq_wait, &iowq.wq);
7677 restore_saved_sigmask_unless(ret == -EINTR);
7679 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7682 static void io_free_page_table(void **table, size_t size)
7684 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7686 for (i = 0; i < nr_tables; i++)
7691 static void **io_alloc_page_table(size_t size)
7693 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7694 size_t init_size = size;
7697 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7701 for (i = 0; i < nr_tables; i++) {
7702 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7704 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7706 io_free_page_table(table, init_size);
7714 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7716 percpu_ref_exit(&ref_node->refs);
7720 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7722 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7723 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7724 unsigned long flags;
7725 bool first_add = false;
7726 unsigned long delay = HZ;
7728 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7731 /* if we are mid-quiesce then do not delay */
7732 if (node->rsrc_data->quiesce)
7735 while (!list_empty(&ctx->rsrc_ref_list)) {
7736 node = list_first_entry(&ctx->rsrc_ref_list,
7737 struct io_rsrc_node, node);
7738 /* recycle ref nodes in order */
7741 list_del(&node->node);
7742 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7744 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7747 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7750 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7752 struct io_rsrc_node *ref_node;
7754 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7758 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7763 INIT_LIST_HEAD(&ref_node->node);
7764 INIT_LIST_HEAD(&ref_node->rsrc_list);
7765 ref_node->done = false;
7769 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7770 struct io_rsrc_data *data_to_kill)
7772 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7773 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7776 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7778 rsrc_node->rsrc_data = data_to_kill;
7779 spin_lock_irq(&ctx->rsrc_ref_lock);
7780 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7781 spin_unlock_irq(&ctx->rsrc_ref_lock);
7783 atomic_inc(&data_to_kill->refs);
7784 percpu_ref_kill(&rsrc_node->refs);
7785 ctx->rsrc_node = NULL;
7788 if (!ctx->rsrc_node) {
7789 ctx->rsrc_node = ctx->rsrc_backup_node;
7790 ctx->rsrc_backup_node = NULL;
7794 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7796 if (ctx->rsrc_backup_node)
7798 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7799 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7802 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7806 /* As we may drop ->uring_lock, other task may have started quiesce */
7810 data->quiesce = true;
7812 ret = io_rsrc_node_switch_start(ctx);
7815 io_rsrc_node_switch(ctx, data);
7817 /* kill initial ref, already quiesced if zero */
7818 if (atomic_dec_and_test(&data->refs))
7820 mutex_unlock(&ctx->uring_lock);
7821 flush_delayed_work(&ctx->rsrc_put_work);
7822 ret = wait_for_completion_interruptible(&data->done);
7824 mutex_lock(&ctx->uring_lock);
7825 if (atomic_read(&data->refs) > 0) {
7827 * it has been revived by another thread while
7830 mutex_unlock(&ctx->uring_lock);
7836 atomic_inc(&data->refs);
7837 /* wait for all works potentially completing data->done */
7838 flush_delayed_work(&ctx->rsrc_put_work);
7839 reinit_completion(&data->done);
7841 ret = io_run_task_work_sig();
7842 mutex_lock(&ctx->uring_lock);
7844 data->quiesce = false;
7849 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7851 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7852 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7854 return &data->tags[table_idx][off];
7857 static void io_rsrc_data_free(struct io_rsrc_data *data)
7859 size_t size = data->nr * sizeof(data->tags[0][0]);
7862 io_free_page_table((void **)data->tags, size);
7866 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7867 u64 __user *utags, unsigned nr,
7868 struct io_rsrc_data **pdata)
7870 struct io_rsrc_data *data;
7874 data = kzalloc(sizeof(*data), GFP_KERNEL);
7877 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7885 data->do_put = do_put;
7888 for (i = 0; i < nr; i++) {
7889 u64 *tag_slot = io_get_tag_slot(data, i);
7891 if (copy_from_user(tag_slot, &utags[i],
7897 atomic_set(&data->refs, 1);
7898 init_completion(&data->done);
7902 io_rsrc_data_free(data);
7906 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7908 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7909 GFP_KERNEL_ACCOUNT);
7910 return !!table->files;
7913 static void io_free_file_tables(struct io_file_table *table)
7915 kvfree(table->files);
7916 table->files = NULL;
7919 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7921 #if defined(CONFIG_UNIX)
7922 if (ctx->ring_sock) {
7923 struct sock *sock = ctx->ring_sock->sk;
7924 struct sk_buff *skb;
7926 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7932 for (i = 0; i < ctx->nr_user_files; i++) {
7935 file = io_file_from_index(ctx, i);
7940 io_free_file_tables(&ctx->file_table);
7941 io_rsrc_data_free(ctx->file_data);
7942 ctx->file_data = NULL;
7943 ctx->nr_user_files = 0;
7946 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7950 if (!ctx->file_data)
7952 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7954 __io_sqe_files_unregister(ctx);
7958 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7959 __releases(&sqd->lock)
7961 WARN_ON_ONCE(sqd->thread == current);
7964 * Do the dance but not conditional clear_bit() because it'd race with
7965 * other threads incrementing park_pending and setting the bit.
7967 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7968 if (atomic_dec_return(&sqd->park_pending))
7969 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7970 mutex_unlock(&sqd->lock);
7973 static void io_sq_thread_park(struct io_sq_data *sqd)
7974 __acquires(&sqd->lock)
7976 WARN_ON_ONCE(sqd->thread == current);
7978 atomic_inc(&sqd->park_pending);
7979 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7980 mutex_lock(&sqd->lock);
7982 wake_up_process(sqd->thread);
7985 static void io_sq_thread_stop(struct io_sq_data *sqd)
7987 WARN_ON_ONCE(sqd->thread == current);
7988 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7990 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7991 mutex_lock(&sqd->lock);
7993 wake_up_process(sqd->thread);
7994 mutex_unlock(&sqd->lock);
7995 wait_for_completion(&sqd->exited);
7998 static void io_put_sq_data(struct io_sq_data *sqd)
8000 if (refcount_dec_and_test(&sqd->refs)) {
8001 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
8003 io_sq_thread_stop(sqd);
8008 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
8010 struct io_sq_data *sqd = ctx->sq_data;
8013 io_sq_thread_park(sqd);
8014 list_del_init(&ctx->sqd_list);
8015 io_sqd_update_thread_idle(sqd);
8016 io_sq_thread_unpark(sqd);
8018 io_put_sq_data(sqd);
8019 ctx->sq_data = NULL;
8023 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
8025 struct io_ring_ctx *ctx_attach;
8026 struct io_sq_data *sqd;
8029 f = fdget(p->wq_fd);
8031 return ERR_PTR(-ENXIO);
8032 if (f.file->f_op != &io_uring_fops) {
8034 return ERR_PTR(-EINVAL);
8037 ctx_attach = f.file->private_data;
8038 sqd = ctx_attach->sq_data;
8041 return ERR_PTR(-EINVAL);
8043 if (sqd->task_tgid != current->tgid) {
8045 return ERR_PTR(-EPERM);
8048 refcount_inc(&sqd->refs);
8053 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
8056 struct io_sq_data *sqd;
8059 if (p->flags & IORING_SETUP_ATTACH_WQ) {
8060 sqd = io_attach_sq_data(p);
8065 /* fall through for EPERM case, setup new sqd/task */
8066 if (PTR_ERR(sqd) != -EPERM)
8070 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8072 return ERR_PTR(-ENOMEM);
8074 atomic_set(&sqd->park_pending, 0);
8075 refcount_set(&sqd->refs, 1);
8076 INIT_LIST_HEAD(&sqd->ctx_list);
8077 mutex_init(&sqd->lock);
8078 init_waitqueue_head(&sqd->wait);
8079 init_completion(&sqd->exited);
8083 #if defined(CONFIG_UNIX)
8085 * Ensure the UNIX gc is aware of our file set, so we are certain that
8086 * the io_uring can be safely unregistered on process exit, even if we have
8087 * loops in the file referencing.
8089 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8091 struct sock *sk = ctx->ring_sock->sk;
8092 struct scm_fp_list *fpl;
8093 struct sk_buff *skb;
8096 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8100 skb = alloc_skb(0, GFP_KERNEL);
8109 fpl->user = get_uid(current_user());
8110 for (i = 0; i < nr; i++) {
8111 struct file *file = io_file_from_index(ctx, i + offset);
8115 fpl->fp[nr_files] = get_file(file);
8116 unix_inflight(fpl->user, fpl->fp[nr_files]);
8121 fpl->max = SCM_MAX_FD;
8122 fpl->count = nr_files;
8123 UNIXCB(skb).fp = fpl;
8124 skb->destructor = unix_destruct_scm;
8125 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8126 skb_queue_head(&sk->sk_receive_queue, skb);
8128 for (i = 0; i < nr_files; i++)
8139 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8140 * causes regular reference counting to break down. We rely on the UNIX
8141 * garbage collection to take care of this problem for us.
8143 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8145 unsigned left, total;
8149 left = ctx->nr_user_files;
8151 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8153 ret = __io_sqe_files_scm(ctx, this_files, total);
8157 total += this_files;
8163 while (total < ctx->nr_user_files) {
8164 struct file *file = io_file_from_index(ctx, total);
8174 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8180 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8182 struct file *file = prsrc->file;
8183 #if defined(CONFIG_UNIX)
8184 struct sock *sock = ctx->ring_sock->sk;
8185 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8186 struct sk_buff *skb;
8189 __skb_queue_head_init(&list);
8192 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8193 * remove this entry and rearrange the file array.
8195 skb = skb_dequeue(head);
8197 struct scm_fp_list *fp;
8199 fp = UNIXCB(skb).fp;
8200 for (i = 0; i < fp->count; i++) {
8203 if (fp->fp[i] != file)
8206 unix_notinflight(fp->user, fp->fp[i]);
8207 left = fp->count - 1 - i;
8209 memmove(&fp->fp[i], &fp->fp[i + 1],
8210 left * sizeof(struct file *));
8217 __skb_queue_tail(&list, skb);
8227 __skb_queue_tail(&list, skb);
8229 skb = skb_dequeue(head);
8232 if (skb_peek(&list)) {
8233 spin_lock_irq(&head->lock);
8234 while ((skb = __skb_dequeue(&list)) != NULL)
8235 __skb_queue_tail(head, skb);
8236 spin_unlock_irq(&head->lock);
8243 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8245 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8246 struct io_ring_ctx *ctx = rsrc_data->ctx;
8247 struct io_rsrc_put *prsrc, *tmp;
8249 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8250 list_del(&prsrc->list);
8253 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8255 io_ring_submit_lock(ctx, lock_ring);
8256 spin_lock(&ctx->completion_lock);
8257 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8259 io_commit_cqring(ctx);
8260 spin_unlock(&ctx->completion_lock);
8261 io_cqring_ev_posted(ctx);
8262 io_ring_submit_unlock(ctx, lock_ring);
8265 rsrc_data->do_put(ctx, prsrc);
8269 io_rsrc_node_destroy(ref_node);
8270 if (atomic_dec_and_test(&rsrc_data->refs))
8271 complete(&rsrc_data->done);
8274 static void io_rsrc_put_work(struct work_struct *work)
8276 struct io_ring_ctx *ctx;
8277 struct llist_node *node;
8279 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8280 node = llist_del_all(&ctx->rsrc_put_llist);
8283 struct io_rsrc_node *ref_node;
8284 struct llist_node *next = node->next;
8286 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8287 __io_rsrc_put_work(ref_node);
8292 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8293 unsigned nr_args, u64 __user *tags)
8295 __s32 __user *fds = (__s32 __user *) arg;
8304 if (nr_args > IORING_MAX_FIXED_FILES)
8306 if (nr_args > rlimit(RLIMIT_NOFILE))
8308 ret = io_rsrc_node_switch_start(ctx);
8311 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8317 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8320 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8321 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8325 /* allow sparse sets */
8328 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8335 if (unlikely(!file))
8339 * Don't allow io_uring instances to be registered. If UNIX
8340 * isn't enabled, then this causes a reference cycle and this
8341 * instance can never get freed. If UNIX is enabled we'll
8342 * handle it just fine, but there's still no point in allowing
8343 * a ring fd as it doesn't support regular read/write anyway.
8345 if (file->f_op == &io_uring_fops) {
8349 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8352 ret = io_sqe_files_scm(ctx);
8354 __io_sqe_files_unregister(ctx);
8358 io_rsrc_node_switch(ctx, NULL);
8361 for (i = 0; i < ctx->nr_user_files; i++) {
8362 file = io_file_from_index(ctx, i);
8366 io_free_file_tables(&ctx->file_table);
8367 ctx->nr_user_files = 0;
8369 io_rsrc_data_free(ctx->file_data);
8370 ctx->file_data = NULL;
8374 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8377 #if defined(CONFIG_UNIX)
8378 struct sock *sock = ctx->ring_sock->sk;
8379 struct sk_buff_head *head = &sock->sk_receive_queue;
8380 struct sk_buff *skb;
8383 * See if we can merge this file into an existing skb SCM_RIGHTS
8384 * file set. If there's no room, fall back to allocating a new skb
8385 * and filling it in.
8387 spin_lock_irq(&head->lock);
8388 skb = skb_peek(head);
8390 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8392 if (fpl->count < SCM_MAX_FD) {
8393 __skb_unlink(skb, head);
8394 spin_unlock_irq(&head->lock);
8395 fpl->fp[fpl->count] = get_file(file);
8396 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8398 spin_lock_irq(&head->lock);
8399 __skb_queue_head(head, skb);
8404 spin_unlock_irq(&head->lock);
8411 return __io_sqe_files_scm(ctx, 1, index);
8417 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8418 struct io_rsrc_node *node, void *rsrc)
8420 struct io_rsrc_put *prsrc;
8422 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8426 prsrc->tag = *io_get_tag_slot(data, idx);
8428 list_add(&prsrc->list, &node->rsrc_list);
8432 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8433 unsigned int issue_flags, u32 slot_index)
8435 struct io_ring_ctx *ctx = req->ctx;
8436 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8437 bool needs_switch = false;
8438 struct io_fixed_file *file_slot;
8441 io_ring_submit_lock(ctx, !force_nonblock);
8442 if (file->f_op == &io_uring_fops)
8445 if (!ctx->file_data)
8448 if (slot_index >= ctx->nr_user_files)
8451 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8452 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8454 if (file_slot->file_ptr) {
8455 struct file *old_file;
8457 ret = io_rsrc_node_switch_start(ctx);
8461 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8462 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8463 ctx->rsrc_node, old_file);
8466 file_slot->file_ptr = 0;
8467 needs_switch = true;
8470 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8471 io_fixed_file_set(file_slot, file);
8472 ret = io_sqe_file_register(ctx, file, slot_index);
8474 file_slot->file_ptr = 0;
8481 io_rsrc_node_switch(ctx, ctx->file_data);
8482 io_ring_submit_unlock(ctx, !force_nonblock);
8488 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8490 unsigned int offset = req->close.file_slot - 1;
8491 struct io_ring_ctx *ctx = req->ctx;
8492 struct io_fixed_file *file_slot;
8496 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8498 if (unlikely(!ctx->file_data))
8501 if (offset >= ctx->nr_user_files)
8503 ret = io_rsrc_node_switch_start(ctx);
8507 i = array_index_nospec(offset, ctx->nr_user_files);
8508 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8510 if (!file_slot->file_ptr)
8513 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8514 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8518 file_slot->file_ptr = 0;
8519 io_rsrc_node_switch(ctx, ctx->file_data);
8522 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8526 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8527 struct io_uring_rsrc_update2 *up,
8530 u64 __user *tags = u64_to_user_ptr(up->tags);
8531 __s32 __user *fds = u64_to_user_ptr(up->data);
8532 struct io_rsrc_data *data = ctx->file_data;
8533 struct io_fixed_file *file_slot;
8537 bool needs_switch = false;
8539 if (!ctx->file_data)
8541 if (up->offset + nr_args > ctx->nr_user_files)
8544 for (done = 0; done < nr_args; done++) {
8547 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8548 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8552 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8556 if (fd == IORING_REGISTER_FILES_SKIP)
8559 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8560 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8562 if (file_slot->file_ptr) {
8563 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8564 err = io_queue_rsrc_removal(data, up->offset + done,
8565 ctx->rsrc_node, file);
8568 file_slot->file_ptr = 0;
8569 needs_switch = true;
8578 * Don't allow io_uring instances to be registered. If
8579 * UNIX isn't enabled, then this causes a reference
8580 * cycle and this instance can never get freed. If UNIX
8581 * is enabled we'll handle it just fine, but there's
8582 * still no point in allowing a ring fd as it doesn't
8583 * support regular read/write anyway.
8585 if (file->f_op == &io_uring_fops) {
8590 *io_get_tag_slot(data, up->offset + done) = tag;
8591 io_fixed_file_set(file_slot, file);
8592 err = io_sqe_file_register(ctx, file, i);
8594 file_slot->file_ptr = 0;
8602 io_rsrc_node_switch(ctx, data);
8603 return done ? done : err;
8606 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8607 struct task_struct *task)
8609 struct io_wq_hash *hash;
8610 struct io_wq_data data;
8611 unsigned int concurrency;
8613 mutex_lock(&ctx->uring_lock);
8614 hash = ctx->hash_map;
8616 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8618 mutex_unlock(&ctx->uring_lock);
8619 return ERR_PTR(-ENOMEM);
8621 refcount_set(&hash->refs, 1);
8622 init_waitqueue_head(&hash->wait);
8623 ctx->hash_map = hash;
8625 mutex_unlock(&ctx->uring_lock);
8629 data.free_work = io_wq_free_work;
8630 data.do_work = io_wq_submit_work;
8632 /* Do QD, or 4 * CPUS, whatever is smallest */
8633 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8635 return io_wq_create(concurrency, &data);
8638 static int io_uring_alloc_task_context(struct task_struct *task,
8639 struct io_ring_ctx *ctx)
8641 struct io_uring_task *tctx;
8644 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8645 if (unlikely(!tctx))
8648 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8649 if (unlikely(ret)) {
8654 tctx->io_wq = io_init_wq_offload(ctx, task);
8655 if (IS_ERR(tctx->io_wq)) {
8656 ret = PTR_ERR(tctx->io_wq);
8657 percpu_counter_destroy(&tctx->inflight);
8663 init_waitqueue_head(&tctx->wait);
8664 atomic_set(&tctx->in_idle, 0);
8665 atomic_set(&tctx->inflight_tracked, 0);
8666 task->io_uring = tctx;
8667 spin_lock_init(&tctx->task_lock);
8668 INIT_WQ_LIST(&tctx->task_list);
8669 init_task_work(&tctx->task_work, tctx_task_work);
8673 void __io_uring_free(struct task_struct *tsk)
8675 struct io_uring_task *tctx = tsk->io_uring;
8677 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8678 WARN_ON_ONCE(tctx->io_wq);
8679 WARN_ON_ONCE(tctx->cached_refs);
8681 percpu_counter_destroy(&tctx->inflight);
8683 tsk->io_uring = NULL;
8686 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8687 struct io_uring_params *p)
8691 /* Retain compatibility with failing for an invalid attach attempt */
8692 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8693 IORING_SETUP_ATTACH_WQ) {
8696 f = fdget(p->wq_fd);
8699 if (f.file->f_op != &io_uring_fops) {
8705 if (ctx->flags & IORING_SETUP_SQPOLL) {
8706 struct task_struct *tsk;
8707 struct io_sq_data *sqd;
8710 sqd = io_get_sq_data(p, &attached);
8716 ctx->sq_creds = get_current_cred();
8718 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8719 if (!ctx->sq_thread_idle)
8720 ctx->sq_thread_idle = HZ;
8722 io_sq_thread_park(sqd);
8723 list_add(&ctx->sqd_list, &sqd->ctx_list);
8724 io_sqd_update_thread_idle(sqd);
8725 /* don't attach to a dying SQPOLL thread, would be racy */
8726 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8727 io_sq_thread_unpark(sqd);
8734 if (p->flags & IORING_SETUP_SQ_AFF) {
8735 int cpu = p->sq_thread_cpu;
8738 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8745 sqd->task_pid = current->pid;
8746 sqd->task_tgid = current->tgid;
8747 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8754 ret = io_uring_alloc_task_context(tsk, ctx);
8755 wake_up_new_task(tsk);
8758 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8759 /* Can't have SQ_AFF without SQPOLL */
8766 complete(&ctx->sq_data->exited);
8768 io_sq_thread_finish(ctx);
8772 static inline void __io_unaccount_mem(struct user_struct *user,
8773 unsigned long nr_pages)
8775 atomic_long_sub(nr_pages, &user->locked_vm);
8778 static inline int __io_account_mem(struct user_struct *user,
8779 unsigned long nr_pages)
8781 unsigned long page_limit, cur_pages, new_pages;
8783 /* Don't allow more pages than we can safely lock */
8784 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8787 cur_pages = atomic_long_read(&user->locked_vm);
8788 new_pages = cur_pages + nr_pages;
8789 if (new_pages > page_limit)
8791 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8792 new_pages) != cur_pages);
8797 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8800 __io_unaccount_mem(ctx->user, nr_pages);
8802 if (ctx->mm_account)
8803 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8806 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8811 ret = __io_account_mem(ctx->user, nr_pages);
8816 if (ctx->mm_account)
8817 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8822 static void io_mem_free(void *ptr)
8829 page = virt_to_head_page(ptr);
8830 if (put_page_testzero(page))
8831 free_compound_page(page);
8834 static void *io_mem_alloc(size_t size)
8836 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
8838 return (void *) __get_free_pages(gfp, get_order(size));
8841 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8844 struct io_rings *rings;
8845 size_t off, sq_array_size;
8847 off = struct_size(rings, cqes, cq_entries);
8848 if (off == SIZE_MAX)
8852 off = ALIGN(off, SMP_CACHE_BYTES);
8860 sq_array_size = array_size(sizeof(u32), sq_entries);
8861 if (sq_array_size == SIZE_MAX)
8864 if (check_add_overflow(off, sq_array_size, &off))
8870 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8872 struct io_mapped_ubuf *imu = *slot;
8875 if (imu != ctx->dummy_ubuf) {
8876 for (i = 0; i < imu->nr_bvecs; i++)
8877 unpin_user_page(imu->bvec[i].bv_page);
8878 if (imu->acct_pages)
8879 io_unaccount_mem(ctx, imu->acct_pages);
8885 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8887 io_buffer_unmap(ctx, &prsrc->buf);
8891 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8895 for (i = 0; i < ctx->nr_user_bufs; i++)
8896 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8897 kfree(ctx->user_bufs);
8898 io_rsrc_data_free(ctx->buf_data);
8899 ctx->user_bufs = NULL;
8900 ctx->buf_data = NULL;
8901 ctx->nr_user_bufs = 0;
8904 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8911 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8913 __io_sqe_buffers_unregister(ctx);
8917 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8918 void __user *arg, unsigned index)
8920 struct iovec __user *src;
8922 #ifdef CONFIG_COMPAT
8924 struct compat_iovec __user *ciovs;
8925 struct compat_iovec ciov;
8927 ciovs = (struct compat_iovec __user *) arg;
8928 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8931 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8932 dst->iov_len = ciov.iov_len;
8936 src = (struct iovec __user *) arg;
8937 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8943 * Not super efficient, but this is just a registration time. And we do cache
8944 * the last compound head, so generally we'll only do a full search if we don't
8947 * We check if the given compound head page has already been accounted, to
8948 * avoid double accounting it. This allows us to account the full size of the
8949 * page, not just the constituent pages of a huge page.
8951 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8952 int nr_pages, struct page *hpage)
8956 /* check current page array */
8957 for (i = 0; i < nr_pages; i++) {
8958 if (!PageCompound(pages[i]))
8960 if (compound_head(pages[i]) == hpage)
8964 /* check previously registered pages */
8965 for (i = 0; i < ctx->nr_user_bufs; i++) {
8966 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8968 for (j = 0; j < imu->nr_bvecs; j++) {
8969 if (!PageCompound(imu->bvec[j].bv_page))
8971 if (compound_head(imu->bvec[j].bv_page) == hpage)
8979 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8980 int nr_pages, struct io_mapped_ubuf *imu,
8981 struct page **last_hpage)
8985 imu->acct_pages = 0;
8986 for (i = 0; i < nr_pages; i++) {
8987 if (!PageCompound(pages[i])) {
8992 hpage = compound_head(pages[i]);
8993 if (hpage == *last_hpage)
8995 *last_hpage = hpage;
8996 if (headpage_already_acct(ctx, pages, i, hpage))
8998 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
9002 if (!imu->acct_pages)
9005 ret = io_account_mem(ctx, imu->acct_pages);
9007 imu->acct_pages = 0;
9011 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
9012 struct io_mapped_ubuf **pimu,
9013 struct page **last_hpage)
9015 struct io_mapped_ubuf *imu = NULL;
9016 struct vm_area_struct **vmas = NULL;
9017 struct page **pages = NULL;
9018 unsigned long off, start, end, ubuf;
9020 int ret, pret, nr_pages, i;
9022 if (!iov->iov_base) {
9023 *pimu = ctx->dummy_ubuf;
9027 ubuf = (unsigned long) iov->iov_base;
9028 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9029 start = ubuf >> PAGE_SHIFT;
9030 nr_pages = end - start;
9035 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9039 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9044 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9049 mmap_read_lock(current->mm);
9050 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9052 if (pret == nr_pages) {
9053 /* don't support file backed memory */
9054 for (i = 0; i < nr_pages; i++) {
9055 struct vm_area_struct *vma = vmas[i];
9057 if (vma_is_shmem(vma))
9060 !is_file_hugepages(vma->vm_file)) {
9066 ret = pret < 0 ? pret : -EFAULT;
9068 mmap_read_unlock(current->mm);
9071 * if we did partial map, or found file backed vmas,
9072 * release any pages we did get
9075 unpin_user_pages(pages, pret);
9079 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9081 unpin_user_pages(pages, pret);
9085 off = ubuf & ~PAGE_MASK;
9086 size = iov->iov_len;
9087 for (i = 0; i < nr_pages; i++) {
9090 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9091 imu->bvec[i].bv_page = pages[i];
9092 imu->bvec[i].bv_len = vec_len;
9093 imu->bvec[i].bv_offset = off;
9097 /* store original address for later verification */
9099 imu->ubuf_end = ubuf + iov->iov_len;
9100 imu->nr_bvecs = nr_pages;
9111 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9113 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9114 return ctx->user_bufs ? 0 : -ENOMEM;
9117 static int io_buffer_validate(struct iovec *iov)
9119 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9122 * Don't impose further limits on the size and buffer
9123 * constraints here, we'll -EINVAL later when IO is
9124 * submitted if they are wrong.
9127 return iov->iov_len ? -EFAULT : 0;
9131 /* arbitrary limit, but we need something */
9132 if (iov->iov_len > SZ_1G)
9135 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9141 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9142 unsigned int nr_args, u64 __user *tags)
9144 struct page *last_hpage = NULL;
9145 struct io_rsrc_data *data;
9151 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9153 ret = io_rsrc_node_switch_start(ctx);
9156 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9159 ret = io_buffers_map_alloc(ctx, nr_args);
9161 io_rsrc_data_free(data);
9165 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9166 ret = io_copy_iov(ctx, &iov, arg, i);
9169 ret = io_buffer_validate(&iov);
9172 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9177 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9183 WARN_ON_ONCE(ctx->buf_data);
9185 ctx->buf_data = data;
9187 __io_sqe_buffers_unregister(ctx);
9189 io_rsrc_node_switch(ctx, NULL);
9193 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9194 struct io_uring_rsrc_update2 *up,
9195 unsigned int nr_args)
9197 u64 __user *tags = u64_to_user_ptr(up->tags);
9198 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9199 struct page *last_hpage = NULL;
9200 bool needs_switch = false;
9206 if (up->offset + nr_args > ctx->nr_user_bufs)
9209 for (done = 0; done < nr_args; done++) {
9210 struct io_mapped_ubuf *imu;
9211 int offset = up->offset + done;
9214 err = io_copy_iov(ctx, &iov, iovs, done);
9217 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9221 err = io_buffer_validate(&iov);
9224 if (!iov.iov_base && tag) {
9228 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9232 i = array_index_nospec(offset, ctx->nr_user_bufs);
9233 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9234 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9235 ctx->rsrc_node, ctx->user_bufs[i]);
9236 if (unlikely(err)) {
9237 io_buffer_unmap(ctx, &imu);
9240 ctx->user_bufs[i] = NULL;
9241 needs_switch = true;
9244 ctx->user_bufs[i] = imu;
9245 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9249 io_rsrc_node_switch(ctx, ctx->buf_data);
9250 return done ? done : err;
9253 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9255 __s32 __user *fds = arg;
9261 if (copy_from_user(&fd, fds, sizeof(*fds)))
9264 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9265 if (IS_ERR(ctx->cq_ev_fd)) {
9266 int ret = PTR_ERR(ctx->cq_ev_fd);
9268 ctx->cq_ev_fd = NULL;
9275 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9277 if (ctx->cq_ev_fd) {
9278 eventfd_ctx_put(ctx->cq_ev_fd);
9279 ctx->cq_ev_fd = NULL;
9286 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9288 struct io_buffer *buf;
9289 unsigned long index;
9291 xa_for_each(&ctx->io_buffers, index, buf)
9292 __io_remove_buffers(ctx, buf, index, -1U);
9295 static void io_req_cache_free(struct list_head *list)
9297 struct io_kiocb *req, *nxt;
9299 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9300 list_del(&req->inflight_entry);
9301 kmem_cache_free(req_cachep, req);
9305 static void io_req_caches_free(struct io_ring_ctx *ctx)
9307 struct io_submit_state *state = &ctx->submit_state;
9309 mutex_lock(&ctx->uring_lock);
9311 if (state->free_reqs) {
9312 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9313 state->free_reqs = 0;
9316 io_flush_cached_locked_reqs(ctx, state);
9317 io_req_cache_free(&state->free_list);
9318 mutex_unlock(&ctx->uring_lock);
9321 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9323 if (data && !atomic_dec_and_test(&data->refs))
9324 wait_for_completion(&data->done);
9327 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9329 io_sq_thread_finish(ctx);
9331 if (ctx->mm_account) {
9332 mmdrop(ctx->mm_account);
9333 ctx->mm_account = NULL;
9336 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9337 io_wait_rsrc_data(ctx->buf_data);
9338 io_wait_rsrc_data(ctx->file_data);
9340 mutex_lock(&ctx->uring_lock);
9342 __io_sqe_buffers_unregister(ctx);
9344 __io_sqe_files_unregister(ctx);
9346 __io_cqring_overflow_flush(ctx, true);
9347 mutex_unlock(&ctx->uring_lock);
9348 io_eventfd_unregister(ctx);
9349 io_destroy_buffers(ctx);
9351 put_cred(ctx->sq_creds);
9353 /* there are no registered resources left, nobody uses it */
9355 io_rsrc_node_destroy(ctx->rsrc_node);
9356 if (ctx->rsrc_backup_node)
9357 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9358 flush_delayed_work(&ctx->rsrc_put_work);
9360 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9361 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9363 #if defined(CONFIG_UNIX)
9364 if (ctx->ring_sock) {
9365 ctx->ring_sock->file = NULL; /* so that iput() is called */
9366 sock_release(ctx->ring_sock);
9369 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9371 io_mem_free(ctx->rings);
9372 io_mem_free(ctx->sq_sqes);
9374 percpu_ref_exit(&ctx->refs);
9375 free_uid(ctx->user);
9376 io_req_caches_free(ctx);
9378 io_wq_put_hash(ctx->hash_map);
9379 kfree(ctx->cancel_hash);
9380 kfree(ctx->dummy_ubuf);
9384 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9386 struct io_ring_ctx *ctx = file->private_data;
9389 poll_wait(file, &ctx->poll_wait, wait);
9391 * synchronizes with barrier from wq_has_sleeper call in
9395 if (!io_sqring_full(ctx))
9396 mask |= EPOLLOUT | EPOLLWRNORM;
9399 * Don't flush cqring overflow list here, just do a simple check.
9400 * Otherwise there could possible be ABBA deadlock:
9403 * lock(&ctx->uring_lock);
9405 * lock(&ctx->uring_lock);
9408 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9409 * pushs them to do the flush.
9411 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9412 mask |= EPOLLIN | EPOLLRDNORM;
9417 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9419 const struct cred *creds;
9421 creds = xa_erase(&ctx->personalities, id);
9430 struct io_tctx_exit {
9431 struct callback_head task_work;
9432 struct completion completion;
9433 struct io_ring_ctx *ctx;
9436 static void io_tctx_exit_cb(struct callback_head *cb)
9438 struct io_uring_task *tctx = current->io_uring;
9439 struct io_tctx_exit *work;
9441 work = container_of(cb, struct io_tctx_exit, task_work);
9443 * When @in_idle, we're in cancellation and it's racy to remove the
9444 * node. It'll be removed by the end of cancellation, just ignore it.
9446 if (!atomic_read(&tctx->in_idle))
9447 io_uring_del_tctx_node((unsigned long)work->ctx);
9448 complete(&work->completion);
9451 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9453 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9455 return req->ctx == data;
9458 static void io_ring_exit_work(struct work_struct *work)
9460 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9461 unsigned long timeout = jiffies + HZ * 60 * 5;
9462 unsigned long interval = HZ / 20;
9463 struct io_tctx_exit exit;
9464 struct io_tctx_node *node;
9468 * If we're doing polled IO and end up having requests being
9469 * submitted async (out-of-line), then completions can come in while
9470 * we're waiting for refs to drop. We need to reap these manually,
9471 * as nobody else will be looking for them.
9474 io_uring_try_cancel_requests(ctx, NULL, true);
9476 struct io_sq_data *sqd = ctx->sq_data;
9477 struct task_struct *tsk;
9479 io_sq_thread_park(sqd);
9481 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9482 io_wq_cancel_cb(tsk->io_uring->io_wq,
9483 io_cancel_ctx_cb, ctx, true);
9484 io_sq_thread_unpark(sqd);
9487 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9488 /* there is little hope left, don't run it too often */
9491 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9493 init_completion(&exit.completion);
9494 init_task_work(&exit.task_work, io_tctx_exit_cb);
9497 * Some may use context even when all refs and requests have been put,
9498 * and they are free to do so while still holding uring_lock or
9499 * completion_lock, see io_req_task_submit(). Apart from other work,
9500 * this lock/unlock section also waits them to finish.
9502 mutex_lock(&ctx->uring_lock);
9503 while (!list_empty(&ctx->tctx_list)) {
9504 WARN_ON_ONCE(time_after(jiffies, timeout));
9506 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9508 /* don't spin on a single task if cancellation failed */
9509 list_rotate_left(&ctx->tctx_list);
9510 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9511 if (WARN_ON_ONCE(ret))
9513 wake_up_process(node->task);
9515 mutex_unlock(&ctx->uring_lock);
9516 wait_for_completion(&exit.completion);
9517 mutex_lock(&ctx->uring_lock);
9519 mutex_unlock(&ctx->uring_lock);
9520 spin_lock(&ctx->completion_lock);
9521 spin_unlock(&ctx->completion_lock);
9523 io_ring_ctx_free(ctx);
9526 /* Returns true if we found and killed one or more timeouts */
9527 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9530 struct io_kiocb *req, *tmp;
9533 spin_lock(&ctx->completion_lock);
9534 spin_lock_irq(&ctx->timeout_lock);
9535 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9536 if (io_match_task(req, tsk, cancel_all)) {
9537 io_kill_timeout(req, -ECANCELED);
9541 spin_unlock_irq(&ctx->timeout_lock);
9543 io_commit_cqring(ctx);
9544 spin_unlock(&ctx->completion_lock);
9546 io_cqring_ev_posted(ctx);
9547 return canceled != 0;
9550 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9552 unsigned long index;
9553 struct creds *creds;
9555 mutex_lock(&ctx->uring_lock);
9556 percpu_ref_kill(&ctx->refs);
9558 __io_cqring_overflow_flush(ctx, true);
9559 xa_for_each(&ctx->personalities, index, creds)
9560 io_unregister_personality(ctx, index);
9561 mutex_unlock(&ctx->uring_lock);
9563 io_kill_timeouts(ctx, NULL, true);
9564 io_poll_remove_all(ctx, NULL, true);
9566 /* if we failed setting up the ctx, we might not have any rings */
9567 io_iopoll_try_reap_events(ctx);
9569 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9571 * Use system_unbound_wq to avoid spawning tons of event kworkers
9572 * if we're exiting a ton of rings at the same time. It just adds
9573 * noise and overhead, there's no discernable change in runtime
9574 * over using system_wq.
9576 queue_work(system_unbound_wq, &ctx->exit_work);
9579 static int io_uring_release(struct inode *inode, struct file *file)
9581 struct io_ring_ctx *ctx = file->private_data;
9583 file->private_data = NULL;
9584 io_ring_ctx_wait_and_kill(ctx);
9588 struct io_task_cancel {
9589 struct task_struct *task;
9593 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9595 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9596 struct io_task_cancel *cancel = data;
9598 return io_match_task_safe(req, cancel->task, cancel->all);
9601 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9602 struct task_struct *task, bool cancel_all)
9604 struct io_defer_entry *de;
9607 spin_lock(&ctx->completion_lock);
9608 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9609 if (io_match_task_safe(de->req, task, cancel_all)) {
9610 list_cut_position(&list, &ctx->defer_list, &de->list);
9614 spin_unlock(&ctx->completion_lock);
9615 if (list_empty(&list))
9618 while (!list_empty(&list)) {
9619 de = list_first_entry(&list, struct io_defer_entry, list);
9620 list_del_init(&de->list);
9621 io_req_complete_failed(de->req, -ECANCELED);
9627 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9629 struct io_tctx_node *node;
9630 enum io_wq_cancel cret;
9633 mutex_lock(&ctx->uring_lock);
9634 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9635 struct io_uring_task *tctx = node->task->io_uring;
9638 * io_wq will stay alive while we hold uring_lock, because it's
9639 * killed after ctx nodes, which requires to take the lock.
9641 if (!tctx || !tctx->io_wq)
9643 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9644 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9646 mutex_unlock(&ctx->uring_lock);
9651 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9652 struct task_struct *task,
9655 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9656 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9659 enum io_wq_cancel cret;
9663 ret |= io_uring_try_cancel_iowq(ctx);
9664 } else if (tctx && tctx->io_wq) {
9666 * Cancels requests of all rings, not only @ctx, but
9667 * it's fine as the task is in exit/exec.
9669 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9671 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9674 /* SQPOLL thread does its own polling */
9675 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9676 (ctx->sq_data && ctx->sq_data->thread == current)) {
9677 while (!list_empty_careful(&ctx->iopoll_list)) {
9678 io_iopoll_try_reap_events(ctx);
9683 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9684 ret |= io_poll_remove_all(ctx, task, cancel_all);
9685 ret |= io_kill_timeouts(ctx, task, cancel_all);
9687 ret |= io_run_task_work();
9694 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9696 struct io_uring_task *tctx = current->io_uring;
9697 struct io_tctx_node *node;
9700 if (unlikely(!tctx)) {
9701 ret = io_uring_alloc_task_context(current, ctx);
9705 tctx = current->io_uring;
9706 if (ctx->iowq_limits_set) {
9707 unsigned int limits[2] = { ctx->iowq_limits[0],
9708 ctx->iowq_limits[1], };
9710 ret = io_wq_max_workers(tctx->io_wq, limits);
9715 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9716 node = kmalloc(sizeof(*node), GFP_KERNEL);
9720 node->task = current;
9722 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9729 mutex_lock(&ctx->uring_lock);
9730 list_add(&node->ctx_node, &ctx->tctx_list);
9731 mutex_unlock(&ctx->uring_lock);
9738 * Note that this task has used io_uring. We use it for cancelation purposes.
9740 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9742 struct io_uring_task *tctx = current->io_uring;
9744 if (likely(tctx && tctx->last == ctx))
9746 return __io_uring_add_tctx_node(ctx);
9750 * Remove this io_uring_file -> task mapping.
9752 static void io_uring_del_tctx_node(unsigned long index)
9754 struct io_uring_task *tctx = current->io_uring;
9755 struct io_tctx_node *node;
9759 node = xa_erase(&tctx->xa, index);
9763 WARN_ON_ONCE(current != node->task);
9764 WARN_ON_ONCE(list_empty(&node->ctx_node));
9766 mutex_lock(&node->ctx->uring_lock);
9767 list_del(&node->ctx_node);
9768 mutex_unlock(&node->ctx->uring_lock);
9770 if (tctx->last == node->ctx)
9775 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9777 struct io_wq *wq = tctx->io_wq;
9778 struct io_tctx_node *node;
9779 unsigned long index;
9781 xa_for_each(&tctx->xa, index, node) {
9782 io_uring_del_tctx_node(index);
9787 * Must be after io_uring_del_task_file() (removes nodes under
9788 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9790 io_wq_put_and_exit(wq);
9795 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9798 return atomic_read(&tctx->inflight_tracked);
9799 return percpu_counter_sum(&tctx->inflight);
9803 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9804 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9806 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9808 struct io_uring_task *tctx = current->io_uring;
9809 struct io_ring_ctx *ctx;
9813 WARN_ON_ONCE(sqd && sqd->thread != current);
9815 if (!current->io_uring)
9818 io_wq_exit_start(tctx->io_wq);
9820 atomic_inc(&tctx->in_idle);
9822 io_uring_drop_tctx_refs(current);
9823 /* read completions before cancelations */
9824 inflight = tctx_inflight(tctx, !cancel_all);
9829 struct io_tctx_node *node;
9830 unsigned long index;
9832 xa_for_each(&tctx->xa, index, node) {
9833 /* sqpoll task will cancel all its requests */
9834 if (node->ctx->sq_data)
9836 io_uring_try_cancel_requests(node->ctx, current,
9840 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9841 io_uring_try_cancel_requests(ctx, current,
9845 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9847 io_uring_drop_tctx_refs(current);
9850 * If we've seen completions, retry without waiting. This
9851 * avoids a race where a completion comes in before we did
9852 * prepare_to_wait().
9854 if (inflight == tctx_inflight(tctx, !cancel_all))
9856 finish_wait(&tctx->wait, &wait);
9859 io_uring_clean_tctx(tctx);
9862 * We shouldn't run task_works after cancel, so just leave
9863 * ->in_idle set for normal exit.
9865 atomic_dec(&tctx->in_idle);
9866 /* for exec all current's requests should be gone, kill tctx */
9867 __io_uring_free(current);
9871 void __io_uring_cancel(bool cancel_all)
9873 io_uring_cancel_generic(cancel_all, NULL);
9876 static void *io_uring_validate_mmap_request(struct file *file,
9877 loff_t pgoff, size_t sz)
9879 struct io_ring_ctx *ctx = file->private_data;
9880 loff_t offset = pgoff << PAGE_SHIFT;
9885 case IORING_OFF_SQ_RING:
9886 case IORING_OFF_CQ_RING:
9889 case IORING_OFF_SQES:
9893 return ERR_PTR(-EINVAL);
9896 page = virt_to_head_page(ptr);
9897 if (sz > page_size(page))
9898 return ERR_PTR(-EINVAL);
9905 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9907 size_t sz = vma->vm_end - vma->vm_start;
9911 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9913 return PTR_ERR(ptr);
9915 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9916 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9919 #else /* !CONFIG_MMU */
9921 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9923 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9926 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9928 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9931 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9932 unsigned long addr, unsigned long len,
9933 unsigned long pgoff, unsigned long flags)
9937 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9939 return PTR_ERR(ptr);
9941 return (unsigned long) ptr;
9944 #endif /* !CONFIG_MMU */
9946 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9951 if (!io_sqring_full(ctx))
9953 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9955 if (!io_sqring_full(ctx))
9958 } while (!signal_pending(current));
9960 finish_wait(&ctx->sqo_sq_wait, &wait);
9964 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9965 struct __kernel_timespec __user **ts,
9966 const sigset_t __user **sig)
9968 struct io_uring_getevents_arg arg;
9971 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9972 * is just a pointer to the sigset_t.
9974 if (!(flags & IORING_ENTER_EXT_ARG)) {
9975 *sig = (const sigset_t __user *) argp;
9981 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9982 * timespec and sigset_t pointers if good.
9984 if (*argsz != sizeof(arg))
9986 if (copy_from_user(&arg, argp, sizeof(arg)))
9988 *sig = u64_to_user_ptr(arg.sigmask);
9989 *argsz = arg.sigmask_sz;
9990 *ts = u64_to_user_ptr(arg.ts);
9994 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9995 u32, min_complete, u32, flags, const void __user *, argp,
9998 struct io_ring_ctx *ctx;
10003 io_run_task_work();
10005 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
10006 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
10010 if (unlikely(!f.file))
10014 if (unlikely(f.file->f_op != &io_uring_fops))
10018 ctx = f.file->private_data;
10019 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
10023 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10027 * For SQ polling, the thread will do all submissions and completions.
10028 * Just return the requested submit count, and wake the thread if
10029 * we were asked to.
10032 if (ctx->flags & IORING_SETUP_SQPOLL) {
10033 io_cqring_overflow_flush(ctx);
10035 if (unlikely(ctx->sq_data->thread == NULL)) {
10039 if (flags & IORING_ENTER_SQ_WAKEUP)
10040 wake_up(&ctx->sq_data->wait);
10041 if (flags & IORING_ENTER_SQ_WAIT) {
10042 ret = io_sqpoll_wait_sq(ctx);
10046 submitted = to_submit;
10047 } else if (to_submit) {
10048 ret = io_uring_add_tctx_node(ctx);
10051 mutex_lock(&ctx->uring_lock);
10052 submitted = io_submit_sqes(ctx, to_submit);
10053 mutex_unlock(&ctx->uring_lock);
10055 if (submitted != to_submit)
10058 if (flags & IORING_ENTER_GETEVENTS) {
10059 const sigset_t __user *sig;
10060 struct __kernel_timespec __user *ts;
10062 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10066 min_complete = min(min_complete, ctx->cq_entries);
10069 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10070 * space applications don't need to do io completion events
10071 * polling again, they can rely on io_sq_thread to do polling
10072 * work, which can reduce cpu usage and uring_lock contention.
10074 if (ctx->flags & IORING_SETUP_IOPOLL &&
10075 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10076 ret = io_iopoll_check(ctx, min_complete);
10078 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10083 percpu_ref_put(&ctx->refs);
10086 return submitted ? submitted : ret;
10089 #ifdef CONFIG_PROC_FS
10090 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10091 const struct cred *cred)
10093 struct user_namespace *uns = seq_user_ns(m);
10094 struct group_info *gi;
10099 seq_printf(m, "%5d\n", id);
10100 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10101 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10102 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10103 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10104 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10105 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10106 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10107 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10108 seq_puts(m, "\n\tGroups:\t");
10109 gi = cred->group_info;
10110 for (g = 0; g < gi->ngroups; g++) {
10111 seq_put_decimal_ull(m, g ? " " : "",
10112 from_kgid_munged(uns, gi->gid[g]));
10114 seq_puts(m, "\n\tCapEff:\t");
10115 cap = cred->cap_effective;
10116 CAP_FOR_EACH_U32(__capi)
10117 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10122 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10124 struct io_sq_data *sq = NULL;
10129 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10130 * since fdinfo case grabs it in the opposite direction of normal use
10131 * cases. If we fail to get the lock, we just don't iterate any
10132 * structures that could be going away outside the io_uring mutex.
10134 has_lock = mutex_trylock(&ctx->uring_lock);
10136 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10142 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10143 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10144 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10145 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10146 struct file *f = io_file_from_index(ctx, i);
10149 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10151 seq_printf(m, "%5u: <none>\n", i);
10153 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10154 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10155 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10156 unsigned int len = buf->ubuf_end - buf->ubuf;
10158 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10160 if (has_lock && !xa_empty(&ctx->personalities)) {
10161 unsigned long index;
10162 const struct cred *cred;
10164 seq_printf(m, "Personalities:\n");
10165 xa_for_each(&ctx->personalities, index, cred)
10166 io_uring_show_cred(m, index, cred);
10168 seq_printf(m, "PollList:\n");
10169 spin_lock(&ctx->completion_lock);
10170 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10171 struct hlist_head *list = &ctx->cancel_hash[i];
10172 struct io_kiocb *req;
10174 hlist_for_each_entry(req, list, hash_node)
10175 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10176 req->task->task_works != NULL);
10178 spin_unlock(&ctx->completion_lock);
10180 mutex_unlock(&ctx->uring_lock);
10183 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10185 struct io_ring_ctx *ctx = f->private_data;
10187 if (percpu_ref_tryget(&ctx->refs)) {
10188 __io_uring_show_fdinfo(ctx, m);
10189 percpu_ref_put(&ctx->refs);
10194 static const struct file_operations io_uring_fops = {
10195 .release = io_uring_release,
10196 .mmap = io_uring_mmap,
10198 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10199 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10201 .poll = io_uring_poll,
10202 #ifdef CONFIG_PROC_FS
10203 .show_fdinfo = io_uring_show_fdinfo,
10207 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10208 struct io_uring_params *p)
10210 struct io_rings *rings;
10211 size_t size, sq_array_offset;
10213 /* make sure these are sane, as we already accounted them */
10214 ctx->sq_entries = p->sq_entries;
10215 ctx->cq_entries = p->cq_entries;
10217 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10218 if (size == SIZE_MAX)
10221 rings = io_mem_alloc(size);
10225 ctx->rings = rings;
10226 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10227 rings->sq_ring_mask = p->sq_entries - 1;
10228 rings->cq_ring_mask = p->cq_entries - 1;
10229 rings->sq_ring_entries = p->sq_entries;
10230 rings->cq_ring_entries = p->cq_entries;
10232 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10233 if (size == SIZE_MAX) {
10234 io_mem_free(ctx->rings);
10239 ctx->sq_sqes = io_mem_alloc(size);
10240 if (!ctx->sq_sqes) {
10241 io_mem_free(ctx->rings);
10249 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10253 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10257 ret = io_uring_add_tctx_node(ctx);
10262 fd_install(fd, file);
10267 * Allocate an anonymous fd, this is what constitutes the application
10268 * visible backing of an io_uring instance. The application mmaps this
10269 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10270 * we have to tie this fd to a socket for file garbage collection purposes.
10272 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10275 #if defined(CONFIG_UNIX)
10278 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10281 return ERR_PTR(ret);
10284 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10285 O_RDWR | O_CLOEXEC);
10286 #if defined(CONFIG_UNIX)
10287 if (IS_ERR(file)) {
10288 sock_release(ctx->ring_sock);
10289 ctx->ring_sock = NULL;
10291 ctx->ring_sock->file = file;
10297 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10298 struct io_uring_params __user *params)
10300 struct io_ring_ctx *ctx;
10306 if (entries > IORING_MAX_ENTRIES) {
10307 if (!(p->flags & IORING_SETUP_CLAMP))
10309 entries = IORING_MAX_ENTRIES;
10313 * Use twice as many entries for the CQ ring. It's possible for the
10314 * application to drive a higher depth than the size of the SQ ring,
10315 * since the sqes are only used at submission time. This allows for
10316 * some flexibility in overcommitting a bit. If the application has
10317 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10318 * of CQ ring entries manually.
10320 p->sq_entries = roundup_pow_of_two(entries);
10321 if (p->flags & IORING_SETUP_CQSIZE) {
10323 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10324 * to a power-of-two, if it isn't already. We do NOT impose
10325 * any cq vs sq ring sizing.
10327 if (!p->cq_entries)
10329 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10330 if (!(p->flags & IORING_SETUP_CLAMP))
10332 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10334 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10335 if (p->cq_entries < p->sq_entries)
10338 p->cq_entries = 2 * p->sq_entries;
10341 ctx = io_ring_ctx_alloc(p);
10344 ctx->compat = in_compat_syscall();
10345 if (!capable(CAP_IPC_LOCK))
10346 ctx->user = get_uid(current_user());
10349 * This is just grabbed for accounting purposes. When a process exits,
10350 * the mm is exited and dropped before the files, hence we need to hang
10351 * on to this mm purely for the purposes of being able to unaccount
10352 * memory (locked/pinned vm). It's not used for anything else.
10354 mmgrab(current->mm);
10355 ctx->mm_account = current->mm;
10357 ret = io_allocate_scq_urings(ctx, p);
10361 ret = io_sq_offload_create(ctx, p);
10364 /* always set a rsrc node */
10365 ret = io_rsrc_node_switch_start(ctx);
10368 io_rsrc_node_switch(ctx, NULL);
10370 memset(&p->sq_off, 0, sizeof(p->sq_off));
10371 p->sq_off.head = offsetof(struct io_rings, sq.head);
10372 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10373 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10374 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10375 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10376 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10377 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10379 memset(&p->cq_off, 0, sizeof(p->cq_off));
10380 p->cq_off.head = offsetof(struct io_rings, cq.head);
10381 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10382 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10383 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10384 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10385 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10386 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10388 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10389 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10390 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10391 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10392 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10393 IORING_FEAT_RSRC_TAGS;
10395 if (copy_to_user(params, p, sizeof(*p))) {
10400 file = io_uring_get_file(ctx);
10401 if (IS_ERR(file)) {
10402 ret = PTR_ERR(file);
10407 * Install ring fd as the very last thing, so we don't risk someone
10408 * having closed it before we finish setup
10410 ret = io_uring_install_fd(ctx, file);
10412 /* fput will clean it up */
10417 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10420 io_ring_ctx_wait_and_kill(ctx);
10425 * Sets up an aio uring context, and returns the fd. Applications asks for a
10426 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10427 * params structure passed in.
10429 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10431 struct io_uring_params p;
10434 if (copy_from_user(&p, params, sizeof(p)))
10436 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10441 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10442 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10443 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10444 IORING_SETUP_R_DISABLED))
10447 return io_uring_create(entries, &p, params);
10450 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10451 struct io_uring_params __user *, params)
10453 return io_uring_setup(entries, params);
10456 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10458 struct io_uring_probe *p;
10462 size = struct_size(p, ops, nr_args);
10463 if (size == SIZE_MAX)
10465 p = kzalloc(size, GFP_KERNEL);
10470 if (copy_from_user(p, arg, size))
10473 if (memchr_inv(p, 0, size))
10476 p->last_op = IORING_OP_LAST - 1;
10477 if (nr_args > IORING_OP_LAST)
10478 nr_args = IORING_OP_LAST;
10480 for (i = 0; i < nr_args; i++) {
10482 if (!io_op_defs[i].not_supported)
10483 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10488 if (copy_to_user(arg, p, size))
10495 static int io_register_personality(struct io_ring_ctx *ctx)
10497 const struct cred *creds;
10501 creds = get_current_cred();
10503 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10504 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10512 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10513 unsigned int nr_args)
10515 struct io_uring_restriction *res;
10519 /* Restrictions allowed only if rings started disabled */
10520 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10523 /* We allow only a single restrictions registration */
10524 if (ctx->restrictions.registered)
10527 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10530 size = array_size(nr_args, sizeof(*res));
10531 if (size == SIZE_MAX)
10534 res = memdup_user(arg, size);
10536 return PTR_ERR(res);
10540 for (i = 0; i < nr_args; i++) {
10541 switch (res[i].opcode) {
10542 case IORING_RESTRICTION_REGISTER_OP:
10543 if (res[i].register_op >= IORING_REGISTER_LAST) {
10548 __set_bit(res[i].register_op,
10549 ctx->restrictions.register_op);
10551 case IORING_RESTRICTION_SQE_OP:
10552 if (res[i].sqe_op >= IORING_OP_LAST) {
10557 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10559 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10560 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10562 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10563 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10572 /* Reset all restrictions if an error happened */
10574 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10576 ctx->restrictions.registered = true;
10582 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10584 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10587 if (ctx->restrictions.registered)
10588 ctx->restricted = 1;
10590 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10591 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10592 wake_up(&ctx->sq_data->wait);
10596 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10597 struct io_uring_rsrc_update2 *up,
10605 if (check_add_overflow(up->offset, nr_args, &tmp))
10607 err = io_rsrc_node_switch_start(ctx);
10612 case IORING_RSRC_FILE:
10613 return __io_sqe_files_update(ctx, up, nr_args);
10614 case IORING_RSRC_BUFFER:
10615 return __io_sqe_buffers_update(ctx, up, nr_args);
10620 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10623 struct io_uring_rsrc_update2 up;
10627 memset(&up, 0, sizeof(up));
10628 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10630 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10633 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10634 unsigned size, unsigned type)
10636 struct io_uring_rsrc_update2 up;
10638 if (size != sizeof(up))
10640 if (copy_from_user(&up, arg, sizeof(up)))
10642 if (!up.nr || up.resv)
10644 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10647 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10648 unsigned int size, unsigned int type)
10650 struct io_uring_rsrc_register rr;
10652 /* keep it extendible */
10653 if (size != sizeof(rr))
10656 memset(&rr, 0, sizeof(rr));
10657 if (copy_from_user(&rr, arg, size))
10659 if (!rr.nr || rr.resv || rr.resv2)
10663 case IORING_RSRC_FILE:
10664 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10665 rr.nr, u64_to_user_ptr(rr.tags));
10666 case IORING_RSRC_BUFFER:
10667 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10668 rr.nr, u64_to_user_ptr(rr.tags));
10673 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10676 struct io_uring_task *tctx = current->io_uring;
10677 cpumask_var_t new_mask;
10680 if (!tctx || !tctx->io_wq)
10683 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10686 cpumask_clear(new_mask);
10687 if (len > cpumask_size())
10688 len = cpumask_size();
10690 if (copy_from_user(new_mask, arg, len)) {
10691 free_cpumask_var(new_mask);
10695 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10696 free_cpumask_var(new_mask);
10700 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10702 struct io_uring_task *tctx = current->io_uring;
10704 if (!tctx || !tctx->io_wq)
10707 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10710 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10712 __must_hold(&ctx->uring_lock)
10714 struct io_tctx_node *node;
10715 struct io_uring_task *tctx = NULL;
10716 struct io_sq_data *sqd = NULL;
10717 __u32 new_count[2];
10720 if (copy_from_user(new_count, arg, sizeof(new_count)))
10722 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10723 if (new_count[i] > INT_MAX)
10726 if (ctx->flags & IORING_SETUP_SQPOLL) {
10727 sqd = ctx->sq_data;
10730 * Observe the correct sqd->lock -> ctx->uring_lock
10731 * ordering. Fine to drop uring_lock here, we hold
10732 * a ref to the ctx.
10734 refcount_inc(&sqd->refs);
10735 mutex_unlock(&ctx->uring_lock);
10736 mutex_lock(&sqd->lock);
10737 mutex_lock(&ctx->uring_lock);
10739 tctx = sqd->thread->io_uring;
10742 tctx = current->io_uring;
10745 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10747 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10749 ctx->iowq_limits[i] = new_count[i];
10750 ctx->iowq_limits_set = true;
10753 if (tctx && tctx->io_wq) {
10754 ret = io_wq_max_workers(tctx->io_wq, new_count);
10758 memset(new_count, 0, sizeof(new_count));
10762 mutex_unlock(&sqd->lock);
10763 io_put_sq_data(sqd);
10766 if (copy_to_user(arg, new_count, sizeof(new_count)))
10769 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10773 /* now propagate the restriction to all registered users */
10774 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10775 struct io_uring_task *tctx = node->task->io_uring;
10777 if (WARN_ON_ONCE(!tctx->io_wq))
10780 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10781 new_count[i] = ctx->iowq_limits[i];
10782 /* ignore errors, it always returns zero anyway */
10783 (void)io_wq_max_workers(tctx->io_wq, new_count);
10788 mutex_unlock(&sqd->lock);
10789 io_put_sq_data(sqd);
10794 static bool io_register_op_must_quiesce(int op)
10797 case IORING_REGISTER_BUFFERS:
10798 case IORING_UNREGISTER_BUFFERS:
10799 case IORING_REGISTER_FILES:
10800 case IORING_UNREGISTER_FILES:
10801 case IORING_REGISTER_FILES_UPDATE:
10802 case IORING_REGISTER_PROBE:
10803 case IORING_REGISTER_PERSONALITY:
10804 case IORING_UNREGISTER_PERSONALITY:
10805 case IORING_REGISTER_FILES2:
10806 case IORING_REGISTER_FILES_UPDATE2:
10807 case IORING_REGISTER_BUFFERS2:
10808 case IORING_REGISTER_BUFFERS_UPDATE:
10809 case IORING_REGISTER_IOWQ_AFF:
10810 case IORING_UNREGISTER_IOWQ_AFF:
10811 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10818 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10822 percpu_ref_kill(&ctx->refs);
10825 * Drop uring mutex before waiting for references to exit. If another
10826 * thread is currently inside io_uring_enter() it might need to grab the
10827 * uring_lock to make progress. If we hold it here across the drain
10828 * wait, then we can deadlock. It's safe to drop the mutex here, since
10829 * no new references will come in after we've killed the percpu ref.
10831 mutex_unlock(&ctx->uring_lock);
10833 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10836 ret = io_run_task_work_sig();
10837 } while (ret >= 0);
10838 mutex_lock(&ctx->uring_lock);
10841 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10845 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10846 void __user *arg, unsigned nr_args)
10847 __releases(ctx->uring_lock)
10848 __acquires(ctx->uring_lock)
10853 * We're inside the ring mutex, if the ref is already dying, then
10854 * someone else killed the ctx or is already going through
10855 * io_uring_register().
10857 if (percpu_ref_is_dying(&ctx->refs))
10860 if (ctx->restricted) {
10861 if (opcode >= IORING_REGISTER_LAST)
10863 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10864 if (!test_bit(opcode, ctx->restrictions.register_op))
10868 if (io_register_op_must_quiesce(opcode)) {
10869 ret = io_ctx_quiesce(ctx);
10875 case IORING_REGISTER_BUFFERS:
10876 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10878 case IORING_UNREGISTER_BUFFERS:
10880 if (arg || nr_args)
10882 ret = io_sqe_buffers_unregister(ctx);
10884 case IORING_REGISTER_FILES:
10885 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10887 case IORING_UNREGISTER_FILES:
10889 if (arg || nr_args)
10891 ret = io_sqe_files_unregister(ctx);
10893 case IORING_REGISTER_FILES_UPDATE:
10894 ret = io_register_files_update(ctx, arg, nr_args);
10896 case IORING_REGISTER_EVENTFD:
10897 case IORING_REGISTER_EVENTFD_ASYNC:
10901 ret = io_eventfd_register(ctx, arg);
10904 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10905 ctx->eventfd_async = 1;
10907 ctx->eventfd_async = 0;
10909 case IORING_UNREGISTER_EVENTFD:
10911 if (arg || nr_args)
10913 ret = io_eventfd_unregister(ctx);
10915 case IORING_REGISTER_PROBE:
10917 if (!arg || nr_args > 256)
10919 ret = io_probe(ctx, arg, nr_args);
10921 case IORING_REGISTER_PERSONALITY:
10923 if (arg || nr_args)
10925 ret = io_register_personality(ctx);
10927 case IORING_UNREGISTER_PERSONALITY:
10931 ret = io_unregister_personality(ctx, nr_args);
10933 case IORING_REGISTER_ENABLE_RINGS:
10935 if (arg || nr_args)
10937 ret = io_register_enable_rings(ctx);
10939 case IORING_REGISTER_RESTRICTIONS:
10940 ret = io_register_restrictions(ctx, arg, nr_args);
10942 case IORING_REGISTER_FILES2:
10943 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10945 case IORING_REGISTER_FILES_UPDATE2:
10946 ret = io_register_rsrc_update(ctx, arg, nr_args,
10949 case IORING_REGISTER_BUFFERS2:
10950 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10952 case IORING_REGISTER_BUFFERS_UPDATE:
10953 ret = io_register_rsrc_update(ctx, arg, nr_args,
10954 IORING_RSRC_BUFFER);
10956 case IORING_REGISTER_IOWQ_AFF:
10958 if (!arg || !nr_args)
10960 ret = io_register_iowq_aff(ctx, arg, nr_args);
10962 case IORING_UNREGISTER_IOWQ_AFF:
10964 if (arg || nr_args)
10966 ret = io_unregister_iowq_aff(ctx);
10968 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10970 if (!arg || nr_args != 2)
10972 ret = io_register_iowq_max_workers(ctx, arg);
10979 if (io_register_op_must_quiesce(opcode)) {
10980 /* bring the ctx back to life */
10981 percpu_ref_reinit(&ctx->refs);
10982 reinit_completion(&ctx->ref_comp);
10987 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10988 void __user *, arg, unsigned int, nr_args)
10990 struct io_ring_ctx *ctx;
10999 if (f.file->f_op != &io_uring_fops)
11002 ctx = f.file->private_data;
11004 io_run_task_work();
11006 mutex_lock(&ctx->uring_lock);
11007 ret = __io_uring_register(ctx, opcode, arg, nr_args);
11008 mutex_unlock(&ctx->uring_lock);
11009 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
11010 ctx->cq_ev_fd != NULL, ret);
11016 static int __init io_uring_init(void)
11018 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11019 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11020 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11023 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11024 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11025 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11026 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11027 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11028 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11029 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11030 BUILD_BUG_SQE_ELEM(8, __u64, off);
11031 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11032 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11033 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11034 BUILD_BUG_SQE_ELEM(24, __u32, len);
11035 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11036 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11037 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11038 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11039 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11040 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11041 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11042 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11043 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11044 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11045 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11046 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11047 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11048 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11049 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11050 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11051 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11052 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11053 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11054 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11055 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11057 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11058 sizeof(struct io_uring_rsrc_update));
11059 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11060 sizeof(struct io_uring_rsrc_update2));
11062 /* ->buf_index is u16 */
11063 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11065 /* should fit into one byte */
11066 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11068 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11069 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11071 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11075 __initcall(io_uring_init);