1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head ltimeout_list;
379 struct list_head cq_overflow_list;
380 struct xarray io_buffers;
381 struct xarray personalities;
383 unsigned sq_thread_idle;
384 } ____cacheline_aligned_in_smp;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list;
388 unsigned int locked_free_nr;
390 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
396 unsigned long check_cq_overflow;
399 unsigned cached_cq_tail;
401 struct eventfd_ctx *cq_ev_fd;
402 struct wait_queue_head poll_wait;
403 struct wait_queue_head cq_wait;
405 atomic_t cq_timeouts;
406 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
460 bool iowq_limits_set;
464 struct io_uring_task {
465 /* submission side */
468 struct wait_queue_head wait;
469 const struct io_ring_ctx *last;
471 struct percpu_counter inflight;
472 atomic_t inflight_tracked;
475 spinlock_t task_lock;
476 struct io_wq_work_list task_list;
477 struct callback_head task_work;
482 * First field must be the file pointer in all the
483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
485 struct io_poll_iocb {
487 struct wait_queue_head *head;
491 struct wait_queue_entry wait;
494 struct io_poll_update {
500 bool update_user_data;
509 struct io_timeout_data {
510 struct io_kiocb *req;
511 struct hrtimer timer;
512 struct timespec64 ts;
513 enum hrtimer_mode mode;
519 struct sockaddr __user *addr;
520 int __user *addr_len;
523 unsigned long nofile;
543 struct list_head list;
544 /* head of the link, used by linked timeouts only */
545 struct io_kiocb *head;
546 /* for linked completions */
547 struct io_kiocb *prev;
550 struct io_timeout_rem {
555 struct timespec64 ts;
561 /* NOTE: kiocb has the file as the first member, so don't do it here */
569 struct sockaddr __user *addr;
576 struct compat_msghdr __user *umsg_compat;
577 struct user_msghdr __user *umsg;
583 struct io_buffer *kbuf;
590 struct filename *filename;
592 unsigned long nofile;
595 struct io_rsrc_update {
621 struct epoll_event event;
625 struct file *file_out;
633 struct io_provide_buf {
647 const char __user *filename;
648 struct statx __user *buffer;
660 struct filename *oldpath;
661 struct filename *newpath;
669 struct filename *filename;
676 struct filename *filename;
682 struct filename *oldpath;
683 struct filename *newpath;
690 struct filename *oldpath;
691 struct filename *newpath;
695 struct io_completion {
700 struct io_async_connect {
701 struct sockaddr_storage address;
704 struct io_async_msghdr {
705 struct iovec fast_iov[UIO_FASTIOV];
706 /* points to an allocated iov, if NULL we use fast_iov instead */
707 struct iovec *free_iov;
708 struct sockaddr __user *uaddr;
710 struct sockaddr_storage addr;
714 struct iovec fast_iov[UIO_FASTIOV];
715 const struct iovec *free_iovec;
716 struct iov_iter iter;
717 struct iov_iter_state iter_state;
719 struct wait_page_queue wpq;
723 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
724 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
725 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
726 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
727 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
728 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
730 /* first byte is taken by user flags, shift it to not overlap */
735 REQ_F_LINK_TIMEOUT_BIT,
736 REQ_F_NEED_CLEANUP_BIT,
738 REQ_F_BUFFER_SELECTED_BIT,
739 REQ_F_COMPLETE_INLINE_BIT,
743 REQ_F_ARM_LTIMEOUT_BIT,
744 /* keep async read/write and isreg together and in order */
745 REQ_F_NOWAIT_READ_BIT,
746 REQ_F_NOWAIT_WRITE_BIT,
749 /* not a real bit, just to check we're not overflowing the space */
755 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
756 /* drain existing IO first */
757 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
759 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
760 /* doesn't sever on completion < 0 */
761 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
763 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
764 /* IOSQE_BUFFER_SELECT */
765 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
767 /* fail rest of links */
768 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
769 /* on inflight list, should be cancelled and waited on exit reliably */
770 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
771 /* read/write uses file position */
772 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
773 /* must not punt to workers */
774 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
775 /* has or had linked timeout */
776 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
778 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
779 /* already went through poll handler */
780 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
781 /* buffer already selected */
782 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
783 /* completion is deferred through io_comp_state */
784 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
785 /* caller should reissue async */
786 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
787 /* supports async reads */
788 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
789 /* supports async writes */
790 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
792 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
793 /* has creds assigned */
794 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
795 /* skip refcounting if not set */
796 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
797 /* there is a linked timeout that has to be armed */
798 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
802 struct io_poll_iocb poll;
803 struct io_poll_iocb *double_poll;
806 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
808 struct io_task_work {
810 struct io_wq_work_node node;
811 struct llist_node fallback_node;
813 io_req_tw_func_t func;
817 IORING_RSRC_FILE = 0,
818 IORING_RSRC_BUFFER = 1,
822 * NOTE! Each of the iocb union members has the file pointer
823 * as the first entry in their struct definition. So you can
824 * access the file pointer through any of the sub-structs,
825 * or directly as just 'ki_filp' in this struct.
831 struct io_poll_iocb poll;
832 struct io_poll_update poll_update;
833 struct io_accept accept;
835 struct io_cancel cancel;
836 struct io_timeout timeout;
837 struct io_timeout_rem timeout_rem;
838 struct io_connect connect;
839 struct io_sr_msg sr_msg;
841 struct io_close close;
842 struct io_rsrc_update rsrc_update;
843 struct io_fadvise fadvise;
844 struct io_madvise madvise;
845 struct io_epoll epoll;
846 struct io_splice splice;
847 struct io_provide_buf pbuf;
848 struct io_statx statx;
849 struct io_shutdown shutdown;
850 struct io_rename rename;
851 struct io_unlink unlink;
852 struct io_mkdir mkdir;
853 struct io_symlink symlink;
854 struct io_hardlink hardlink;
855 /* use only after cleaning per-op data, see io_clean_op() */
856 struct io_completion compl;
859 /* opcode allocated if it needs to store data for async defer */
862 /* polled IO has completed */
868 struct io_ring_ctx *ctx;
871 struct task_struct *task;
874 struct io_kiocb *link;
875 struct percpu_ref *fixed_rsrc_refs;
877 /* used with ctx->iopoll_list with reads/writes */
878 struct list_head inflight_entry;
879 struct io_task_work io_task_work;
880 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
881 struct hlist_node hash_node;
882 struct async_poll *apoll;
883 struct io_wq_work work;
884 const struct cred *creds;
886 /* store used ubuf, so we can prevent reloading */
887 struct io_mapped_ubuf *imu;
890 struct io_tctx_node {
891 struct list_head ctx_node;
892 struct task_struct *task;
893 struct io_ring_ctx *ctx;
896 struct io_defer_entry {
897 struct list_head list;
898 struct io_kiocb *req;
903 /* needs req->file assigned */
904 unsigned needs_file : 1;
905 /* hash wq insertion if file is a regular file */
906 unsigned hash_reg_file : 1;
907 /* unbound wq insertion if file is a non-regular file */
908 unsigned unbound_nonreg_file : 1;
909 /* opcode is not supported by this kernel */
910 unsigned not_supported : 1;
911 /* set if opcode supports polled "wait" */
913 unsigned pollout : 1;
914 /* op supports buffer selection */
915 unsigned buffer_select : 1;
916 /* do prep async if is going to be punted */
917 unsigned needs_async_setup : 1;
918 /* should block plug */
920 /* size of async data needed, if any */
921 unsigned short async_size;
924 static const struct io_op_def io_op_defs[] = {
925 [IORING_OP_NOP] = {},
926 [IORING_OP_READV] = {
928 .unbound_nonreg_file = 1,
931 .needs_async_setup = 1,
933 .async_size = sizeof(struct io_async_rw),
935 [IORING_OP_WRITEV] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_setup = 1,
942 .async_size = sizeof(struct io_async_rw),
944 [IORING_OP_FSYNC] = {
947 [IORING_OP_READ_FIXED] = {
949 .unbound_nonreg_file = 1,
952 .async_size = sizeof(struct io_async_rw),
954 [IORING_OP_WRITE_FIXED] = {
957 .unbound_nonreg_file = 1,
960 .async_size = sizeof(struct io_async_rw),
962 [IORING_OP_POLL_ADD] = {
964 .unbound_nonreg_file = 1,
966 [IORING_OP_POLL_REMOVE] = {},
967 [IORING_OP_SYNC_FILE_RANGE] = {
970 [IORING_OP_SENDMSG] = {
972 .unbound_nonreg_file = 1,
974 .needs_async_setup = 1,
975 .async_size = sizeof(struct io_async_msghdr),
977 [IORING_OP_RECVMSG] = {
979 .unbound_nonreg_file = 1,
982 .needs_async_setup = 1,
983 .async_size = sizeof(struct io_async_msghdr),
985 [IORING_OP_TIMEOUT] = {
986 .async_size = sizeof(struct io_timeout_data),
988 [IORING_OP_TIMEOUT_REMOVE] = {
989 /* used by timeout updates' prep() */
991 [IORING_OP_ACCEPT] = {
993 .unbound_nonreg_file = 1,
996 [IORING_OP_ASYNC_CANCEL] = {},
997 [IORING_OP_LINK_TIMEOUT] = {
998 .async_size = sizeof(struct io_timeout_data),
1000 [IORING_OP_CONNECT] = {
1002 .unbound_nonreg_file = 1,
1004 .needs_async_setup = 1,
1005 .async_size = sizeof(struct io_async_connect),
1007 [IORING_OP_FALLOCATE] = {
1010 [IORING_OP_OPENAT] = {},
1011 [IORING_OP_CLOSE] = {},
1012 [IORING_OP_FILES_UPDATE] = {},
1013 [IORING_OP_STATX] = {},
1014 [IORING_OP_READ] = {
1016 .unbound_nonreg_file = 1,
1020 .async_size = sizeof(struct io_async_rw),
1022 [IORING_OP_WRITE] = {
1025 .unbound_nonreg_file = 1,
1028 .async_size = sizeof(struct io_async_rw),
1030 [IORING_OP_FADVISE] = {
1033 [IORING_OP_MADVISE] = {},
1034 [IORING_OP_SEND] = {
1036 .unbound_nonreg_file = 1,
1039 [IORING_OP_RECV] = {
1041 .unbound_nonreg_file = 1,
1045 [IORING_OP_OPENAT2] = {
1047 [IORING_OP_EPOLL_CTL] = {
1048 .unbound_nonreg_file = 1,
1050 [IORING_OP_SPLICE] = {
1053 .unbound_nonreg_file = 1,
1055 [IORING_OP_PROVIDE_BUFFERS] = {},
1056 [IORING_OP_REMOVE_BUFFERS] = {},
1060 .unbound_nonreg_file = 1,
1062 [IORING_OP_SHUTDOWN] = {
1065 [IORING_OP_RENAMEAT] = {},
1066 [IORING_OP_UNLINKAT] = {},
1067 [IORING_OP_MKDIRAT] = {},
1068 [IORING_OP_SYMLINKAT] = {},
1069 [IORING_OP_LINKAT] = {},
1072 /* requests with any of those set should undergo io_disarm_next() */
1073 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1075 static bool io_disarm_next(struct io_kiocb *req);
1076 static void io_uring_del_tctx_node(unsigned long index);
1077 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1078 struct task_struct *task,
1080 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1082 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1083 long res, unsigned int cflags);
1084 static void io_put_req(struct io_kiocb *req);
1085 static void io_put_req_deferred(struct io_kiocb *req);
1086 static void io_dismantle_req(struct io_kiocb *req);
1087 static void io_queue_linked_timeout(struct io_kiocb *req);
1088 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1089 struct io_uring_rsrc_update2 *up,
1091 static void io_clean_op(struct io_kiocb *req);
1092 static struct file *io_file_get(struct io_ring_ctx *ctx,
1093 struct io_kiocb *req, int fd, bool fixed);
1094 static void __io_queue_sqe(struct io_kiocb *req);
1095 static void io_rsrc_put_work(struct work_struct *work);
1097 static void io_req_task_queue(struct io_kiocb *req);
1098 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1099 static int io_req_prep_async(struct io_kiocb *req);
1101 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1102 unsigned int issue_flags, u32 slot_index);
1103 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1105 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1107 static struct kmem_cache *req_cachep;
1109 static const struct file_operations io_uring_fops;
1111 struct sock *io_uring_get_socket(struct file *file)
1113 #if defined(CONFIG_UNIX)
1114 if (file->f_op == &io_uring_fops) {
1115 struct io_ring_ctx *ctx = file->private_data;
1117 return ctx->ring_sock->sk;
1122 EXPORT_SYMBOL(io_uring_get_socket);
1124 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1127 mutex_lock(&ctx->uring_lock);
1132 #define io_for_each_link(pos, head) \
1133 for (pos = (head); pos; pos = pos->link)
1136 * Shamelessly stolen from the mm implementation of page reference checking,
1137 * see commit f958d7b528b1 for details.
1139 #define req_ref_zero_or_close_to_overflow(req) \
1140 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1142 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1144 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1145 return atomic_inc_not_zero(&req->refs);
1148 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1150 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1153 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1154 return atomic_dec_and_test(&req->refs);
1157 static inline void req_ref_put(struct io_kiocb *req)
1159 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1160 WARN_ON_ONCE(req_ref_put_and_test(req));
1163 static inline void req_ref_get(struct io_kiocb *req)
1165 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1166 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1167 atomic_inc(&req->refs);
1170 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1172 if (!(req->flags & REQ_F_REFCOUNT)) {
1173 req->flags |= REQ_F_REFCOUNT;
1174 atomic_set(&req->refs, nr);
1178 static inline void io_req_set_refcount(struct io_kiocb *req)
1180 __io_req_set_refcount(req, 1);
1183 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1185 struct io_ring_ctx *ctx = req->ctx;
1187 if (!req->fixed_rsrc_refs) {
1188 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1189 percpu_ref_get(req->fixed_rsrc_refs);
1193 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1195 bool got = percpu_ref_tryget(ref);
1197 /* already at zero, wait for ->release() */
1199 wait_for_completion(compl);
1200 percpu_ref_resurrect(ref);
1202 percpu_ref_put(ref);
1205 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1207 __must_hold(&req->ctx->timeout_lock)
1209 struct io_kiocb *req;
1211 if (task && head->task != task)
1216 io_for_each_link(req, head) {
1217 if (req->flags & REQ_F_INFLIGHT)
1223 static bool io_match_linked(struct io_kiocb *head)
1225 struct io_kiocb *req;
1227 io_for_each_link(req, head) {
1228 if (req->flags & REQ_F_INFLIGHT)
1235 * As io_match_task() but protected against racing with linked timeouts.
1236 * User must not hold timeout_lock.
1238 static bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
1243 if (task && head->task != task)
1248 if (head->flags & REQ_F_LINK_TIMEOUT) {
1249 struct io_ring_ctx *ctx = head->ctx;
1251 /* protect against races with linked timeouts */
1252 spin_lock_irq(&ctx->timeout_lock);
1253 matched = io_match_linked(head);
1254 spin_unlock_irq(&ctx->timeout_lock);
1256 matched = io_match_linked(head);
1261 static inline void req_set_fail(struct io_kiocb *req)
1263 req->flags |= REQ_F_FAIL;
1266 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1272 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1274 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1276 complete(&ctx->ref_comp);
1279 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1281 return !req->timeout.off;
1284 static void io_fallback_req_func(struct work_struct *work)
1286 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1287 fallback_work.work);
1288 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1289 struct io_kiocb *req, *tmp;
1290 bool locked = false;
1292 percpu_ref_get(&ctx->refs);
1293 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1294 req->io_task_work.func(req, &locked);
1297 if (ctx->submit_state.compl_nr)
1298 io_submit_flush_completions(ctx);
1299 mutex_unlock(&ctx->uring_lock);
1301 percpu_ref_put(&ctx->refs);
1305 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1307 struct io_ring_ctx *ctx;
1310 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1315 * Use 5 bits less than the max cq entries, that should give us around
1316 * 32 entries per hash list if totally full and uniformly spread.
1318 hash_bits = ilog2(p->cq_entries);
1322 ctx->cancel_hash_bits = hash_bits;
1323 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1325 if (!ctx->cancel_hash)
1327 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1329 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1330 if (!ctx->dummy_ubuf)
1332 /* set invalid range, so io_import_fixed() fails meeting it */
1333 ctx->dummy_ubuf->ubuf = -1UL;
1335 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1336 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1339 ctx->flags = p->flags;
1340 init_waitqueue_head(&ctx->sqo_sq_wait);
1341 INIT_LIST_HEAD(&ctx->sqd_list);
1342 init_waitqueue_head(&ctx->poll_wait);
1343 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1344 init_completion(&ctx->ref_comp);
1345 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1346 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1347 mutex_init(&ctx->uring_lock);
1348 init_waitqueue_head(&ctx->cq_wait);
1349 spin_lock_init(&ctx->completion_lock);
1350 spin_lock_init(&ctx->timeout_lock);
1351 INIT_LIST_HEAD(&ctx->iopoll_list);
1352 INIT_LIST_HEAD(&ctx->defer_list);
1353 INIT_LIST_HEAD(&ctx->timeout_list);
1354 INIT_LIST_HEAD(&ctx->ltimeout_list);
1355 spin_lock_init(&ctx->rsrc_ref_lock);
1356 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1357 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1358 init_llist_head(&ctx->rsrc_put_llist);
1359 INIT_LIST_HEAD(&ctx->tctx_list);
1360 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1361 INIT_LIST_HEAD(&ctx->locked_free_list);
1362 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1365 kfree(ctx->dummy_ubuf);
1366 kfree(ctx->cancel_hash);
1371 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1373 struct io_rings *r = ctx->rings;
1375 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1379 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1381 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1382 struct io_ring_ctx *ctx = req->ctx;
1384 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1390 #define FFS_ASYNC_READ 0x1UL
1391 #define FFS_ASYNC_WRITE 0x2UL
1393 #define FFS_ISREG 0x4UL
1395 #define FFS_ISREG 0x0UL
1397 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1399 static inline bool io_req_ffs_set(struct io_kiocb *req)
1401 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1404 static void io_req_track_inflight(struct io_kiocb *req)
1406 if (!(req->flags & REQ_F_INFLIGHT)) {
1407 req->flags |= REQ_F_INFLIGHT;
1408 atomic_inc(¤t->io_uring->inflight_tracked);
1412 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1414 if (WARN_ON_ONCE(!req->link))
1417 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1418 req->flags |= REQ_F_LINK_TIMEOUT;
1420 /* linked timeouts should have two refs once prep'ed */
1421 io_req_set_refcount(req);
1422 __io_req_set_refcount(req->link, 2);
1426 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1428 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1430 return __io_prep_linked_timeout(req);
1433 static void io_prep_async_work(struct io_kiocb *req)
1435 const struct io_op_def *def = &io_op_defs[req->opcode];
1436 struct io_ring_ctx *ctx = req->ctx;
1438 if (!(req->flags & REQ_F_CREDS)) {
1439 req->flags |= REQ_F_CREDS;
1440 req->creds = get_current_cred();
1443 req->work.list.next = NULL;
1444 req->work.flags = 0;
1445 if (req->flags & REQ_F_FORCE_ASYNC)
1446 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1448 if (req->flags & REQ_F_ISREG) {
1449 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1450 io_wq_hash_work(&req->work, file_inode(req->file));
1451 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1452 if (def->unbound_nonreg_file)
1453 req->work.flags |= IO_WQ_WORK_UNBOUND;
1457 static void io_prep_async_link(struct io_kiocb *req)
1459 struct io_kiocb *cur;
1461 if (req->flags & REQ_F_LINK_TIMEOUT) {
1462 struct io_ring_ctx *ctx = req->ctx;
1464 spin_lock_irq(&ctx->timeout_lock);
1465 io_for_each_link(cur, req)
1466 io_prep_async_work(cur);
1467 spin_unlock_irq(&ctx->timeout_lock);
1469 io_for_each_link(cur, req)
1470 io_prep_async_work(cur);
1474 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1476 struct io_ring_ctx *ctx = req->ctx;
1477 struct io_kiocb *link = io_prep_linked_timeout(req);
1478 struct io_uring_task *tctx = req->task->io_uring;
1480 /* must not take the lock, NULL it as a precaution */
1484 BUG_ON(!tctx->io_wq);
1486 /* init ->work of the whole link before punting */
1487 io_prep_async_link(req);
1490 * Not expected to happen, but if we do have a bug where this _can_
1491 * happen, catch it here and ensure the request is marked as
1492 * canceled. That will make io-wq go through the usual work cancel
1493 * procedure rather than attempt to run this request (or create a new
1496 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1497 req->work.flags |= IO_WQ_WORK_CANCEL;
1499 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1500 &req->work, req->flags);
1501 io_wq_enqueue(tctx->io_wq, &req->work);
1503 io_queue_linked_timeout(link);
1506 static void io_kill_timeout(struct io_kiocb *req, int status)
1507 __must_hold(&req->ctx->completion_lock)
1508 __must_hold(&req->ctx->timeout_lock)
1510 struct io_timeout_data *io = req->async_data;
1512 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1515 atomic_set(&req->ctx->cq_timeouts,
1516 atomic_read(&req->ctx->cq_timeouts) + 1);
1517 list_del_init(&req->timeout.list);
1518 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1519 io_put_req_deferred(req);
1523 static void io_queue_deferred(struct io_ring_ctx *ctx)
1525 while (!list_empty(&ctx->defer_list)) {
1526 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1527 struct io_defer_entry, list);
1529 if (req_need_defer(de->req, de->seq))
1531 list_del_init(&de->list);
1532 io_req_task_queue(de->req);
1537 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1538 __must_hold(&ctx->completion_lock)
1540 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1541 struct io_kiocb *req, *tmp;
1543 spin_lock_irq(&ctx->timeout_lock);
1544 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1545 u32 events_needed, events_got;
1547 if (io_is_timeout_noseq(req))
1551 * Since seq can easily wrap around over time, subtract
1552 * the last seq at which timeouts were flushed before comparing.
1553 * Assuming not more than 2^31-1 events have happened since,
1554 * these subtractions won't have wrapped, so we can check if
1555 * target is in [last_seq, current_seq] by comparing the two.
1557 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1558 events_got = seq - ctx->cq_last_tm_flush;
1559 if (events_got < events_needed)
1562 io_kill_timeout(req, 0);
1564 ctx->cq_last_tm_flush = seq;
1565 spin_unlock_irq(&ctx->timeout_lock);
1568 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1570 if (ctx->off_timeout_used)
1571 io_flush_timeouts(ctx);
1572 if (ctx->drain_active)
1573 io_queue_deferred(ctx);
1576 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1578 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1579 __io_commit_cqring_flush(ctx);
1580 /* order cqe stores with ring update */
1581 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1584 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1586 struct io_rings *r = ctx->rings;
1588 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1591 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1593 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1596 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1598 struct io_rings *rings = ctx->rings;
1599 unsigned tail, mask = ctx->cq_entries - 1;
1602 * writes to the cq entry need to come after reading head; the
1603 * control dependency is enough as we're using WRITE_ONCE to
1606 if (__io_cqring_events(ctx) == ctx->cq_entries)
1609 tail = ctx->cached_cq_tail++;
1610 return &rings->cqes[tail & mask];
1613 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1615 if (likely(!ctx->cq_ev_fd))
1617 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1619 return !ctx->eventfd_async || io_wq_current_is_worker();
1623 * This should only get called when at least one event has been posted.
1624 * Some applications rely on the eventfd notification count only changing
1625 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1626 * 1:1 relationship between how many times this function is called (and
1627 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1629 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1632 * wake_up_all() may seem excessive, but io_wake_function() and
1633 * io_should_wake() handle the termination of the loop and only
1634 * wake as many waiters as we need to.
1636 if (wq_has_sleeper(&ctx->cq_wait))
1637 wake_up_all(&ctx->cq_wait);
1638 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1639 wake_up(&ctx->sq_data->wait);
1640 if (io_should_trigger_evfd(ctx))
1641 eventfd_signal(ctx->cq_ev_fd, 1);
1642 if (waitqueue_active(&ctx->poll_wait))
1643 wake_up_interruptible(&ctx->poll_wait);
1646 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1648 /* see waitqueue_active() comment */
1651 if (ctx->flags & IORING_SETUP_SQPOLL) {
1652 if (waitqueue_active(&ctx->cq_wait))
1653 wake_up_all(&ctx->cq_wait);
1655 if (io_should_trigger_evfd(ctx))
1656 eventfd_signal(ctx->cq_ev_fd, 1);
1657 if (waitqueue_active(&ctx->poll_wait))
1658 wake_up_interruptible(&ctx->poll_wait);
1661 /* Returns true if there are no backlogged entries after the flush */
1662 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1664 bool all_flushed, posted;
1666 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1670 spin_lock(&ctx->completion_lock);
1671 while (!list_empty(&ctx->cq_overflow_list)) {
1672 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1673 struct io_overflow_cqe *ocqe;
1677 ocqe = list_first_entry(&ctx->cq_overflow_list,
1678 struct io_overflow_cqe, list);
1680 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1682 io_account_cq_overflow(ctx);
1685 list_del(&ocqe->list);
1689 all_flushed = list_empty(&ctx->cq_overflow_list);
1691 clear_bit(0, &ctx->check_cq_overflow);
1692 WRITE_ONCE(ctx->rings->sq_flags,
1693 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1697 io_commit_cqring(ctx);
1698 spin_unlock(&ctx->completion_lock);
1700 io_cqring_ev_posted(ctx);
1704 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1708 if (test_bit(0, &ctx->check_cq_overflow)) {
1709 /* iopoll syncs against uring_lock, not completion_lock */
1710 if (ctx->flags & IORING_SETUP_IOPOLL)
1711 mutex_lock(&ctx->uring_lock);
1712 ret = __io_cqring_overflow_flush(ctx, false);
1713 if (ctx->flags & IORING_SETUP_IOPOLL)
1714 mutex_unlock(&ctx->uring_lock);
1720 /* must to be called somewhat shortly after putting a request */
1721 static inline void io_put_task(struct task_struct *task, int nr)
1723 struct io_uring_task *tctx = task->io_uring;
1725 if (likely(task == current)) {
1726 tctx->cached_refs += nr;
1728 percpu_counter_sub(&tctx->inflight, nr);
1729 if (unlikely(atomic_read(&tctx->in_idle)))
1730 wake_up(&tctx->wait);
1731 put_task_struct_many(task, nr);
1735 static void io_task_refs_refill(struct io_uring_task *tctx)
1737 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1739 percpu_counter_add(&tctx->inflight, refill);
1740 refcount_add(refill, ¤t->usage);
1741 tctx->cached_refs += refill;
1744 static inline void io_get_task_refs(int nr)
1746 struct io_uring_task *tctx = current->io_uring;
1748 tctx->cached_refs -= nr;
1749 if (unlikely(tctx->cached_refs < 0))
1750 io_task_refs_refill(tctx);
1753 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
1755 struct io_uring_task *tctx = task->io_uring;
1756 unsigned int refs = tctx->cached_refs;
1759 tctx->cached_refs = 0;
1760 percpu_counter_sub(&tctx->inflight, refs);
1761 put_task_struct_many(task, refs);
1765 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1766 long res, unsigned int cflags)
1768 struct io_overflow_cqe *ocqe;
1770 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1773 * If we're in ring overflow flush mode, or in task cancel mode,
1774 * or cannot allocate an overflow entry, then we need to drop it
1777 io_account_cq_overflow(ctx);
1780 if (list_empty(&ctx->cq_overflow_list)) {
1781 set_bit(0, &ctx->check_cq_overflow);
1782 WRITE_ONCE(ctx->rings->sq_flags,
1783 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1786 ocqe->cqe.user_data = user_data;
1787 ocqe->cqe.res = res;
1788 ocqe->cqe.flags = cflags;
1789 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1793 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1794 long res, unsigned int cflags)
1796 struct io_uring_cqe *cqe;
1798 trace_io_uring_complete(ctx, user_data, res, cflags);
1801 * If we can't get a cq entry, userspace overflowed the
1802 * submission (by quite a lot). Increment the overflow count in
1805 cqe = io_get_cqe(ctx);
1807 WRITE_ONCE(cqe->user_data, user_data);
1808 WRITE_ONCE(cqe->res, res);
1809 WRITE_ONCE(cqe->flags, cflags);
1812 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1815 /* not as hot to bloat with inlining */
1816 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1817 long res, unsigned int cflags)
1819 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1822 static void io_req_complete_post(struct io_kiocb *req, long res,
1823 unsigned int cflags)
1825 struct io_ring_ctx *ctx = req->ctx;
1827 spin_lock(&ctx->completion_lock);
1828 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1830 * If we're the last reference to this request, add to our locked
1833 if (req_ref_put_and_test(req)) {
1834 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1835 if (req->flags & IO_DISARM_MASK)
1836 io_disarm_next(req);
1838 io_req_task_queue(req->link);
1842 io_dismantle_req(req);
1843 io_put_task(req->task, 1);
1844 list_add(&req->inflight_entry, &ctx->locked_free_list);
1845 ctx->locked_free_nr++;
1847 if (!percpu_ref_tryget(&ctx->refs))
1850 io_commit_cqring(ctx);
1851 spin_unlock(&ctx->completion_lock);
1854 io_cqring_ev_posted(ctx);
1855 percpu_ref_put(&ctx->refs);
1859 static inline bool io_req_needs_clean(struct io_kiocb *req)
1861 return req->flags & IO_REQ_CLEAN_FLAGS;
1864 static void io_req_complete_state(struct io_kiocb *req, long res,
1865 unsigned int cflags)
1867 if (io_req_needs_clean(req))
1870 req->compl.cflags = cflags;
1871 req->flags |= REQ_F_COMPLETE_INLINE;
1874 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1875 long res, unsigned cflags)
1877 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1878 io_req_complete_state(req, res, cflags);
1880 io_req_complete_post(req, res, cflags);
1883 static inline void io_req_complete(struct io_kiocb *req, long res)
1885 __io_req_complete(req, 0, res, 0);
1888 static void io_req_complete_failed(struct io_kiocb *req, long res)
1891 io_req_complete_post(req, res, 0);
1894 static void io_req_complete_fail_submit(struct io_kiocb *req)
1897 * We don't submit, fail them all, for that replace hardlinks with
1898 * normal links. Extra REQ_F_LINK is tolerated.
1900 req->flags &= ~REQ_F_HARDLINK;
1901 req->flags |= REQ_F_LINK;
1902 io_req_complete_failed(req, req->result);
1906 * Don't initialise the fields below on every allocation, but do that in
1907 * advance and keep them valid across allocations.
1909 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1913 req->async_data = NULL;
1914 /* not necessary, but safer to zero */
1918 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1919 struct io_submit_state *state)
1921 spin_lock(&ctx->completion_lock);
1922 list_splice_init(&ctx->locked_free_list, &state->free_list);
1923 ctx->locked_free_nr = 0;
1924 spin_unlock(&ctx->completion_lock);
1927 /* Returns true IFF there are requests in the cache */
1928 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1930 struct io_submit_state *state = &ctx->submit_state;
1934 * If we have more than a batch's worth of requests in our IRQ side
1935 * locked cache, grab the lock and move them over to our submission
1938 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1939 io_flush_cached_locked_reqs(ctx, state);
1941 nr = state->free_reqs;
1942 while (!list_empty(&state->free_list)) {
1943 struct io_kiocb *req = list_first_entry(&state->free_list,
1944 struct io_kiocb, inflight_entry);
1946 list_del(&req->inflight_entry);
1947 state->reqs[nr++] = req;
1948 if (nr == ARRAY_SIZE(state->reqs))
1952 state->free_reqs = nr;
1957 * A request might get retired back into the request caches even before opcode
1958 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1959 * Because of that, io_alloc_req() should be called only under ->uring_lock
1960 * and with extra caution to not get a request that is still worked on.
1962 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1963 __must_hold(&ctx->uring_lock)
1965 struct io_submit_state *state = &ctx->submit_state;
1966 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1969 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1971 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1974 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1978 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1979 * retry single alloc to be on the safe side.
1981 if (unlikely(ret <= 0)) {
1982 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1983 if (!state->reqs[0])
1988 for (i = 0; i < ret; i++)
1989 io_preinit_req(state->reqs[i], ctx);
1990 state->free_reqs = ret;
1993 return state->reqs[state->free_reqs];
1996 static inline void io_put_file(struct file *file)
2002 static void io_dismantle_req(struct io_kiocb *req)
2004 unsigned int flags = req->flags;
2006 if (io_req_needs_clean(req))
2008 if (!(flags & REQ_F_FIXED_FILE))
2009 io_put_file(req->file);
2010 if (req->fixed_rsrc_refs)
2011 percpu_ref_put(req->fixed_rsrc_refs);
2012 if (req->async_data) {
2013 kfree(req->async_data);
2014 req->async_data = NULL;
2018 static void __io_free_req(struct io_kiocb *req)
2020 struct io_ring_ctx *ctx = req->ctx;
2022 io_dismantle_req(req);
2023 io_put_task(req->task, 1);
2025 spin_lock(&ctx->completion_lock);
2026 list_add(&req->inflight_entry, &ctx->locked_free_list);
2027 ctx->locked_free_nr++;
2028 spin_unlock(&ctx->completion_lock);
2030 percpu_ref_put(&ctx->refs);
2033 static inline void io_remove_next_linked(struct io_kiocb *req)
2035 struct io_kiocb *nxt = req->link;
2037 req->link = nxt->link;
2041 static bool io_kill_linked_timeout(struct io_kiocb *req)
2042 __must_hold(&req->ctx->completion_lock)
2043 __must_hold(&req->ctx->timeout_lock)
2045 struct io_kiocb *link = req->link;
2047 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2048 struct io_timeout_data *io = link->async_data;
2050 io_remove_next_linked(req);
2051 link->timeout.head = NULL;
2052 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2053 list_del(&link->timeout.list);
2054 io_cqring_fill_event(link->ctx, link->user_data,
2056 io_put_req_deferred(link);
2063 static void io_fail_links(struct io_kiocb *req)
2064 __must_hold(&req->ctx->completion_lock)
2066 struct io_kiocb *nxt, *link = req->link;
2070 long res = -ECANCELED;
2072 if (link->flags & REQ_F_FAIL)
2078 trace_io_uring_fail_link(req, link);
2079 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2080 io_put_req_deferred(link);
2085 static bool io_disarm_next(struct io_kiocb *req)
2086 __must_hold(&req->ctx->completion_lock)
2088 bool posted = false;
2090 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2091 struct io_kiocb *link = req->link;
2093 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2094 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2095 io_remove_next_linked(req);
2096 io_cqring_fill_event(link->ctx, link->user_data,
2098 io_put_req_deferred(link);
2101 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2102 struct io_ring_ctx *ctx = req->ctx;
2104 spin_lock_irq(&ctx->timeout_lock);
2105 posted = io_kill_linked_timeout(req);
2106 spin_unlock_irq(&ctx->timeout_lock);
2108 if (unlikely((req->flags & REQ_F_FAIL) &&
2109 !(req->flags & REQ_F_HARDLINK))) {
2110 posted |= (req->link != NULL);
2116 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2118 struct io_kiocb *nxt;
2121 * If LINK is set, we have dependent requests in this chain. If we
2122 * didn't fail this request, queue the first one up, moving any other
2123 * dependencies to the next request. In case of failure, fail the rest
2126 if (req->flags & IO_DISARM_MASK) {
2127 struct io_ring_ctx *ctx = req->ctx;
2130 spin_lock(&ctx->completion_lock);
2131 posted = io_disarm_next(req);
2133 io_commit_cqring(req->ctx);
2134 spin_unlock(&ctx->completion_lock);
2136 io_cqring_ev_posted(ctx);
2143 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2145 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2147 return __io_req_find_next(req);
2150 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2155 if (ctx->submit_state.compl_nr)
2156 io_submit_flush_completions(ctx);
2157 mutex_unlock(&ctx->uring_lock);
2160 percpu_ref_put(&ctx->refs);
2163 static void tctx_task_work(struct callback_head *cb)
2165 bool locked = false;
2166 struct io_ring_ctx *ctx = NULL;
2167 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2171 struct io_wq_work_node *node;
2173 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2174 io_submit_flush_completions(ctx);
2176 spin_lock_irq(&tctx->task_lock);
2177 node = tctx->task_list.first;
2178 INIT_WQ_LIST(&tctx->task_list);
2180 tctx->task_running = false;
2181 spin_unlock_irq(&tctx->task_lock);
2186 struct io_wq_work_node *next = node->next;
2187 struct io_kiocb *req = container_of(node, struct io_kiocb,
2190 if (req->ctx != ctx) {
2191 ctx_flush_and_put(ctx, &locked);
2193 /* if not contended, grab and improve batching */
2194 locked = mutex_trylock(&ctx->uring_lock);
2195 percpu_ref_get(&ctx->refs);
2197 req->io_task_work.func(req, &locked);
2204 ctx_flush_and_put(ctx, &locked);
2206 /* relaxed read is enough as only the task itself sets ->in_idle */
2207 if (unlikely(atomic_read(&tctx->in_idle)))
2208 io_uring_drop_tctx_refs(current);
2211 static void io_req_task_work_add(struct io_kiocb *req)
2213 struct task_struct *tsk = req->task;
2214 struct io_uring_task *tctx = tsk->io_uring;
2215 enum task_work_notify_mode notify;
2216 struct io_wq_work_node *node;
2217 unsigned long flags;
2220 WARN_ON_ONCE(!tctx);
2222 spin_lock_irqsave(&tctx->task_lock, flags);
2223 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2224 running = tctx->task_running;
2226 tctx->task_running = true;
2227 spin_unlock_irqrestore(&tctx->task_lock, flags);
2229 /* task_work already pending, we're done */
2234 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2235 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2236 * processing task_work. There's no reliable way to tell if TWA_RESUME
2239 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2240 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2241 wake_up_process(tsk);
2245 spin_lock_irqsave(&tctx->task_lock, flags);
2246 tctx->task_running = false;
2247 node = tctx->task_list.first;
2248 INIT_WQ_LIST(&tctx->task_list);
2249 spin_unlock_irqrestore(&tctx->task_lock, flags);
2252 req = container_of(node, struct io_kiocb, io_task_work.node);
2254 if (llist_add(&req->io_task_work.fallback_node,
2255 &req->ctx->fallback_llist))
2256 schedule_delayed_work(&req->ctx->fallback_work, 1);
2260 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2262 struct io_ring_ctx *ctx = req->ctx;
2264 /* not needed for normal modes, but SQPOLL depends on it */
2265 io_tw_lock(ctx, locked);
2266 io_req_complete_failed(req, req->result);
2269 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2271 struct io_ring_ctx *ctx = req->ctx;
2273 io_tw_lock(ctx, locked);
2274 /* req->task == current here, checking PF_EXITING is safe */
2275 if (likely(!(req->task->flags & PF_EXITING)))
2276 __io_queue_sqe(req);
2278 io_req_complete_failed(req, -EFAULT);
2281 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2284 req->io_task_work.func = io_req_task_cancel;
2285 io_req_task_work_add(req);
2288 static void io_req_task_queue(struct io_kiocb *req)
2290 req->io_task_work.func = io_req_task_submit;
2291 io_req_task_work_add(req);
2294 static void io_req_task_queue_reissue(struct io_kiocb *req)
2296 req->io_task_work.func = io_queue_async_work;
2297 io_req_task_work_add(req);
2300 static inline void io_queue_next(struct io_kiocb *req)
2302 struct io_kiocb *nxt = io_req_find_next(req);
2305 io_req_task_queue(nxt);
2308 static void io_free_req(struct io_kiocb *req)
2314 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2320 struct task_struct *task;
2325 static inline void io_init_req_batch(struct req_batch *rb)
2332 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2333 struct req_batch *rb)
2336 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2338 io_put_task(rb->task, rb->task_refs);
2341 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2342 struct io_submit_state *state)
2345 io_dismantle_req(req);
2347 if (req->task != rb->task) {
2349 io_put_task(rb->task, rb->task_refs);
2350 rb->task = req->task;
2356 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2357 state->reqs[state->free_reqs++] = req;
2359 list_add(&req->inflight_entry, &state->free_list);
2362 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2363 __must_hold(&ctx->uring_lock)
2365 struct io_submit_state *state = &ctx->submit_state;
2366 int i, nr = state->compl_nr;
2367 struct req_batch rb;
2369 spin_lock(&ctx->completion_lock);
2370 for (i = 0; i < nr; i++) {
2371 struct io_kiocb *req = state->compl_reqs[i];
2373 __io_cqring_fill_event(ctx, req->user_data, req->result,
2376 io_commit_cqring(ctx);
2377 spin_unlock(&ctx->completion_lock);
2378 io_cqring_ev_posted(ctx);
2380 io_init_req_batch(&rb);
2381 for (i = 0; i < nr; i++) {
2382 struct io_kiocb *req = state->compl_reqs[i];
2384 if (req_ref_put_and_test(req))
2385 io_req_free_batch(&rb, req, &ctx->submit_state);
2388 io_req_free_batch_finish(ctx, &rb);
2389 state->compl_nr = 0;
2393 * Drop reference to request, return next in chain (if there is one) if this
2394 * was the last reference to this request.
2396 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2398 struct io_kiocb *nxt = NULL;
2400 if (req_ref_put_and_test(req)) {
2401 nxt = io_req_find_next(req);
2407 static inline void io_put_req(struct io_kiocb *req)
2409 if (req_ref_put_and_test(req))
2413 static inline void io_put_req_deferred(struct io_kiocb *req)
2415 if (req_ref_put_and_test(req)) {
2416 req->io_task_work.func = io_free_req_work;
2417 io_req_task_work_add(req);
2421 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2423 /* See comment at the top of this file */
2425 return __io_cqring_events(ctx);
2428 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2430 struct io_rings *rings = ctx->rings;
2432 /* make sure SQ entry isn't read before tail */
2433 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2436 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2438 unsigned int cflags;
2440 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2441 cflags |= IORING_CQE_F_BUFFER;
2442 req->flags &= ~REQ_F_BUFFER_SELECTED;
2447 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2449 struct io_buffer *kbuf;
2451 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2453 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2454 return io_put_kbuf(req, kbuf);
2457 static inline bool io_run_task_work(void)
2459 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2460 __set_current_state(TASK_RUNNING);
2461 tracehook_notify_signal();
2469 * Find and free completed poll iocbs
2471 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2472 struct list_head *done)
2474 struct req_batch rb;
2475 struct io_kiocb *req;
2477 /* order with ->result store in io_complete_rw_iopoll() */
2480 io_init_req_batch(&rb);
2481 while (!list_empty(done)) {
2482 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2483 list_del(&req->inflight_entry);
2485 __io_cqring_fill_event(ctx, req->user_data, req->result,
2486 io_put_rw_kbuf(req));
2489 if (req_ref_put_and_test(req))
2490 io_req_free_batch(&rb, req, &ctx->submit_state);
2493 io_commit_cqring(ctx);
2494 io_cqring_ev_posted_iopoll(ctx);
2495 io_req_free_batch_finish(ctx, &rb);
2498 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2501 struct io_kiocb *req, *tmp;
2506 * Only spin for completions if we don't have multiple devices hanging
2507 * off our complete list, and we're under the requested amount.
2509 spin = !ctx->poll_multi_queue && *nr_events < min;
2511 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2512 struct kiocb *kiocb = &req->rw.kiocb;
2516 * Move completed and retryable entries to our local lists.
2517 * If we find a request that requires polling, break out
2518 * and complete those lists first, if we have entries there.
2520 if (READ_ONCE(req->iopoll_completed)) {
2521 list_move_tail(&req->inflight_entry, &done);
2524 if (!list_empty(&done))
2527 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2528 if (unlikely(ret < 0))
2533 /* iopoll may have completed current req */
2534 if (READ_ONCE(req->iopoll_completed))
2535 list_move_tail(&req->inflight_entry, &done);
2538 if (!list_empty(&done))
2539 io_iopoll_complete(ctx, nr_events, &done);
2545 * We can't just wait for polled events to come to us, we have to actively
2546 * find and complete them.
2548 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2550 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2553 mutex_lock(&ctx->uring_lock);
2554 while (!list_empty(&ctx->iopoll_list)) {
2555 unsigned int nr_events = 0;
2557 io_do_iopoll(ctx, &nr_events, 0);
2559 /* let it sleep and repeat later if can't complete a request */
2563 * Ensure we allow local-to-the-cpu processing to take place,
2564 * in this case we need to ensure that we reap all events.
2565 * Also let task_work, etc. to progress by releasing the mutex
2567 if (need_resched()) {
2568 mutex_unlock(&ctx->uring_lock);
2570 mutex_lock(&ctx->uring_lock);
2573 mutex_unlock(&ctx->uring_lock);
2576 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2578 unsigned int nr_events = 0;
2582 * We disallow the app entering submit/complete with polling, but we
2583 * still need to lock the ring to prevent racing with polled issue
2584 * that got punted to a workqueue.
2586 mutex_lock(&ctx->uring_lock);
2588 * Don't enter poll loop if we already have events pending.
2589 * If we do, we can potentially be spinning for commands that
2590 * already triggered a CQE (eg in error).
2592 if (test_bit(0, &ctx->check_cq_overflow))
2593 __io_cqring_overflow_flush(ctx, false);
2594 if (io_cqring_events(ctx))
2598 * If a submit got punted to a workqueue, we can have the
2599 * application entering polling for a command before it gets
2600 * issued. That app will hold the uring_lock for the duration
2601 * of the poll right here, so we need to take a breather every
2602 * now and then to ensure that the issue has a chance to add
2603 * the poll to the issued list. Otherwise we can spin here
2604 * forever, while the workqueue is stuck trying to acquire the
2607 if (list_empty(&ctx->iopoll_list)) {
2608 u32 tail = ctx->cached_cq_tail;
2610 mutex_unlock(&ctx->uring_lock);
2612 mutex_lock(&ctx->uring_lock);
2614 /* some requests don't go through iopoll_list */
2615 if (tail != ctx->cached_cq_tail ||
2616 list_empty(&ctx->iopoll_list))
2619 ret = io_do_iopoll(ctx, &nr_events, min);
2620 } while (!ret && nr_events < min && !need_resched());
2622 mutex_unlock(&ctx->uring_lock);
2626 static void kiocb_end_write(struct io_kiocb *req)
2629 * Tell lockdep we inherited freeze protection from submission
2632 if (req->flags & REQ_F_ISREG) {
2633 struct super_block *sb = file_inode(req->file)->i_sb;
2635 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2641 static bool io_resubmit_prep(struct io_kiocb *req)
2643 struct io_async_rw *rw = req->async_data;
2646 return !io_req_prep_async(req);
2647 iov_iter_restore(&rw->iter, &rw->iter_state);
2651 static bool io_rw_should_reissue(struct io_kiocb *req)
2653 umode_t mode = file_inode(req->file)->i_mode;
2654 struct io_ring_ctx *ctx = req->ctx;
2656 if (!S_ISBLK(mode) && !S_ISREG(mode))
2658 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2659 !(ctx->flags & IORING_SETUP_IOPOLL)))
2662 * If ref is dying, we might be running poll reap from the exit work.
2663 * Don't attempt to reissue from that path, just let it fail with
2666 if (percpu_ref_is_dying(&ctx->refs))
2669 * Play it safe and assume not safe to re-import and reissue if we're
2670 * not in the original thread group (or in task context).
2672 if (!same_thread_group(req->task, current) || !in_task())
2677 static bool io_resubmit_prep(struct io_kiocb *req)
2681 static bool io_rw_should_reissue(struct io_kiocb *req)
2687 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2689 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2690 kiocb_end_write(req);
2691 if (res != req->result) {
2692 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2693 io_rw_should_reissue(req)) {
2694 req->flags |= REQ_F_REISSUE;
2703 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2705 unsigned int cflags = io_put_rw_kbuf(req);
2706 long res = req->result;
2709 struct io_ring_ctx *ctx = req->ctx;
2710 struct io_submit_state *state = &ctx->submit_state;
2712 io_req_complete_state(req, res, cflags);
2713 state->compl_reqs[state->compl_nr++] = req;
2714 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2715 io_submit_flush_completions(ctx);
2717 io_req_complete_post(req, res, cflags);
2721 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2722 unsigned int issue_flags)
2724 if (__io_complete_rw_common(req, res))
2726 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2729 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2731 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2733 if (__io_complete_rw_common(req, res))
2736 req->io_task_work.func = io_req_task_complete;
2737 io_req_task_work_add(req);
2740 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2742 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2744 if (kiocb->ki_flags & IOCB_WRITE)
2745 kiocb_end_write(req);
2746 if (unlikely(res != req->result)) {
2747 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2748 req->flags |= REQ_F_REISSUE;
2753 WRITE_ONCE(req->result, res);
2754 /* order with io_iopoll_complete() checking ->result */
2756 WRITE_ONCE(req->iopoll_completed, 1);
2760 * After the iocb has been issued, it's safe to be found on the poll list.
2761 * Adding the kiocb to the list AFTER submission ensures that we don't
2762 * find it from a io_do_iopoll() thread before the issuer is done
2763 * accessing the kiocb cookie.
2765 static void io_iopoll_req_issued(struct io_kiocb *req)
2767 struct io_ring_ctx *ctx = req->ctx;
2768 const bool in_async = io_wq_current_is_worker();
2770 /* workqueue context doesn't hold uring_lock, grab it now */
2771 if (unlikely(in_async))
2772 mutex_lock(&ctx->uring_lock);
2775 * Track whether we have multiple files in our lists. This will impact
2776 * how we do polling eventually, not spinning if we're on potentially
2777 * different devices.
2779 if (list_empty(&ctx->iopoll_list)) {
2780 ctx->poll_multi_queue = false;
2781 } else if (!ctx->poll_multi_queue) {
2782 struct io_kiocb *list_req;
2783 unsigned int queue_num0, queue_num1;
2785 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2788 if (list_req->file != req->file) {
2789 ctx->poll_multi_queue = true;
2791 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2792 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2793 if (queue_num0 != queue_num1)
2794 ctx->poll_multi_queue = true;
2799 * For fast devices, IO may have already completed. If it has, add
2800 * it to the front so we find it first.
2802 if (READ_ONCE(req->iopoll_completed))
2803 list_add(&req->inflight_entry, &ctx->iopoll_list);
2805 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2807 if (unlikely(in_async)) {
2809 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2810 * in sq thread task context or in io worker task context. If
2811 * current task context is sq thread, we don't need to check
2812 * whether should wake up sq thread.
2814 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2815 wq_has_sleeper(&ctx->sq_data->wait))
2816 wake_up(&ctx->sq_data->wait);
2818 mutex_unlock(&ctx->uring_lock);
2822 static bool io_bdev_nowait(struct block_device *bdev)
2824 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2828 * If we tracked the file through the SCM inflight mechanism, we could support
2829 * any file. For now, just ensure that anything potentially problematic is done
2832 static bool __io_file_supports_nowait(struct file *file, int rw)
2834 umode_t mode = file_inode(file)->i_mode;
2836 if (S_ISBLK(mode)) {
2837 if (IS_ENABLED(CONFIG_BLOCK) &&
2838 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2844 if (S_ISREG(mode)) {
2845 if (IS_ENABLED(CONFIG_BLOCK) &&
2846 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2847 file->f_op != &io_uring_fops)
2852 /* any ->read/write should understand O_NONBLOCK */
2853 if (file->f_flags & O_NONBLOCK)
2856 if (!(file->f_mode & FMODE_NOWAIT))
2860 return file->f_op->read_iter != NULL;
2862 return file->f_op->write_iter != NULL;
2865 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2867 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2869 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2872 return __io_file_supports_nowait(req->file, rw);
2875 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2878 struct io_ring_ctx *ctx = req->ctx;
2879 struct kiocb *kiocb = &req->rw.kiocb;
2880 struct file *file = req->file;
2884 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2885 req->flags |= REQ_F_ISREG;
2887 kiocb->ki_pos = READ_ONCE(sqe->off);
2888 if (kiocb->ki_pos == -1) {
2889 if (!(file->f_mode & FMODE_STREAM)) {
2890 req->flags |= REQ_F_CUR_POS;
2891 kiocb->ki_pos = file->f_pos;
2896 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2897 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2898 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2903 * If the file is marked O_NONBLOCK, still allow retry for it if it
2904 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2905 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2907 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2908 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2909 req->flags |= REQ_F_NOWAIT;
2911 ioprio = READ_ONCE(sqe->ioprio);
2913 ret = ioprio_check_cap(ioprio);
2917 kiocb->ki_ioprio = ioprio;
2919 kiocb->ki_ioprio = get_current_ioprio();
2921 if (ctx->flags & IORING_SETUP_IOPOLL) {
2922 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2923 !kiocb->ki_filp->f_op->iopoll)
2926 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2927 kiocb->ki_complete = io_complete_rw_iopoll;
2928 req->iopoll_completed = 0;
2930 if (kiocb->ki_flags & IOCB_HIPRI)
2932 kiocb->ki_complete = io_complete_rw;
2935 /* used for fixed read/write too - just read unconditionally */
2936 req->buf_index = READ_ONCE(sqe->buf_index);
2939 if (req->opcode == IORING_OP_READ_FIXED ||
2940 req->opcode == IORING_OP_WRITE_FIXED) {
2941 struct io_ring_ctx *ctx = req->ctx;
2944 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
2946 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
2947 req->imu = ctx->user_bufs[index];
2948 io_req_set_rsrc_node(req);
2951 req->rw.addr = READ_ONCE(sqe->addr);
2952 req->rw.len = READ_ONCE(sqe->len);
2956 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2962 case -ERESTARTNOINTR:
2963 case -ERESTARTNOHAND:
2964 case -ERESTART_RESTARTBLOCK:
2966 * We can't just restart the syscall, since previously
2967 * submitted sqes may already be in progress. Just fail this
2973 kiocb->ki_complete(kiocb, ret, 0);
2977 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2978 unsigned int issue_flags)
2980 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2981 struct io_async_rw *io = req->async_data;
2983 /* add previously done IO, if any */
2984 if (io && io->bytes_done > 0) {
2986 ret = io->bytes_done;
2988 ret += io->bytes_done;
2991 if (req->flags & REQ_F_CUR_POS)
2992 req->file->f_pos = kiocb->ki_pos;
2993 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2994 __io_complete_rw(req, ret, 0, issue_flags);
2996 io_rw_done(kiocb, ret);
2998 if (req->flags & REQ_F_REISSUE) {
2999 req->flags &= ~REQ_F_REISSUE;
3000 if (io_resubmit_prep(req)) {
3001 io_req_task_queue_reissue(req);
3003 unsigned int cflags = io_put_rw_kbuf(req);
3004 struct io_ring_ctx *ctx = req->ctx;
3007 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
3008 mutex_lock(&ctx->uring_lock);
3009 __io_req_complete(req, issue_flags, ret, cflags);
3010 mutex_unlock(&ctx->uring_lock);
3012 __io_req_complete(req, issue_flags, ret, cflags);
3018 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3019 struct io_mapped_ubuf *imu)
3021 size_t len = req->rw.len;
3022 u64 buf_end, buf_addr = req->rw.addr;
3025 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
3027 /* not inside the mapped region */
3028 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
3032 * May not be a start of buffer, set size appropriately
3033 * and advance us to the beginning.
3035 offset = buf_addr - imu->ubuf;
3036 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3040 * Don't use iov_iter_advance() here, as it's really slow for
3041 * using the latter parts of a big fixed buffer - it iterates
3042 * over each segment manually. We can cheat a bit here, because
3045 * 1) it's a BVEC iter, we set it up
3046 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3047 * first and last bvec
3049 * So just find our index, and adjust the iterator afterwards.
3050 * If the offset is within the first bvec (or the whole first
3051 * bvec, just use iov_iter_advance(). This makes it easier
3052 * since we can just skip the first segment, which may not
3053 * be PAGE_SIZE aligned.
3055 const struct bio_vec *bvec = imu->bvec;
3057 if (offset <= bvec->bv_len) {
3058 iov_iter_advance(iter, offset);
3060 unsigned long seg_skip;
3062 /* skip first vec */
3063 offset -= bvec->bv_len;
3064 seg_skip = 1 + (offset >> PAGE_SHIFT);
3066 iter->bvec = bvec + seg_skip;
3067 iter->nr_segs -= seg_skip;
3068 iter->count -= bvec->bv_len + offset;
3069 iter->iov_offset = offset & ~PAGE_MASK;
3076 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3078 if (WARN_ON_ONCE(!req->imu))
3080 return __io_import_fixed(req, rw, iter, req->imu);
3083 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3086 mutex_unlock(&ctx->uring_lock);
3089 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3092 * "Normal" inline submissions always hold the uring_lock, since we
3093 * grab it from the system call. Same is true for the SQPOLL offload.
3094 * The only exception is when we've detached the request and issue it
3095 * from an async worker thread, grab the lock for that case.
3098 mutex_lock(&ctx->uring_lock);
3101 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3102 int bgid, struct io_buffer *kbuf,
3105 struct io_buffer *head;
3107 if (req->flags & REQ_F_BUFFER_SELECTED)
3110 io_ring_submit_lock(req->ctx, needs_lock);
3112 lockdep_assert_held(&req->ctx->uring_lock);
3114 head = xa_load(&req->ctx->io_buffers, bgid);
3116 if (!list_empty(&head->list)) {
3117 kbuf = list_last_entry(&head->list, struct io_buffer,
3119 list_del(&kbuf->list);
3122 xa_erase(&req->ctx->io_buffers, bgid);
3124 if (*len > kbuf->len)
3127 kbuf = ERR_PTR(-ENOBUFS);
3130 io_ring_submit_unlock(req->ctx, needs_lock);
3135 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3138 struct io_buffer *kbuf;
3141 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3142 bgid = req->buf_index;
3143 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3146 req->rw.addr = (u64) (unsigned long) kbuf;
3147 req->flags |= REQ_F_BUFFER_SELECTED;
3148 return u64_to_user_ptr(kbuf->addr);
3151 #ifdef CONFIG_COMPAT
3152 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3155 struct compat_iovec __user *uiov;
3156 compat_ssize_t clen;
3160 uiov = u64_to_user_ptr(req->rw.addr);
3161 if (!access_ok(uiov, sizeof(*uiov)))
3163 if (__get_user(clen, &uiov->iov_len))
3169 buf = io_rw_buffer_select(req, &len, needs_lock);
3171 return PTR_ERR(buf);
3172 iov[0].iov_base = buf;
3173 iov[0].iov_len = (compat_size_t) len;
3178 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3181 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3185 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3188 len = iov[0].iov_len;
3191 buf = io_rw_buffer_select(req, &len, needs_lock);
3193 return PTR_ERR(buf);
3194 iov[0].iov_base = buf;
3195 iov[0].iov_len = len;
3199 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3202 if (req->flags & REQ_F_BUFFER_SELECTED) {
3203 struct io_buffer *kbuf;
3205 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3206 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3207 iov[0].iov_len = kbuf->len;
3210 if (req->rw.len != 1)
3213 #ifdef CONFIG_COMPAT
3214 if (req->ctx->compat)
3215 return io_compat_import(req, iov, needs_lock);
3218 return __io_iov_buffer_select(req, iov, needs_lock);
3221 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3222 struct iov_iter *iter, bool needs_lock)
3224 void __user *buf = u64_to_user_ptr(req->rw.addr);
3225 size_t sqe_len = req->rw.len;
3226 u8 opcode = req->opcode;
3229 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3231 return io_import_fixed(req, rw, iter);
3234 /* buffer index only valid with fixed read/write, or buffer select */
3235 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3238 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3239 if (req->flags & REQ_F_BUFFER_SELECT) {
3240 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3242 return PTR_ERR(buf);
3243 req->rw.len = sqe_len;
3246 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3251 if (req->flags & REQ_F_BUFFER_SELECT) {
3252 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3254 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3259 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3263 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3265 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3269 * For files that don't have ->read_iter() and ->write_iter(), handle them
3270 * by looping over ->read() or ->write() manually.
3272 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3274 struct kiocb *kiocb = &req->rw.kiocb;
3275 struct file *file = req->file;
3279 * Don't support polled IO through this interface, and we can't
3280 * support non-blocking either. For the latter, this just causes
3281 * the kiocb to be handled from an async context.
3283 if (kiocb->ki_flags & IOCB_HIPRI)
3285 if (kiocb->ki_flags & IOCB_NOWAIT)
3288 while (iov_iter_count(iter)) {
3292 if (!iov_iter_is_bvec(iter)) {
3293 iovec = iov_iter_iovec(iter);
3295 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3296 iovec.iov_len = req->rw.len;
3300 nr = file->f_op->read(file, iovec.iov_base,
3301 iovec.iov_len, io_kiocb_ppos(kiocb));
3303 nr = file->f_op->write(file, iovec.iov_base,
3304 iovec.iov_len, io_kiocb_ppos(kiocb));
3313 if (!iov_iter_is_bvec(iter)) {
3314 iov_iter_advance(iter, nr);
3321 if (nr != iovec.iov_len)
3328 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3329 const struct iovec *fast_iov, struct iov_iter *iter)
3331 struct io_async_rw *rw = req->async_data;
3333 memcpy(&rw->iter, iter, sizeof(*iter));
3334 rw->free_iovec = iovec;
3336 /* can only be fixed buffers, no need to do anything */
3337 if (iov_iter_is_bvec(iter))
3340 unsigned iov_off = 0;
3342 rw->iter.iov = rw->fast_iov;
3343 if (iter->iov != fast_iov) {
3344 iov_off = iter->iov - fast_iov;
3345 rw->iter.iov += iov_off;
3347 if (rw->fast_iov != fast_iov)
3348 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3349 sizeof(struct iovec) * iter->nr_segs);
3351 req->flags |= REQ_F_NEED_CLEANUP;
3355 static inline int io_alloc_async_data(struct io_kiocb *req)
3357 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3358 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3359 return req->async_data == NULL;
3362 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3363 const struct iovec *fast_iov,
3364 struct iov_iter *iter, bool force)
3366 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3368 if (!req->async_data) {
3369 struct io_async_rw *iorw;
3371 if (io_alloc_async_data(req)) {
3376 io_req_map_rw(req, iovec, fast_iov, iter);
3377 iorw = req->async_data;
3378 /* we've copied and mapped the iter, ensure state is saved */
3379 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3384 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3386 struct io_async_rw *iorw = req->async_data;
3387 struct iovec *iov = iorw->fast_iov;
3390 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3391 if (unlikely(ret < 0))
3394 iorw->bytes_done = 0;
3395 iorw->free_iovec = iov;
3397 req->flags |= REQ_F_NEED_CLEANUP;
3398 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3402 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3404 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3406 return io_prep_rw(req, sqe, READ);
3410 * This is our waitqueue callback handler, registered through lock_page_async()
3411 * when we initially tried to do the IO with the iocb armed our waitqueue.
3412 * This gets called when the page is unlocked, and we generally expect that to
3413 * happen when the page IO is completed and the page is now uptodate. This will
3414 * queue a task_work based retry of the operation, attempting to copy the data
3415 * again. If the latter fails because the page was NOT uptodate, then we will
3416 * do a thread based blocking retry of the operation. That's the unexpected
3419 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3420 int sync, void *arg)
3422 struct wait_page_queue *wpq;
3423 struct io_kiocb *req = wait->private;
3424 struct wait_page_key *key = arg;
3426 wpq = container_of(wait, struct wait_page_queue, wait);
3428 if (!wake_page_match(wpq, key))
3431 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3432 list_del_init(&wait->entry);
3433 io_req_task_queue(req);
3438 * This controls whether a given IO request should be armed for async page
3439 * based retry. If we return false here, the request is handed to the async
3440 * worker threads for retry. If we're doing buffered reads on a regular file,
3441 * we prepare a private wait_page_queue entry and retry the operation. This
3442 * will either succeed because the page is now uptodate and unlocked, or it
3443 * will register a callback when the page is unlocked at IO completion. Through
3444 * that callback, io_uring uses task_work to setup a retry of the operation.
3445 * That retry will attempt the buffered read again. The retry will generally
3446 * succeed, or in rare cases where it fails, we then fall back to using the
3447 * async worker threads for a blocking retry.
3449 static bool io_rw_should_retry(struct io_kiocb *req)
3451 struct io_async_rw *rw = req->async_data;
3452 struct wait_page_queue *wait = &rw->wpq;
3453 struct kiocb *kiocb = &req->rw.kiocb;
3455 /* never retry for NOWAIT, we just complete with -EAGAIN */
3456 if (req->flags & REQ_F_NOWAIT)
3459 /* Only for buffered IO */
3460 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3464 * just use poll if we can, and don't attempt if the fs doesn't
3465 * support callback based unlocks
3467 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3470 wait->wait.func = io_async_buf_func;
3471 wait->wait.private = req;
3472 wait->wait.flags = 0;
3473 INIT_LIST_HEAD(&wait->wait.entry);
3474 kiocb->ki_flags |= IOCB_WAITQ;
3475 kiocb->ki_flags &= ~IOCB_NOWAIT;
3476 kiocb->ki_waitq = wait;
3480 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3482 if (req->file->f_op->read_iter)
3483 return call_read_iter(req->file, &req->rw.kiocb, iter);
3484 else if (req->file->f_op->read)
3485 return loop_rw_iter(READ, req, iter);
3490 static bool need_read_all(struct io_kiocb *req)
3492 return req->flags & REQ_F_ISREG ||
3493 S_ISBLK(file_inode(req->file)->i_mode);
3496 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3498 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3499 struct kiocb *kiocb = &req->rw.kiocb;
3500 struct iov_iter __iter, *iter = &__iter;
3501 struct io_async_rw *rw = req->async_data;
3502 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3503 struct iov_iter_state __state, *state;
3508 state = &rw->iter_state;
3510 * We come here from an earlier attempt, restore our state to
3511 * match in case it doesn't. It's cheap enough that we don't
3512 * need to make this conditional.
3514 iov_iter_restore(iter, state);
3517 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3521 iov_iter_save_state(iter, state);
3523 req->result = iov_iter_count(iter);
3525 /* Ensure we clear previously set non-block flag */
3526 if (!force_nonblock)
3527 kiocb->ki_flags &= ~IOCB_NOWAIT;
3529 kiocb->ki_flags |= IOCB_NOWAIT;
3531 /* If the file doesn't support async, just async punt */
3532 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3533 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3534 return ret ?: -EAGAIN;
3537 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3538 if (unlikely(ret)) {
3543 ret = io_iter_do_read(req, iter);
3545 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3546 req->flags &= ~REQ_F_REISSUE;
3547 /* IOPOLL retry should happen for io-wq threads */
3548 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3550 /* no retry on NONBLOCK nor RWF_NOWAIT */
3551 if (req->flags & REQ_F_NOWAIT)
3554 } else if (ret == -EIOCBQUEUED) {
3556 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3557 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3558 /* read all, failed, already did sync or don't want to retry */
3563 * Don't depend on the iter state matching what was consumed, or being
3564 * untouched in case of error. Restore it and we'll advance it
3565 * manually if we need to.
3567 iov_iter_restore(iter, state);
3569 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3574 rw = req->async_data;
3576 * Now use our persistent iterator and state, if we aren't already.
3577 * We've restored and mapped the iter to match.
3579 if (iter != &rw->iter) {
3581 state = &rw->iter_state;
3586 * We end up here because of a partial read, either from
3587 * above or inside this loop. Advance the iter by the bytes
3588 * that were consumed.
3590 iov_iter_advance(iter, ret);
3591 if (!iov_iter_count(iter))
3593 rw->bytes_done += ret;
3594 iov_iter_save_state(iter, state);
3596 /* if we can retry, do so with the callbacks armed */
3597 if (!io_rw_should_retry(req)) {
3598 kiocb->ki_flags &= ~IOCB_WAITQ;
3603 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3604 * we get -EIOCBQUEUED, then we'll get a notification when the
3605 * desired page gets unlocked. We can also get a partial read
3606 * here, and if we do, then just retry at the new offset.
3608 ret = io_iter_do_read(req, iter);
3609 if (ret == -EIOCBQUEUED)
3611 /* we got some bytes, but not all. retry. */
3612 kiocb->ki_flags &= ~IOCB_WAITQ;
3613 iov_iter_restore(iter, state);
3616 kiocb_done(kiocb, ret, issue_flags);
3618 /* it's faster to check here then delegate to kfree */
3624 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3626 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3628 return io_prep_rw(req, sqe, WRITE);
3631 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3633 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3634 struct kiocb *kiocb = &req->rw.kiocb;
3635 struct iov_iter __iter, *iter = &__iter;
3636 struct io_async_rw *rw = req->async_data;
3637 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3638 struct iov_iter_state __state, *state;
3643 state = &rw->iter_state;
3644 iov_iter_restore(iter, state);
3647 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3651 iov_iter_save_state(iter, state);
3653 req->result = iov_iter_count(iter);
3655 /* Ensure we clear previously set non-block flag */
3656 if (!force_nonblock)
3657 kiocb->ki_flags &= ~IOCB_NOWAIT;
3659 kiocb->ki_flags |= IOCB_NOWAIT;
3661 /* If the file doesn't support async, just async punt */
3662 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3665 /* file path doesn't support NOWAIT for non-direct_IO */
3666 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3667 (req->flags & REQ_F_ISREG))
3670 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3675 * Open-code file_start_write here to grab freeze protection,
3676 * which will be released by another thread in
3677 * io_complete_rw(). Fool lockdep by telling it the lock got
3678 * released so that it doesn't complain about the held lock when
3679 * we return to userspace.
3681 if (req->flags & REQ_F_ISREG) {
3682 sb_start_write(file_inode(req->file)->i_sb);
3683 __sb_writers_release(file_inode(req->file)->i_sb,
3686 kiocb->ki_flags |= IOCB_WRITE;
3688 if (req->file->f_op->write_iter)
3689 ret2 = call_write_iter(req->file, kiocb, iter);
3690 else if (req->file->f_op->write)
3691 ret2 = loop_rw_iter(WRITE, req, iter);
3695 if (req->flags & REQ_F_REISSUE) {
3696 req->flags &= ~REQ_F_REISSUE;
3701 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3702 * retry them without IOCB_NOWAIT.
3704 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3706 /* no retry on NONBLOCK nor RWF_NOWAIT */
3707 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3709 if (!force_nonblock || ret2 != -EAGAIN) {
3710 /* IOPOLL retry should happen for io-wq threads */
3711 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3714 kiocb_done(kiocb, ret2, issue_flags);
3717 iov_iter_restore(iter, state);
3718 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3719 return ret ?: -EAGAIN;
3722 /* it's reportedly faster than delegating the null check to kfree() */
3728 static int io_renameat_prep(struct io_kiocb *req,
3729 const struct io_uring_sqe *sqe)
3731 struct io_rename *ren = &req->rename;
3732 const char __user *oldf, *newf;
3734 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3736 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3738 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3741 ren->old_dfd = READ_ONCE(sqe->fd);
3742 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3743 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3744 ren->new_dfd = READ_ONCE(sqe->len);
3745 ren->flags = READ_ONCE(sqe->rename_flags);
3747 ren->oldpath = getname(oldf);
3748 if (IS_ERR(ren->oldpath))
3749 return PTR_ERR(ren->oldpath);
3751 ren->newpath = getname(newf);
3752 if (IS_ERR(ren->newpath)) {
3753 putname(ren->oldpath);
3754 return PTR_ERR(ren->newpath);
3757 req->flags |= REQ_F_NEED_CLEANUP;
3761 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3763 struct io_rename *ren = &req->rename;
3766 if (issue_flags & IO_URING_F_NONBLOCK)
3769 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3770 ren->newpath, ren->flags);
3772 req->flags &= ~REQ_F_NEED_CLEANUP;
3775 io_req_complete(req, ret);
3779 static int io_unlinkat_prep(struct io_kiocb *req,
3780 const struct io_uring_sqe *sqe)
3782 struct io_unlink *un = &req->unlink;
3783 const char __user *fname;
3785 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3787 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3790 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3793 un->dfd = READ_ONCE(sqe->fd);
3795 un->flags = READ_ONCE(sqe->unlink_flags);
3796 if (un->flags & ~AT_REMOVEDIR)
3799 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3800 un->filename = getname(fname);
3801 if (IS_ERR(un->filename))
3802 return PTR_ERR(un->filename);
3804 req->flags |= REQ_F_NEED_CLEANUP;
3808 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3810 struct io_unlink *un = &req->unlink;
3813 if (issue_flags & IO_URING_F_NONBLOCK)
3816 if (un->flags & AT_REMOVEDIR)
3817 ret = do_rmdir(un->dfd, un->filename);
3819 ret = do_unlinkat(un->dfd, un->filename);
3821 req->flags &= ~REQ_F_NEED_CLEANUP;
3824 io_req_complete(req, ret);
3828 static int io_mkdirat_prep(struct io_kiocb *req,
3829 const struct io_uring_sqe *sqe)
3831 struct io_mkdir *mkd = &req->mkdir;
3832 const char __user *fname;
3834 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3836 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3839 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3842 mkd->dfd = READ_ONCE(sqe->fd);
3843 mkd->mode = READ_ONCE(sqe->len);
3845 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3846 mkd->filename = getname(fname);
3847 if (IS_ERR(mkd->filename))
3848 return PTR_ERR(mkd->filename);
3850 req->flags |= REQ_F_NEED_CLEANUP;
3854 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3856 struct io_mkdir *mkd = &req->mkdir;
3859 if (issue_flags & IO_URING_F_NONBLOCK)
3862 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3864 req->flags &= ~REQ_F_NEED_CLEANUP;
3867 io_req_complete(req, ret);
3871 static int io_symlinkat_prep(struct io_kiocb *req,
3872 const struct io_uring_sqe *sqe)
3874 struct io_symlink *sl = &req->symlink;
3875 const char __user *oldpath, *newpath;
3877 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3879 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3882 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3885 sl->new_dfd = READ_ONCE(sqe->fd);
3886 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3887 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3889 sl->oldpath = getname(oldpath);
3890 if (IS_ERR(sl->oldpath))
3891 return PTR_ERR(sl->oldpath);
3893 sl->newpath = getname(newpath);
3894 if (IS_ERR(sl->newpath)) {
3895 putname(sl->oldpath);
3896 return PTR_ERR(sl->newpath);
3899 req->flags |= REQ_F_NEED_CLEANUP;
3903 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3905 struct io_symlink *sl = &req->symlink;
3908 if (issue_flags & IO_URING_F_NONBLOCK)
3911 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3913 req->flags &= ~REQ_F_NEED_CLEANUP;
3916 io_req_complete(req, ret);
3920 static int io_linkat_prep(struct io_kiocb *req,
3921 const struct io_uring_sqe *sqe)
3923 struct io_hardlink *lnk = &req->hardlink;
3924 const char __user *oldf, *newf;
3926 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3928 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3930 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3933 lnk->old_dfd = READ_ONCE(sqe->fd);
3934 lnk->new_dfd = READ_ONCE(sqe->len);
3935 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3936 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3937 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3939 lnk->oldpath = getname(oldf);
3940 if (IS_ERR(lnk->oldpath))
3941 return PTR_ERR(lnk->oldpath);
3943 lnk->newpath = getname(newf);
3944 if (IS_ERR(lnk->newpath)) {
3945 putname(lnk->oldpath);
3946 return PTR_ERR(lnk->newpath);
3949 req->flags |= REQ_F_NEED_CLEANUP;
3953 static int io_linkat(struct io_kiocb *req, int issue_flags)
3955 struct io_hardlink *lnk = &req->hardlink;
3958 if (issue_flags & IO_URING_F_NONBLOCK)
3961 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3962 lnk->newpath, lnk->flags);
3964 req->flags &= ~REQ_F_NEED_CLEANUP;
3967 io_req_complete(req, ret);
3971 static int io_shutdown_prep(struct io_kiocb *req,
3972 const struct io_uring_sqe *sqe)
3974 #if defined(CONFIG_NET)
3975 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3977 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3978 sqe->buf_index || sqe->splice_fd_in))
3981 req->shutdown.how = READ_ONCE(sqe->len);
3988 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3990 #if defined(CONFIG_NET)
3991 struct socket *sock;
3994 if (issue_flags & IO_URING_F_NONBLOCK)
3997 sock = sock_from_file(req->file);
3998 if (unlikely(!sock))
4001 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4004 io_req_complete(req, ret);
4011 static int __io_splice_prep(struct io_kiocb *req,
4012 const struct io_uring_sqe *sqe)
4014 struct io_splice *sp = &req->splice;
4015 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4017 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4020 sp->len = READ_ONCE(sqe->len);
4021 sp->flags = READ_ONCE(sqe->splice_flags);
4022 if (unlikely(sp->flags & ~valid_flags))
4024 sp->splice_fd_in = READ_ONCE(sqe->splice_fd_in);
4028 static int io_tee_prep(struct io_kiocb *req,
4029 const struct io_uring_sqe *sqe)
4031 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4033 return __io_splice_prep(req, sqe);
4036 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4038 struct io_splice *sp = &req->splice;
4039 struct file *out = sp->file_out;
4040 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4044 if (issue_flags & IO_URING_F_NONBLOCK)
4047 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4048 (sp->flags & SPLICE_F_FD_IN_FIXED));
4055 ret = do_tee(in, out, sp->len, flags);
4057 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4062 io_req_complete(req, ret);
4066 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4068 struct io_splice *sp = &req->splice;
4070 sp->off_in = READ_ONCE(sqe->splice_off_in);
4071 sp->off_out = READ_ONCE(sqe->off);
4072 return __io_splice_prep(req, sqe);
4075 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4077 struct io_splice *sp = &req->splice;
4078 struct file *out = sp->file_out;
4079 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4080 loff_t *poff_in, *poff_out;
4084 if (issue_flags & IO_URING_F_NONBLOCK)
4087 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4088 (sp->flags & SPLICE_F_FD_IN_FIXED));
4094 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4095 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4098 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4100 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4105 io_req_complete(req, ret);
4110 * IORING_OP_NOP just posts a completion event, nothing else.
4112 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4114 struct io_ring_ctx *ctx = req->ctx;
4116 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4119 __io_req_complete(req, issue_flags, 0, 0);
4123 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4125 struct io_ring_ctx *ctx = req->ctx;
4127 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4129 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4133 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4134 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4137 req->sync.off = READ_ONCE(sqe->off);
4138 req->sync.len = READ_ONCE(sqe->len);
4142 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4144 loff_t end = req->sync.off + req->sync.len;
4147 /* fsync always requires a blocking context */
4148 if (issue_flags & IO_URING_F_NONBLOCK)
4151 ret = vfs_fsync_range(req->file, req->sync.off,
4152 end > 0 ? end : LLONG_MAX,
4153 req->sync.flags & IORING_FSYNC_DATASYNC);
4156 io_req_complete(req, ret);
4160 static int io_fallocate_prep(struct io_kiocb *req,
4161 const struct io_uring_sqe *sqe)
4163 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4166 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4169 req->sync.off = READ_ONCE(sqe->off);
4170 req->sync.len = READ_ONCE(sqe->addr);
4171 req->sync.mode = READ_ONCE(sqe->len);
4175 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4179 /* fallocate always requiring blocking context */
4180 if (issue_flags & IO_URING_F_NONBLOCK)
4182 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4186 io_req_complete(req, ret);
4190 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4192 const char __user *fname;
4195 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4197 if (unlikely(sqe->ioprio || sqe->buf_index))
4199 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4202 /* open.how should be already initialised */
4203 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4204 req->open.how.flags |= O_LARGEFILE;
4206 req->open.dfd = READ_ONCE(sqe->fd);
4207 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4208 req->open.filename = getname(fname);
4209 if (IS_ERR(req->open.filename)) {
4210 ret = PTR_ERR(req->open.filename);
4211 req->open.filename = NULL;
4215 req->open.file_slot = READ_ONCE(sqe->file_index);
4216 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4219 req->open.nofile = rlimit(RLIMIT_NOFILE);
4220 req->flags |= REQ_F_NEED_CLEANUP;
4224 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4226 u64 mode = READ_ONCE(sqe->len);
4227 u64 flags = READ_ONCE(sqe->open_flags);
4229 req->open.how = build_open_how(flags, mode);
4230 return __io_openat_prep(req, sqe);
4233 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4235 struct open_how __user *how;
4239 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4240 len = READ_ONCE(sqe->len);
4241 if (len < OPEN_HOW_SIZE_VER0)
4244 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4249 return __io_openat_prep(req, sqe);
4252 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4254 struct open_flags op;
4256 bool resolve_nonblock, nonblock_set;
4257 bool fixed = !!req->open.file_slot;
4260 ret = build_open_flags(&req->open.how, &op);
4263 nonblock_set = op.open_flag & O_NONBLOCK;
4264 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4265 if (issue_flags & IO_URING_F_NONBLOCK) {
4267 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4268 * it'll always -EAGAIN
4270 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4272 op.lookup_flags |= LOOKUP_CACHED;
4273 op.open_flag |= O_NONBLOCK;
4277 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4282 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4285 * We could hang on to this 'fd' on retrying, but seems like
4286 * marginal gain for something that is now known to be a slower
4287 * path. So just put it, and we'll get a new one when we retry.
4292 ret = PTR_ERR(file);
4293 /* only retry if RESOLVE_CACHED wasn't already set by application */
4294 if (ret == -EAGAIN &&
4295 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4300 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4301 file->f_flags &= ~O_NONBLOCK;
4302 fsnotify_open(file);
4305 fd_install(ret, file);
4307 ret = io_install_fixed_file(req, file, issue_flags,
4308 req->open.file_slot - 1);
4310 putname(req->open.filename);
4311 req->flags &= ~REQ_F_NEED_CLEANUP;
4314 __io_req_complete(req, issue_flags, ret, 0);
4318 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4320 return io_openat2(req, issue_flags);
4323 static int io_remove_buffers_prep(struct io_kiocb *req,
4324 const struct io_uring_sqe *sqe)
4326 struct io_provide_buf *p = &req->pbuf;
4329 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4333 tmp = READ_ONCE(sqe->fd);
4334 if (!tmp || tmp > USHRT_MAX)
4337 memset(p, 0, sizeof(*p));
4339 p->bgid = READ_ONCE(sqe->buf_group);
4343 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4344 int bgid, unsigned nbufs)
4348 /* shouldn't happen */
4352 /* the head kbuf is the list itself */
4353 while (!list_empty(&buf->list)) {
4354 struct io_buffer *nxt;
4356 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4357 list_del(&nxt->list);
4365 xa_erase(&ctx->io_buffers, bgid);
4370 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4372 struct io_provide_buf *p = &req->pbuf;
4373 struct io_ring_ctx *ctx = req->ctx;
4374 struct io_buffer *head;
4376 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4378 io_ring_submit_lock(ctx, !force_nonblock);
4380 lockdep_assert_held(&ctx->uring_lock);
4383 head = xa_load(&ctx->io_buffers, p->bgid);
4385 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4389 /* complete before unlock, IOPOLL may need the lock */
4390 __io_req_complete(req, issue_flags, ret, 0);
4391 io_ring_submit_unlock(ctx, !force_nonblock);
4395 static int io_provide_buffers_prep(struct io_kiocb *req,
4396 const struct io_uring_sqe *sqe)
4398 unsigned long size, tmp_check;
4399 struct io_provide_buf *p = &req->pbuf;
4402 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4405 tmp = READ_ONCE(sqe->fd);
4406 if (!tmp || tmp > USHRT_MAX)
4409 p->addr = READ_ONCE(sqe->addr);
4410 p->len = READ_ONCE(sqe->len);
4412 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4415 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4418 size = (unsigned long)p->len * p->nbufs;
4419 if (!access_ok(u64_to_user_ptr(p->addr), size))
4422 p->bgid = READ_ONCE(sqe->buf_group);
4423 tmp = READ_ONCE(sqe->off);
4424 if (tmp > USHRT_MAX)
4430 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4432 struct io_buffer *buf;
4433 u64 addr = pbuf->addr;
4434 int i, bid = pbuf->bid;
4436 for (i = 0; i < pbuf->nbufs; i++) {
4437 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4442 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4447 INIT_LIST_HEAD(&buf->list);
4450 list_add_tail(&buf->list, &(*head)->list);
4455 return i ? i : -ENOMEM;
4458 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4460 struct io_provide_buf *p = &req->pbuf;
4461 struct io_ring_ctx *ctx = req->ctx;
4462 struct io_buffer *head, *list;
4464 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4466 io_ring_submit_lock(ctx, !force_nonblock);
4468 lockdep_assert_held(&ctx->uring_lock);
4470 list = head = xa_load(&ctx->io_buffers, p->bgid);
4472 ret = io_add_buffers(p, &head);
4473 if (ret >= 0 && !list) {
4474 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4476 __io_remove_buffers(ctx, head, p->bgid, -1U);
4480 /* complete before unlock, IOPOLL may need the lock */
4481 __io_req_complete(req, issue_flags, ret, 0);
4482 io_ring_submit_unlock(ctx, !force_nonblock);
4486 static int io_epoll_ctl_prep(struct io_kiocb *req,
4487 const struct io_uring_sqe *sqe)
4489 #if defined(CONFIG_EPOLL)
4490 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4492 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4495 req->epoll.epfd = READ_ONCE(sqe->fd);
4496 req->epoll.op = READ_ONCE(sqe->len);
4497 req->epoll.fd = READ_ONCE(sqe->off);
4499 if (ep_op_has_event(req->epoll.op)) {
4500 struct epoll_event __user *ev;
4502 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4503 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4513 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4515 #if defined(CONFIG_EPOLL)
4516 struct io_epoll *ie = &req->epoll;
4518 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4520 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4521 if (force_nonblock && ret == -EAGAIN)
4526 __io_req_complete(req, issue_flags, ret, 0);
4533 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4535 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4536 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4538 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4541 req->madvise.addr = READ_ONCE(sqe->addr);
4542 req->madvise.len = READ_ONCE(sqe->len);
4543 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4550 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4552 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4553 struct io_madvise *ma = &req->madvise;
4556 if (issue_flags & IO_URING_F_NONBLOCK)
4559 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4562 io_req_complete(req, ret);
4569 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4571 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4573 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4576 req->fadvise.offset = READ_ONCE(sqe->off);
4577 req->fadvise.len = READ_ONCE(sqe->len);
4578 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4582 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4584 struct io_fadvise *fa = &req->fadvise;
4587 if (issue_flags & IO_URING_F_NONBLOCK) {
4588 switch (fa->advice) {
4589 case POSIX_FADV_NORMAL:
4590 case POSIX_FADV_RANDOM:
4591 case POSIX_FADV_SEQUENTIAL:
4598 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4601 __io_req_complete(req, issue_flags, ret, 0);
4605 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4607 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4609 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4611 if (req->flags & REQ_F_FIXED_FILE)
4614 req->statx.dfd = READ_ONCE(sqe->fd);
4615 req->statx.mask = READ_ONCE(sqe->len);
4616 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4617 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4618 req->statx.flags = READ_ONCE(sqe->statx_flags);
4623 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4625 struct io_statx *ctx = &req->statx;
4628 if (issue_flags & IO_URING_F_NONBLOCK)
4631 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4636 io_req_complete(req, ret);
4640 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4642 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4644 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4645 sqe->rw_flags || sqe->buf_index)
4647 if (req->flags & REQ_F_FIXED_FILE)
4650 req->close.fd = READ_ONCE(sqe->fd);
4651 req->close.file_slot = READ_ONCE(sqe->file_index);
4652 if (req->close.file_slot && req->close.fd)
4658 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4660 struct files_struct *files = current->files;
4661 struct io_close *close = &req->close;
4662 struct fdtable *fdt;
4663 struct file *file = NULL;
4666 if (req->close.file_slot) {
4667 ret = io_close_fixed(req, issue_flags);
4671 spin_lock(&files->file_lock);
4672 fdt = files_fdtable(files);
4673 if (close->fd >= fdt->max_fds) {
4674 spin_unlock(&files->file_lock);
4677 file = fdt->fd[close->fd];
4678 if (!file || file->f_op == &io_uring_fops) {
4679 spin_unlock(&files->file_lock);
4684 /* if the file has a flush method, be safe and punt to async */
4685 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4686 spin_unlock(&files->file_lock);
4690 ret = __close_fd_get_file(close->fd, &file);
4691 spin_unlock(&files->file_lock);
4698 /* No ->flush() or already async, safely close from here */
4699 ret = filp_close(file, current->files);
4705 __io_req_complete(req, issue_flags, ret, 0);
4709 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4711 struct io_ring_ctx *ctx = req->ctx;
4713 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4715 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4719 req->sync.off = READ_ONCE(sqe->off);
4720 req->sync.len = READ_ONCE(sqe->len);
4721 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4725 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4729 /* sync_file_range always requires a blocking context */
4730 if (issue_flags & IO_URING_F_NONBLOCK)
4733 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4737 io_req_complete(req, ret);
4741 #if defined(CONFIG_NET)
4742 static int io_setup_async_msg(struct io_kiocb *req,
4743 struct io_async_msghdr *kmsg)
4745 struct io_async_msghdr *async_msg = req->async_data;
4749 if (io_alloc_async_data(req)) {
4750 kfree(kmsg->free_iov);
4753 async_msg = req->async_data;
4754 req->flags |= REQ_F_NEED_CLEANUP;
4755 memcpy(async_msg, kmsg, sizeof(*kmsg));
4756 async_msg->msg.msg_name = &async_msg->addr;
4757 /* if were using fast_iov, set it to the new one */
4758 if (!async_msg->free_iov)
4759 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4764 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4765 struct io_async_msghdr *iomsg)
4767 iomsg->msg.msg_name = &iomsg->addr;
4768 iomsg->free_iov = iomsg->fast_iov;
4769 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4770 req->sr_msg.msg_flags, &iomsg->free_iov);
4773 static int io_sendmsg_prep_async(struct io_kiocb *req)
4777 ret = io_sendmsg_copy_hdr(req, req->async_data);
4779 req->flags |= REQ_F_NEED_CLEANUP;
4783 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4785 struct io_sr_msg *sr = &req->sr_msg;
4787 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4789 if (unlikely(sqe->addr2 || sqe->file_index))
4791 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
4794 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4795 sr->len = READ_ONCE(sqe->len);
4796 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4797 if (sr->msg_flags & MSG_DONTWAIT)
4798 req->flags |= REQ_F_NOWAIT;
4800 #ifdef CONFIG_COMPAT
4801 if (req->ctx->compat)
4802 sr->msg_flags |= MSG_CMSG_COMPAT;
4807 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4809 struct io_async_msghdr iomsg, *kmsg;
4810 struct socket *sock;
4815 sock = sock_from_file(req->file);
4816 if (unlikely(!sock))
4819 kmsg = req->async_data;
4821 ret = io_sendmsg_copy_hdr(req, &iomsg);
4827 flags = req->sr_msg.msg_flags;
4828 if (issue_flags & IO_URING_F_NONBLOCK)
4829 flags |= MSG_DONTWAIT;
4830 if (flags & MSG_WAITALL)
4831 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4833 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4834 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4835 return io_setup_async_msg(req, kmsg);
4836 if (ret == -ERESTARTSYS)
4839 /* fast path, check for non-NULL to avoid function call */
4841 kfree(kmsg->free_iov);
4842 req->flags &= ~REQ_F_NEED_CLEANUP;
4845 __io_req_complete(req, issue_flags, ret, 0);
4849 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4851 struct io_sr_msg *sr = &req->sr_msg;
4854 struct socket *sock;
4859 sock = sock_from_file(req->file);
4860 if (unlikely(!sock))
4863 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4867 msg.msg_name = NULL;
4868 msg.msg_control = NULL;
4869 msg.msg_controllen = 0;
4870 msg.msg_namelen = 0;
4872 flags = req->sr_msg.msg_flags;
4873 if (issue_flags & IO_URING_F_NONBLOCK)
4874 flags |= MSG_DONTWAIT;
4875 if (flags & MSG_WAITALL)
4876 min_ret = iov_iter_count(&msg.msg_iter);
4878 msg.msg_flags = flags;
4879 ret = sock_sendmsg(sock, &msg);
4880 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4882 if (ret == -ERESTARTSYS)
4887 __io_req_complete(req, issue_flags, ret, 0);
4891 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4892 struct io_async_msghdr *iomsg)
4894 struct io_sr_msg *sr = &req->sr_msg;
4895 struct iovec __user *uiov;
4899 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4900 &iomsg->uaddr, &uiov, &iov_len);
4904 if (req->flags & REQ_F_BUFFER_SELECT) {
4907 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4909 sr->len = iomsg->fast_iov[0].iov_len;
4910 iomsg->free_iov = NULL;
4912 iomsg->free_iov = iomsg->fast_iov;
4913 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4914 &iomsg->free_iov, &iomsg->msg.msg_iter,
4923 #ifdef CONFIG_COMPAT
4924 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4925 struct io_async_msghdr *iomsg)
4927 struct io_sr_msg *sr = &req->sr_msg;
4928 struct compat_iovec __user *uiov;
4933 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4938 uiov = compat_ptr(ptr);
4939 if (req->flags & REQ_F_BUFFER_SELECT) {
4940 compat_ssize_t clen;
4944 if (!access_ok(uiov, sizeof(*uiov)))
4946 if (__get_user(clen, &uiov->iov_len))
4951 iomsg->free_iov = NULL;
4953 iomsg->free_iov = iomsg->fast_iov;
4954 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4955 UIO_FASTIOV, &iomsg->free_iov,
4956 &iomsg->msg.msg_iter, true);
4965 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4966 struct io_async_msghdr *iomsg)
4968 iomsg->msg.msg_name = &iomsg->addr;
4970 #ifdef CONFIG_COMPAT
4971 if (req->ctx->compat)
4972 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4975 return __io_recvmsg_copy_hdr(req, iomsg);
4978 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4981 struct io_sr_msg *sr = &req->sr_msg;
4982 struct io_buffer *kbuf;
4984 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4989 req->flags |= REQ_F_BUFFER_SELECTED;
4993 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4995 return io_put_kbuf(req, req->sr_msg.kbuf);
4998 static int io_recvmsg_prep_async(struct io_kiocb *req)
5002 ret = io_recvmsg_copy_hdr(req, req->async_data);
5004 req->flags |= REQ_F_NEED_CLEANUP;
5008 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5010 struct io_sr_msg *sr = &req->sr_msg;
5012 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5014 if (unlikely(sqe->addr2 || sqe->file_index))
5016 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
5019 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5020 sr->len = READ_ONCE(sqe->len);
5021 sr->bgid = READ_ONCE(sqe->buf_group);
5022 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5023 if (sr->msg_flags & MSG_DONTWAIT)
5024 req->flags |= REQ_F_NOWAIT;
5026 #ifdef CONFIG_COMPAT
5027 if (req->ctx->compat)
5028 sr->msg_flags |= MSG_CMSG_COMPAT;
5033 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5035 struct io_async_msghdr iomsg, *kmsg;
5036 struct socket *sock;
5037 struct io_buffer *kbuf;
5040 int ret, cflags = 0;
5041 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5043 sock = sock_from_file(req->file);
5044 if (unlikely(!sock))
5047 kmsg = req->async_data;
5049 ret = io_recvmsg_copy_hdr(req, &iomsg);
5055 if (req->flags & REQ_F_BUFFER_SELECT) {
5056 kbuf = io_recv_buffer_select(req, !force_nonblock);
5058 return PTR_ERR(kbuf);
5059 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5060 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5061 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5062 1, req->sr_msg.len);
5065 flags = req->sr_msg.msg_flags;
5067 flags |= MSG_DONTWAIT;
5068 if (flags & MSG_WAITALL)
5069 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5071 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5072 kmsg->uaddr, flags);
5073 if (force_nonblock && ret == -EAGAIN)
5074 return io_setup_async_msg(req, kmsg);
5075 if (ret == -ERESTARTSYS)
5078 if (req->flags & REQ_F_BUFFER_SELECTED)
5079 cflags = io_put_recv_kbuf(req);
5080 /* fast path, check for non-NULL to avoid function call */
5082 kfree(kmsg->free_iov);
5083 req->flags &= ~REQ_F_NEED_CLEANUP;
5084 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5086 __io_req_complete(req, issue_flags, ret, cflags);
5090 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5092 struct io_buffer *kbuf;
5093 struct io_sr_msg *sr = &req->sr_msg;
5095 void __user *buf = sr->buf;
5096 struct socket *sock;
5100 int ret, cflags = 0;
5101 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5103 sock = sock_from_file(req->file);
5104 if (unlikely(!sock))
5107 if (req->flags & REQ_F_BUFFER_SELECT) {
5108 kbuf = io_recv_buffer_select(req, !force_nonblock);
5110 return PTR_ERR(kbuf);
5111 buf = u64_to_user_ptr(kbuf->addr);
5114 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5118 msg.msg_name = NULL;
5119 msg.msg_control = NULL;
5120 msg.msg_controllen = 0;
5121 msg.msg_namelen = 0;
5122 msg.msg_iocb = NULL;
5125 flags = req->sr_msg.msg_flags;
5127 flags |= MSG_DONTWAIT;
5128 if (flags & MSG_WAITALL)
5129 min_ret = iov_iter_count(&msg.msg_iter);
5131 ret = sock_recvmsg(sock, &msg, flags);
5132 if (force_nonblock && ret == -EAGAIN)
5134 if (ret == -ERESTARTSYS)
5137 if (req->flags & REQ_F_BUFFER_SELECTED)
5138 cflags = io_put_recv_kbuf(req);
5139 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5141 __io_req_complete(req, issue_flags, ret, cflags);
5145 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5147 struct io_accept *accept = &req->accept;
5149 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5151 if (sqe->ioprio || sqe->len || sqe->buf_index)
5154 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5155 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5156 accept->flags = READ_ONCE(sqe->accept_flags);
5157 accept->nofile = rlimit(RLIMIT_NOFILE);
5159 accept->file_slot = READ_ONCE(sqe->file_index);
5160 if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))
5162 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5164 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5165 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5169 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5171 struct io_accept *accept = &req->accept;
5172 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5173 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5174 bool fixed = !!accept->file_slot;
5178 if (req->file->f_flags & O_NONBLOCK)
5179 req->flags |= REQ_F_NOWAIT;
5182 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5183 if (unlikely(fd < 0))
5186 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5191 ret = PTR_ERR(file);
5192 if (ret == -EAGAIN && force_nonblock)
5194 if (ret == -ERESTARTSYS)
5197 } else if (!fixed) {
5198 fd_install(fd, file);
5201 ret = io_install_fixed_file(req, file, issue_flags,
5202 accept->file_slot - 1);
5204 __io_req_complete(req, issue_flags, ret, 0);
5208 static int io_connect_prep_async(struct io_kiocb *req)
5210 struct io_async_connect *io = req->async_data;
5211 struct io_connect *conn = &req->connect;
5213 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5216 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5218 struct io_connect *conn = &req->connect;
5220 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5222 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5226 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5227 conn->addr_len = READ_ONCE(sqe->addr2);
5231 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5233 struct io_async_connect __io, *io;
5234 unsigned file_flags;
5236 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5238 if (req->async_data) {
5239 io = req->async_data;
5241 ret = move_addr_to_kernel(req->connect.addr,
5242 req->connect.addr_len,
5249 file_flags = force_nonblock ? O_NONBLOCK : 0;
5251 ret = __sys_connect_file(req->file, &io->address,
5252 req->connect.addr_len, file_flags);
5253 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5254 if (req->async_data)
5256 if (io_alloc_async_data(req)) {
5260 memcpy(req->async_data, &__io, sizeof(__io));
5263 if (ret == -ERESTARTSYS)
5268 __io_req_complete(req, issue_flags, ret, 0);
5271 #else /* !CONFIG_NET */
5272 #define IO_NETOP_FN(op) \
5273 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5275 return -EOPNOTSUPP; \
5278 #define IO_NETOP_PREP(op) \
5280 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5282 return -EOPNOTSUPP; \
5285 #define IO_NETOP_PREP_ASYNC(op) \
5287 static int io_##op##_prep_async(struct io_kiocb *req) \
5289 return -EOPNOTSUPP; \
5292 IO_NETOP_PREP_ASYNC(sendmsg);
5293 IO_NETOP_PREP_ASYNC(recvmsg);
5294 IO_NETOP_PREP_ASYNC(connect);
5295 IO_NETOP_PREP(accept);
5298 #endif /* CONFIG_NET */
5300 struct io_poll_table {
5301 struct poll_table_struct pt;
5302 struct io_kiocb *req;
5307 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5308 __poll_t mask, io_req_tw_func_t func)
5310 /* for instances that support it check for an event match first: */
5311 if (mask && !(mask & poll->events))
5314 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5316 list_del_init(&poll->wait.entry);
5319 req->io_task_work.func = func;
5322 * If this fails, then the task is exiting. When a task exits, the
5323 * work gets canceled, so just cancel this request as well instead
5324 * of executing it. We can't safely execute it anyway, as we may not
5325 * have the needed state needed for it anyway.
5327 io_req_task_work_add(req);
5331 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5332 __acquires(&req->ctx->completion_lock)
5334 struct io_ring_ctx *ctx = req->ctx;
5336 /* req->task == current here, checking PF_EXITING is safe */
5337 if (unlikely(req->task->flags & PF_EXITING))
5338 WRITE_ONCE(poll->canceled, true);
5340 if (!req->result && !READ_ONCE(poll->canceled)) {
5341 struct poll_table_struct pt = { ._key = poll->events };
5343 req->result = vfs_poll(req->file, &pt) & poll->events;
5346 spin_lock(&ctx->completion_lock);
5347 if (!req->result && !READ_ONCE(poll->canceled)) {
5348 add_wait_queue(poll->head, &poll->wait);
5355 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5357 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5358 if (req->opcode == IORING_OP_POLL_ADD)
5359 return req->async_data;
5360 return req->apoll->double_poll;
5363 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5365 if (req->opcode == IORING_OP_POLL_ADD)
5367 return &req->apoll->poll;
5370 static void io_poll_remove_double(struct io_kiocb *req)
5371 __must_hold(&req->ctx->completion_lock)
5373 struct io_poll_iocb *poll = io_poll_get_double(req);
5375 lockdep_assert_held(&req->ctx->completion_lock);
5377 if (poll && poll->head) {
5378 struct wait_queue_head *head = poll->head;
5380 spin_lock_irq(&head->lock);
5381 list_del_init(&poll->wait.entry);
5382 if (poll->wait.private)
5385 spin_unlock_irq(&head->lock);
5389 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5390 __must_hold(&req->ctx->completion_lock)
5392 struct io_ring_ctx *ctx = req->ctx;
5393 unsigned flags = IORING_CQE_F_MORE;
5396 if (READ_ONCE(req->poll.canceled)) {
5398 req->poll.events |= EPOLLONESHOT;
5400 error = mangle_poll(mask);
5402 if (req->poll.events & EPOLLONESHOT)
5404 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5405 req->poll.events |= EPOLLONESHOT;
5408 if (flags & IORING_CQE_F_MORE)
5411 return !(flags & IORING_CQE_F_MORE);
5414 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5415 __must_hold(&req->ctx->completion_lock)
5419 done = __io_poll_complete(req, mask);
5420 io_commit_cqring(req->ctx);
5424 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5426 struct io_ring_ctx *ctx = req->ctx;
5427 struct io_kiocb *nxt;
5429 if (io_poll_rewait(req, &req->poll)) {
5430 spin_unlock(&ctx->completion_lock);
5434 if (req->poll.done) {
5435 spin_unlock(&ctx->completion_lock);
5438 done = __io_poll_complete(req, req->result);
5440 io_poll_remove_double(req);
5441 hash_del(&req->hash_node);
5442 req->poll.done = true;
5445 add_wait_queue(req->poll.head, &req->poll.wait);
5447 io_commit_cqring(ctx);
5448 spin_unlock(&ctx->completion_lock);
5449 io_cqring_ev_posted(ctx);
5452 nxt = io_put_req_find_next(req);
5454 io_req_task_submit(nxt, locked);
5459 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5460 int sync, void *key)
5462 struct io_kiocb *req = wait->private;
5463 struct io_poll_iocb *poll = io_poll_get_single(req);
5464 __poll_t mask = key_to_poll(key);
5465 unsigned long flags;
5467 /* for instances that support it check for an event match first: */
5468 if (mask && !(mask & poll->events))
5470 if (!(poll->events & EPOLLONESHOT))
5471 return poll->wait.func(&poll->wait, mode, sync, key);
5473 list_del_init(&wait->entry);
5478 spin_lock_irqsave(&poll->head->lock, flags);
5479 done = list_empty(&poll->wait.entry);
5481 list_del_init(&poll->wait.entry);
5482 /* make sure double remove sees this as being gone */
5483 wait->private = NULL;
5484 spin_unlock_irqrestore(&poll->head->lock, flags);
5486 /* use wait func handler, so it matches the rq type */
5487 poll->wait.func(&poll->wait, mode, sync, key);
5494 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5495 wait_queue_func_t wake_func)
5499 poll->canceled = false;
5500 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5501 /* mask in events that we always want/need */
5502 poll->events = events | IO_POLL_UNMASK;
5503 INIT_LIST_HEAD(&poll->wait.entry);
5504 init_waitqueue_func_entry(&poll->wait, wake_func);
5507 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5508 struct wait_queue_head *head,
5509 struct io_poll_iocb **poll_ptr)
5511 struct io_kiocb *req = pt->req;
5514 * The file being polled uses multiple waitqueues for poll handling
5515 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5518 if (unlikely(pt->nr_entries)) {
5519 struct io_poll_iocb *poll_one = poll;
5521 /* double add on the same waitqueue head, ignore */
5522 if (poll_one->head == head)
5524 /* already have a 2nd entry, fail a third attempt */
5526 if ((*poll_ptr)->head == head)
5528 pt->error = -EINVAL;
5532 * Can't handle multishot for double wait for now, turn it
5533 * into one-shot mode.
5535 if (!(poll_one->events & EPOLLONESHOT))
5536 poll_one->events |= EPOLLONESHOT;
5537 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5539 pt->error = -ENOMEM;
5542 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5544 poll->wait.private = req;
5551 if (poll->events & EPOLLEXCLUSIVE)
5552 add_wait_queue_exclusive(head, &poll->wait);
5554 add_wait_queue(head, &poll->wait);
5557 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5558 struct poll_table_struct *p)
5560 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5561 struct async_poll *apoll = pt->req->apoll;
5563 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5566 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5568 struct async_poll *apoll = req->apoll;
5569 struct io_ring_ctx *ctx = req->ctx;
5571 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5573 if (io_poll_rewait(req, &apoll->poll)) {
5574 spin_unlock(&ctx->completion_lock);
5578 hash_del(&req->hash_node);
5579 io_poll_remove_double(req);
5580 apoll->poll.done = true;
5581 spin_unlock(&ctx->completion_lock);
5583 if (!READ_ONCE(apoll->poll.canceled))
5584 io_req_task_submit(req, locked);
5586 io_req_complete_failed(req, -ECANCELED);
5589 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5592 struct io_kiocb *req = wait->private;
5593 struct io_poll_iocb *poll = &req->apoll->poll;
5595 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5598 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5601 static void io_poll_req_insert(struct io_kiocb *req)
5603 struct io_ring_ctx *ctx = req->ctx;
5604 struct hlist_head *list;
5606 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5607 hlist_add_head(&req->hash_node, list);
5610 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5611 struct io_poll_iocb *poll,
5612 struct io_poll_table *ipt, __poll_t mask,
5613 wait_queue_func_t wake_func)
5614 __acquires(&ctx->completion_lock)
5616 struct io_ring_ctx *ctx = req->ctx;
5617 bool cancel = false;
5619 INIT_HLIST_NODE(&req->hash_node);
5620 io_init_poll_iocb(poll, mask, wake_func);
5621 poll->file = req->file;
5622 poll->wait.private = req;
5624 ipt->pt._key = mask;
5627 ipt->nr_entries = 0;
5629 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5630 if (unlikely(!ipt->nr_entries) && !ipt->error)
5631 ipt->error = -EINVAL;
5633 spin_lock(&ctx->completion_lock);
5634 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5635 io_poll_remove_double(req);
5636 if (likely(poll->head)) {
5637 spin_lock_irq(&poll->head->lock);
5638 if (unlikely(list_empty(&poll->wait.entry))) {
5644 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5645 list_del_init(&poll->wait.entry);
5647 WRITE_ONCE(poll->canceled, true);
5648 else if (!poll->done) /* actually waiting for an event */
5649 io_poll_req_insert(req);
5650 spin_unlock_irq(&poll->head->lock);
5662 static int io_arm_poll_handler(struct io_kiocb *req)
5664 const struct io_op_def *def = &io_op_defs[req->opcode];
5665 struct io_ring_ctx *ctx = req->ctx;
5666 struct async_poll *apoll;
5667 struct io_poll_table ipt;
5668 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5670 if (!req->file || !file_can_poll(req->file))
5671 return IO_APOLL_ABORTED;
5672 if (req->flags & REQ_F_POLLED)
5673 return IO_APOLL_ABORTED;
5674 if (!def->pollin && !def->pollout)
5675 return IO_APOLL_ABORTED;
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))
5685 mask |= POLLOUT | POLLWRNORM;
5688 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5689 if (unlikely(!apoll))
5690 return IO_APOLL_ABORTED;
5691 apoll->double_poll = NULL;
5693 req->flags |= REQ_F_POLLED;
5694 ipt.pt._qproc = io_async_queue_proc;
5695 io_req_set_refcount(req);
5697 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5699 spin_unlock(&ctx->completion_lock);
5700 if (ret || ipt.error)
5701 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5703 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5704 mask, apoll->poll.events);
5708 static bool __io_poll_remove_one(struct io_kiocb *req,
5709 struct io_poll_iocb *poll, bool do_cancel)
5710 __must_hold(&req->ctx->completion_lock)
5712 bool do_complete = false;
5716 spin_lock_irq(&poll->head->lock);
5718 WRITE_ONCE(poll->canceled, true);
5719 if (!list_empty(&poll->wait.entry)) {
5720 list_del_init(&poll->wait.entry);
5723 spin_unlock_irq(&poll->head->lock);
5724 hash_del(&req->hash_node);
5728 static bool io_poll_remove_one(struct io_kiocb *req)
5729 __must_hold(&req->ctx->completion_lock)
5733 io_poll_remove_double(req);
5734 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5737 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5738 io_commit_cqring(req->ctx);
5740 io_put_req_deferred(req);
5746 * Returns true if we found and killed one or more poll requests
5748 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5751 struct hlist_node *tmp;
5752 struct io_kiocb *req;
5755 spin_lock(&ctx->completion_lock);
5756 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5757 struct hlist_head *list;
5759 list = &ctx->cancel_hash[i];
5760 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5761 if (io_match_task_safe(req, tsk, cancel_all))
5762 posted += io_poll_remove_one(req);
5765 spin_unlock(&ctx->completion_lock);
5768 io_cqring_ev_posted(ctx);
5773 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5775 __must_hold(&ctx->completion_lock)
5777 struct hlist_head *list;
5778 struct io_kiocb *req;
5780 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5781 hlist_for_each_entry(req, list, hash_node) {
5782 if (sqe_addr != req->user_data)
5784 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5791 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5793 __must_hold(&ctx->completion_lock)
5795 struct io_kiocb *req;
5797 req = io_poll_find(ctx, sqe_addr, poll_only);
5800 if (io_poll_remove_one(req))
5806 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5811 events = READ_ONCE(sqe->poll32_events);
5813 events = swahw32(events);
5815 if (!(flags & IORING_POLL_ADD_MULTI))
5816 events |= EPOLLONESHOT;
5817 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5820 static int io_poll_update_prep(struct io_kiocb *req,
5821 const struct io_uring_sqe *sqe)
5823 struct io_poll_update *upd = &req->poll_update;
5826 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5828 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5830 flags = READ_ONCE(sqe->len);
5831 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5832 IORING_POLL_ADD_MULTI))
5834 /* meaningless without update */
5835 if (flags == IORING_POLL_ADD_MULTI)
5838 upd->old_user_data = READ_ONCE(sqe->addr);
5839 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5840 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5842 upd->new_user_data = READ_ONCE(sqe->off);
5843 if (!upd->update_user_data && upd->new_user_data)
5845 if (upd->update_events)
5846 upd->events = io_poll_parse_events(sqe, flags);
5847 else if (sqe->poll32_events)
5853 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5856 struct io_kiocb *req = wait->private;
5857 struct io_poll_iocb *poll = &req->poll;
5859 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5862 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5863 struct poll_table_struct *p)
5865 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5867 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5870 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5872 struct io_poll_iocb *poll = &req->poll;
5875 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5877 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5879 flags = READ_ONCE(sqe->len);
5880 if (flags & ~IORING_POLL_ADD_MULTI)
5883 io_req_set_refcount(req);
5884 poll->events = io_poll_parse_events(sqe, flags);
5888 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5890 struct io_poll_iocb *poll = &req->poll;
5891 struct io_ring_ctx *ctx = req->ctx;
5892 struct io_poll_table ipt;
5896 ipt.pt._qproc = io_poll_queue_proc;
5898 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5901 if (mask) { /* no async, we'd stolen it */
5903 done = io_poll_complete(req, mask);
5905 spin_unlock(&ctx->completion_lock);
5908 io_cqring_ev_posted(ctx);
5915 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5917 struct io_ring_ctx *ctx = req->ctx;
5918 struct io_kiocb *preq;
5922 spin_lock(&ctx->completion_lock);
5923 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5929 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5931 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5936 * Don't allow racy completion with singleshot, as we cannot safely
5937 * update those. For multishot, if we're racing with completion, just
5938 * let completion re-add it.
5940 io_poll_remove_double(preq);
5941 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5942 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5946 /* we now have a detached poll request. reissue. */
5950 spin_unlock(&ctx->completion_lock);
5952 io_req_complete(req, ret);
5955 /* only mask one event flags, keep behavior flags */
5956 if (req->poll_update.update_events) {
5957 preq->poll.events &= ~0xffff;
5958 preq->poll.events |= req->poll_update.events & 0xffff;
5959 preq->poll.events |= IO_POLL_UNMASK;
5961 if (req->poll_update.update_user_data)
5962 preq->user_data = req->poll_update.new_user_data;
5963 spin_unlock(&ctx->completion_lock);
5965 /* complete update request, we're done with it */
5966 io_req_complete(req, ret);
5969 ret = io_poll_add(preq, issue_flags);
5972 io_req_complete(preq, ret);
5978 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5981 io_req_complete_post(req, -ETIME, 0);
5984 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5986 struct io_timeout_data *data = container_of(timer,
5987 struct io_timeout_data, timer);
5988 struct io_kiocb *req = data->req;
5989 struct io_ring_ctx *ctx = req->ctx;
5990 unsigned long flags;
5992 spin_lock_irqsave(&ctx->timeout_lock, flags);
5993 list_del_init(&req->timeout.list);
5994 atomic_set(&req->ctx->cq_timeouts,
5995 atomic_read(&req->ctx->cq_timeouts) + 1);
5996 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5998 req->io_task_work.func = io_req_task_timeout;
5999 io_req_task_work_add(req);
6000 return HRTIMER_NORESTART;
6003 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
6005 __must_hold(&ctx->timeout_lock)
6007 struct io_timeout_data *io;
6008 struct io_kiocb *req;
6011 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
6012 found = user_data == req->user_data;
6017 return ERR_PTR(-ENOENT);
6019 io = req->async_data;
6020 if (hrtimer_try_to_cancel(&io->timer) == -1)
6021 return ERR_PTR(-EALREADY);
6022 list_del_init(&req->timeout.list);
6026 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
6027 __must_hold(&ctx->completion_lock)
6028 __must_hold(&ctx->timeout_lock)
6030 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6033 return PTR_ERR(req);
6036 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
6037 io_put_req_deferred(req);
6041 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6043 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6044 case IORING_TIMEOUT_BOOTTIME:
6045 return CLOCK_BOOTTIME;
6046 case IORING_TIMEOUT_REALTIME:
6047 return CLOCK_REALTIME;
6049 /* can't happen, vetted at prep time */
6053 return CLOCK_MONOTONIC;
6057 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6058 struct timespec64 *ts, enum hrtimer_mode mode)
6059 __must_hold(&ctx->timeout_lock)
6061 struct io_timeout_data *io;
6062 struct io_kiocb *req;
6065 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6066 found = user_data == req->user_data;
6073 io = req->async_data;
6074 if (hrtimer_try_to_cancel(&io->timer) == -1)
6076 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6077 io->timer.function = io_link_timeout_fn;
6078 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6082 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6083 struct timespec64 *ts, enum hrtimer_mode mode)
6084 __must_hold(&ctx->timeout_lock)
6086 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6087 struct io_timeout_data *data;
6090 return PTR_ERR(req);
6092 req->timeout.off = 0; /* noseq */
6093 data = req->async_data;
6094 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6095 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6096 data->timer.function = io_timeout_fn;
6097 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6101 static int io_timeout_remove_prep(struct io_kiocb *req,
6102 const struct io_uring_sqe *sqe)
6104 struct io_timeout_rem *tr = &req->timeout_rem;
6106 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6108 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6110 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6113 tr->ltimeout = false;
6114 tr->addr = READ_ONCE(sqe->addr);
6115 tr->flags = READ_ONCE(sqe->timeout_flags);
6116 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6117 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6119 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6120 tr->ltimeout = true;
6121 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6123 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6125 } else if (tr->flags) {
6126 /* timeout removal doesn't support flags */
6133 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6135 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6140 * Remove or update an existing timeout command
6142 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6144 struct io_timeout_rem *tr = &req->timeout_rem;
6145 struct io_ring_ctx *ctx = req->ctx;
6148 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6149 spin_lock(&ctx->completion_lock);
6150 spin_lock_irq(&ctx->timeout_lock);
6151 ret = io_timeout_cancel(ctx, tr->addr);
6152 spin_unlock_irq(&ctx->timeout_lock);
6153 spin_unlock(&ctx->completion_lock);
6155 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6157 spin_lock_irq(&ctx->timeout_lock);
6159 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6161 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6162 spin_unlock_irq(&ctx->timeout_lock);
6167 io_req_complete_post(req, ret, 0);
6171 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6172 bool is_timeout_link)
6174 struct io_timeout_data *data;
6176 u32 off = READ_ONCE(sqe->off);
6178 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6180 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6183 if (off && is_timeout_link)
6185 flags = READ_ONCE(sqe->timeout_flags);
6186 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6188 /* more than one clock specified is invalid, obviously */
6189 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6192 INIT_LIST_HEAD(&req->timeout.list);
6193 req->timeout.off = off;
6194 if (unlikely(off && !req->ctx->off_timeout_used))
6195 req->ctx->off_timeout_used = true;
6197 if (!req->async_data && io_alloc_async_data(req))
6200 data = req->async_data;
6202 data->flags = flags;
6204 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6207 INIT_LIST_HEAD(&req->timeout.list);
6208 data->mode = io_translate_timeout_mode(flags);
6209 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6211 if (is_timeout_link) {
6212 struct io_submit_link *link = &req->ctx->submit_state.link;
6216 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6218 req->timeout.head = link->last;
6219 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6224 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6226 struct io_ring_ctx *ctx = req->ctx;
6227 struct io_timeout_data *data = req->async_data;
6228 struct list_head *entry;
6229 u32 tail, off = req->timeout.off;
6231 spin_lock_irq(&ctx->timeout_lock);
6234 * sqe->off holds how many events that need to occur for this
6235 * timeout event to be satisfied. If it isn't set, then this is
6236 * a pure timeout request, sequence isn't used.
6238 if (io_is_timeout_noseq(req)) {
6239 entry = ctx->timeout_list.prev;
6243 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6244 req->timeout.target_seq = tail + off;
6246 /* Update the last seq here in case io_flush_timeouts() hasn't.
6247 * This is safe because ->completion_lock is held, and submissions
6248 * and completions are never mixed in the same ->completion_lock section.
6250 ctx->cq_last_tm_flush = tail;
6253 * Insertion sort, ensuring the first entry in the list is always
6254 * the one we need first.
6256 list_for_each_prev(entry, &ctx->timeout_list) {
6257 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6260 if (io_is_timeout_noseq(nxt))
6262 /* nxt.seq is behind @tail, otherwise would've been completed */
6263 if (off >= nxt->timeout.target_seq - tail)
6267 list_add(&req->timeout.list, entry);
6268 data->timer.function = io_timeout_fn;
6269 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6270 spin_unlock_irq(&ctx->timeout_lock);
6274 struct io_cancel_data {
6275 struct io_ring_ctx *ctx;
6279 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6281 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6282 struct io_cancel_data *cd = data;
6284 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6287 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6288 struct io_ring_ctx *ctx)
6290 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6291 enum io_wq_cancel cancel_ret;
6294 if (!tctx || !tctx->io_wq)
6297 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6298 switch (cancel_ret) {
6299 case IO_WQ_CANCEL_OK:
6302 case IO_WQ_CANCEL_RUNNING:
6305 case IO_WQ_CANCEL_NOTFOUND:
6313 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6315 struct io_ring_ctx *ctx = req->ctx;
6318 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6320 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6324 spin_lock(&ctx->completion_lock);
6325 spin_lock_irq(&ctx->timeout_lock);
6326 ret = io_timeout_cancel(ctx, sqe_addr);
6327 spin_unlock_irq(&ctx->timeout_lock);
6330 ret = io_poll_cancel(ctx, sqe_addr, false);
6332 spin_unlock(&ctx->completion_lock);
6336 static int io_async_cancel_prep(struct io_kiocb *req,
6337 const struct io_uring_sqe *sqe)
6339 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6341 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6343 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6347 req->cancel.addr = READ_ONCE(sqe->addr);
6351 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6353 struct io_ring_ctx *ctx = req->ctx;
6354 u64 sqe_addr = req->cancel.addr;
6355 struct io_tctx_node *node;
6358 ret = io_try_cancel_userdata(req, sqe_addr);
6362 /* slow path, try all io-wq's */
6363 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6365 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6366 struct io_uring_task *tctx = node->task->io_uring;
6368 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6372 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6376 io_req_complete_post(req, ret, 0);
6380 static int io_rsrc_update_prep(struct io_kiocb *req,
6381 const struct io_uring_sqe *sqe)
6383 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6385 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6388 req->rsrc_update.offset = READ_ONCE(sqe->off);
6389 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6390 if (!req->rsrc_update.nr_args)
6392 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6396 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6398 struct io_ring_ctx *ctx = req->ctx;
6399 struct io_uring_rsrc_update2 up;
6402 up.offset = req->rsrc_update.offset;
6403 up.data = req->rsrc_update.arg;
6409 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6410 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6411 &up, req->rsrc_update.nr_args);
6412 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6416 __io_req_complete(req, issue_flags, ret, 0);
6420 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6422 switch (req->opcode) {
6425 case IORING_OP_READV:
6426 case IORING_OP_READ_FIXED:
6427 case IORING_OP_READ:
6428 return io_read_prep(req, sqe);
6429 case IORING_OP_WRITEV:
6430 case IORING_OP_WRITE_FIXED:
6431 case IORING_OP_WRITE:
6432 return io_write_prep(req, sqe);
6433 case IORING_OP_POLL_ADD:
6434 return io_poll_add_prep(req, sqe);
6435 case IORING_OP_POLL_REMOVE:
6436 return io_poll_update_prep(req, sqe);
6437 case IORING_OP_FSYNC:
6438 return io_fsync_prep(req, sqe);
6439 case IORING_OP_SYNC_FILE_RANGE:
6440 return io_sfr_prep(req, sqe);
6441 case IORING_OP_SENDMSG:
6442 case IORING_OP_SEND:
6443 return io_sendmsg_prep(req, sqe);
6444 case IORING_OP_RECVMSG:
6445 case IORING_OP_RECV:
6446 return io_recvmsg_prep(req, sqe);
6447 case IORING_OP_CONNECT:
6448 return io_connect_prep(req, sqe);
6449 case IORING_OP_TIMEOUT:
6450 return io_timeout_prep(req, sqe, false);
6451 case IORING_OP_TIMEOUT_REMOVE:
6452 return io_timeout_remove_prep(req, sqe);
6453 case IORING_OP_ASYNC_CANCEL:
6454 return io_async_cancel_prep(req, sqe);
6455 case IORING_OP_LINK_TIMEOUT:
6456 return io_timeout_prep(req, sqe, true);
6457 case IORING_OP_ACCEPT:
6458 return io_accept_prep(req, sqe);
6459 case IORING_OP_FALLOCATE:
6460 return io_fallocate_prep(req, sqe);
6461 case IORING_OP_OPENAT:
6462 return io_openat_prep(req, sqe);
6463 case IORING_OP_CLOSE:
6464 return io_close_prep(req, sqe);
6465 case IORING_OP_FILES_UPDATE:
6466 return io_rsrc_update_prep(req, sqe);
6467 case IORING_OP_STATX:
6468 return io_statx_prep(req, sqe);
6469 case IORING_OP_FADVISE:
6470 return io_fadvise_prep(req, sqe);
6471 case IORING_OP_MADVISE:
6472 return io_madvise_prep(req, sqe);
6473 case IORING_OP_OPENAT2:
6474 return io_openat2_prep(req, sqe);
6475 case IORING_OP_EPOLL_CTL:
6476 return io_epoll_ctl_prep(req, sqe);
6477 case IORING_OP_SPLICE:
6478 return io_splice_prep(req, sqe);
6479 case IORING_OP_PROVIDE_BUFFERS:
6480 return io_provide_buffers_prep(req, sqe);
6481 case IORING_OP_REMOVE_BUFFERS:
6482 return io_remove_buffers_prep(req, sqe);
6484 return io_tee_prep(req, sqe);
6485 case IORING_OP_SHUTDOWN:
6486 return io_shutdown_prep(req, sqe);
6487 case IORING_OP_RENAMEAT:
6488 return io_renameat_prep(req, sqe);
6489 case IORING_OP_UNLINKAT:
6490 return io_unlinkat_prep(req, sqe);
6491 case IORING_OP_MKDIRAT:
6492 return io_mkdirat_prep(req, sqe);
6493 case IORING_OP_SYMLINKAT:
6494 return io_symlinkat_prep(req, sqe);
6495 case IORING_OP_LINKAT:
6496 return io_linkat_prep(req, sqe);
6499 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6504 static int io_req_prep_async(struct io_kiocb *req)
6506 if (!io_op_defs[req->opcode].needs_async_setup)
6508 if (WARN_ON_ONCE(req->async_data))
6510 if (io_alloc_async_data(req))
6513 switch (req->opcode) {
6514 case IORING_OP_READV:
6515 return io_rw_prep_async(req, READ);
6516 case IORING_OP_WRITEV:
6517 return io_rw_prep_async(req, WRITE);
6518 case IORING_OP_SENDMSG:
6519 return io_sendmsg_prep_async(req);
6520 case IORING_OP_RECVMSG:
6521 return io_recvmsg_prep_async(req);
6522 case IORING_OP_CONNECT:
6523 return io_connect_prep_async(req);
6525 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6530 static u32 io_get_sequence(struct io_kiocb *req)
6532 u32 seq = req->ctx->cached_sq_head;
6534 /* need original cached_sq_head, but it was increased for each req */
6535 io_for_each_link(req, req)
6540 static bool io_drain_req(struct io_kiocb *req)
6542 struct io_kiocb *pos;
6543 struct io_ring_ctx *ctx = req->ctx;
6544 struct io_defer_entry *de;
6548 if (req->flags & REQ_F_FAIL) {
6549 io_req_complete_fail_submit(req);
6554 * If we need to drain a request in the middle of a link, drain the
6555 * head request and the next request/link after the current link.
6556 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6557 * maintained for every request of our link.
6559 if (ctx->drain_next) {
6560 req->flags |= REQ_F_IO_DRAIN;
6561 ctx->drain_next = false;
6563 /* not interested in head, start from the first linked */
6564 io_for_each_link(pos, req->link) {
6565 if (pos->flags & REQ_F_IO_DRAIN) {
6566 ctx->drain_next = true;
6567 req->flags |= REQ_F_IO_DRAIN;
6572 /* Still need defer if there is pending req in defer list. */
6573 spin_lock(&ctx->completion_lock);
6574 if (likely(list_empty_careful(&ctx->defer_list) &&
6575 !(req->flags & REQ_F_IO_DRAIN))) {
6576 spin_unlock(&ctx->completion_lock);
6577 ctx->drain_active = false;
6580 spin_unlock(&ctx->completion_lock);
6582 seq = io_get_sequence(req);
6583 /* Still a chance to pass the sequence check */
6584 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6587 ret = io_req_prep_async(req);
6590 io_prep_async_link(req);
6591 de = kmalloc(sizeof(*de), GFP_KERNEL);
6595 io_req_complete_failed(req, ret);
6599 spin_lock(&ctx->completion_lock);
6600 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6601 spin_unlock(&ctx->completion_lock);
6603 io_queue_async_work(req, NULL);
6607 trace_io_uring_defer(ctx, req, req->user_data);
6610 list_add_tail(&de->list, &ctx->defer_list);
6611 spin_unlock(&ctx->completion_lock);
6615 static void io_clean_op(struct io_kiocb *req)
6617 if (req->flags & REQ_F_BUFFER_SELECTED) {
6618 switch (req->opcode) {
6619 case IORING_OP_READV:
6620 case IORING_OP_READ_FIXED:
6621 case IORING_OP_READ:
6622 kfree((void *)(unsigned long)req->rw.addr);
6624 case IORING_OP_RECVMSG:
6625 case IORING_OP_RECV:
6626 kfree(req->sr_msg.kbuf);
6631 if (req->flags & REQ_F_NEED_CLEANUP) {
6632 switch (req->opcode) {
6633 case IORING_OP_READV:
6634 case IORING_OP_READ_FIXED:
6635 case IORING_OP_READ:
6636 case IORING_OP_WRITEV:
6637 case IORING_OP_WRITE_FIXED:
6638 case IORING_OP_WRITE: {
6639 struct io_async_rw *io = req->async_data;
6641 kfree(io->free_iovec);
6644 case IORING_OP_RECVMSG:
6645 case IORING_OP_SENDMSG: {
6646 struct io_async_msghdr *io = req->async_data;
6648 kfree(io->free_iov);
6651 case IORING_OP_OPENAT:
6652 case IORING_OP_OPENAT2:
6653 if (req->open.filename)
6654 putname(req->open.filename);
6656 case IORING_OP_RENAMEAT:
6657 putname(req->rename.oldpath);
6658 putname(req->rename.newpath);
6660 case IORING_OP_UNLINKAT:
6661 putname(req->unlink.filename);
6663 case IORING_OP_MKDIRAT:
6664 putname(req->mkdir.filename);
6666 case IORING_OP_SYMLINKAT:
6667 putname(req->symlink.oldpath);
6668 putname(req->symlink.newpath);
6670 case IORING_OP_LINKAT:
6671 putname(req->hardlink.oldpath);
6672 putname(req->hardlink.newpath);
6676 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6677 kfree(req->apoll->double_poll);
6681 if (req->flags & REQ_F_INFLIGHT) {
6682 struct io_uring_task *tctx = req->task->io_uring;
6684 atomic_dec(&tctx->inflight_tracked);
6686 if (req->flags & REQ_F_CREDS)
6687 put_cred(req->creds);
6689 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6692 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6694 struct io_ring_ctx *ctx = req->ctx;
6695 const struct cred *creds = NULL;
6698 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6699 creds = override_creds(req->creds);
6701 switch (req->opcode) {
6703 ret = io_nop(req, issue_flags);
6705 case IORING_OP_READV:
6706 case IORING_OP_READ_FIXED:
6707 case IORING_OP_READ:
6708 ret = io_read(req, issue_flags);
6710 case IORING_OP_WRITEV:
6711 case IORING_OP_WRITE_FIXED:
6712 case IORING_OP_WRITE:
6713 ret = io_write(req, issue_flags);
6715 case IORING_OP_FSYNC:
6716 ret = io_fsync(req, issue_flags);
6718 case IORING_OP_POLL_ADD:
6719 ret = io_poll_add(req, issue_flags);
6721 case IORING_OP_POLL_REMOVE:
6722 ret = io_poll_update(req, issue_flags);
6724 case IORING_OP_SYNC_FILE_RANGE:
6725 ret = io_sync_file_range(req, issue_flags);
6727 case IORING_OP_SENDMSG:
6728 ret = io_sendmsg(req, issue_flags);
6730 case IORING_OP_SEND:
6731 ret = io_send(req, issue_flags);
6733 case IORING_OP_RECVMSG:
6734 ret = io_recvmsg(req, issue_flags);
6736 case IORING_OP_RECV:
6737 ret = io_recv(req, issue_flags);
6739 case IORING_OP_TIMEOUT:
6740 ret = io_timeout(req, issue_flags);
6742 case IORING_OP_TIMEOUT_REMOVE:
6743 ret = io_timeout_remove(req, issue_flags);
6745 case IORING_OP_ACCEPT:
6746 ret = io_accept(req, issue_flags);
6748 case IORING_OP_CONNECT:
6749 ret = io_connect(req, issue_flags);
6751 case IORING_OP_ASYNC_CANCEL:
6752 ret = io_async_cancel(req, issue_flags);
6754 case IORING_OP_FALLOCATE:
6755 ret = io_fallocate(req, issue_flags);
6757 case IORING_OP_OPENAT:
6758 ret = io_openat(req, issue_flags);
6760 case IORING_OP_CLOSE:
6761 ret = io_close(req, issue_flags);
6763 case IORING_OP_FILES_UPDATE:
6764 ret = io_files_update(req, issue_flags);
6766 case IORING_OP_STATX:
6767 ret = io_statx(req, issue_flags);
6769 case IORING_OP_FADVISE:
6770 ret = io_fadvise(req, issue_flags);
6772 case IORING_OP_MADVISE:
6773 ret = io_madvise(req, issue_flags);
6775 case IORING_OP_OPENAT2:
6776 ret = io_openat2(req, issue_flags);
6778 case IORING_OP_EPOLL_CTL:
6779 ret = io_epoll_ctl(req, issue_flags);
6781 case IORING_OP_SPLICE:
6782 ret = io_splice(req, issue_flags);
6784 case IORING_OP_PROVIDE_BUFFERS:
6785 ret = io_provide_buffers(req, issue_flags);
6787 case IORING_OP_REMOVE_BUFFERS:
6788 ret = io_remove_buffers(req, issue_flags);
6791 ret = io_tee(req, issue_flags);
6793 case IORING_OP_SHUTDOWN:
6794 ret = io_shutdown(req, issue_flags);
6796 case IORING_OP_RENAMEAT:
6797 ret = io_renameat(req, issue_flags);
6799 case IORING_OP_UNLINKAT:
6800 ret = io_unlinkat(req, issue_flags);
6802 case IORING_OP_MKDIRAT:
6803 ret = io_mkdirat(req, issue_flags);
6805 case IORING_OP_SYMLINKAT:
6806 ret = io_symlinkat(req, issue_flags);
6808 case IORING_OP_LINKAT:
6809 ret = io_linkat(req, issue_flags);
6817 revert_creds(creds);
6820 /* If the op doesn't have a file, we're not polling for it */
6821 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6822 io_iopoll_req_issued(req);
6827 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6829 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6831 req = io_put_req_find_next(req);
6832 return req ? &req->work : NULL;
6835 static void io_wq_submit_work(struct io_wq_work *work)
6837 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6838 struct io_kiocb *timeout;
6841 /* one will be dropped by ->io_free_work() after returning to io-wq */
6842 if (!(req->flags & REQ_F_REFCOUNT))
6843 __io_req_set_refcount(req, 2);
6847 timeout = io_prep_linked_timeout(req);
6849 io_queue_linked_timeout(timeout);
6851 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6852 if (work->flags & IO_WQ_WORK_CANCEL)
6857 ret = io_issue_sqe(req, 0);
6859 * We can get EAGAIN for polled IO even though we're
6860 * forcing a sync submission from here, since we can't
6861 * wait for request slots on the block side.
6869 /* avoid locking problems by failing it from a clean context */
6871 io_req_task_queue_fail(req, ret);
6874 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6877 return &table->files[i];
6880 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6883 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6885 return (struct file *) (slot->file_ptr & FFS_MASK);
6888 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6890 unsigned long file_ptr = (unsigned long) file;
6892 if (__io_file_supports_nowait(file, READ))
6893 file_ptr |= FFS_ASYNC_READ;
6894 if (__io_file_supports_nowait(file, WRITE))
6895 file_ptr |= FFS_ASYNC_WRITE;
6896 if (S_ISREG(file_inode(file)->i_mode))
6897 file_ptr |= FFS_ISREG;
6898 file_slot->file_ptr = file_ptr;
6901 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6902 struct io_kiocb *req, int fd)
6905 unsigned long file_ptr;
6907 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6909 fd = array_index_nospec(fd, ctx->nr_user_files);
6910 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6911 file = (struct file *) (file_ptr & FFS_MASK);
6912 file_ptr &= ~FFS_MASK;
6913 /* mask in overlapping REQ_F and FFS bits */
6914 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6915 io_req_set_rsrc_node(req);
6919 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6920 struct io_kiocb *req, int fd)
6922 struct file *file = fget(fd);
6924 trace_io_uring_file_get(ctx, fd);
6926 /* we don't allow fixed io_uring files */
6927 if (file && unlikely(file->f_op == &io_uring_fops))
6928 io_req_track_inflight(req);
6932 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6933 struct io_kiocb *req, int fd, bool fixed)
6936 return io_file_get_fixed(ctx, req, fd);
6938 return io_file_get_normal(ctx, req, fd);
6941 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6943 struct io_kiocb *prev = req->timeout.prev;
6947 if (!(req->task->flags & PF_EXITING))
6948 ret = io_try_cancel_userdata(req, prev->user_data);
6949 io_req_complete_post(req, ret ?: -ETIME, 0);
6952 io_req_complete_post(req, -ETIME, 0);
6956 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6958 struct io_timeout_data *data = container_of(timer,
6959 struct io_timeout_data, timer);
6960 struct io_kiocb *prev, *req = data->req;
6961 struct io_ring_ctx *ctx = req->ctx;
6962 unsigned long flags;
6964 spin_lock_irqsave(&ctx->timeout_lock, flags);
6965 prev = req->timeout.head;
6966 req->timeout.head = NULL;
6969 * We don't expect the list to be empty, that will only happen if we
6970 * race with the completion of the linked work.
6973 io_remove_next_linked(prev);
6974 if (!req_ref_inc_not_zero(prev))
6977 list_del(&req->timeout.list);
6978 req->timeout.prev = prev;
6979 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6981 req->io_task_work.func = io_req_task_link_timeout;
6982 io_req_task_work_add(req);
6983 return HRTIMER_NORESTART;
6986 static void io_queue_linked_timeout(struct io_kiocb *req)
6988 struct io_ring_ctx *ctx = req->ctx;
6990 spin_lock_irq(&ctx->timeout_lock);
6992 * If the back reference is NULL, then our linked request finished
6993 * before we got a chance to setup the timer
6995 if (req->timeout.head) {
6996 struct io_timeout_data *data = req->async_data;
6998 data->timer.function = io_link_timeout_fn;
6999 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
7001 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
7003 spin_unlock_irq(&ctx->timeout_lock);
7004 /* drop submission reference */
7008 static void __io_queue_sqe(struct io_kiocb *req)
7009 __must_hold(&req->ctx->uring_lock)
7011 struct io_kiocb *linked_timeout;
7015 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
7018 * We async punt it if the file wasn't marked NOWAIT, or if the file
7019 * doesn't support non-blocking read/write attempts
7022 if (req->flags & REQ_F_COMPLETE_INLINE) {
7023 struct io_ring_ctx *ctx = req->ctx;
7024 struct io_submit_state *state = &ctx->submit_state;
7026 state->compl_reqs[state->compl_nr++] = req;
7027 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
7028 io_submit_flush_completions(ctx);
7032 linked_timeout = io_prep_linked_timeout(req);
7034 io_queue_linked_timeout(linked_timeout);
7035 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
7036 linked_timeout = io_prep_linked_timeout(req);
7038 switch (io_arm_poll_handler(req)) {
7039 case IO_APOLL_READY:
7041 io_queue_linked_timeout(linked_timeout);
7043 case IO_APOLL_ABORTED:
7045 * Queued up for async execution, worker will release
7046 * submit reference when the iocb is actually submitted.
7048 io_queue_async_work(req, NULL);
7053 io_queue_linked_timeout(linked_timeout);
7055 io_req_complete_failed(req, ret);
7059 static inline void io_queue_sqe(struct io_kiocb *req)
7060 __must_hold(&req->ctx->uring_lock)
7062 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7065 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7066 __io_queue_sqe(req);
7067 } else if (req->flags & REQ_F_FAIL) {
7068 io_req_complete_fail_submit(req);
7070 int ret = io_req_prep_async(req);
7073 io_req_complete_failed(req, ret);
7075 io_queue_async_work(req, NULL);
7080 * Check SQE restrictions (opcode and flags).
7082 * Returns 'true' if SQE is allowed, 'false' otherwise.
7084 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7085 struct io_kiocb *req,
7086 unsigned int sqe_flags)
7088 if (likely(!ctx->restricted))
7091 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7094 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7095 ctx->restrictions.sqe_flags_required)
7098 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7099 ctx->restrictions.sqe_flags_required))
7105 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7106 const struct io_uring_sqe *sqe)
7107 __must_hold(&ctx->uring_lock)
7109 struct io_submit_state *state;
7110 unsigned int sqe_flags;
7111 int personality, ret = 0;
7113 /* req is partially pre-initialised, see io_preinit_req() */
7114 req->opcode = READ_ONCE(sqe->opcode);
7115 /* same numerical values with corresponding REQ_F_*, safe to copy */
7116 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7117 req->user_data = READ_ONCE(sqe->user_data);
7119 req->fixed_rsrc_refs = NULL;
7120 req->task = current;
7122 /* enforce forwards compatibility on users */
7123 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7125 if (unlikely(req->opcode >= IORING_OP_LAST))
7127 if (!io_check_restriction(ctx, req, sqe_flags))
7130 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7131 !io_op_defs[req->opcode].buffer_select)
7133 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7134 ctx->drain_active = true;
7136 personality = READ_ONCE(sqe->personality);
7138 req->creds = xa_load(&ctx->personalities, personality);
7141 get_cred(req->creds);
7142 req->flags |= REQ_F_CREDS;
7144 state = &ctx->submit_state;
7147 * Plug now if we have more than 1 IO left after this, and the target
7148 * is potentially a read/write to block based storage.
7150 if (!state->plug_started && state->ios_left > 1 &&
7151 io_op_defs[req->opcode].plug) {
7152 blk_start_plug(&state->plug);
7153 state->plug_started = true;
7156 if (io_op_defs[req->opcode].needs_file) {
7157 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7158 (sqe_flags & IOSQE_FIXED_FILE));
7159 if (unlikely(!req->file))
7167 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7168 const struct io_uring_sqe *sqe)
7169 __must_hold(&ctx->uring_lock)
7171 struct io_submit_link *link = &ctx->submit_state.link;
7174 ret = io_init_req(ctx, req, sqe);
7175 if (unlikely(ret)) {
7177 /* fail even hard links since we don't submit */
7180 * we can judge a link req is failed or cancelled by if
7181 * REQ_F_FAIL is set, but the head is an exception since
7182 * it may be set REQ_F_FAIL because of other req's failure
7183 * so let's leverage req->result to distinguish if a head
7184 * is set REQ_F_FAIL because of its failure or other req's
7185 * failure so that we can set the correct ret code for it.
7186 * init result here to avoid affecting the normal path.
7188 if (!(link->head->flags & REQ_F_FAIL))
7189 req_fail_link_node(link->head, -ECANCELED);
7190 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7192 * the current req is a normal req, we should return
7193 * error and thus break the submittion loop.
7195 io_req_complete_failed(req, ret);
7198 req_fail_link_node(req, ret);
7200 ret = io_req_prep(req, sqe);
7205 /* don't need @sqe from now on */
7206 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7208 ctx->flags & IORING_SETUP_SQPOLL);
7211 * If we already have a head request, queue this one for async
7212 * submittal once the head completes. If we don't have a head but
7213 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7214 * submitted sync once the chain is complete. If none of those
7215 * conditions are true (normal request), then just queue it.
7218 struct io_kiocb *head = link->head;
7220 if (!(req->flags & REQ_F_FAIL)) {
7221 ret = io_req_prep_async(req);
7222 if (unlikely(ret)) {
7223 req_fail_link_node(req, ret);
7224 if (!(head->flags & REQ_F_FAIL))
7225 req_fail_link_node(head, -ECANCELED);
7228 trace_io_uring_link(ctx, req, head);
7229 link->last->link = req;
7232 /* last request of a link, enqueue the link */
7233 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7238 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7250 * Batched submission is done, ensure local IO is flushed out.
7252 static void io_submit_state_end(struct io_submit_state *state,
7253 struct io_ring_ctx *ctx)
7255 if (state->link.head)
7256 io_queue_sqe(state->link.head);
7257 if (state->compl_nr)
7258 io_submit_flush_completions(ctx);
7259 if (state->plug_started)
7260 blk_finish_plug(&state->plug);
7264 * Start submission side cache.
7266 static void io_submit_state_start(struct io_submit_state *state,
7267 unsigned int max_ios)
7269 state->plug_started = false;
7270 state->ios_left = max_ios;
7271 /* set only head, no need to init link_last in advance */
7272 state->link.head = NULL;
7275 static void io_commit_sqring(struct io_ring_ctx *ctx)
7277 struct io_rings *rings = ctx->rings;
7280 * Ensure any loads from the SQEs are done at this point,
7281 * since once we write the new head, the application could
7282 * write new data to them.
7284 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7288 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7289 * that is mapped by userspace. This means that care needs to be taken to
7290 * ensure that reads are stable, as we cannot rely on userspace always
7291 * being a good citizen. If members of the sqe are validated and then later
7292 * used, it's important that those reads are done through READ_ONCE() to
7293 * prevent a re-load down the line.
7295 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7297 unsigned head, mask = ctx->sq_entries - 1;
7298 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7301 * The cached sq head (or cq tail) serves two purposes:
7303 * 1) allows us to batch the cost of updating the user visible
7305 * 2) allows the kernel side to track the head on its own, even
7306 * though the application is the one updating it.
7308 head = READ_ONCE(ctx->sq_array[sq_idx]);
7309 if (likely(head < ctx->sq_entries))
7310 return &ctx->sq_sqes[head];
7312 /* drop invalid entries */
7314 WRITE_ONCE(ctx->rings->sq_dropped,
7315 READ_ONCE(ctx->rings->sq_dropped) + 1);
7319 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7320 __must_hold(&ctx->uring_lock)
7324 /* make sure SQ entry isn't read before tail */
7325 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7326 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7328 io_get_task_refs(nr);
7330 io_submit_state_start(&ctx->submit_state, nr);
7331 while (submitted < nr) {
7332 const struct io_uring_sqe *sqe;
7333 struct io_kiocb *req;
7335 req = io_alloc_req(ctx);
7336 if (unlikely(!req)) {
7338 submitted = -EAGAIN;
7341 sqe = io_get_sqe(ctx);
7342 if (unlikely(!sqe)) {
7343 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7346 /* will complete beyond this point, count as submitted */
7348 if (io_submit_sqe(ctx, req, sqe))
7352 if (unlikely(submitted != nr)) {
7353 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7354 int unused = nr - ref_used;
7356 current->io_uring->cached_refs += unused;
7357 percpu_ref_put_many(&ctx->refs, unused);
7360 io_submit_state_end(&ctx->submit_state, ctx);
7361 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7362 io_commit_sqring(ctx);
7367 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7369 return READ_ONCE(sqd->state);
7372 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7374 /* Tell userspace we may need a wakeup call */
7375 spin_lock(&ctx->completion_lock);
7376 WRITE_ONCE(ctx->rings->sq_flags,
7377 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7378 spin_unlock(&ctx->completion_lock);
7381 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
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 int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7391 unsigned int to_submit;
7394 to_submit = io_sqring_entries(ctx);
7395 /* if we're handling multiple rings, cap submit size for fairness */
7396 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7397 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7399 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7400 unsigned nr_events = 0;
7401 const struct cred *creds = NULL;
7403 if (ctx->sq_creds != current_cred())
7404 creds = override_creds(ctx->sq_creds);
7406 mutex_lock(&ctx->uring_lock);
7407 if (!list_empty(&ctx->iopoll_list))
7408 io_do_iopoll(ctx, &nr_events, 0);
7411 * Don't submit if refs are dying, good for io_uring_register(),
7412 * but also it is relied upon by io_ring_exit_work()
7414 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7415 !(ctx->flags & IORING_SETUP_R_DISABLED))
7416 ret = io_submit_sqes(ctx, to_submit);
7417 mutex_unlock(&ctx->uring_lock);
7419 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7420 wake_up(&ctx->sqo_sq_wait);
7422 revert_creds(creds);
7428 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7430 struct io_ring_ctx *ctx;
7431 unsigned sq_thread_idle = 0;
7433 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7434 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7435 sqd->sq_thread_idle = sq_thread_idle;
7438 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7440 bool did_sig = false;
7441 struct ksignal ksig;
7443 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7444 signal_pending(current)) {
7445 mutex_unlock(&sqd->lock);
7446 if (signal_pending(current))
7447 did_sig = get_signal(&ksig);
7449 mutex_lock(&sqd->lock);
7451 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7454 static int io_sq_thread(void *data)
7456 struct io_sq_data *sqd = data;
7457 struct io_ring_ctx *ctx;
7458 unsigned long timeout = 0;
7459 char buf[TASK_COMM_LEN];
7462 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7463 set_task_comm(current, buf);
7465 if (sqd->sq_cpu != -1)
7466 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7468 set_cpus_allowed_ptr(current, cpu_online_mask);
7469 current->flags |= PF_NO_SETAFFINITY;
7471 mutex_lock(&sqd->lock);
7473 bool cap_entries, sqt_spin = false;
7475 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7476 if (io_sqd_handle_event(sqd))
7478 timeout = jiffies + sqd->sq_thread_idle;
7481 cap_entries = !list_is_singular(&sqd->ctx_list);
7482 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7483 int ret = __io_sq_thread(ctx, cap_entries);
7485 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7488 if (io_run_task_work())
7491 if (sqt_spin || !time_after(jiffies, timeout)) {
7494 timeout = jiffies + sqd->sq_thread_idle;
7498 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7499 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7500 bool needs_sched = true;
7502 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7503 io_ring_set_wakeup_flag(ctx);
7505 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7506 !list_empty_careful(&ctx->iopoll_list)) {
7507 needs_sched = false;
7510 if (io_sqring_entries(ctx)) {
7511 needs_sched = false;
7517 mutex_unlock(&sqd->lock);
7519 mutex_lock(&sqd->lock);
7521 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7522 io_ring_clear_wakeup_flag(ctx);
7525 finish_wait(&sqd->wait, &wait);
7526 timeout = jiffies + sqd->sq_thread_idle;
7529 io_uring_cancel_generic(true, sqd);
7531 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7532 io_ring_set_wakeup_flag(ctx);
7534 mutex_unlock(&sqd->lock);
7536 complete(&sqd->exited);
7540 struct io_wait_queue {
7541 struct wait_queue_entry wq;
7542 struct io_ring_ctx *ctx;
7544 unsigned nr_timeouts;
7547 static inline bool io_should_wake(struct io_wait_queue *iowq)
7549 struct io_ring_ctx *ctx = iowq->ctx;
7550 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7553 * Wake up if we have enough events, or if a timeout occurred since we
7554 * started waiting. For timeouts, we always want to return to userspace,
7555 * regardless of event count.
7557 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7560 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7561 int wake_flags, void *key)
7563 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7567 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7568 * the task, and the next invocation will do it.
7570 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7571 return autoremove_wake_function(curr, mode, wake_flags, key);
7575 static int io_run_task_work_sig(void)
7577 if (io_run_task_work())
7579 if (!signal_pending(current))
7581 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7582 return -ERESTARTSYS;
7586 /* when returns >0, the caller should retry */
7587 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7588 struct io_wait_queue *iowq,
7593 /* make sure we run task_work before checking for signals */
7594 ret = io_run_task_work_sig();
7595 if (ret || io_should_wake(iowq))
7597 /* let the caller flush overflows, retry */
7598 if (test_bit(0, &ctx->check_cq_overflow))
7601 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
7607 * Wait until events become available, if we don't already have some. The
7608 * application must reap them itself, as they reside on the shared cq ring.
7610 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7611 const sigset_t __user *sig, size_t sigsz,
7612 struct __kernel_timespec __user *uts)
7614 struct io_wait_queue iowq;
7615 struct io_rings *rings = ctx->rings;
7616 ktime_t timeout = KTIME_MAX;
7620 io_cqring_overflow_flush(ctx);
7621 if (io_cqring_events(ctx) >= min_events)
7623 if (!io_run_task_work())
7628 struct timespec64 ts;
7630 if (get_timespec64(&ts, uts))
7632 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
7636 #ifdef CONFIG_COMPAT
7637 if (in_compat_syscall())
7638 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7642 ret = set_user_sigmask(sig, sigsz);
7648 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7649 iowq.wq.private = current;
7650 INIT_LIST_HEAD(&iowq.wq.entry);
7652 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7653 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7655 trace_io_uring_cqring_wait(ctx, min_events);
7657 /* if we can't even flush overflow, don't wait for more */
7658 if (!io_cqring_overflow_flush(ctx)) {
7662 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7663 TASK_INTERRUPTIBLE);
7664 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
7665 finish_wait(&ctx->cq_wait, &iowq.wq);
7669 restore_saved_sigmask_unless(ret == -EINTR);
7671 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7674 static void io_free_page_table(void **table, size_t size)
7676 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7678 for (i = 0; i < nr_tables; i++)
7683 static void **io_alloc_page_table(size_t size)
7685 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7686 size_t init_size = size;
7689 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7693 for (i = 0; i < nr_tables; i++) {
7694 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7696 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7698 io_free_page_table(table, init_size);
7706 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7708 percpu_ref_exit(&ref_node->refs);
7712 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7714 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7715 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7716 unsigned long flags;
7717 bool first_add = false;
7718 unsigned long delay = HZ;
7720 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7723 /* if we are mid-quiesce then do not delay */
7724 if (node->rsrc_data->quiesce)
7727 while (!list_empty(&ctx->rsrc_ref_list)) {
7728 node = list_first_entry(&ctx->rsrc_ref_list,
7729 struct io_rsrc_node, node);
7730 /* recycle ref nodes in order */
7733 list_del(&node->node);
7734 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7736 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7739 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7742 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7744 struct io_rsrc_node *ref_node;
7746 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7750 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7755 INIT_LIST_HEAD(&ref_node->node);
7756 INIT_LIST_HEAD(&ref_node->rsrc_list);
7757 ref_node->done = false;
7761 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7762 struct io_rsrc_data *data_to_kill)
7764 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7765 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7768 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7770 rsrc_node->rsrc_data = data_to_kill;
7771 spin_lock_irq(&ctx->rsrc_ref_lock);
7772 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7773 spin_unlock_irq(&ctx->rsrc_ref_lock);
7775 atomic_inc(&data_to_kill->refs);
7776 percpu_ref_kill(&rsrc_node->refs);
7777 ctx->rsrc_node = NULL;
7780 if (!ctx->rsrc_node) {
7781 ctx->rsrc_node = ctx->rsrc_backup_node;
7782 ctx->rsrc_backup_node = NULL;
7786 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7788 if (ctx->rsrc_backup_node)
7790 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7791 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7794 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7798 /* As we may drop ->uring_lock, other task may have started quiesce */
7802 data->quiesce = true;
7804 ret = io_rsrc_node_switch_start(ctx);
7807 io_rsrc_node_switch(ctx, data);
7809 /* kill initial ref, already quiesced if zero */
7810 if (atomic_dec_and_test(&data->refs))
7812 mutex_unlock(&ctx->uring_lock);
7813 flush_delayed_work(&ctx->rsrc_put_work);
7814 ret = wait_for_completion_interruptible(&data->done);
7816 mutex_lock(&ctx->uring_lock);
7817 if (atomic_read(&data->refs) > 0) {
7819 * it has been revived by another thread while
7822 mutex_unlock(&ctx->uring_lock);
7828 atomic_inc(&data->refs);
7829 /* wait for all works potentially completing data->done */
7830 flush_delayed_work(&ctx->rsrc_put_work);
7831 reinit_completion(&data->done);
7833 ret = io_run_task_work_sig();
7834 mutex_lock(&ctx->uring_lock);
7836 data->quiesce = false;
7841 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7843 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7844 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7846 return &data->tags[table_idx][off];
7849 static void io_rsrc_data_free(struct io_rsrc_data *data)
7851 size_t size = data->nr * sizeof(data->tags[0][0]);
7854 io_free_page_table((void **)data->tags, size);
7858 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7859 u64 __user *utags, unsigned nr,
7860 struct io_rsrc_data **pdata)
7862 struct io_rsrc_data *data;
7866 data = kzalloc(sizeof(*data), GFP_KERNEL);
7869 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7877 data->do_put = do_put;
7880 for (i = 0; i < nr; i++) {
7881 u64 *tag_slot = io_get_tag_slot(data, i);
7883 if (copy_from_user(tag_slot, &utags[i],
7889 atomic_set(&data->refs, 1);
7890 init_completion(&data->done);
7894 io_rsrc_data_free(data);
7898 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7900 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7901 GFP_KERNEL_ACCOUNT);
7902 return !!table->files;
7905 static void io_free_file_tables(struct io_file_table *table)
7907 kvfree(table->files);
7908 table->files = NULL;
7911 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7913 #if defined(CONFIG_UNIX)
7914 if (ctx->ring_sock) {
7915 struct sock *sock = ctx->ring_sock->sk;
7916 struct sk_buff *skb;
7918 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7924 for (i = 0; i < ctx->nr_user_files; i++) {
7927 file = io_file_from_index(ctx, i);
7932 io_free_file_tables(&ctx->file_table);
7933 io_rsrc_data_free(ctx->file_data);
7934 ctx->file_data = NULL;
7935 ctx->nr_user_files = 0;
7938 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7940 unsigned nr = ctx->nr_user_files;
7943 if (!ctx->file_data)
7947 * Quiesce may unlock ->uring_lock, and while it's not held
7948 * prevent new requests using the table.
7950 ctx->nr_user_files = 0;
7951 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7952 ctx->nr_user_files = nr;
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; i++) {
8129 struct file *file = io_file_from_index(ctx, i + offset);
8136 free_uid(fpl->user);
8144 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8145 * causes regular reference counting to break down. We rely on the UNIX
8146 * garbage collection to take care of this problem for us.
8148 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8150 unsigned left, total;
8154 left = ctx->nr_user_files;
8156 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8158 ret = __io_sqe_files_scm(ctx, this_files, total);
8162 total += this_files;
8168 while (total < ctx->nr_user_files) {
8169 struct file *file = io_file_from_index(ctx, total);
8179 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8185 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8187 struct file *file = prsrc->file;
8188 #if defined(CONFIG_UNIX)
8189 struct sock *sock = ctx->ring_sock->sk;
8190 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8191 struct sk_buff *skb;
8194 __skb_queue_head_init(&list);
8197 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8198 * remove this entry and rearrange the file array.
8200 skb = skb_dequeue(head);
8202 struct scm_fp_list *fp;
8204 fp = UNIXCB(skb).fp;
8205 for (i = 0; i < fp->count; i++) {
8208 if (fp->fp[i] != file)
8211 unix_notinflight(fp->user, fp->fp[i]);
8212 left = fp->count - 1 - i;
8214 memmove(&fp->fp[i], &fp->fp[i + 1],
8215 left * sizeof(struct file *));
8222 __skb_queue_tail(&list, skb);
8232 __skb_queue_tail(&list, skb);
8234 skb = skb_dequeue(head);
8237 if (skb_peek(&list)) {
8238 spin_lock_irq(&head->lock);
8239 while ((skb = __skb_dequeue(&list)) != NULL)
8240 __skb_queue_tail(head, skb);
8241 spin_unlock_irq(&head->lock);
8248 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8250 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8251 struct io_ring_ctx *ctx = rsrc_data->ctx;
8252 struct io_rsrc_put *prsrc, *tmp;
8254 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8255 list_del(&prsrc->list);
8258 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8260 io_ring_submit_lock(ctx, lock_ring);
8261 spin_lock(&ctx->completion_lock);
8262 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8264 io_commit_cqring(ctx);
8265 spin_unlock(&ctx->completion_lock);
8266 io_cqring_ev_posted(ctx);
8267 io_ring_submit_unlock(ctx, lock_ring);
8270 rsrc_data->do_put(ctx, prsrc);
8274 io_rsrc_node_destroy(ref_node);
8275 if (atomic_dec_and_test(&rsrc_data->refs))
8276 complete(&rsrc_data->done);
8279 static void io_rsrc_put_work(struct work_struct *work)
8281 struct io_ring_ctx *ctx;
8282 struct llist_node *node;
8284 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8285 node = llist_del_all(&ctx->rsrc_put_llist);
8288 struct io_rsrc_node *ref_node;
8289 struct llist_node *next = node->next;
8291 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8292 __io_rsrc_put_work(ref_node);
8297 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8298 unsigned nr_args, u64 __user *tags)
8300 __s32 __user *fds = (__s32 __user *) arg;
8309 if (nr_args > IORING_MAX_FIXED_FILES)
8311 if (nr_args > rlimit(RLIMIT_NOFILE))
8313 ret = io_rsrc_node_switch_start(ctx);
8316 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8322 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8325 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8326 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8330 /* allow sparse sets */
8333 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8340 if (unlikely(!file))
8344 * Don't allow io_uring instances to be registered. If UNIX
8345 * isn't enabled, then this causes a reference cycle and this
8346 * instance can never get freed. If UNIX is enabled we'll
8347 * handle it just fine, but there's still no point in allowing
8348 * a ring fd as it doesn't support regular read/write anyway.
8350 if (file->f_op == &io_uring_fops) {
8354 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8357 ret = io_sqe_files_scm(ctx);
8359 __io_sqe_files_unregister(ctx);
8363 io_rsrc_node_switch(ctx, NULL);
8366 for (i = 0; i < ctx->nr_user_files; i++) {
8367 file = io_file_from_index(ctx, i);
8371 io_free_file_tables(&ctx->file_table);
8372 ctx->nr_user_files = 0;
8374 io_rsrc_data_free(ctx->file_data);
8375 ctx->file_data = NULL;
8379 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8382 #if defined(CONFIG_UNIX)
8383 struct sock *sock = ctx->ring_sock->sk;
8384 struct sk_buff_head *head = &sock->sk_receive_queue;
8385 struct sk_buff *skb;
8388 * See if we can merge this file into an existing skb SCM_RIGHTS
8389 * file set. If there's no room, fall back to allocating a new skb
8390 * and filling it in.
8392 spin_lock_irq(&head->lock);
8393 skb = skb_peek(head);
8395 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8397 if (fpl->count < SCM_MAX_FD) {
8398 __skb_unlink(skb, head);
8399 spin_unlock_irq(&head->lock);
8400 fpl->fp[fpl->count] = get_file(file);
8401 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8403 spin_lock_irq(&head->lock);
8404 __skb_queue_head(head, skb);
8409 spin_unlock_irq(&head->lock);
8416 return __io_sqe_files_scm(ctx, 1, index);
8422 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8423 struct io_rsrc_node *node, void *rsrc)
8425 u64 *tag_slot = io_get_tag_slot(data, idx);
8426 struct io_rsrc_put *prsrc;
8428 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8432 prsrc->tag = *tag_slot;
8435 list_add(&prsrc->list, &node->rsrc_list);
8439 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8440 unsigned int issue_flags, u32 slot_index)
8442 struct io_ring_ctx *ctx = req->ctx;
8443 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8444 bool needs_switch = false;
8445 struct io_fixed_file *file_slot;
8448 io_ring_submit_lock(ctx, !force_nonblock);
8449 if (file->f_op == &io_uring_fops)
8452 if (!ctx->file_data)
8455 if (slot_index >= ctx->nr_user_files)
8458 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8459 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8461 if (file_slot->file_ptr) {
8462 struct file *old_file;
8464 ret = io_rsrc_node_switch_start(ctx);
8468 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8469 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8470 ctx->rsrc_node, old_file);
8473 file_slot->file_ptr = 0;
8474 needs_switch = true;
8477 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8478 io_fixed_file_set(file_slot, file);
8479 ret = io_sqe_file_register(ctx, file, slot_index);
8481 file_slot->file_ptr = 0;
8488 io_rsrc_node_switch(ctx, ctx->file_data);
8489 io_ring_submit_unlock(ctx, !force_nonblock);
8495 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8497 unsigned int offset = req->close.file_slot - 1;
8498 struct io_ring_ctx *ctx = req->ctx;
8499 struct io_fixed_file *file_slot;
8503 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8505 if (unlikely(!ctx->file_data))
8508 if (offset >= ctx->nr_user_files)
8510 ret = io_rsrc_node_switch_start(ctx);
8514 offset = array_index_nospec(offset, ctx->nr_user_files);
8515 file_slot = io_fixed_file_slot(&ctx->file_table, offset);
8517 if (!file_slot->file_ptr)
8520 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8521 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8525 file_slot->file_ptr = 0;
8526 io_rsrc_node_switch(ctx, ctx->file_data);
8529 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8533 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8534 struct io_uring_rsrc_update2 *up,
8537 u64 __user *tags = u64_to_user_ptr(up->tags);
8538 __s32 __user *fds = u64_to_user_ptr(up->data);
8539 struct io_rsrc_data *data = ctx->file_data;
8540 struct io_fixed_file *file_slot;
8544 bool needs_switch = false;
8546 if (!ctx->file_data)
8548 if (up->offset + nr_args > ctx->nr_user_files)
8551 for (done = 0; done < nr_args; done++) {
8554 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8555 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8559 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8563 if (fd == IORING_REGISTER_FILES_SKIP)
8566 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8567 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8569 if (file_slot->file_ptr) {
8570 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8571 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
8574 file_slot->file_ptr = 0;
8575 needs_switch = true;
8584 * Don't allow io_uring instances to be registered. If
8585 * UNIX isn't enabled, then this causes a reference
8586 * cycle and this instance can never get freed. If UNIX
8587 * is enabled we'll handle it just fine, but there's
8588 * still no point in allowing a ring fd as it doesn't
8589 * support regular read/write anyway.
8591 if (file->f_op == &io_uring_fops) {
8596 *io_get_tag_slot(data, i) = tag;
8597 io_fixed_file_set(file_slot, file);
8598 err = io_sqe_file_register(ctx, file, i);
8600 file_slot->file_ptr = 0;
8608 io_rsrc_node_switch(ctx, data);
8609 return done ? done : err;
8612 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8613 struct task_struct *task)
8615 struct io_wq_hash *hash;
8616 struct io_wq_data data;
8617 unsigned int concurrency;
8619 mutex_lock(&ctx->uring_lock);
8620 hash = ctx->hash_map;
8622 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8624 mutex_unlock(&ctx->uring_lock);
8625 return ERR_PTR(-ENOMEM);
8627 refcount_set(&hash->refs, 1);
8628 init_waitqueue_head(&hash->wait);
8629 ctx->hash_map = hash;
8631 mutex_unlock(&ctx->uring_lock);
8635 data.free_work = io_wq_free_work;
8636 data.do_work = io_wq_submit_work;
8638 /* Do QD, or 4 * CPUS, whatever is smallest */
8639 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8641 return io_wq_create(concurrency, &data);
8644 static int io_uring_alloc_task_context(struct task_struct *task,
8645 struct io_ring_ctx *ctx)
8647 struct io_uring_task *tctx;
8650 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8651 if (unlikely(!tctx))
8654 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8655 if (unlikely(ret)) {
8660 tctx->io_wq = io_init_wq_offload(ctx, task);
8661 if (IS_ERR(tctx->io_wq)) {
8662 ret = PTR_ERR(tctx->io_wq);
8663 percpu_counter_destroy(&tctx->inflight);
8669 init_waitqueue_head(&tctx->wait);
8670 atomic_set(&tctx->in_idle, 0);
8671 atomic_set(&tctx->inflight_tracked, 0);
8672 task->io_uring = tctx;
8673 spin_lock_init(&tctx->task_lock);
8674 INIT_WQ_LIST(&tctx->task_list);
8675 init_task_work(&tctx->task_work, tctx_task_work);
8679 void __io_uring_free(struct task_struct *tsk)
8681 struct io_uring_task *tctx = tsk->io_uring;
8683 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8684 WARN_ON_ONCE(tctx->io_wq);
8685 WARN_ON_ONCE(tctx->cached_refs);
8687 percpu_counter_destroy(&tctx->inflight);
8689 tsk->io_uring = NULL;
8692 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8693 struct io_uring_params *p)
8697 /* Retain compatibility with failing for an invalid attach attempt */
8698 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8699 IORING_SETUP_ATTACH_WQ) {
8702 f = fdget(p->wq_fd);
8705 if (f.file->f_op != &io_uring_fops) {
8711 if (ctx->flags & IORING_SETUP_SQPOLL) {
8712 struct task_struct *tsk;
8713 struct io_sq_data *sqd;
8716 sqd = io_get_sq_data(p, &attached);
8722 ctx->sq_creds = get_current_cred();
8724 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8725 if (!ctx->sq_thread_idle)
8726 ctx->sq_thread_idle = HZ;
8728 io_sq_thread_park(sqd);
8729 list_add(&ctx->sqd_list, &sqd->ctx_list);
8730 io_sqd_update_thread_idle(sqd);
8731 /* don't attach to a dying SQPOLL thread, would be racy */
8732 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8733 io_sq_thread_unpark(sqd);
8740 if (p->flags & IORING_SETUP_SQ_AFF) {
8741 int cpu = p->sq_thread_cpu;
8744 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8751 sqd->task_pid = current->pid;
8752 sqd->task_tgid = current->tgid;
8753 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8760 ret = io_uring_alloc_task_context(tsk, ctx);
8761 wake_up_new_task(tsk);
8764 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8765 /* Can't have SQ_AFF without SQPOLL */
8772 complete(&ctx->sq_data->exited);
8774 io_sq_thread_finish(ctx);
8778 static inline void __io_unaccount_mem(struct user_struct *user,
8779 unsigned long nr_pages)
8781 atomic_long_sub(nr_pages, &user->locked_vm);
8784 static inline int __io_account_mem(struct user_struct *user,
8785 unsigned long nr_pages)
8787 unsigned long page_limit, cur_pages, new_pages;
8789 /* Don't allow more pages than we can safely lock */
8790 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8793 cur_pages = atomic_long_read(&user->locked_vm);
8794 new_pages = cur_pages + nr_pages;
8795 if (new_pages > page_limit)
8797 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8798 new_pages) != cur_pages);
8803 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8806 __io_unaccount_mem(ctx->user, nr_pages);
8808 if (ctx->mm_account)
8809 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8812 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8817 ret = __io_account_mem(ctx->user, nr_pages);
8822 if (ctx->mm_account)
8823 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8828 static void io_mem_free(void *ptr)
8835 page = virt_to_head_page(ptr);
8836 if (put_page_testzero(page))
8837 free_compound_page(page);
8840 static void *io_mem_alloc(size_t size)
8842 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
8844 return (void *) __get_free_pages(gfp, get_order(size));
8847 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8850 struct io_rings *rings;
8851 size_t off, sq_array_size;
8853 off = struct_size(rings, cqes, cq_entries);
8854 if (off == SIZE_MAX)
8858 off = ALIGN(off, SMP_CACHE_BYTES);
8866 sq_array_size = array_size(sizeof(u32), sq_entries);
8867 if (sq_array_size == SIZE_MAX)
8870 if (check_add_overflow(off, sq_array_size, &off))
8876 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8878 struct io_mapped_ubuf *imu = *slot;
8881 if (imu != ctx->dummy_ubuf) {
8882 for (i = 0; i < imu->nr_bvecs; i++)
8883 unpin_user_page(imu->bvec[i].bv_page);
8884 if (imu->acct_pages)
8885 io_unaccount_mem(ctx, imu->acct_pages);
8891 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8893 io_buffer_unmap(ctx, &prsrc->buf);
8897 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8901 for (i = 0; i < ctx->nr_user_bufs; i++)
8902 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8903 kfree(ctx->user_bufs);
8904 io_rsrc_data_free(ctx->buf_data);
8905 ctx->user_bufs = NULL;
8906 ctx->buf_data = NULL;
8907 ctx->nr_user_bufs = 0;
8910 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8912 unsigned nr = ctx->nr_user_bufs;
8919 * Quiesce may unlock ->uring_lock, and while it's not held
8920 * prevent new requests using the table.
8922 ctx->nr_user_bufs = 0;
8923 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8924 ctx->nr_user_bufs = nr;
8926 __io_sqe_buffers_unregister(ctx);
8930 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8931 void __user *arg, unsigned index)
8933 struct iovec __user *src;
8935 #ifdef CONFIG_COMPAT
8937 struct compat_iovec __user *ciovs;
8938 struct compat_iovec ciov;
8940 ciovs = (struct compat_iovec __user *) arg;
8941 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8944 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8945 dst->iov_len = ciov.iov_len;
8949 src = (struct iovec __user *) arg;
8950 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8956 * Not super efficient, but this is just a registration time. And we do cache
8957 * the last compound head, so generally we'll only do a full search if we don't
8960 * We check if the given compound head page has already been accounted, to
8961 * avoid double accounting it. This allows us to account the full size of the
8962 * page, not just the constituent pages of a huge page.
8964 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8965 int nr_pages, struct page *hpage)
8969 /* check current page array */
8970 for (i = 0; i < nr_pages; i++) {
8971 if (!PageCompound(pages[i]))
8973 if (compound_head(pages[i]) == hpage)
8977 /* check previously registered pages */
8978 for (i = 0; i < ctx->nr_user_bufs; i++) {
8979 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8981 for (j = 0; j < imu->nr_bvecs; j++) {
8982 if (!PageCompound(imu->bvec[j].bv_page))
8984 if (compound_head(imu->bvec[j].bv_page) == hpage)
8992 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8993 int nr_pages, struct io_mapped_ubuf *imu,
8994 struct page **last_hpage)
8998 imu->acct_pages = 0;
8999 for (i = 0; i < nr_pages; i++) {
9000 if (!PageCompound(pages[i])) {
9005 hpage = compound_head(pages[i]);
9006 if (hpage == *last_hpage)
9008 *last_hpage = hpage;
9009 if (headpage_already_acct(ctx, pages, i, hpage))
9011 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
9015 if (!imu->acct_pages)
9018 ret = io_account_mem(ctx, imu->acct_pages);
9020 imu->acct_pages = 0;
9024 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
9025 struct io_mapped_ubuf **pimu,
9026 struct page **last_hpage)
9028 struct io_mapped_ubuf *imu = NULL;
9029 struct vm_area_struct **vmas = NULL;
9030 struct page **pages = NULL;
9031 unsigned long off, start, end, ubuf;
9033 int ret, pret, nr_pages, i;
9035 if (!iov->iov_base) {
9036 *pimu = ctx->dummy_ubuf;
9040 ubuf = (unsigned long) iov->iov_base;
9041 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9042 start = ubuf >> PAGE_SHIFT;
9043 nr_pages = end - start;
9048 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9052 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9057 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9062 mmap_read_lock(current->mm);
9063 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9065 if (pret == nr_pages) {
9066 /* don't support file backed memory */
9067 for (i = 0; i < nr_pages; i++) {
9068 struct vm_area_struct *vma = vmas[i];
9070 if (vma_is_shmem(vma))
9073 !is_file_hugepages(vma->vm_file)) {
9079 ret = pret < 0 ? pret : -EFAULT;
9081 mmap_read_unlock(current->mm);
9084 * if we did partial map, or found file backed vmas,
9085 * release any pages we did get
9088 unpin_user_pages(pages, pret);
9092 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9094 unpin_user_pages(pages, pret);
9098 off = ubuf & ~PAGE_MASK;
9099 size = iov->iov_len;
9100 for (i = 0; i < nr_pages; i++) {
9103 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9104 imu->bvec[i].bv_page = pages[i];
9105 imu->bvec[i].bv_len = vec_len;
9106 imu->bvec[i].bv_offset = off;
9110 /* store original address for later verification */
9112 imu->ubuf_end = ubuf + iov->iov_len;
9113 imu->nr_bvecs = nr_pages;
9124 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9126 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9127 return ctx->user_bufs ? 0 : -ENOMEM;
9130 static int io_buffer_validate(struct iovec *iov)
9132 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9135 * Don't impose further limits on the size and buffer
9136 * constraints here, we'll -EINVAL later when IO is
9137 * submitted if they are wrong.
9140 return iov->iov_len ? -EFAULT : 0;
9144 /* arbitrary limit, but we need something */
9145 if (iov->iov_len > SZ_1G)
9148 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9154 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9155 unsigned int nr_args, u64 __user *tags)
9157 struct page *last_hpage = NULL;
9158 struct io_rsrc_data *data;
9164 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9166 ret = io_rsrc_node_switch_start(ctx);
9169 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9172 ret = io_buffers_map_alloc(ctx, nr_args);
9174 io_rsrc_data_free(data);
9178 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9179 ret = io_copy_iov(ctx, &iov, arg, i);
9182 ret = io_buffer_validate(&iov);
9185 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9190 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9196 WARN_ON_ONCE(ctx->buf_data);
9198 ctx->buf_data = data;
9200 __io_sqe_buffers_unregister(ctx);
9202 io_rsrc_node_switch(ctx, NULL);
9206 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9207 struct io_uring_rsrc_update2 *up,
9208 unsigned int nr_args)
9210 u64 __user *tags = u64_to_user_ptr(up->tags);
9211 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9212 struct page *last_hpage = NULL;
9213 bool needs_switch = false;
9219 if (up->offset + nr_args > ctx->nr_user_bufs)
9222 for (done = 0; done < nr_args; done++) {
9223 struct io_mapped_ubuf *imu;
9224 int offset = up->offset + done;
9227 err = io_copy_iov(ctx, &iov, iovs, done);
9230 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9234 err = io_buffer_validate(&iov);
9237 if (!iov.iov_base && tag) {
9241 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9245 i = array_index_nospec(offset, ctx->nr_user_bufs);
9246 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9247 err = io_queue_rsrc_removal(ctx->buf_data, i,
9248 ctx->rsrc_node, ctx->user_bufs[i]);
9249 if (unlikely(err)) {
9250 io_buffer_unmap(ctx, &imu);
9253 ctx->user_bufs[i] = NULL;
9254 needs_switch = true;
9257 ctx->user_bufs[i] = imu;
9258 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9262 io_rsrc_node_switch(ctx, ctx->buf_data);
9263 return done ? done : err;
9266 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9268 __s32 __user *fds = arg;
9274 if (copy_from_user(&fd, fds, sizeof(*fds)))
9277 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9278 if (IS_ERR(ctx->cq_ev_fd)) {
9279 int ret = PTR_ERR(ctx->cq_ev_fd);
9281 ctx->cq_ev_fd = NULL;
9288 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9290 if (ctx->cq_ev_fd) {
9291 eventfd_ctx_put(ctx->cq_ev_fd);
9292 ctx->cq_ev_fd = NULL;
9299 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9301 struct io_buffer *buf;
9302 unsigned long index;
9304 xa_for_each(&ctx->io_buffers, index, buf)
9305 __io_remove_buffers(ctx, buf, index, -1U);
9308 static void io_req_cache_free(struct list_head *list)
9310 struct io_kiocb *req, *nxt;
9312 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9313 list_del(&req->inflight_entry);
9314 kmem_cache_free(req_cachep, req);
9318 static void io_req_caches_free(struct io_ring_ctx *ctx)
9320 struct io_submit_state *state = &ctx->submit_state;
9322 mutex_lock(&ctx->uring_lock);
9324 if (state->free_reqs) {
9325 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9326 state->free_reqs = 0;
9329 io_flush_cached_locked_reqs(ctx, state);
9330 io_req_cache_free(&state->free_list);
9331 mutex_unlock(&ctx->uring_lock);
9334 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9336 if (data && !atomic_dec_and_test(&data->refs))
9337 wait_for_completion(&data->done);
9340 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9342 io_sq_thread_finish(ctx);
9344 if (ctx->mm_account) {
9345 mmdrop(ctx->mm_account);
9346 ctx->mm_account = NULL;
9349 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9350 io_wait_rsrc_data(ctx->buf_data);
9351 io_wait_rsrc_data(ctx->file_data);
9353 mutex_lock(&ctx->uring_lock);
9355 __io_sqe_buffers_unregister(ctx);
9357 __io_sqe_files_unregister(ctx);
9359 __io_cqring_overflow_flush(ctx, true);
9360 mutex_unlock(&ctx->uring_lock);
9361 io_eventfd_unregister(ctx);
9362 io_destroy_buffers(ctx);
9364 put_cred(ctx->sq_creds);
9366 /* there are no registered resources left, nobody uses it */
9368 io_rsrc_node_destroy(ctx->rsrc_node);
9369 if (ctx->rsrc_backup_node)
9370 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9371 flush_delayed_work(&ctx->rsrc_put_work);
9373 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9374 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9376 #if defined(CONFIG_UNIX)
9377 if (ctx->ring_sock) {
9378 ctx->ring_sock->file = NULL; /* so that iput() is called */
9379 sock_release(ctx->ring_sock);
9382 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9384 io_mem_free(ctx->rings);
9385 io_mem_free(ctx->sq_sqes);
9387 percpu_ref_exit(&ctx->refs);
9388 free_uid(ctx->user);
9389 io_req_caches_free(ctx);
9391 io_wq_put_hash(ctx->hash_map);
9392 kfree(ctx->cancel_hash);
9393 kfree(ctx->dummy_ubuf);
9397 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9399 struct io_ring_ctx *ctx = file->private_data;
9402 poll_wait(file, &ctx->poll_wait, wait);
9404 * synchronizes with barrier from wq_has_sleeper call in
9408 if (!io_sqring_full(ctx))
9409 mask |= EPOLLOUT | EPOLLWRNORM;
9412 * Don't flush cqring overflow list here, just do a simple check.
9413 * Otherwise there could possible be ABBA deadlock:
9416 * lock(&ctx->uring_lock);
9418 * lock(&ctx->uring_lock);
9421 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9422 * pushs them to do the flush.
9424 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9425 mask |= EPOLLIN | EPOLLRDNORM;
9430 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9432 const struct cred *creds;
9434 creds = xa_erase(&ctx->personalities, id);
9443 struct io_tctx_exit {
9444 struct callback_head task_work;
9445 struct completion completion;
9446 struct io_ring_ctx *ctx;
9449 static void io_tctx_exit_cb(struct callback_head *cb)
9451 struct io_uring_task *tctx = current->io_uring;
9452 struct io_tctx_exit *work;
9454 work = container_of(cb, struct io_tctx_exit, task_work);
9456 * When @in_idle, we're in cancellation and it's racy to remove the
9457 * node. It'll be removed by the end of cancellation, just ignore it.
9459 if (!atomic_read(&tctx->in_idle))
9460 io_uring_del_tctx_node((unsigned long)work->ctx);
9461 complete(&work->completion);
9464 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9466 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9468 return req->ctx == data;
9471 static void io_ring_exit_work(struct work_struct *work)
9473 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9474 unsigned long timeout = jiffies + HZ * 60 * 5;
9475 unsigned long interval = HZ / 20;
9476 struct io_tctx_exit exit;
9477 struct io_tctx_node *node;
9481 * If we're doing polled IO and end up having requests being
9482 * submitted async (out-of-line), then completions can come in while
9483 * we're waiting for refs to drop. We need to reap these manually,
9484 * as nobody else will be looking for them.
9487 io_uring_try_cancel_requests(ctx, NULL, true);
9489 struct io_sq_data *sqd = ctx->sq_data;
9490 struct task_struct *tsk;
9492 io_sq_thread_park(sqd);
9494 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9495 io_wq_cancel_cb(tsk->io_uring->io_wq,
9496 io_cancel_ctx_cb, ctx, true);
9497 io_sq_thread_unpark(sqd);
9500 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9501 /* there is little hope left, don't run it too often */
9504 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9506 init_completion(&exit.completion);
9507 init_task_work(&exit.task_work, io_tctx_exit_cb);
9510 * Some may use context even when all refs and requests have been put,
9511 * and they are free to do so while still holding uring_lock or
9512 * completion_lock, see io_req_task_submit(). Apart from other work,
9513 * this lock/unlock section also waits them to finish.
9515 mutex_lock(&ctx->uring_lock);
9516 while (!list_empty(&ctx->tctx_list)) {
9517 WARN_ON_ONCE(time_after(jiffies, timeout));
9519 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9521 /* don't spin on a single task if cancellation failed */
9522 list_rotate_left(&ctx->tctx_list);
9523 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9524 if (WARN_ON_ONCE(ret))
9526 wake_up_process(node->task);
9528 mutex_unlock(&ctx->uring_lock);
9529 wait_for_completion(&exit.completion);
9530 mutex_lock(&ctx->uring_lock);
9532 mutex_unlock(&ctx->uring_lock);
9533 spin_lock(&ctx->completion_lock);
9534 spin_unlock(&ctx->completion_lock);
9536 io_ring_ctx_free(ctx);
9539 /* Returns true if we found and killed one or more timeouts */
9540 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9543 struct io_kiocb *req, *tmp;
9546 spin_lock(&ctx->completion_lock);
9547 spin_lock_irq(&ctx->timeout_lock);
9548 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9549 if (io_match_task(req, tsk, cancel_all)) {
9550 io_kill_timeout(req, -ECANCELED);
9554 spin_unlock_irq(&ctx->timeout_lock);
9556 io_commit_cqring(ctx);
9557 spin_unlock(&ctx->completion_lock);
9559 io_cqring_ev_posted(ctx);
9560 return canceled != 0;
9563 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9565 unsigned long index;
9566 struct creds *creds;
9568 mutex_lock(&ctx->uring_lock);
9569 percpu_ref_kill(&ctx->refs);
9571 __io_cqring_overflow_flush(ctx, true);
9572 xa_for_each(&ctx->personalities, index, creds)
9573 io_unregister_personality(ctx, index);
9574 mutex_unlock(&ctx->uring_lock);
9576 io_kill_timeouts(ctx, NULL, true);
9577 io_poll_remove_all(ctx, NULL, true);
9579 /* if we failed setting up the ctx, we might not have any rings */
9580 io_iopoll_try_reap_events(ctx);
9582 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9584 * Use system_unbound_wq to avoid spawning tons of event kworkers
9585 * if we're exiting a ton of rings at the same time. It just adds
9586 * noise and overhead, there's no discernable change in runtime
9587 * over using system_wq.
9589 queue_work(system_unbound_wq, &ctx->exit_work);
9592 static int io_uring_release(struct inode *inode, struct file *file)
9594 struct io_ring_ctx *ctx = file->private_data;
9596 file->private_data = NULL;
9597 io_ring_ctx_wait_and_kill(ctx);
9601 struct io_task_cancel {
9602 struct task_struct *task;
9606 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9608 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9609 struct io_task_cancel *cancel = data;
9611 return io_match_task_safe(req, cancel->task, cancel->all);
9614 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9615 struct task_struct *task, bool cancel_all)
9617 struct io_defer_entry *de;
9620 spin_lock(&ctx->completion_lock);
9621 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9622 if (io_match_task_safe(de->req, task, cancel_all)) {
9623 list_cut_position(&list, &ctx->defer_list, &de->list);
9627 spin_unlock(&ctx->completion_lock);
9628 if (list_empty(&list))
9631 while (!list_empty(&list)) {
9632 de = list_first_entry(&list, struct io_defer_entry, list);
9633 list_del_init(&de->list);
9634 io_req_complete_failed(de->req, -ECANCELED);
9640 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9642 struct io_tctx_node *node;
9643 enum io_wq_cancel cret;
9646 mutex_lock(&ctx->uring_lock);
9647 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9648 struct io_uring_task *tctx = node->task->io_uring;
9651 * io_wq will stay alive while we hold uring_lock, because it's
9652 * killed after ctx nodes, which requires to take the lock.
9654 if (!tctx || !tctx->io_wq)
9656 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9657 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9659 mutex_unlock(&ctx->uring_lock);
9664 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9665 struct task_struct *task,
9668 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9669 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9672 enum io_wq_cancel cret;
9676 ret |= io_uring_try_cancel_iowq(ctx);
9677 } else if (tctx && tctx->io_wq) {
9679 * Cancels requests of all rings, not only @ctx, but
9680 * it's fine as the task is in exit/exec.
9682 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9684 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9687 /* SQPOLL thread does its own polling */
9688 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9689 (ctx->sq_data && ctx->sq_data->thread == current)) {
9690 while (!list_empty_careful(&ctx->iopoll_list)) {
9691 io_iopoll_try_reap_events(ctx);
9696 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9697 ret |= io_poll_remove_all(ctx, task, cancel_all);
9698 ret |= io_kill_timeouts(ctx, task, cancel_all);
9700 ret |= io_run_task_work();
9707 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9709 struct io_uring_task *tctx = current->io_uring;
9710 struct io_tctx_node *node;
9713 if (unlikely(!tctx)) {
9714 ret = io_uring_alloc_task_context(current, ctx);
9718 tctx = current->io_uring;
9719 if (ctx->iowq_limits_set) {
9720 unsigned int limits[2] = { ctx->iowq_limits[0],
9721 ctx->iowq_limits[1], };
9723 ret = io_wq_max_workers(tctx->io_wq, limits);
9728 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9729 node = kmalloc(sizeof(*node), GFP_KERNEL);
9733 node->task = current;
9735 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9742 mutex_lock(&ctx->uring_lock);
9743 list_add(&node->ctx_node, &ctx->tctx_list);
9744 mutex_unlock(&ctx->uring_lock);
9751 * Note that this task has used io_uring. We use it for cancelation purposes.
9753 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9755 struct io_uring_task *tctx = current->io_uring;
9757 if (likely(tctx && tctx->last == ctx))
9759 return __io_uring_add_tctx_node(ctx);
9763 * Remove this io_uring_file -> task mapping.
9765 static void io_uring_del_tctx_node(unsigned long index)
9767 struct io_uring_task *tctx = current->io_uring;
9768 struct io_tctx_node *node;
9772 node = xa_erase(&tctx->xa, index);
9776 WARN_ON_ONCE(current != node->task);
9777 WARN_ON_ONCE(list_empty(&node->ctx_node));
9779 mutex_lock(&node->ctx->uring_lock);
9780 list_del(&node->ctx_node);
9781 mutex_unlock(&node->ctx->uring_lock);
9783 if (tctx->last == node->ctx)
9788 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9790 struct io_wq *wq = tctx->io_wq;
9791 struct io_tctx_node *node;
9792 unsigned long index;
9794 xa_for_each(&tctx->xa, index, node) {
9795 io_uring_del_tctx_node(index);
9800 * Must be after io_uring_del_task_file() (removes nodes under
9801 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9803 io_wq_put_and_exit(wq);
9808 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9811 return atomic_read(&tctx->inflight_tracked);
9812 return percpu_counter_sum(&tctx->inflight);
9816 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9817 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9819 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9821 struct io_uring_task *tctx = current->io_uring;
9822 struct io_ring_ctx *ctx;
9826 WARN_ON_ONCE(sqd && sqd->thread != current);
9828 if (!current->io_uring)
9831 io_wq_exit_start(tctx->io_wq);
9833 atomic_inc(&tctx->in_idle);
9835 io_uring_drop_tctx_refs(current);
9836 /* read completions before cancelations */
9837 inflight = tctx_inflight(tctx, !cancel_all);
9842 struct io_tctx_node *node;
9843 unsigned long index;
9845 xa_for_each(&tctx->xa, index, node) {
9846 /* sqpoll task will cancel all its requests */
9847 if (node->ctx->sq_data)
9849 io_uring_try_cancel_requests(node->ctx, current,
9853 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9854 io_uring_try_cancel_requests(ctx, current,
9858 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9860 io_uring_drop_tctx_refs(current);
9863 * If we've seen completions, retry without waiting. This
9864 * avoids a race where a completion comes in before we did
9865 * prepare_to_wait().
9867 if (inflight == tctx_inflight(tctx, !cancel_all))
9869 finish_wait(&tctx->wait, &wait);
9872 io_uring_clean_tctx(tctx);
9875 * We shouldn't run task_works after cancel, so just leave
9876 * ->in_idle set for normal exit.
9878 atomic_dec(&tctx->in_idle);
9879 /* for exec all current's requests should be gone, kill tctx */
9880 __io_uring_free(current);
9884 void __io_uring_cancel(bool cancel_all)
9886 io_uring_cancel_generic(cancel_all, NULL);
9889 static void *io_uring_validate_mmap_request(struct file *file,
9890 loff_t pgoff, size_t sz)
9892 struct io_ring_ctx *ctx = file->private_data;
9893 loff_t offset = pgoff << PAGE_SHIFT;
9898 case IORING_OFF_SQ_RING:
9899 case IORING_OFF_CQ_RING:
9902 case IORING_OFF_SQES:
9906 return ERR_PTR(-EINVAL);
9909 page = virt_to_head_page(ptr);
9910 if (sz > page_size(page))
9911 return ERR_PTR(-EINVAL);
9918 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9920 size_t sz = vma->vm_end - vma->vm_start;
9924 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9926 return PTR_ERR(ptr);
9928 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9929 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9932 #else /* !CONFIG_MMU */
9934 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9936 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9939 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9941 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9944 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9945 unsigned long addr, unsigned long len,
9946 unsigned long pgoff, unsigned long flags)
9950 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9952 return PTR_ERR(ptr);
9954 return (unsigned long) ptr;
9957 #endif /* !CONFIG_MMU */
9959 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9964 if (!io_sqring_full(ctx))
9966 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9968 if (!io_sqring_full(ctx))
9971 } while (!signal_pending(current));
9973 finish_wait(&ctx->sqo_sq_wait, &wait);
9977 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9978 struct __kernel_timespec __user **ts,
9979 const sigset_t __user **sig)
9981 struct io_uring_getevents_arg arg;
9984 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9985 * is just a pointer to the sigset_t.
9987 if (!(flags & IORING_ENTER_EXT_ARG)) {
9988 *sig = (const sigset_t __user *) argp;
9994 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9995 * timespec and sigset_t pointers if good.
9997 if (*argsz != sizeof(arg))
9999 if (copy_from_user(&arg, argp, sizeof(arg)))
10003 *sig = u64_to_user_ptr(arg.sigmask);
10004 *argsz = arg.sigmask_sz;
10005 *ts = u64_to_user_ptr(arg.ts);
10009 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
10010 u32, min_complete, u32, flags, const void __user *, argp,
10013 struct io_ring_ctx *ctx;
10018 io_run_task_work();
10020 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
10021 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
10025 if (unlikely(!f.file))
10029 if (unlikely(f.file->f_op != &io_uring_fops))
10033 ctx = f.file->private_data;
10034 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
10038 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10042 * For SQ polling, the thread will do all submissions and completions.
10043 * Just return the requested submit count, and wake the thread if
10044 * we were asked to.
10047 if (ctx->flags & IORING_SETUP_SQPOLL) {
10048 io_cqring_overflow_flush(ctx);
10050 if (unlikely(ctx->sq_data->thread == NULL)) {
10054 if (flags & IORING_ENTER_SQ_WAKEUP)
10055 wake_up(&ctx->sq_data->wait);
10056 if (flags & IORING_ENTER_SQ_WAIT) {
10057 ret = io_sqpoll_wait_sq(ctx);
10061 submitted = to_submit;
10062 } else if (to_submit) {
10063 ret = io_uring_add_tctx_node(ctx);
10066 mutex_lock(&ctx->uring_lock);
10067 submitted = io_submit_sqes(ctx, to_submit);
10068 mutex_unlock(&ctx->uring_lock);
10070 if (submitted != to_submit)
10073 if (flags & IORING_ENTER_GETEVENTS) {
10074 const sigset_t __user *sig;
10075 struct __kernel_timespec __user *ts;
10077 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10081 min_complete = min(min_complete, ctx->cq_entries);
10084 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10085 * space applications don't need to do io completion events
10086 * polling again, they can rely on io_sq_thread to do polling
10087 * work, which can reduce cpu usage and uring_lock contention.
10089 if (ctx->flags & IORING_SETUP_IOPOLL &&
10090 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10091 ret = io_iopoll_check(ctx, min_complete);
10093 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10098 percpu_ref_put(&ctx->refs);
10101 return submitted ? submitted : ret;
10104 #ifdef CONFIG_PROC_FS
10105 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10106 const struct cred *cred)
10108 struct user_namespace *uns = seq_user_ns(m);
10109 struct group_info *gi;
10114 seq_printf(m, "%5d\n", id);
10115 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10116 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10117 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10118 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10119 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10120 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10121 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10122 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10123 seq_puts(m, "\n\tGroups:\t");
10124 gi = cred->group_info;
10125 for (g = 0; g < gi->ngroups; g++) {
10126 seq_put_decimal_ull(m, g ? " " : "",
10127 from_kgid_munged(uns, gi->gid[g]));
10129 seq_puts(m, "\n\tCapEff:\t");
10130 cap = cred->cap_effective;
10131 CAP_FOR_EACH_U32(__capi)
10132 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10137 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10139 struct io_sq_data *sq = NULL;
10144 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10145 * since fdinfo case grabs it in the opposite direction of normal use
10146 * cases. If we fail to get the lock, we just don't iterate any
10147 * structures that could be going away outside the io_uring mutex.
10149 has_lock = mutex_trylock(&ctx->uring_lock);
10151 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10157 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10158 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10159 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10160 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10161 struct file *f = io_file_from_index(ctx, i);
10164 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10166 seq_printf(m, "%5u: <none>\n", i);
10168 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10169 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10170 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10171 unsigned int len = buf->ubuf_end - buf->ubuf;
10173 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10175 if (has_lock && !xa_empty(&ctx->personalities)) {
10176 unsigned long index;
10177 const struct cred *cred;
10179 seq_printf(m, "Personalities:\n");
10180 xa_for_each(&ctx->personalities, index, cred)
10181 io_uring_show_cred(m, index, cred);
10183 seq_printf(m, "PollList:\n");
10184 spin_lock(&ctx->completion_lock);
10185 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10186 struct hlist_head *list = &ctx->cancel_hash[i];
10187 struct io_kiocb *req;
10189 hlist_for_each_entry(req, list, hash_node)
10190 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10191 req->task->task_works != NULL);
10193 spin_unlock(&ctx->completion_lock);
10195 mutex_unlock(&ctx->uring_lock);
10198 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10200 struct io_ring_ctx *ctx = f->private_data;
10202 if (percpu_ref_tryget(&ctx->refs)) {
10203 __io_uring_show_fdinfo(ctx, m);
10204 percpu_ref_put(&ctx->refs);
10209 static const struct file_operations io_uring_fops = {
10210 .release = io_uring_release,
10211 .mmap = io_uring_mmap,
10213 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10214 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10216 .poll = io_uring_poll,
10217 #ifdef CONFIG_PROC_FS
10218 .show_fdinfo = io_uring_show_fdinfo,
10222 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10223 struct io_uring_params *p)
10225 struct io_rings *rings;
10226 size_t size, sq_array_offset;
10228 /* make sure these are sane, as we already accounted them */
10229 ctx->sq_entries = p->sq_entries;
10230 ctx->cq_entries = p->cq_entries;
10232 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10233 if (size == SIZE_MAX)
10236 rings = io_mem_alloc(size);
10240 ctx->rings = rings;
10241 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10242 rings->sq_ring_mask = p->sq_entries - 1;
10243 rings->cq_ring_mask = p->cq_entries - 1;
10244 rings->sq_ring_entries = p->sq_entries;
10245 rings->cq_ring_entries = p->cq_entries;
10247 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10248 if (size == SIZE_MAX) {
10249 io_mem_free(ctx->rings);
10254 ctx->sq_sqes = io_mem_alloc(size);
10255 if (!ctx->sq_sqes) {
10256 io_mem_free(ctx->rings);
10264 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10268 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10272 ret = io_uring_add_tctx_node(ctx);
10277 fd_install(fd, file);
10282 * Allocate an anonymous fd, this is what constitutes the application
10283 * visible backing of an io_uring instance. The application mmaps this
10284 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10285 * we have to tie this fd to a socket for file garbage collection purposes.
10287 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10290 #if defined(CONFIG_UNIX)
10293 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10296 return ERR_PTR(ret);
10299 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10300 O_RDWR | O_CLOEXEC);
10301 #if defined(CONFIG_UNIX)
10302 if (IS_ERR(file)) {
10303 sock_release(ctx->ring_sock);
10304 ctx->ring_sock = NULL;
10306 ctx->ring_sock->file = file;
10312 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10313 struct io_uring_params __user *params)
10315 struct io_ring_ctx *ctx;
10321 if (entries > IORING_MAX_ENTRIES) {
10322 if (!(p->flags & IORING_SETUP_CLAMP))
10324 entries = IORING_MAX_ENTRIES;
10328 * Use twice as many entries for the CQ ring. It's possible for the
10329 * application to drive a higher depth than the size of the SQ ring,
10330 * since the sqes are only used at submission time. This allows for
10331 * some flexibility in overcommitting a bit. If the application has
10332 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10333 * of CQ ring entries manually.
10335 p->sq_entries = roundup_pow_of_two(entries);
10336 if (p->flags & IORING_SETUP_CQSIZE) {
10338 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10339 * to a power-of-two, if it isn't already. We do NOT impose
10340 * any cq vs sq ring sizing.
10342 if (!p->cq_entries)
10344 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10345 if (!(p->flags & IORING_SETUP_CLAMP))
10347 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10349 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10350 if (p->cq_entries < p->sq_entries)
10353 p->cq_entries = 2 * p->sq_entries;
10356 ctx = io_ring_ctx_alloc(p);
10359 ctx->compat = in_compat_syscall();
10360 if (!capable(CAP_IPC_LOCK))
10361 ctx->user = get_uid(current_user());
10364 * This is just grabbed for accounting purposes. When a process exits,
10365 * the mm is exited and dropped before the files, hence we need to hang
10366 * on to this mm purely for the purposes of being able to unaccount
10367 * memory (locked/pinned vm). It's not used for anything else.
10369 mmgrab(current->mm);
10370 ctx->mm_account = current->mm;
10372 ret = io_allocate_scq_urings(ctx, p);
10376 ret = io_sq_offload_create(ctx, p);
10379 /* always set a rsrc node */
10380 ret = io_rsrc_node_switch_start(ctx);
10383 io_rsrc_node_switch(ctx, NULL);
10385 memset(&p->sq_off, 0, sizeof(p->sq_off));
10386 p->sq_off.head = offsetof(struct io_rings, sq.head);
10387 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10388 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10389 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10390 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10391 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10392 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10394 memset(&p->cq_off, 0, sizeof(p->cq_off));
10395 p->cq_off.head = offsetof(struct io_rings, cq.head);
10396 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10397 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10398 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10399 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10400 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10401 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10403 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10404 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10405 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10406 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10407 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10408 IORING_FEAT_RSRC_TAGS;
10410 if (copy_to_user(params, p, sizeof(*p))) {
10415 file = io_uring_get_file(ctx);
10416 if (IS_ERR(file)) {
10417 ret = PTR_ERR(file);
10422 * Install ring fd as the very last thing, so we don't risk someone
10423 * having closed it before we finish setup
10425 ret = io_uring_install_fd(ctx, file);
10427 /* fput will clean it up */
10432 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10435 io_ring_ctx_wait_and_kill(ctx);
10440 * Sets up an aio uring context, and returns the fd. Applications asks for a
10441 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10442 * params structure passed in.
10444 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10446 struct io_uring_params p;
10449 if (copy_from_user(&p, params, sizeof(p)))
10451 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10456 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10457 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10458 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10459 IORING_SETUP_R_DISABLED))
10462 return io_uring_create(entries, &p, params);
10465 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10466 struct io_uring_params __user *, params)
10468 return io_uring_setup(entries, params);
10471 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10473 struct io_uring_probe *p;
10477 size = struct_size(p, ops, nr_args);
10478 if (size == SIZE_MAX)
10480 p = kzalloc(size, GFP_KERNEL);
10485 if (copy_from_user(p, arg, size))
10488 if (memchr_inv(p, 0, size))
10491 p->last_op = IORING_OP_LAST - 1;
10492 if (nr_args > IORING_OP_LAST)
10493 nr_args = IORING_OP_LAST;
10495 for (i = 0; i < nr_args; i++) {
10497 if (!io_op_defs[i].not_supported)
10498 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10503 if (copy_to_user(arg, p, size))
10510 static int io_register_personality(struct io_ring_ctx *ctx)
10512 const struct cred *creds;
10516 creds = get_current_cred();
10518 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10519 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10527 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10528 unsigned int nr_args)
10530 struct io_uring_restriction *res;
10534 /* Restrictions allowed only if rings started disabled */
10535 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10538 /* We allow only a single restrictions registration */
10539 if (ctx->restrictions.registered)
10542 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10545 size = array_size(nr_args, sizeof(*res));
10546 if (size == SIZE_MAX)
10549 res = memdup_user(arg, size);
10551 return PTR_ERR(res);
10555 for (i = 0; i < nr_args; i++) {
10556 switch (res[i].opcode) {
10557 case IORING_RESTRICTION_REGISTER_OP:
10558 if (res[i].register_op >= IORING_REGISTER_LAST) {
10563 __set_bit(res[i].register_op,
10564 ctx->restrictions.register_op);
10566 case IORING_RESTRICTION_SQE_OP:
10567 if (res[i].sqe_op >= IORING_OP_LAST) {
10572 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10574 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10575 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10577 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10578 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10587 /* Reset all restrictions if an error happened */
10589 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10591 ctx->restrictions.registered = true;
10597 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10599 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10602 if (ctx->restrictions.registered)
10603 ctx->restricted = 1;
10605 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10606 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10607 wake_up(&ctx->sq_data->wait);
10611 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10612 struct io_uring_rsrc_update2 *up,
10618 if (check_add_overflow(up->offset, nr_args, &tmp))
10620 err = io_rsrc_node_switch_start(ctx);
10625 case IORING_RSRC_FILE:
10626 return __io_sqe_files_update(ctx, up, nr_args);
10627 case IORING_RSRC_BUFFER:
10628 return __io_sqe_buffers_update(ctx, up, nr_args);
10633 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10636 struct io_uring_rsrc_update2 up;
10640 memset(&up, 0, sizeof(up));
10641 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10643 if (up.resv || up.resv2)
10645 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10648 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10649 unsigned size, unsigned type)
10651 struct io_uring_rsrc_update2 up;
10653 if (size != sizeof(up))
10655 if (copy_from_user(&up, arg, sizeof(up)))
10657 if (!up.nr || up.resv || up.resv2)
10659 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10662 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10663 unsigned int size, unsigned int type)
10665 struct io_uring_rsrc_register rr;
10667 /* keep it extendible */
10668 if (size != sizeof(rr))
10671 memset(&rr, 0, sizeof(rr));
10672 if (copy_from_user(&rr, arg, size))
10674 if (!rr.nr || rr.resv || rr.resv2)
10678 case IORING_RSRC_FILE:
10679 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10680 rr.nr, u64_to_user_ptr(rr.tags));
10681 case IORING_RSRC_BUFFER:
10682 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10683 rr.nr, u64_to_user_ptr(rr.tags));
10688 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10691 struct io_uring_task *tctx = current->io_uring;
10692 cpumask_var_t new_mask;
10695 if (!tctx || !tctx->io_wq)
10698 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10701 cpumask_clear(new_mask);
10702 if (len > cpumask_size())
10703 len = cpumask_size();
10705 if (in_compat_syscall()) {
10706 ret = compat_get_bitmap(cpumask_bits(new_mask),
10707 (const compat_ulong_t __user *)arg,
10708 len * 8 /* CHAR_BIT */);
10710 ret = copy_from_user(new_mask, arg, len);
10714 free_cpumask_var(new_mask);
10718 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10719 free_cpumask_var(new_mask);
10723 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10725 struct io_uring_task *tctx = current->io_uring;
10727 if (!tctx || !tctx->io_wq)
10730 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10733 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10735 __must_hold(&ctx->uring_lock)
10737 struct io_tctx_node *node;
10738 struct io_uring_task *tctx = NULL;
10739 struct io_sq_data *sqd = NULL;
10740 __u32 new_count[2];
10743 if (copy_from_user(new_count, arg, sizeof(new_count)))
10745 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10746 if (new_count[i] > INT_MAX)
10749 if (ctx->flags & IORING_SETUP_SQPOLL) {
10750 sqd = ctx->sq_data;
10753 * Observe the correct sqd->lock -> ctx->uring_lock
10754 * ordering. Fine to drop uring_lock here, we hold
10755 * a ref to the ctx.
10757 refcount_inc(&sqd->refs);
10758 mutex_unlock(&ctx->uring_lock);
10759 mutex_lock(&sqd->lock);
10760 mutex_lock(&ctx->uring_lock);
10762 tctx = sqd->thread->io_uring;
10765 tctx = current->io_uring;
10768 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10770 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10772 ctx->iowq_limits[i] = new_count[i];
10773 ctx->iowq_limits_set = true;
10776 if (tctx && tctx->io_wq) {
10777 ret = io_wq_max_workers(tctx->io_wq, new_count);
10781 memset(new_count, 0, sizeof(new_count));
10785 mutex_unlock(&sqd->lock);
10786 io_put_sq_data(sqd);
10789 if (copy_to_user(arg, new_count, sizeof(new_count)))
10792 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10796 /* now propagate the restriction to all registered users */
10797 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10798 struct io_uring_task *tctx = node->task->io_uring;
10800 if (WARN_ON_ONCE(!tctx->io_wq))
10803 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10804 new_count[i] = ctx->iowq_limits[i];
10805 /* ignore errors, it always returns zero anyway */
10806 (void)io_wq_max_workers(tctx->io_wq, new_count);
10811 mutex_unlock(&sqd->lock);
10812 io_put_sq_data(sqd);
10817 static bool io_register_op_must_quiesce(int op)
10820 case IORING_REGISTER_BUFFERS:
10821 case IORING_UNREGISTER_BUFFERS:
10822 case IORING_REGISTER_FILES:
10823 case IORING_UNREGISTER_FILES:
10824 case IORING_REGISTER_FILES_UPDATE:
10825 case IORING_REGISTER_PROBE:
10826 case IORING_REGISTER_PERSONALITY:
10827 case IORING_UNREGISTER_PERSONALITY:
10828 case IORING_REGISTER_FILES2:
10829 case IORING_REGISTER_FILES_UPDATE2:
10830 case IORING_REGISTER_BUFFERS2:
10831 case IORING_REGISTER_BUFFERS_UPDATE:
10832 case IORING_REGISTER_IOWQ_AFF:
10833 case IORING_UNREGISTER_IOWQ_AFF:
10834 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10841 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10845 percpu_ref_kill(&ctx->refs);
10848 * Drop uring mutex before waiting for references to exit. If another
10849 * thread is currently inside io_uring_enter() it might need to grab the
10850 * uring_lock to make progress. If we hold it here across the drain
10851 * wait, then we can deadlock. It's safe to drop the mutex here, since
10852 * no new references will come in after we've killed the percpu ref.
10854 mutex_unlock(&ctx->uring_lock);
10856 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10859 ret = io_run_task_work_sig();
10860 } while (ret >= 0);
10861 mutex_lock(&ctx->uring_lock);
10864 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10868 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10869 void __user *arg, unsigned nr_args)
10870 __releases(ctx->uring_lock)
10871 __acquires(ctx->uring_lock)
10876 * We're inside the ring mutex, if the ref is already dying, then
10877 * someone else killed the ctx or is already going through
10878 * io_uring_register().
10880 if (percpu_ref_is_dying(&ctx->refs))
10883 if (ctx->restricted) {
10884 if (opcode >= IORING_REGISTER_LAST)
10886 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10887 if (!test_bit(opcode, ctx->restrictions.register_op))
10891 if (io_register_op_must_quiesce(opcode)) {
10892 ret = io_ctx_quiesce(ctx);
10898 case IORING_REGISTER_BUFFERS:
10899 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10901 case IORING_UNREGISTER_BUFFERS:
10903 if (arg || nr_args)
10905 ret = io_sqe_buffers_unregister(ctx);
10907 case IORING_REGISTER_FILES:
10908 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10910 case IORING_UNREGISTER_FILES:
10912 if (arg || nr_args)
10914 ret = io_sqe_files_unregister(ctx);
10916 case IORING_REGISTER_FILES_UPDATE:
10917 ret = io_register_files_update(ctx, arg, nr_args);
10919 case IORING_REGISTER_EVENTFD:
10920 case IORING_REGISTER_EVENTFD_ASYNC:
10924 ret = io_eventfd_register(ctx, arg);
10927 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10928 ctx->eventfd_async = 1;
10930 ctx->eventfd_async = 0;
10932 case IORING_UNREGISTER_EVENTFD:
10934 if (arg || nr_args)
10936 ret = io_eventfd_unregister(ctx);
10938 case IORING_REGISTER_PROBE:
10940 if (!arg || nr_args > 256)
10942 ret = io_probe(ctx, arg, nr_args);
10944 case IORING_REGISTER_PERSONALITY:
10946 if (arg || nr_args)
10948 ret = io_register_personality(ctx);
10950 case IORING_UNREGISTER_PERSONALITY:
10954 ret = io_unregister_personality(ctx, nr_args);
10956 case IORING_REGISTER_ENABLE_RINGS:
10958 if (arg || nr_args)
10960 ret = io_register_enable_rings(ctx);
10962 case IORING_REGISTER_RESTRICTIONS:
10963 ret = io_register_restrictions(ctx, arg, nr_args);
10965 case IORING_REGISTER_FILES2:
10966 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10968 case IORING_REGISTER_FILES_UPDATE2:
10969 ret = io_register_rsrc_update(ctx, arg, nr_args,
10972 case IORING_REGISTER_BUFFERS2:
10973 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10975 case IORING_REGISTER_BUFFERS_UPDATE:
10976 ret = io_register_rsrc_update(ctx, arg, nr_args,
10977 IORING_RSRC_BUFFER);
10979 case IORING_REGISTER_IOWQ_AFF:
10981 if (!arg || !nr_args)
10983 ret = io_register_iowq_aff(ctx, arg, nr_args);
10985 case IORING_UNREGISTER_IOWQ_AFF:
10987 if (arg || nr_args)
10989 ret = io_unregister_iowq_aff(ctx);
10991 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10993 if (!arg || nr_args != 2)
10995 ret = io_register_iowq_max_workers(ctx, arg);
11002 if (io_register_op_must_quiesce(opcode)) {
11003 /* bring the ctx back to life */
11004 percpu_ref_reinit(&ctx->refs);
11005 reinit_completion(&ctx->ref_comp);
11010 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
11011 void __user *, arg, unsigned int, nr_args)
11013 struct io_ring_ctx *ctx;
11022 if (f.file->f_op != &io_uring_fops)
11025 ctx = f.file->private_data;
11027 io_run_task_work();
11029 mutex_lock(&ctx->uring_lock);
11030 ret = __io_uring_register(ctx, opcode, arg, nr_args);
11031 mutex_unlock(&ctx->uring_lock);
11032 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
11033 ctx->cq_ev_fd != NULL, ret);
11039 static int __init io_uring_init(void)
11041 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11042 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11043 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11046 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11047 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11048 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11049 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11050 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11051 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11052 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11053 BUILD_BUG_SQE_ELEM(8, __u64, off);
11054 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11055 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11056 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11057 BUILD_BUG_SQE_ELEM(24, __u32, len);
11058 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11059 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11060 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11061 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11062 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11063 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11064 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11065 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11066 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11067 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11068 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11069 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11070 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11071 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11072 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11073 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11074 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11075 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11076 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11077 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11078 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11080 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11081 sizeof(struct io_uring_rsrc_update));
11082 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11083 sizeof(struct io_uring_rsrc_update2));
11085 /* ->buf_index is u16 */
11086 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11088 /* should fit into one byte */
11089 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11091 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11092 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11094 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11098 __initcall(io_uring_init);