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 void io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 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(&req->task->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_fill_cqe_req(req, 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 s32 res, u32 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_fill_cqe(struct io_ring_ctx *ctx, u64 user_data,
1794 s32 res, u32 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 static noinline void io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 cflags)
1817 __io_fill_cqe(req->ctx, req->user_data, res, cflags);
1820 static noinline bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data,
1821 s32 res, u32 cflags)
1824 return __io_fill_cqe(ctx, user_data, res, cflags);
1827 static void io_req_complete_post(struct io_kiocb *req, s32 res,
1830 struct io_ring_ctx *ctx = req->ctx;
1832 spin_lock(&ctx->completion_lock);
1833 __io_fill_cqe(ctx, req->user_data, res, cflags);
1835 * If we're the last reference to this request, add to our locked
1838 if (req_ref_put_and_test(req)) {
1839 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1840 if (req->flags & IO_DISARM_MASK)
1841 io_disarm_next(req);
1843 io_req_task_queue(req->link);
1847 io_dismantle_req(req);
1848 io_put_task(req->task, 1);
1849 list_add(&req->inflight_entry, &ctx->locked_free_list);
1850 ctx->locked_free_nr++;
1852 if (!percpu_ref_tryget(&ctx->refs))
1855 io_commit_cqring(ctx);
1856 spin_unlock(&ctx->completion_lock);
1859 io_cqring_ev_posted(ctx);
1860 percpu_ref_put(&ctx->refs);
1864 static inline bool io_req_needs_clean(struct io_kiocb *req)
1866 return req->flags & IO_REQ_CLEAN_FLAGS;
1869 static inline void io_req_complete_state(struct io_kiocb *req, s32 res,
1872 if (io_req_needs_clean(req))
1875 req->compl.cflags = cflags;
1876 req->flags |= REQ_F_COMPLETE_INLINE;
1879 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1880 s32 res, u32 cflags)
1882 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1883 io_req_complete_state(req, res, cflags);
1885 io_req_complete_post(req, res, cflags);
1888 static inline void io_req_complete(struct io_kiocb *req, s32 res)
1890 __io_req_complete(req, 0, res, 0);
1893 static void io_req_complete_failed(struct io_kiocb *req, s32 res)
1896 io_req_complete_post(req, res, 0);
1899 static void io_req_complete_fail_submit(struct io_kiocb *req)
1902 * We don't submit, fail them all, for that replace hardlinks with
1903 * normal links. Extra REQ_F_LINK is tolerated.
1905 req->flags &= ~REQ_F_HARDLINK;
1906 req->flags |= REQ_F_LINK;
1907 io_req_complete_failed(req, req->result);
1911 * Don't initialise the fields below on every allocation, but do that in
1912 * advance and keep them valid across allocations.
1914 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1918 req->async_data = NULL;
1919 /* not necessary, but safer to zero */
1923 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1924 struct io_submit_state *state)
1926 spin_lock(&ctx->completion_lock);
1927 list_splice_init(&ctx->locked_free_list, &state->free_list);
1928 ctx->locked_free_nr = 0;
1929 spin_unlock(&ctx->completion_lock);
1932 /* Returns true IFF there are requests in the cache */
1933 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1935 struct io_submit_state *state = &ctx->submit_state;
1939 * If we have more than a batch's worth of requests in our IRQ side
1940 * locked cache, grab the lock and move them over to our submission
1943 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1944 io_flush_cached_locked_reqs(ctx, state);
1946 nr = state->free_reqs;
1947 while (!list_empty(&state->free_list)) {
1948 struct io_kiocb *req = list_first_entry(&state->free_list,
1949 struct io_kiocb, inflight_entry);
1951 list_del(&req->inflight_entry);
1952 state->reqs[nr++] = req;
1953 if (nr == ARRAY_SIZE(state->reqs))
1957 state->free_reqs = nr;
1962 * A request might get retired back into the request caches even before opcode
1963 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1964 * Because of that, io_alloc_req() should be called only under ->uring_lock
1965 * and with extra caution to not get a request that is still worked on.
1967 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1968 __must_hold(&ctx->uring_lock)
1970 struct io_submit_state *state = &ctx->submit_state;
1971 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1974 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1976 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1979 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1983 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1984 * retry single alloc to be on the safe side.
1986 if (unlikely(ret <= 0)) {
1987 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1988 if (!state->reqs[0])
1993 for (i = 0; i < ret; i++)
1994 io_preinit_req(state->reqs[i], ctx);
1995 state->free_reqs = ret;
1998 return state->reqs[state->free_reqs];
2001 static inline void io_put_file(struct file *file)
2007 static void io_dismantle_req(struct io_kiocb *req)
2009 unsigned int flags = req->flags;
2011 if (io_req_needs_clean(req))
2013 if (!(flags & REQ_F_FIXED_FILE))
2014 io_put_file(req->file);
2015 if (req->fixed_rsrc_refs)
2016 percpu_ref_put(req->fixed_rsrc_refs);
2017 if (req->async_data) {
2018 kfree(req->async_data);
2019 req->async_data = NULL;
2023 static void __io_free_req(struct io_kiocb *req)
2025 struct io_ring_ctx *ctx = req->ctx;
2027 io_dismantle_req(req);
2028 io_put_task(req->task, 1);
2030 spin_lock(&ctx->completion_lock);
2031 list_add(&req->inflight_entry, &ctx->locked_free_list);
2032 ctx->locked_free_nr++;
2033 spin_unlock(&ctx->completion_lock);
2035 percpu_ref_put(&ctx->refs);
2038 static inline void io_remove_next_linked(struct io_kiocb *req)
2040 struct io_kiocb *nxt = req->link;
2042 req->link = nxt->link;
2046 static bool io_kill_linked_timeout(struct io_kiocb *req)
2047 __must_hold(&req->ctx->completion_lock)
2048 __must_hold(&req->ctx->timeout_lock)
2050 struct io_kiocb *link = req->link;
2052 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2053 struct io_timeout_data *io = link->async_data;
2055 io_remove_next_linked(req);
2056 link->timeout.head = NULL;
2057 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2058 list_del(&link->timeout.list);
2059 io_fill_cqe_req(link, -ECANCELED, 0);
2060 io_put_req_deferred(link);
2067 static void io_fail_links(struct io_kiocb *req)
2068 __must_hold(&req->ctx->completion_lock)
2070 struct io_kiocb *nxt, *link = req->link;
2074 long res = -ECANCELED;
2076 if (link->flags & REQ_F_FAIL)
2082 trace_io_uring_fail_link(req, link);
2083 io_fill_cqe_req(link, res, 0);
2084 io_put_req_deferred(link);
2089 static bool io_disarm_next(struct io_kiocb *req)
2090 __must_hold(&req->ctx->completion_lock)
2092 bool posted = false;
2094 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2095 struct io_kiocb *link = req->link;
2097 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2098 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2099 io_remove_next_linked(req);
2100 io_fill_cqe_req(link, -ECANCELED, 0);
2101 io_put_req_deferred(link);
2104 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2105 struct io_ring_ctx *ctx = req->ctx;
2107 spin_lock_irq(&ctx->timeout_lock);
2108 posted = io_kill_linked_timeout(req);
2109 spin_unlock_irq(&ctx->timeout_lock);
2111 if (unlikely((req->flags & REQ_F_FAIL) &&
2112 !(req->flags & REQ_F_HARDLINK))) {
2113 posted |= (req->link != NULL);
2119 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2121 struct io_kiocb *nxt;
2124 * If LINK is set, we have dependent requests in this chain. If we
2125 * didn't fail this request, queue the first one up, moving any other
2126 * dependencies to the next request. In case of failure, fail the rest
2129 if (req->flags & IO_DISARM_MASK) {
2130 struct io_ring_ctx *ctx = req->ctx;
2133 spin_lock(&ctx->completion_lock);
2134 posted = io_disarm_next(req);
2136 io_commit_cqring(req->ctx);
2137 spin_unlock(&ctx->completion_lock);
2139 io_cqring_ev_posted(ctx);
2146 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2148 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2150 return __io_req_find_next(req);
2153 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2158 if (ctx->submit_state.compl_nr)
2159 io_submit_flush_completions(ctx);
2160 mutex_unlock(&ctx->uring_lock);
2163 percpu_ref_put(&ctx->refs);
2166 static void tctx_task_work(struct callback_head *cb)
2168 bool locked = false;
2169 struct io_ring_ctx *ctx = NULL;
2170 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2174 struct io_wq_work_node *node;
2176 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2177 io_submit_flush_completions(ctx);
2179 spin_lock_irq(&tctx->task_lock);
2180 node = tctx->task_list.first;
2181 INIT_WQ_LIST(&tctx->task_list);
2183 tctx->task_running = false;
2184 spin_unlock_irq(&tctx->task_lock);
2189 struct io_wq_work_node *next = node->next;
2190 struct io_kiocb *req = container_of(node, struct io_kiocb,
2193 if (req->ctx != ctx) {
2194 ctx_flush_and_put(ctx, &locked);
2196 /* if not contended, grab and improve batching */
2197 locked = mutex_trylock(&ctx->uring_lock);
2198 percpu_ref_get(&ctx->refs);
2200 req->io_task_work.func(req, &locked);
2207 ctx_flush_and_put(ctx, &locked);
2209 /* relaxed read is enough as only the task itself sets ->in_idle */
2210 if (unlikely(atomic_read(&tctx->in_idle)))
2211 io_uring_drop_tctx_refs(current);
2214 static void io_req_task_work_add(struct io_kiocb *req)
2216 struct task_struct *tsk = req->task;
2217 struct io_uring_task *tctx = tsk->io_uring;
2218 enum task_work_notify_mode notify;
2219 struct io_wq_work_node *node;
2220 unsigned long flags;
2223 WARN_ON_ONCE(!tctx);
2225 spin_lock_irqsave(&tctx->task_lock, flags);
2226 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2227 running = tctx->task_running;
2229 tctx->task_running = true;
2230 spin_unlock_irqrestore(&tctx->task_lock, flags);
2232 /* task_work already pending, we're done */
2237 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2238 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2239 * processing task_work. There's no reliable way to tell if TWA_RESUME
2242 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2243 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2244 wake_up_process(tsk);
2248 spin_lock_irqsave(&tctx->task_lock, flags);
2249 tctx->task_running = false;
2250 node = tctx->task_list.first;
2251 INIT_WQ_LIST(&tctx->task_list);
2252 spin_unlock_irqrestore(&tctx->task_lock, flags);
2255 req = container_of(node, struct io_kiocb, io_task_work.node);
2257 if (llist_add(&req->io_task_work.fallback_node,
2258 &req->ctx->fallback_llist))
2259 schedule_delayed_work(&req->ctx->fallback_work, 1);
2263 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2265 struct io_ring_ctx *ctx = req->ctx;
2267 /* not needed for normal modes, but SQPOLL depends on it */
2268 io_tw_lock(ctx, locked);
2269 io_req_complete_failed(req, req->result);
2272 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2274 struct io_ring_ctx *ctx = req->ctx;
2276 io_tw_lock(ctx, locked);
2277 /* req->task == current here, checking PF_EXITING is safe */
2278 if (likely(!(req->task->flags & PF_EXITING)))
2279 __io_queue_sqe(req);
2281 io_req_complete_failed(req, -EFAULT);
2284 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2287 req->io_task_work.func = io_req_task_cancel;
2288 io_req_task_work_add(req);
2291 static void io_req_task_queue(struct io_kiocb *req)
2293 req->io_task_work.func = io_req_task_submit;
2294 io_req_task_work_add(req);
2297 static void io_req_task_queue_reissue(struct io_kiocb *req)
2299 req->io_task_work.func = io_queue_async_work;
2300 io_req_task_work_add(req);
2303 static inline void io_queue_next(struct io_kiocb *req)
2305 struct io_kiocb *nxt = io_req_find_next(req);
2308 io_req_task_queue(nxt);
2311 static void io_free_req(struct io_kiocb *req)
2317 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2323 struct task_struct *task;
2328 static inline void io_init_req_batch(struct req_batch *rb)
2335 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2336 struct req_batch *rb)
2339 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2341 io_put_task(rb->task, rb->task_refs);
2344 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2345 struct io_submit_state *state)
2348 io_dismantle_req(req);
2350 if (req->task != rb->task) {
2352 io_put_task(rb->task, rb->task_refs);
2353 rb->task = req->task;
2359 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2360 state->reqs[state->free_reqs++] = req;
2362 list_add(&req->inflight_entry, &state->free_list);
2365 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2366 __must_hold(&ctx->uring_lock)
2368 struct io_submit_state *state = &ctx->submit_state;
2369 int i, nr = state->compl_nr;
2370 struct req_batch rb;
2372 spin_lock(&ctx->completion_lock);
2373 for (i = 0; i < nr; i++) {
2374 struct io_kiocb *req = state->compl_reqs[i];
2376 __io_fill_cqe(ctx, req->user_data, req->result,
2379 io_commit_cqring(ctx);
2380 spin_unlock(&ctx->completion_lock);
2381 io_cqring_ev_posted(ctx);
2383 io_init_req_batch(&rb);
2384 for (i = 0; i < nr; i++) {
2385 struct io_kiocb *req = state->compl_reqs[i];
2387 if (req_ref_put_and_test(req))
2388 io_req_free_batch(&rb, req, &ctx->submit_state);
2391 io_req_free_batch_finish(ctx, &rb);
2392 state->compl_nr = 0;
2396 * Drop reference to request, return next in chain (if there is one) if this
2397 * was the last reference to this request.
2399 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2401 struct io_kiocb *nxt = NULL;
2403 if (req_ref_put_and_test(req)) {
2404 nxt = io_req_find_next(req);
2410 static inline void io_put_req(struct io_kiocb *req)
2412 if (req_ref_put_and_test(req))
2416 static inline void io_put_req_deferred(struct io_kiocb *req)
2418 if (req_ref_put_and_test(req)) {
2419 req->io_task_work.func = io_free_req_work;
2420 io_req_task_work_add(req);
2424 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2426 /* See comment at the top of this file */
2428 return __io_cqring_events(ctx);
2431 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2433 struct io_rings *rings = ctx->rings;
2435 /* make sure SQ entry isn't read before tail */
2436 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2439 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2441 unsigned int cflags;
2443 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2444 cflags |= IORING_CQE_F_BUFFER;
2445 req->flags &= ~REQ_F_BUFFER_SELECTED;
2450 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2452 struct io_buffer *kbuf;
2454 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2456 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2457 return io_put_kbuf(req, kbuf);
2460 static inline bool io_run_task_work(void)
2462 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2463 __set_current_state(TASK_RUNNING);
2464 tracehook_notify_signal();
2472 * Find and free completed poll iocbs
2474 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2475 struct list_head *done)
2477 struct req_batch rb;
2478 struct io_kiocb *req;
2480 /* order with ->result store in io_complete_rw_iopoll() */
2483 io_init_req_batch(&rb);
2484 while (!list_empty(done)) {
2485 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2486 list_del(&req->inflight_entry);
2488 io_fill_cqe_req(req, req->result, io_put_rw_kbuf(req));
2491 if (req_ref_put_and_test(req))
2492 io_req_free_batch(&rb, req, &ctx->submit_state);
2495 io_commit_cqring(ctx);
2496 io_cqring_ev_posted_iopoll(ctx);
2497 io_req_free_batch_finish(ctx, &rb);
2500 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2503 struct io_kiocb *req, *tmp;
2508 * Only spin for completions if we don't have multiple devices hanging
2509 * off our complete list, and we're under the requested amount.
2511 spin = !ctx->poll_multi_queue && *nr_events < min;
2513 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2514 struct kiocb *kiocb = &req->rw.kiocb;
2518 * Move completed and retryable entries to our local lists.
2519 * If we find a request that requires polling, break out
2520 * and complete those lists first, if we have entries there.
2522 if (READ_ONCE(req->iopoll_completed)) {
2523 list_move_tail(&req->inflight_entry, &done);
2526 if (!list_empty(&done))
2529 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2530 if (unlikely(ret < 0))
2535 /* iopoll may have completed current req */
2536 if (READ_ONCE(req->iopoll_completed))
2537 list_move_tail(&req->inflight_entry, &done);
2540 if (!list_empty(&done))
2541 io_iopoll_complete(ctx, nr_events, &done);
2547 * We can't just wait for polled events to come to us, we have to actively
2548 * find and complete them.
2550 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2552 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2555 mutex_lock(&ctx->uring_lock);
2556 while (!list_empty(&ctx->iopoll_list)) {
2557 unsigned int nr_events = 0;
2559 io_do_iopoll(ctx, &nr_events, 0);
2561 /* let it sleep and repeat later if can't complete a request */
2565 * Ensure we allow local-to-the-cpu processing to take place,
2566 * in this case we need to ensure that we reap all events.
2567 * Also let task_work, etc. to progress by releasing the mutex
2569 if (need_resched()) {
2570 mutex_unlock(&ctx->uring_lock);
2572 mutex_lock(&ctx->uring_lock);
2575 mutex_unlock(&ctx->uring_lock);
2578 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2580 unsigned int nr_events = 0;
2584 * We disallow the app entering submit/complete with polling, but we
2585 * still need to lock the ring to prevent racing with polled issue
2586 * that got punted to a workqueue.
2588 mutex_lock(&ctx->uring_lock);
2590 * Don't enter poll loop if we already have events pending.
2591 * If we do, we can potentially be spinning for commands that
2592 * already triggered a CQE (eg in error).
2594 if (test_bit(0, &ctx->check_cq_overflow))
2595 __io_cqring_overflow_flush(ctx, false);
2596 if (io_cqring_events(ctx))
2600 * If a submit got punted to a workqueue, we can have the
2601 * application entering polling for a command before it gets
2602 * issued. That app will hold the uring_lock for the duration
2603 * of the poll right here, so we need to take a breather every
2604 * now and then to ensure that the issue has a chance to add
2605 * the poll to the issued list. Otherwise we can spin here
2606 * forever, while the workqueue is stuck trying to acquire the
2609 if (list_empty(&ctx->iopoll_list)) {
2610 u32 tail = ctx->cached_cq_tail;
2612 mutex_unlock(&ctx->uring_lock);
2614 mutex_lock(&ctx->uring_lock);
2616 /* some requests don't go through iopoll_list */
2617 if (tail != ctx->cached_cq_tail ||
2618 list_empty(&ctx->iopoll_list))
2621 ret = io_do_iopoll(ctx, &nr_events, min);
2622 } while (!ret && nr_events < min && !need_resched());
2624 mutex_unlock(&ctx->uring_lock);
2628 static void kiocb_end_write(struct io_kiocb *req)
2631 * Tell lockdep we inherited freeze protection from submission
2634 if (req->flags & REQ_F_ISREG) {
2635 struct super_block *sb = file_inode(req->file)->i_sb;
2637 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2643 static bool io_resubmit_prep(struct io_kiocb *req)
2645 struct io_async_rw *rw = req->async_data;
2648 return !io_req_prep_async(req);
2649 iov_iter_restore(&rw->iter, &rw->iter_state);
2653 static bool io_rw_should_reissue(struct io_kiocb *req)
2655 umode_t mode = file_inode(req->file)->i_mode;
2656 struct io_ring_ctx *ctx = req->ctx;
2658 if (!S_ISBLK(mode) && !S_ISREG(mode))
2660 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2661 !(ctx->flags & IORING_SETUP_IOPOLL)))
2664 * If ref is dying, we might be running poll reap from the exit work.
2665 * Don't attempt to reissue from that path, just let it fail with
2668 if (percpu_ref_is_dying(&ctx->refs))
2671 * Play it safe and assume not safe to re-import and reissue if we're
2672 * not in the original thread group (or in task context).
2674 if (!same_thread_group(req->task, current) || !in_task())
2679 static bool io_resubmit_prep(struct io_kiocb *req)
2683 static bool io_rw_should_reissue(struct io_kiocb *req)
2689 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2691 if (req->rw.kiocb.ki_flags & IOCB_WRITE) {
2692 kiocb_end_write(req);
2693 fsnotify_modify(req->file);
2695 fsnotify_access(req->file);
2697 if (res != req->result) {
2698 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2699 io_rw_should_reissue(req)) {
2700 req->flags |= REQ_F_REISSUE;
2709 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2711 unsigned int cflags = io_put_rw_kbuf(req);
2712 int res = req->result;
2715 struct io_ring_ctx *ctx = req->ctx;
2716 struct io_submit_state *state = &ctx->submit_state;
2718 io_req_complete_state(req, res, cflags);
2719 state->compl_reqs[state->compl_nr++] = req;
2720 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2721 io_submit_flush_completions(ctx);
2723 io_req_complete_post(req, res, cflags);
2727 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2728 unsigned int issue_flags)
2730 if (__io_complete_rw_common(req, res))
2732 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2735 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2737 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2739 if (__io_complete_rw_common(req, res))
2742 req->io_task_work.func = io_req_task_complete;
2743 io_req_task_work_add(req);
2746 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2748 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2750 if (kiocb->ki_flags & IOCB_WRITE)
2751 kiocb_end_write(req);
2752 if (unlikely(res != req->result)) {
2753 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2754 req->flags |= REQ_F_REISSUE;
2759 WRITE_ONCE(req->result, res);
2760 /* order with io_iopoll_complete() checking ->result */
2762 WRITE_ONCE(req->iopoll_completed, 1);
2766 * After the iocb has been issued, it's safe to be found on the poll list.
2767 * Adding the kiocb to the list AFTER submission ensures that we don't
2768 * find it from a io_do_iopoll() thread before the issuer is done
2769 * accessing the kiocb cookie.
2771 static void io_iopoll_req_issued(struct io_kiocb *req)
2773 struct io_ring_ctx *ctx = req->ctx;
2774 const bool in_async = io_wq_current_is_worker();
2776 /* workqueue context doesn't hold uring_lock, grab it now */
2777 if (unlikely(in_async))
2778 mutex_lock(&ctx->uring_lock);
2781 * Track whether we have multiple files in our lists. This will impact
2782 * how we do polling eventually, not spinning if we're on potentially
2783 * different devices.
2785 if (list_empty(&ctx->iopoll_list)) {
2786 ctx->poll_multi_queue = false;
2787 } else if (!ctx->poll_multi_queue) {
2788 struct io_kiocb *list_req;
2789 unsigned int queue_num0, queue_num1;
2791 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2794 if (list_req->file != req->file) {
2795 ctx->poll_multi_queue = true;
2797 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2798 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2799 if (queue_num0 != queue_num1)
2800 ctx->poll_multi_queue = true;
2805 * For fast devices, IO may have already completed. If it has, add
2806 * it to the front so we find it first.
2808 if (READ_ONCE(req->iopoll_completed))
2809 list_add(&req->inflight_entry, &ctx->iopoll_list);
2811 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2813 if (unlikely(in_async)) {
2815 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2816 * in sq thread task context or in io worker task context. If
2817 * current task context is sq thread, we don't need to check
2818 * whether should wake up sq thread.
2820 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2821 wq_has_sleeper(&ctx->sq_data->wait))
2822 wake_up(&ctx->sq_data->wait);
2824 mutex_unlock(&ctx->uring_lock);
2828 static bool io_bdev_nowait(struct block_device *bdev)
2830 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2834 * If we tracked the file through the SCM inflight mechanism, we could support
2835 * any file. For now, just ensure that anything potentially problematic is done
2838 static bool __io_file_supports_nowait(struct file *file, int rw)
2840 umode_t mode = file_inode(file)->i_mode;
2842 if (S_ISBLK(mode)) {
2843 if (IS_ENABLED(CONFIG_BLOCK) &&
2844 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2850 if (S_ISREG(mode)) {
2851 if (IS_ENABLED(CONFIG_BLOCK) &&
2852 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2853 file->f_op != &io_uring_fops)
2858 /* any ->read/write should understand O_NONBLOCK */
2859 if (file->f_flags & O_NONBLOCK)
2862 if (!(file->f_mode & FMODE_NOWAIT))
2866 return file->f_op->read_iter != NULL;
2868 return file->f_op->write_iter != NULL;
2871 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2873 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2875 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2878 return __io_file_supports_nowait(req->file, rw);
2881 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2884 struct io_ring_ctx *ctx = req->ctx;
2885 struct kiocb *kiocb = &req->rw.kiocb;
2886 struct file *file = req->file;
2890 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2891 req->flags |= REQ_F_ISREG;
2893 kiocb->ki_pos = READ_ONCE(sqe->off);
2894 if (kiocb->ki_pos == -1) {
2895 if (!(file->f_mode & FMODE_STREAM)) {
2896 req->flags |= REQ_F_CUR_POS;
2897 kiocb->ki_pos = file->f_pos;
2902 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2903 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2904 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2909 * If the file is marked O_NONBLOCK, still allow retry for it if it
2910 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2911 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2913 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2914 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2915 req->flags |= REQ_F_NOWAIT;
2917 ioprio = READ_ONCE(sqe->ioprio);
2919 ret = ioprio_check_cap(ioprio);
2923 kiocb->ki_ioprio = ioprio;
2925 kiocb->ki_ioprio = get_current_ioprio();
2927 if (ctx->flags & IORING_SETUP_IOPOLL) {
2928 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2929 !kiocb->ki_filp->f_op->iopoll)
2932 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2933 kiocb->ki_complete = io_complete_rw_iopoll;
2934 req->iopoll_completed = 0;
2936 if (kiocb->ki_flags & IOCB_HIPRI)
2938 kiocb->ki_complete = io_complete_rw;
2941 /* used for fixed read/write too - just read unconditionally */
2942 req->buf_index = READ_ONCE(sqe->buf_index);
2945 if (req->opcode == IORING_OP_READ_FIXED ||
2946 req->opcode == IORING_OP_WRITE_FIXED) {
2947 struct io_ring_ctx *ctx = req->ctx;
2950 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
2952 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
2953 req->imu = ctx->user_bufs[index];
2954 io_req_set_rsrc_node(req);
2957 req->rw.addr = READ_ONCE(sqe->addr);
2958 req->rw.len = READ_ONCE(sqe->len);
2962 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2968 case -ERESTARTNOINTR:
2969 case -ERESTARTNOHAND:
2970 case -ERESTART_RESTARTBLOCK:
2972 * We can't just restart the syscall, since previously
2973 * submitted sqes may already be in progress. Just fail this
2979 kiocb->ki_complete(kiocb, ret, 0);
2983 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2984 unsigned int issue_flags)
2986 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2987 struct io_async_rw *io = req->async_data;
2989 /* add previously done IO, if any */
2990 if (io && io->bytes_done > 0) {
2992 ret = io->bytes_done;
2994 ret += io->bytes_done;
2997 if (req->flags & REQ_F_CUR_POS)
2998 req->file->f_pos = kiocb->ki_pos;
2999 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
3000 __io_complete_rw(req, ret, 0, issue_flags);
3002 io_rw_done(kiocb, ret);
3004 if (req->flags & REQ_F_REISSUE) {
3005 req->flags &= ~REQ_F_REISSUE;
3006 if (io_resubmit_prep(req)) {
3007 io_req_task_queue_reissue(req);
3009 unsigned int cflags = io_put_rw_kbuf(req);
3010 struct io_ring_ctx *ctx = req->ctx;
3013 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
3014 mutex_lock(&ctx->uring_lock);
3015 __io_req_complete(req, issue_flags, ret, cflags);
3016 mutex_unlock(&ctx->uring_lock);
3018 __io_req_complete(req, issue_flags, ret, cflags);
3024 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3025 struct io_mapped_ubuf *imu)
3027 size_t len = req->rw.len;
3028 u64 buf_end, buf_addr = req->rw.addr;
3031 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
3033 /* not inside the mapped region */
3034 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
3038 * May not be a start of buffer, set size appropriately
3039 * and advance us to the beginning.
3041 offset = buf_addr - imu->ubuf;
3042 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3046 * Don't use iov_iter_advance() here, as it's really slow for
3047 * using the latter parts of a big fixed buffer - it iterates
3048 * over each segment manually. We can cheat a bit here, because
3051 * 1) it's a BVEC iter, we set it up
3052 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3053 * first and last bvec
3055 * So just find our index, and adjust the iterator afterwards.
3056 * If the offset is within the first bvec (or the whole first
3057 * bvec, just use iov_iter_advance(). This makes it easier
3058 * since we can just skip the first segment, which may not
3059 * be PAGE_SIZE aligned.
3061 const struct bio_vec *bvec = imu->bvec;
3063 if (offset <= bvec->bv_len) {
3064 iov_iter_advance(iter, offset);
3066 unsigned long seg_skip;
3068 /* skip first vec */
3069 offset -= bvec->bv_len;
3070 seg_skip = 1 + (offset >> PAGE_SHIFT);
3072 iter->bvec = bvec + seg_skip;
3073 iter->nr_segs -= seg_skip;
3074 iter->count -= bvec->bv_len + offset;
3075 iter->iov_offset = offset & ~PAGE_MASK;
3082 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3084 if (WARN_ON_ONCE(!req->imu))
3086 return __io_import_fixed(req, rw, iter, req->imu);
3089 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3092 mutex_unlock(&ctx->uring_lock);
3095 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3098 * "Normal" inline submissions always hold the uring_lock, since we
3099 * grab it from the system call. Same is true for the SQPOLL offload.
3100 * The only exception is when we've detached the request and issue it
3101 * from an async worker thread, grab the lock for that case.
3104 mutex_lock(&ctx->uring_lock);
3107 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3108 int bgid, struct io_buffer *kbuf,
3111 struct io_buffer *head;
3113 if (req->flags & REQ_F_BUFFER_SELECTED)
3116 io_ring_submit_lock(req->ctx, needs_lock);
3118 lockdep_assert_held(&req->ctx->uring_lock);
3120 head = xa_load(&req->ctx->io_buffers, bgid);
3122 if (!list_empty(&head->list)) {
3123 kbuf = list_last_entry(&head->list, struct io_buffer,
3125 list_del(&kbuf->list);
3128 xa_erase(&req->ctx->io_buffers, bgid);
3130 if (*len > kbuf->len)
3133 kbuf = ERR_PTR(-ENOBUFS);
3136 io_ring_submit_unlock(req->ctx, needs_lock);
3141 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3144 struct io_buffer *kbuf;
3147 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3148 bgid = req->buf_index;
3149 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3152 req->rw.addr = (u64) (unsigned long) kbuf;
3153 req->flags |= REQ_F_BUFFER_SELECTED;
3154 return u64_to_user_ptr(kbuf->addr);
3157 #ifdef CONFIG_COMPAT
3158 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3161 struct compat_iovec __user *uiov;
3162 compat_ssize_t clen;
3166 uiov = u64_to_user_ptr(req->rw.addr);
3167 if (!access_ok(uiov, sizeof(*uiov)))
3169 if (__get_user(clen, &uiov->iov_len))
3175 buf = io_rw_buffer_select(req, &len, needs_lock);
3177 return PTR_ERR(buf);
3178 iov[0].iov_base = buf;
3179 iov[0].iov_len = (compat_size_t) len;
3184 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3187 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3191 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3194 len = iov[0].iov_len;
3197 buf = io_rw_buffer_select(req, &len, needs_lock);
3199 return PTR_ERR(buf);
3200 iov[0].iov_base = buf;
3201 iov[0].iov_len = len;
3205 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3208 if (req->flags & REQ_F_BUFFER_SELECTED) {
3209 struct io_buffer *kbuf;
3211 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3212 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3213 iov[0].iov_len = kbuf->len;
3216 if (req->rw.len != 1)
3219 #ifdef CONFIG_COMPAT
3220 if (req->ctx->compat)
3221 return io_compat_import(req, iov, needs_lock);
3224 return __io_iov_buffer_select(req, iov, needs_lock);
3227 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3228 struct iov_iter *iter, bool needs_lock)
3230 void __user *buf = u64_to_user_ptr(req->rw.addr);
3231 size_t sqe_len = req->rw.len;
3232 u8 opcode = req->opcode;
3235 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3237 return io_import_fixed(req, rw, iter);
3240 /* buffer index only valid with fixed read/write, or buffer select */
3241 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3244 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3245 if (req->flags & REQ_F_BUFFER_SELECT) {
3246 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3248 return PTR_ERR(buf);
3249 req->rw.len = sqe_len;
3252 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3257 if (req->flags & REQ_F_BUFFER_SELECT) {
3258 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3260 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3265 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3269 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3271 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3275 * For files that don't have ->read_iter() and ->write_iter(), handle them
3276 * by looping over ->read() or ->write() manually.
3278 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3280 struct kiocb *kiocb = &req->rw.kiocb;
3281 struct file *file = req->file;
3285 * Don't support polled IO through this interface, and we can't
3286 * support non-blocking either. For the latter, this just causes
3287 * the kiocb to be handled from an async context.
3289 if (kiocb->ki_flags & IOCB_HIPRI)
3291 if (kiocb->ki_flags & IOCB_NOWAIT)
3294 while (iov_iter_count(iter)) {
3298 if (!iov_iter_is_bvec(iter)) {
3299 iovec = iov_iter_iovec(iter);
3301 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3302 iovec.iov_len = req->rw.len;
3306 nr = file->f_op->read(file, iovec.iov_base,
3307 iovec.iov_len, io_kiocb_ppos(kiocb));
3309 nr = file->f_op->write(file, iovec.iov_base,
3310 iovec.iov_len, io_kiocb_ppos(kiocb));
3319 if (!iov_iter_is_bvec(iter)) {
3320 iov_iter_advance(iter, nr);
3327 if (nr != iovec.iov_len)
3334 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3335 const struct iovec *fast_iov, struct iov_iter *iter)
3337 struct io_async_rw *rw = req->async_data;
3339 memcpy(&rw->iter, iter, sizeof(*iter));
3340 rw->free_iovec = iovec;
3342 /* can only be fixed buffers, no need to do anything */
3343 if (iov_iter_is_bvec(iter))
3346 unsigned iov_off = 0;
3348 rw->iter.iov = rw->fast_iov;
3349 if (iter->iov != fast_iov) {
3350 iov_off = iter->iov - fast_iov;
3351 rw->iter.iov += iov_off;
3353 if (rw->fast_iov != fast_iov)
3354 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3355 sizeof(struct iovec) * iter->nr_segs);
3357 req->flags |= REQ_F_NEED_CLEANUP;
3361 static inline int io_alloc_async_data(struct io_kiocb *req)
3363 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3364 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3365 return req->async_data == NULL;
3368 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3369 const struct iovec *fast_iov,
3370 struct iov_iter *iter, bool force)
3372 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3374 if (!req->async_data) {
3375 struct io_async_rw *iorw;
3377 if (io_alloc_async_data(req)) {
3382 io_req_map_rw(req, iovec, fast_iov, iter);
3383 iorw = req->async_data;
3384 /* we've copied and mapped the iter, ensure state is saved */
3385 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3390 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3392 struct io_async_rw *iorw = req->async_data;
3393 struct iovec *iov = iorw->fast_iov;
3396 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3397 if (unlikely(ret < 0))
3400 iorw->bytes_done = 0;
3401 iorw->free_iovec = iov;
3403 req->flags |= REQ_F_NEED_CLEANUP;
3404 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3408 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3410 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3412 return io_prep_rw(req, sqe, READ);
3416 * This is our waitqueue callback handler, registered through lock_page_async()
3417 * when we initially tried to do the IO with the iocb armed our waitqueue.
3418 * This gets called when the page is unlocked, and we generally expect that to
3419 * happen when the page IO is completed and the page is now uptodate. This will
3420 * queue a task_work based retry of the operation, attempting to copy the data
3421 * again. If the latter fails because the page was NOT uptodate, then we will
3422 * do a thread based blocking retry of the operation. That's the unexpected
3425 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3426 int sync, void *arg)
3428 struct wait_page_queue *wpq;
3429 struct io_kiocb *req = wait->private;
3430 struct wait_page_key *key = arg;
3432 wpq = container_of(wait, struct wait_page_queue, wait);
3434 if (!wake_page_match(wpq, key))
3437 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3438 list_del_init(&wait->entry);
3439 io_req_task_queue(req);
3444 * This controls whether a given IO request should be armed for async page
3445 * based retry. If we return false here, the request is handed to the async
3446 * worker threads for retry. If we're doing buffered reads on a regular file,
3447 * we prepare a private wait_page_queue entry and retry the operation. This
3448 * will either succeed because the page is now uptodate and unlocked, or it
3449 * will register a callback when the page is unlocked at IO completion. Through
3450 * that callback, io_uring uses task_work to setup a retry of the operation.
3451 * That retry will attempt the buffered read again. The retry will generally
3452 * succeed, or in rare cases where it fails, we then fall back to using the
3453 * async worker threads for a blocking retry.
3455 static bool io_rw_should_retry(struct io_kiocb *req)
3457 struct io_async_rw *rw = req->async_data;
3458 struct wait_page_queue *wait = &rw->wpq;
3459 struct kiocb *kiocb = &req->rw.kiocb;
3461 /* never retry for NOWAIT, we just complete with -EAGAIN */
3462 if (req->flags & REQ_F_NOWAIT)
3465 /* Only for buffered IO */
3466 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3470 * just use poll if we can, and don't attempt if the fs doesn't
3471 * support callback based unlocks
3473 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3476 wait->wait.func = io_async_buf_func;
3477 wait->wait.private = req;
3478 wait->wait.flags = 0;
3479 INIT_LIST_HEAD(&wait->wait.entry);
3480 kiocb->ki_flags |= IOCB_WAITQ;
3481 kiocb->ki_flags &= ~IOCB_NOWAIT;
3482 kiocb->ki_waitq = wait;
3486 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3488 if (req->file->f_op->read_iter)
3489 return call_read_iter(req->file, &req->rw.kiocb, iter);
3490 else if (req->file->f_op->read)
3491 return loop_rw_iter(READ, req, iter);
3496 static bool need_read_all(struct io_kiocb *req)
3498 return req->flags & REQ_F_ISREG ||
3499 S_ISBLK(file_inode(req->file)->i_mode);
3502 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3504 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3505 struct kiocb *kiocb = &req->rw.kiocb;
3506 struct iov_iter __iter, *iter = &__iter;
3507 struct io_async_rw *rw = req->async_data;
3508 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3509 struct iov_iter_state __state, *state;
3514 state = &rw->iter_state;
3516 * We come here from an earlier attempt, restore our state to
3517 * match in case it doesn't. It's cheap enough that we don't
3518 * need to make this conditional.
3520 iov_iter_restore(iter, state);
3523 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3527 iov_iter_save_state(iter, state);
3529 req->result = iov_iter_count(iter);
3531 /* Ensure we clear previously set non-block flag */
3532 if (!force_nonblock)
3533 kiocb->ki_flags &= ~IOCB_NOWAIT;
3535 kiocb->ki_flags |= IOCB_NOWAIT;
3537 /* If the file doesn't support async, just async punt */
3538 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3539 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3540 return ret ?: -EAGAIN;
3543 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3544 if (unlikely(ret)) {
3549 ret = io_iter_do_read(req, iter);
3551 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3552 req->flags &= ~REQ_F_REISSUE;
3553 /* IOPOLL retry should happen for io-wq threads */
3554 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3556 /* no retry on NONBLOCK nor RWF_NOWAIT */
3557 if (req->flags & REQ_F_NOWAIT)
3560 } else if (ret == -EIOCBQUEUED) {
3562 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3563 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3564 /* read all, failed, already did sync or don't want to retry */
3569 * Don't depend on the iter state matching what was consumed, or being
3570 * untouched in case of error. Restore it and we'll advance it
3571 * manually if we need to.
3573 iov_iter_restore(iter, state);
3575 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3580 rw = req->async_data;
3582 * Now use our persistent iterator and state, if we aren't already.
3583 * We've restored and mapped the iter to match.
3585 if (iter != &rw->iter) {
3587 state = &rw->iter_state;
3592 * We end up here because of a partial read, either from
3593 * above or inside this loop. Advance the iter by the bytes
3594 * that were consumed.
3596 iov_iter_advance(iter, ret);
3597 if (!iov_iter_count(iter))
3599 rw->bytes_done += ret;
3600 iov_iter_save_state(iter, state);
3602 /* if we can retry, do so with the callbacks armed */
3603 if (!io_rw_should_retry(req)) {
3604 kiocb->ki_flags &= ~IOCB_WAITQ;
3609 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3610 * we get -EIOCBQUEUED, then we'll get a notification when the
3611 * desired page gets unlocked. We can also get a partial read
3612 * here, and if we do, then just retry at the new offset.
3614 ret = io_iter_do_read(req, iter);
3615 if (ret == -EIOCBQUEUED)
3617 /* we got some bytes, but not all. retry. */
3618 kiocb->ki_flags &= ~IOCB_WAITQ;
3619 iov_iter_restore(iter, state);
3622 kiocb_done(kiocb, ret, issue_flags);
3624 /* it's faster to check here then delegate to kfree */
3630 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3632 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3634 return io_prep_rw(req, sqe, WRITE);
3637 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3639 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3640 struct kiocb *kiocb = &req->rw.kiocb;
3641 struct iov_iter __iter, *iter = &__iter;
3642 struct io_async_rw *rw = req->async_data;
3643 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3644 struct iov_iter_state __state, *state;
3649 state = &rw->iter_state;
3650 iov_iter_restore(iter, state);
3653 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3657 iov_iter_save_state(iter, state);
3659 req->result = iov_iter_count(iter);
3661 /* Ensure we clear previously set non-block flag */
3662 if (!force_nonblock)
3663 kiocb->ki_flags &= ~IOCB_NOWAIT;
3665 kiocb->ki_flags |= IOCB_NOWAIT;
3667 /* If the file doesn't support async, just async punt */
3668 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3671 /* file path doesn't support NOWAIT for non-direct_IO */
3672 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3673 (req->flags & REQ_F_ISREG))
3676 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3681 * Open-code file_start_write here to grab freeze protection,
3682 * which will be released by another thread in
3683 * io_complete_rw(). Fool lockdep by telling it the lock got
3684 * released so that it doesn't complain about the held lock when
3685 * we return to userspace.
3687 if (req->flags & REQ_F_ISREG) {
3688 sb_start_write(file_inode(req->file)->i_sb);
3689 __sb_writers_release(file_inode(req->file)->i_sb,
3692 kiocb->ki_flags |= IOCB_WRITE;
3694 if (req->file->f_op->write_iter)
3695 ret2 = call_write_iter(req->file, kiocb, iter);
3696 else if (req->file->f_op->write)
3697 ret2 = loop_rw_iter(WRITE, req, iter);
3701 if (req->flags & REQ_F_REISSUE) {
3702 req->flags &= ~REQ_F_REISSUE;
3707 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3708 * retry them without IOCB_NOWAIT.
3710 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3712 /* no retry on NONBLOCK nor RWF_NOWAIT */
3713 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3715 if (!force_nonblock || ret2 != -EAGAIN) {
3716 /* IOPOLL retry should happen for io-wq threads */
3717 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3720 kiocb_done(kiocb, ret2, issue_flags);
3723 iov_iter_restore(iter, state);
3724 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3726 if (kiocb->ki_flags & IOCB_WRITE)
3727 kiocb_end_write(req);
3733 /* it's reportedly faster than delegating the null check to kfree() */
3739 static int io_renameat_prep(struct io_kiocb *req,
3740 const struct io_uring_sqe *sqe)
3742 struct io_rename *ren = &req->rename;
3743 const char __user *oldf, *newf;
3745 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3747 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3749 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3752 ren->old_dfd = READ_ONCE(sqe->fd);
3753 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3754 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3755 ren->new_dfd = READ_ONCE(sqe->len);
3756 ren->flags = READ_ONCE(sqe->rename_flags);
3758 ren->oldpath = getname(oldf);
3759 if (IS_ERR(ren->oldpath))
3760 return PTR_ERR(ren->oldpath);
3762 ren->newpath = getname(newf);
3763 if (IS_ERR(ren->newpath)) {
3764 putname(ren->oldpath);
3765 return PTR_ERR(ren->newpath);
3768 req->flags |= REQ_F_NEED_CLEANUP;
3772 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3774 struct io_rename *ren = &req->rename;
3777 if (issue_flags & IO_URING_F_NONBLOCK)
3780 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3781 ren->newpath, ren->flags);
3783 req->flags &= ~REQ_F_NEED_CLEANUP;
3786 io_req_complete(req, ret);
3790 static int io_unlinkat_prep(struct io_kiocb *req,
3791 const struct io_uring_sqe *sqe)
3793 struct io_unlink *un = &req->unlink;
3794 const char __user *fname;
3796 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3798 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3801 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3804 un->dfd = READ_ONCE(sqe->fd);
3806 un->flags = READ_ONCE(sqe->unlink_flags);
3807 if (un->flags & ~AT_REMOVEDIR)
3810 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3811 un->filename = getname(fname);
3812 if (IS_ERR(un->filename))
3813 return PTR_ERR(un->filename);
3815 req->flags |= REQ_F_NEED_CLEANUP;
3819 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3821 struct io_unlink *un = &req->unlink;
3824 if (issue_flags & IO_URING_F_NONBLOCK)
3827 if (un->flags & AT_REMOVEDIR)
3828 ret = do_rmdir(un->dfd, un->filename);
3830 ret = do_unlinkat(un->dfd, un->filename);
3832 req->flags &= ~REQ_F_NEED_CLEANUP;
3835 io_req_complete(req, ret);
3839 static int io_mkdirat_prep(struct io_kiocb *req,
3840 const struct io_uring_sqe *sqe)
3842 struct io_mkdir *mkd = &req->mkdir;
3843 const char __user *fname;
3845 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3847 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3850 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3853 mkd->dfd = READ_ONCE(sqe->fd);
3854 mkd->mode = READ_ONCE(sqe->len);
3856 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3857 mkd->filename = getname(fname);
3858 if (IS_ERR(mkd->filename))
3859 return PTR_ERR(mkd->filename);
3861 req->flags |= REQ_F_NEED_CLEANUP;
3865 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3867 struct io_mkdir *mkd = &req->mkdir;
3870 if (issue_flags & IO_URING_F_NONBLOCK)
3873 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3875 req->flags &= ~REQ_F_NEED_CLEANUP;
3878 io_req_complete(req, ret);
3882 static int io_symlinkat_prep(struct io_kiocb *req,
3883 const struct io_uring_sqe *sqe)
3885 struct io_symlink *sl = &req->symlink;
3886 const char __user *oldpath, *newpath;
3888 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3890 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3893 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3896 sl->new_dfd = READ_ONCE(sqe->fd);
3897 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3898 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3900 sl->oldpath = getname(oldpath);
3901 if (IS_ERR(sl->oldpath))
3902 return PTR_ERR(sl->oldpath);
3904 sl->newpath = getname(newpath);
3905 if (IS_ERR(sl->newpath)) {
3906 putname(sl->oldpath);
3907 return PTR_ERR(sl->newpath);
3910 req->flags |= REQ_F_NEED_CLEANUP;
3914 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3916 struct io_symlink *sl = &req->symlink;
3919 if (issue_flags & IO_URING_F_NONBLOCK)
3922 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3924 req->flags &= ~REQ_F_NEED_CLEANUP;
3927 io_req_complete(req, ret);
3931 static int io_linkat_prep(struct io_kiocb *req,
3932 const struct io_uring_sqe *sqe)
3934 struct io_hardlink *lnk = &req->hardlink;
3935 const char __user *oldf, *newf;
3937 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3939 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3941 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3944 lnk->old_dfd = READ_ONCE(sqe->fd);
3945 lnk->new_dfd = READ_ONCE(sqe->len);
3946 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3947 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3948 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3950 lnk->oldpath = getname(oldf);
3951 if (IS_ERR(lnk->oldpath))
3952 return PTR_ERR(lnk->oldpath);
3954 lnk->newpath = getname(newf);
3955 if (IS_ERR(lnk->newpath)) {
3956 putname(lnk->oldpath);
3957 return PTR_ERR(lnk->newpath);
3960 req->flags |= REQ_F_NEED_CLEANUP;
3964 static int io_linkat(struct io_kiocb *req, int issue_flags)
3966 struct io_hardlink *lnk = &req->hardlink;
3969 if (issue_flags & IO_URING_F_NONBLOCK)
3972 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3973 lnk->newpath, lnk->flags);
3975 req->flags &= ~REQ_F_NEED_CLEANUP;
3978 io_req_complete(req, ret);
3982 static int io_shutdown_prep(struct io_kiocb *req,
3983 const struct io_uring_sqe *sqe)
3985 #if defined(CONFIG_NET)
3986 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3988 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3989 sqe->buf_index || sqe->splice_fd_in))
3992 req->shutdown.how = READ_ONCE(sqe->len);
3999 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
4001 #if defined(CONFIG_NET)
4002 struct socket *sock;
4005 if (issue_flags & IO_URING_F_NONBLOCK)
4008 sock = sock_from_file(req->file);
4009 if (unlikely(!sock))
4012 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4015 io_req_complete(req, ret);
4022 static int __io_splice_prep(struct io_kiocb *req,
4023 const struct io_uring_sqe *sqe)
4025 struct io_splice *sp = &req->splice;
4026 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4028 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4031 sp->len = READ_ONCE(sqe->len);
4032 sp->flags = READ_ONCE(sqe->splice_flags);
4033 if (unlikely(sp->flags & ~valid_flags))
4035 sp->splice_fd_in = READ_ONCE(sqe->splice_fd_in);
4039 static int io_tee_prep(struct io_kiocb *req,
4040 const struct io_uring_sqe *sqe)
4042 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4044 return __io_splice_prep(req, sqe);
4047 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4049 struct io_splice *sp = &req->splice;
4050 struct file *out = sp->file_out;
4051 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4055 if (issue_flags & IO_URING_F_NONBLOCK)
4058 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4059 (sp->flags & SPLICE_F_FD_IN_FIXED));
4066 ret = do_tee(in, out, sp->len, flags);
4068 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4073 io_req_complete(req, ret);
4077 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4079 struct io_splice *sp = &req->splice;
4081 sp->off_in = READ_ONCE(sqe->splice_off_in);
4082 sp->off_out = READ_ONCE(sqe->off);
4083 return __io_splice_prep(req, sqe);
4086 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4088 struct io_splice *sp = &req->splice;
4089 struct file *out = sp->file_out;
4090 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4091 loff_t *poff_in, *poff_out;
4095 if (issue_flags & IO_URING_F_NONBLOCK)
4098 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4099 (sp->flags & SPLICE_F_FD_IN_FIXED));
4105 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4106 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4109 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4111 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4116 io_req_complete(req, ret);
4121 * IORING_OP_NOP just posts a completion event, nothing else.
4123 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4125 struct io_ring_ctx *ctx = req->ctx;
4127 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4130 __io_req_complete(req, issue_flags, 0, 0);
4134 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4136 struct io_ring_ctx *ctx = req->ctx;
4138 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4140 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4144 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4145 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4148 req->sync.off = READ_ONCE(sqe->off);
4149 req->sync.len = READ_ONCE(sqe->len);
4153 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4155 loff_t end = req->sync.off + req->sync.len;
4158 /* fsync always requires a blocking context */
4159 if (issue_flags & IO_URING_F_NONBLOCK)
4162 ret = vfs_fsync_range(req->file, req->sync.off,
4163 end > 0 ? end : LLONG_MAX,
4164 req->sync.flags & IORING_FSYNC_DATASYNC);
4167 io_req_complete(req, ret);
4171 static int io_fallocate_prep(struct io_kiocb *req,
4172 const struct io_uring_sqe *sqe)
4174 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4177 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4180 req->sync.off = READ_ONCE(sqe->off);
4181 req->sync.len = READ_ONCE(sqe->addr);
4182 req->sync.mode = READ_ONCE(sqe->len);
4186 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4190 /* fallocate always requiring blocking context */
4191 if (issue_flags & IO_URING_F_NONBLOCK)
4193 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4198 fsnotify_modify(req->file);
4199 io_req_complete(req, ret);
4203 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4205 const char __user *fname;
4208 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4210 if (unlikely(sqe->ioprio || sqe->buf_index))
4212 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4215 /* open.how should be already initialised */
4216 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4217 req->open.how.flags |= O_LARGEFILE;
4219 req->open.dfd = READ_ONCE(sqe->fd);
4220 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4221 req->open.filename = getname(fname);
4222 if (IS_ERR(req->open.filename)) {
4223 ret = PTR_ERR(req->open.filename);
4224 req->open.filename = NULL;
4228 req->open.file_slot = READ_ONCE(sqe->file_index);
4229 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4232 req->open.nofile = rlimit(RLIMIT_NOFILE);
4233 req->flags |= REQ_F_NEED_CLEANUP;
4237 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4239 u64 mode = READ_ONCE(sqe->len);
4240 u64 flags = READ_ONCE(sqe->open_flags);
4242 req->open.how = build_open_how(flags, mode);
4243 return __io_openat_prep(req, sqe);
4246 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4248 struct open_how __user *how;
4252 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4253 len = READ_ONCE(sqe->len);
4254 if (len < OPEN_HOW_SIZE_VER0)
4257 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4262 return __io_openat_prep(req, sqe);
4265 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4267 struct open_flags op;
4269 bool resolve_nonblock, nonblock_set;
4270 bool fixed = !!req->open.file_slot;
4273 ret = build_open_flags(&req->open.how, &op);
4276 nonblock_set = op.open_flag & O_NONBLOCK;
4277 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4278 if (issue_flags & IO_URING_F_NONBLOCK) {
4280 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4281 * it'll always -EAGAIN
4283 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4285 op.lookup_flags |= LOOKUP_CACHED;
4286 op.open_flag |= O_NONBLOCK;
4290 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4295 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4298 * We could hang on to this 'fd' on retrying, but seems like
4299 * marginal gain for something that is now known to be a slower
4300 * path. So just put it, and we'll get a new one when we retry.
4305 ret = PTR_ERR(file);
4306 /* only retry if RESOLVE_CACHED wasn't already set by application */
4307 if (ret == -EAGAIN &&
4308 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4313 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4314 file->f_flags &= ~O_NONBLOCK;
4315 fsnotify_open(file);
4318 fd_install(ret, file);
4320 ret = io_install_fixed_file(req, file, issue_flags,
4321 req->open.file_slot - 1);
4323 putname(req->open.filename);
4324 req->flags &= ~REQ_F_NEED_CLEANUP;
4327 __io_req_complete(req, issue_flags, ret, 0);
4331 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4333 return io_openat2(req, issue_flags);
4336 static int io_remove_buffers_prep(struct io_kiocb *req,
4337 const struct io_uring_sqe *sqe)
4339 struct io_provide_buf *p = &req->pbuf;
4342 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4346 tmp = READ_ONCE(sqe->fd);
4347 if (!tmp || tmp > USHRT_MAX)
4350 memset(p, 0, sizeof(*p));
4352 p->bgid = READ_ONCE(sqe->buf_group);
4356 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4357 int bgid, unsigned nbufs)
4361 /* shouldn't happen */
4365 /* the head kbuf is the list itself */
4366 while (!list_empty(&buf->list)) {
4367 struct io_buffer *nxt;
4369 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4370 list_del(&nxt->list);
4378 xa_erase(&ctx->io_buffers, bgid);
4383 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4385 struct io_provide_buf *p = &req->pbuf;
4386 struct io_ring_ctx *ctx = req->ctx;
4387 struct io_buffer *head;
4389 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4391 io_ring_submit_lock(ctx, !force_nonblock);
4393 lockdep_assert_held(&ctx->uring_lock);
4396 head = xa_load(&ctx->io_buffers, p->bgid);
4398 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4402 /* complete before unlock, IOPOLL may need the lock */
4403 __io_req_complete(req, issue_flags, ret, 0);
4404 io_ring_submit_unlock(ctx, !force_nonblock);
4408 static int io_provide_buffers_prep(struct io_kiocb *req,
4409 const struct io_uring_sqe *sqe)
4411 unsigned long size, tmp_check;
4412 struct io_provide_buf *p = &req->pbuf;
4415 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4418 tmp = READ_ONCE(sqe->fd);
4419 if (!tmp || tmp > USHRT_MAX)
4422 p->addr = READ_ONCE(sqe->addr);
4423 p->len = READ_ONCE(sqe->len);
4425 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4428 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4431 size = (unsigned long)p->len * p->nbufs;
4432 if (!access_ok(u64_to_user_ptr(p->addr), size))
4435 p->bgid = READ_ONCE(sqe->buf_group);
4436 tmp = READ_ONCE(sqe->off);
4437 if (tmp > USHRT_MAX)
4443 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4445 struct io_buffer *buf;
4446 u64 addr = pbuf->addr;
4447 int i, bid = pbuf->bid;
4449 for (i = 0; i < pbuf->nbufs; i++) {
4450 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4455 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4460 INIT_LIST_HEAD(&buf->list);
4463 list_add_tail(&buf->list, &(*head)->list);
4468 return i ? i : -ENOMEM;
4471 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4473 struct io_provide_buf *p = &req->pbuf;
4474 struct io_ring_ctx *ctx = req->ctx;
4475 struct io_buffer *head, *list;
4477 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4479 io_ring_submit_lock(ctx, !force_nonblock);
4481 lockdep_assert_held(&ctx->uring_lock);
4483 list = head = xa_load(&ctx->io_buffers, p->bgid);
4485 ret = io_add_buffers(p, &head);
4486 if (ret >= 0 && !list) {
4487 ret = xa_insert(&ctx->io_buffers, p->bgid, head,
4488 GFP_KERNEL_ACCOUNT);
4490 __io_remove_buffers(ctx, head, p->bgid, -1U);
4494 /* complete before unlock, IOPOLL may need the lock */
4495 __io_req_complete(req, issue_flags, ret, 0);
4496 io_ring_submit_unlock(ctx, !force_nonblock);
4500 static int io_epoll_ctl_prep(struct io_kiocb *req,
4501 const struct io_uring_sqe *sqe)
4503 #if defined(CONFIG_EPOLL)
4504 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4506 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4509 req->epoll.epfd = READ_ONCE(sqe->fd);
4510 req->epoll.op = READ_ONCE(sqe->len);
4511 req->epoll.fd = READ_ONCE(sqe->off);
4513 if (ep_op_has_event(req->epoll.op)) {
4514 struct epoll_event __user *ev;
4516 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4517 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4527 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4529 #if defined(CONFIG_EPOLL)
4530 struct io_epoll *ie = &req->epoll;
4532 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4534 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4535 if (force_nonblock && ret == -EAGAIN)
4540 __io_req_complete(req, issue_flags, ret, 0);
4547 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4549 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4550 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4552 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4555 req->madvise.addr = READ_ONCE(sqe->addr);
4556 req->madvise.len = READ_ONCE(sqe->len);
4557 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4564 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4566 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4567 struct io_madvise *ma = &req->madvise;
4570 if (issue_flags & IO_URING_F_NONBLOCK)
4573 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4576 io_req_complete(req, ret);
4583 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4585 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4587 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4590 req->fadvise.offset = READ_ONCE(sqe->off);
4591 req->fadvise.len = READ_ONCE(sqe->len);
4592 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4596 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4598 struct io_fadvise *fa = &req->fadvise;
4601 if (issue_flags & IO_URING_F_NONBLOCK) {
4602 switch (fa->advice) {
4603 case POSIX_FADV_NORMAL:
4604 case POSIX_FADV_RANDOM:
4605 case POSIX_FADV_SEQUENTIAL:
4612 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4615 __io_req_complete(req, issue_flags, ret, 0);
4619 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4621 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4623 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4625 if (req->flags & REQ_F_FIXED_FILE)
4628 req->statx.dfd = READ_ONCE(sqe->fd);
4629 req->statx.mask = READ_ONCE(sqe->len);
4630 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4631 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4632 req->statx.flags = READ_ONCE(sqe->statx_flags);
4637 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4639 struct io_statx *ctx = &req->statx;
4642 if (issue_flags & IO_URING_F_NONBLOCK)
4645 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4650 io_req_complete(req, ret);
4654 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4656 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4658 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4659 sqe->rw_flags || sqe->buf_index)
4661 if (req->flags & REQ_F_FIXED_FILE)
4664 req->close.fd = READ_ONCE(sqe->fd);
4665 req->close.file_slot = READ_ONCE(sqe->file_index);
4666 if (req->close.file_slot && req->close.fd)
4672 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4674 struct files_struct *files = current->files;
4675 struct io_close *close = &req->close;
4676 struct fdtable *fdt;
4677 struct file *file = NULL;
4680 if (req->close.file_slot) {
4681 ret = io_close_fixed(req, issue_flags);
4685 spin_lock(&files->file_lock);
4686 fdt = files_fdtable(files);
4687 if (close->fd >= fdt->max_fds) {
4688 spin_unlock(&files->file_lock);
4691 file = fdt->fd[close->fd];
4692 if (!file || file->f_op == &io_uring_fops) {
4693 spin_unlock(&files->file_lock);
4698 /* if the file has a flush method, be safe and punt to async */
4699 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4700 spin_unlock(&files->file_lock);
4704 ret = __close_fd_get_file(close->fd, &file);
4705 spin_unlock(&files->file_lock);
4712 /* No ->flush() or already async, safely close from here */
4713 ret = filp_close(file, current->files);
4719 __io_req_complete(req, issue_flags, ret, 0);
4723 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4725 struct io_ring_ctx *ctx = req->ctx;
4727 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4729 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4733 req->sync.off = READ_ONCE(sqe->off);
4734 req->sync.len = READ_ONCE(sqe->len);
4735 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4739 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4743 /* sync_file_range always requires a blocking context */
4744 if (issue_flags & IO_URING_F_NONBLOCK)
4747 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4751 io_req_complete(req, ret);
4755 #if defined(CONFIG_NET)
4756 static int io_setup_async_msg(struct io_kiocb *req,
4757 struct io_async_msghdr *kmsg)
4759 struct io_async_msghdr *async_msg = req->async_data;
4763 if (io_alloc_async_data(req)) {
4764 kfree(kmsg->free_iov);
4767 async_msg = req->async_data;
4768 req->flags |= REQ_F_NEED_CLEANUP;
4769 memcpy(async_msg, kmsg, sizeof(*kmsg));
4770 async_msg->msg.msg_name = &async_msg->addr;
4771 /* if were using fast_iov, set it to the new one */
4772 if (!async_msg->free_iov)
4773 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4778 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4779 struct io_async_msghdr *iomsg)
4781 iomsg->msg.msg_name = &iomsg->addr;
4782 iomsg->free_iov = iomsg->fast_iov;
4783 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4784 req->sr_msg.msg_flags, &iomsg->free_iov);
4787 static int io_sendmsg_prep_async(struct io_kiocb *req)
4791 ret = io_sendmsg_copy_hdr(req, req->async_data);
4793 req->flags |= REQ_F_NEED_CLEANUP;
4797 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4799 struct io_sr_msg *sr = &req->sr_msg;
4801 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4803 if (unlikely(sqe->addr2 || sqe->file_index))
4805 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
4808 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4809 sr->len = READ_ONCE(sqe->len);
4810 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4811 if (sr->msg_flags & MSG_DONTWAIT)
4812 req->flags |= REQ_F_NOWAIT;
4814 #ifdef CONFIG_COMPAT
4815 if (req->ctx->compat)
4816 sr->msg_flags |= MSG_CMSG_COMPAT;
4821 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4823 struct io_async_msghdr iomsg, *kmsg;
4824 struct socket *sock;
4829 sock = sock_from_file(req->file);
4830 if (unlikely(!sock))
4833 kmsg = req->async_data;
4835 ret = io_sendmsg_copy_hdr(req, &iomsg);
4841 flags = req->sr_msg.msg_flags;
4842 if (issue_flags & IO_URING_F_NONBLOCK)
4843 flags |= MSG_DONTWAIT;
4844 if (flags & MSG_WAITALL)
4845 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4847 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4848 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4849 return io_setup_async_msg(req, kmsg);
4850 if (ret == -ERESTARTSYS)
4853 /* fast path, check for non-NULL to avoid function call */
4855 kfree(kmsg->free_iov);
4856 req->flags &= ~REQ_F_NEED_CLEANUP;
4859 __io_req_complete(req, issue_flags, ret, 0);
4863 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4865 struct io_sr_msg *sr = &req->sr_msg;
4868 struct socket *sock;
4873 sock = sock_from_file(req->file);
4874 if (unlikely(!sock))
4877 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4881 msg.msg_name = NULL;
4882 msg.msg_control = NULL;
4883 msg.msg_controllen = 0;
4884 msg.msg_namelen = 0;
4886 flags = req->sr_msg.msg_flags;
4887 if (issue_flags & IO_URING_F_NONBLOCK)
4888 flags |= MSG_DONTWAIT;
4889 if (flags & MSG_WAITALL)
4890 min_ret = iov_iter_count(&msg.msg_iter);
4892 msg.msg_flags = flags;
4893 ret = sock_sendmsg(sock, &msg);
4894 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4896 if (ret == -ERESTARTSYS)
4901 __io_req_complete(req, issue_flags, ret, 0);
4905 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4906 struct io_async_msghdr *iomsg)
4908 struct io_sr_msg *sr = &req->sr_msg;
4909 struct iovec __user *uiov;
4913 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4914 &iomsg->uaddr, &uiov, &iov_len);
4918 if (req->flags & REQ_F_BUFFER_SELECT) {
4921 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4923 sr->len = iomsg->fast_iov[0].iov_len;
4924 iomsg->free_iov = NULL;
4926 iomsg->free_iov = iomsg->fast_iov;
4927 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4928 &iomsg->free_iov, &iomsg->msg.msg_iter,
4937 #ifdef CONFIG_COMPAT
4938 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4939 struct io_async_msghdr *iomsg)
4941 struct io_sr_msg *sr = &req->sr_msg;
4942 struct compat_iovec __user *uiov;
4947 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4952 uiov = compat_ptr(ptr);
4953 if (req->flags & REQ_F_BUFFER_SELECT) {
4954 compat_ssize_t clen;
4958 if (!access_ok(uiov, sizeof(*uiov)))
4960 if (__get_user(clen, &uiov->iov_len))
4965 iomsg->free_iov = NULL;
4967 iomsg->free_iov = iomsg->fast_iov;
4968 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4969 UIO_FASTIOV, &iomsg->free_iov,
4970 &iomsg->msg.msg_iter, true);
4979 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4980 struct io_async_msghdr *iomsg)
4982 iomsg->msg.msg_name = &iomsg->addr;
4984 #ifdef CONFIG_COMPAT
4985 if (req->ctx->compat)
4986 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4989 return __io_recvmsg_copy_hdr(req, iomsg);
4992 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4995 struct io_sr_msg *sr = &req->sr_msg;
4996 struct io_buffer *kbuf;
4998 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
5003 req->flags |= REQ_F_BUFFER_SELECTED;
5007 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
5009 return io_put_kbuf(req, req->sr_msg.kbuf);
5012 static int io_recvmsg_prep_async(struct io_kiocb *req)
5016 ret = io_recvmsg_copy_hdr(req, req->async_data);
5018 req->flags |= REQ_F_NEED_CLEANUP;
5022 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5024 struct io_sr_msg *sr = &req->sr_msg;
5026 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5028 if (unlikely(sqe->addr2 || sqe->file_index))
5030 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
5033 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5034 sr->len = READ_ONCE(sqe->len);
5035 sr->bgid = READ_ONCE(sqe->buf_group);
5036 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5037 if (sr->msg_flags & MSG_DONTWAIT)
5038 req->flags |= REQ_F_NOWAIT;
5040 #ifdef CONFIG_COMPAT
5041 if (req->ctx->compat)
5042 sr->msg_flags |= MSG_CMSG_COMPAT;
5047 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5049 struct io_async_msghdr iomsg, *kmsg;
5050 struct socket *sock;
5051 struct io_buffer *kbuf;
5054 int ret, cflags = 0;
5055 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5057 sock = sock_from_file(req->file);
5058 if (unlikely(!sock))
5061 kmsg = req->async_data;
5063 ret = io_recvmsg_copy_hdr(req, &iomsg);
5069 if (req->flags & REQ_F_BUFFER_SELECT) {
5070 kbuf = io_recv_buffer_select(req, !force_nonblock);
5072 return PTR_ERR(kbuf);
5073 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5074 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5075 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5076 1, req->sr_msg.len);
5079 flags = req->sr_msg.msg_flags;
5081 flags |= MSG_DONTWAIT;
5082 if (flags & MSG_WAITALL)
5083 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5085 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5086 kmsg->uaddr, flags);
5087 if (force_nonblock && ret == -EAGAIN)
5088 return io_setup_async_msg(req, kmsg);
5089 if (ret == -ERESTARTSYS)
5092 if (req->flags & REQ_F_BUFFER_SELECTED)
5093 cflags = io_put_recv_kbuf(req);
5094 /* fast path, check for non-NULL to avoid function call */
5096 kfree(kmsg->free_iov);
5097 req->flags &= ~REQ_F_NEED_CLEANUP;
5098 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5100 __io_req_complete(req, issue_flags, ret, cflags);
5104 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5106 struct io_buffer *kbuf;
5107 struct io_sr_msg *sr = &req->sr_msg;
5109 void __user *buf = sr->buf;
5110 struct socket *sock;
5114 int ret, cflags = 0;
5115 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5117 sock = sock_from_file(req->file);
5118 if (unlikely(!sock))
5121 if (req->flags & REQ_F_BUFFER_SELECT) {
5122 kbuf = io_recv_buffer_select(req, !force_nonblock);
5124 return PTR_ERR(kbuf);
5125 buf = u64_to_user_ptr(kbuf->addr);
5128 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5132 msg.msg_name = NULL;
5133 msg.msg_control = NULL;
5134 msg.msg_controllen = 0;
5135 msg.msg_namelen = 0;
5136 msg.msg_iocb = NULL;
5139 flags = req->sr_msg.msg_flags;
5141 flags |= MSG_DONTWAIT;
5142 if (flags & MSG_WAITALL)
5143 min_ret = iov_iter_count(&msg.msg_iter);
5145 ret = sock_recvmsg(sock, &msg, flags);
5146 if (force_nonblock && ret == -EAGAIN)
5148 if (ret == -ERESTARTSYS)
5151 if (req->flags & REQ_F_BUFFER_SELECTED)
5152 cflags = io_put_recv_kbuf(req);
5153 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5155 __io_req_complete(req, issue_flags, ret, cflags);
5159 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5161 struct io_accept *accept = &req->accept;
5163 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5165 if (sqe->ioprio || sqe->len || sqe->buf_index)
5168 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5169 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5170 accept->flags = READ_ONCE(sqe->accept_flags);
5171 accept->nofile = rlimit(RLIMIT_NOFILE);
5173 accept->file_slot = READ_ONCE(sqe->file_index);
5174 if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))
5176 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5178 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5179 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5183 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5185 struct io_accept *accept = &req->accept;
5186 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5187 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5188 bool fixed = !!accept->file_slot;
5192 if (req->file->f_flags & O_NONBLOCK)
5193 req->flags |= REQ_F_NOWAIT;
5196 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5197 if (unlikely(fd < 0))
5200 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5205 ret = PTR_ERR(file);
5206 if (ret == -EAGAIN && force_nonblock)
5208 if (ret == -ERESTARTSYS)
5211 } else if (!fixed) {
5212 fd_install(fd, file);
5215 ret = io_install_fixed_file(req, file, issue_flags,
5216 accept->file_slot - 1);
5218 __io_req_complete(req, issue_flags, ret, 0);
5222 static int io_connect_prep_async(struct io_kiocb *req)
5224 struct io_async_connect *io = req->async_data;
5225 struct io_connect *conn = &req->connect;
5227 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5230 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5232 struct io_connect *conn = &req->connect;
5234 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5236 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5240 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5241 conn->addr_len = READ_ONCE(sqe->addr2);
5245 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5247 struct io_async_connect __io, *io;
5248 unsigned file_flags;
5250 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5252 if (req->async_data) {
5253 io = req->async_data;
5255 ret = move_addr_to_kernel(req->connect.addr,
5256 req->connect.addr_len,
5263 file_flags = force_nonblock ? O_NONBLOCK : 0;
5265 ret = __sys_connect_file(req->file, &io->address,
5266 req->connect.addr_len, file_flags);
5267 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5268 if (req->async_data)
5270 if (io_alloc_async_data(req)) {
5274 memcpy(req->async_data, &__io, sizeof(__io));
5277 if (ret == -ERESTARTSYS)
5282 __io_req_complete(req, issue_flags, ret, 0);
5285 #else /* !CONFIG_NET */
5286 #define IO_NETOP_FN(op) \
5287 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5289 return -EOPNOTSUPP; \
5292 #define IO_NETOP_PREP(op) \
5294 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5296 return -EOPNOTSUPP; \
5299 #define IO_NETOP_PREP_ASYNC(op) \
5301 static int io_##op##_prep_async(struct io_kiocb *req) \
5303 return -EOPNOTSUPP; \
5306 IO_NETOP_PREP_ASYNC(sendmsg);
5307 IO_NETOP_PREP_ASYNC(recvmsg);
5308 IO_NETOP_PREP_ASYNC(connect);
5309 IO_NETOP_PREP(accept);
5312 #endif /* CONFIG_NET */
5314 struct io_poll_table {
5315 struct poll_table_struct pt;
5316 struct io_kiocb *req;
5321 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5323 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5324 if (req->opcode == IORING_OP_POLL_ADD)
5325 return req->async_data;
5326 return req->apoll->double_poll;
5329 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5331 if (req->opcode == IORING_OP_POLL_ADD)
5333 return &req->apoll->poll;
5336 static void io_poll_req_insert(struct io_kiocb *req)
5338 struct io_ring_ctx *ctx = req->ctx;
5339 struct hlist_head *list;
5341 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5342 hlist_add_head(&req->hash_node, list);
5345 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5346 wait_queue_func_t wake_func)
5350 poll->canceled = false;
5351 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5352 /* mask in events that we always want/need */
5353 poll->events = events | IO_POLL_UNMASK;
5354 INIT_LIST_HEAD(&poll->wait.entry);
5355 init_waitqueue_func_entry(&poll->wait, wake_func);
5358 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5359 __poll_t mask, io_req_tw_func_t func)
5361 /* for instances that support it check for an event match first: */
5362 if (mask && !(mask & poll->events))
5365 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5367 list_del_init(&poll->wait.entry);
5370 req->io_task_work.func = func;
5373 * If this fails, then the task is exiting. When a task exits, the
5374 * work gets canceled, so just cancel this request as well instead
5375 * of executing it. We can't safely execute it anyway, as we may not
5376 * have the needed state needed for it anyway.
5378 io_req_task_work_add(req);
5382 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5383 __acquires(&req->ctx->completion_lock)
5385 struct io_ring_ctx *ctx = req->ctx;
5387 /* req->task == current here, checking PF_EXITING is safe */
5388 if (unlikely(req->task->flags & PF_EXITING))
5389 WRITE_ONCE(poll->canceled, true);
5391 if (!req->result && !READ_ONCE(poll->canceled)) {
5392 struct poll_table_struct pt = { ._key = poll->events };
5394 req->result = vfs_poll(req->file, &pt) & poll->events;
5397 spin_lock(&ctx->completion_lock);
5398 if (!req->result && !READ_ONCE(poll->canceled)) {
5399 add_wait_queue(poll->head, &poll->wait);
5406 static void io_poll_remove_double(struct io_kiocb *req)
5407 __must_hold(&req->ctx->completion_lock)
5409 struct io_poll_iocb *poll = io_poll_get_double(req);
5411 lockdep_assert_held(&req->ctx->completion_lock);
5413 if (poll && poll->head) {
5414 struct wait_queue_head *head = poll->head;
5416 spin_lock_irq(&head->lock);
5417 list_del_init(&poll->wait.entry);
5418 if (poll->wait.private)
5421 spin_unlock_irq(&head->lock);
5425 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5426 __must_hold(&req->ctx->completion_lock)
5428 struct io_ring_ctx *ctx = req->ctx;
5429 unsigned flags = IORING_CQE_F_MORE;
5432 if (READ_ONCE(req->poll.canceled)) {
5434 req->poll.events |= EPOLLONESHOT;
5436 error = mangle_poll(mask);
5438 if (req->poll.events & EPOLLONESHOT)
5441 if (!(flags & IORING_CQE_F_MORE)) {
5442 io_fill_cqe_req(req, error, flags);
5443 } else if (!io_fill_cqe_aux(ctx, req->user_data, error, flags)) {
5444 req->poll.events |= EPOLLONESHOT;
5447 return !(flags & IORING_CQE_F_MORE);
5450 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5451 __must_hold(&req->ctx->completion_lock)
5455 done = __io_poll_complete(req, mask);
5456 io_commit_cqring(req->ctx);
5460 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5462 struct io_ring_ctx *ctx = req->ctx;
5464 if (io_poll_rewait(req, &req->poll)) {
5465 spin_unlock(&ctx->completion_lock);
5469 if (req->poll.done) {
5470 spin_unlock(&ctx->completion_lock);
5473 done = __io_poll_complete(req, req->result);
5475 io_poll_remove_double(req);
5476 hash_del(&req->hash_node);
5477 req->poll.done = true;
5480 add_wait_queue(req->poll.head, &req->poll.wait);
5482 io_commit_cqring(ctx);
5483 spin_unlock(&ctx->completion_lock);
5484 io_cqring_ev_posted(ctx);
5491 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5492 int sync, void *key)
5494 struct io_kiocb *req = wait->private;
5495 struct io_poll_iocb *poll = io_poll_get_single(req);
5496 __poll_t mask = key_to_poll(key);
5497 unsigned long flags;
5499 /* for instances that support it check for an event match first: */
5500 if (mask && !(mask & poll->events))
5502 if (!(poll->events & EPOLLONESHOT))
5503 return poll->wait.func(&poll->wait, mode, sync, key);
5505 list_del_init(&wait->entry);
5510 spin_lock_irqsave(&poll->head->lock, flags);
5511 done = list_empty(&poll->wait.entry);
5513 list_del_init(&poll->wait.entry);
5514 /* make sure double remove sees this as being gone */
5515 wait->private = NULL;
5516 spin_unlock_irqrestore(&poll->head->lock, flags);
5518 /* use wait func handler, so it matches the rq type */
5519 poll->wait.func(&poll->wait, mode, sync, key);
5526 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5527 struct wait_queue_head *head,
5528 struct io_poll_iocb **poll_ptr)
5530 struct io_kiocb *req = pt->req;
5533 * The file being polled uses multiple waitqueues for poll handling
5534 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5537 if (unlikely(pt->nr_entries)) {
5538 struct io_poll_iocb *poll_one = poll;
5540 /* double add on the same waitqueue head, ignore */
5541 if (poll_one->head == head)
5543 /* already have a 2nd entry, fail a third attempt */
5545 if ((*poll_ptr)->head == head)
5547 pt->error = -EINVAL;
5551 * Can't handle multishot for double wait for now, turn it
5552 * into one-shot mode.
5554 if (!(poll_one->events & EPOLLONESHOT))
5555 poll_one->events |= EPOLLONESHOT;
5556 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5558 pt->error = -ENOMEM;
5561 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5563 poll->wait.private = req;
5570 if (poll->events & EPOLLEXCLUSIVE)
5571 add_wait_queue_exclusive(head, &poll->wait);
5573 add_wait_queue(head, &poll->wait);
5576 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5577 struct poll_table_struct *p)
5579 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5580 struct async_poll *apoll = pt->req->apoll;
5582 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5585 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5587 struct async_poll *apoll = req->apoll;
5588 struct io_ring_ctx *ctx = req->ctx;
5590 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5592 if (io_poll_rewait(req, &apoll->poll)) {
5593 spin_unlock(&ctx->completion_lock);
5597 hash_del(&req->hash_node);
5598 io_poll_remove_double(req);
5599 apoll->poll.done = true;
5600 spin_unlock(&ctx->completion_lock);
5602 if (!READ_ONCE(apoll->poll.canceled))
5603 io_req_task_submit(req, locked);
5605 io_req_complete_failed(req, -ECANCELED);
5608 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5611 struct io_kiocb *req = wait->private;
5612 struct io_poll_iocb *poll = &req->apoll->poll;
5614 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5617 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5620 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5621 struct io_poll_iocb *poll,
5622 struct io_poll_table *ipt, __poll_t mask,
5623 wait_queue_func_t wake_func)
5624 __acquires(&ctx->completion_lock)
5626 struct io_ring_ctx *ctx = req->ctx;
5627 bool cancel = false;
5629 INIT_HLIST_NODE(&req->hash_node);
5630 io_init_poll_iocb(poll, mask, wake_func);
5631 poll->file = req->file;
5632 poll->wait.private = req;
5634 ipt->pt._key = mask;
5637 ipt->nr_entries = 0;
5639 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5640 if (unlikely(!ipt->nr_entries) && !ipt->error)
5641 ipt->error = -EINVAL;
5643 spin_lock(&ctx->completion_lock);
5644 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5645 io_poll_remove_double(req);
5646 if (likely(poll->head)) {
5647 spin_lock_irq(&poll->head->lock);
5648 if (unlikely(list_empty(&poll->wait.entry))) {
5654 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5655 list_del_init(&poll->wait.entry);
5657 WRITE_ONCE(poll->canceled, true);
5658 else if (!poll->done) /* actually waiting for an event */
5659 io_poll_req_insert(req);
5660 spin_unlock_irq(&poll->head->lock);
5672 static int io_arm_poll_handler(struct io_kiocb *req)
5674 const struct io_op_def *def = &io_op_defs[req->opcode];
5675 struct io_ring_ctx *ctx = req->ctx;
5676 struct async_poll *apoll;
5677 struct io_poll_table ipt;
5678 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5680 if (!req->file || !file_can_poll(req->file))
5681 return IO_APOLL_ABORTED;
5682 if (req->flags & REQ_F_POLLED)
5683 return IO_APOLL_ABORTED;
5684 if (!def->pollin && !def->pollout)
5685 return IO_APOLL_ABORTED;
5688 mask |= POLLIN | POLLRDNORM;
5690 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5691 if ((req->opcode == IORING_OP_RECVMSG) &&
5692 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5695 mask |= POLLOUT | POLLWRNORM;
5698 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5699 if (unlikely(!apoll))
5700 return IO_APOLL_ABORTED;
5701 apoll->double_poll = NULL;
5703 req->flags |= REQ_F_POLLED;
5704 ipt.pt._qproc = io_async_queue_proc;
5705 io_req_set_refcount(req);
5707 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5709 spin_unlock(&ctx->completion_lock);
5710 if (ret || ipt.error)
5711 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5713 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5714 mask, apoll->poll.events);
5718 static bool __io_poll_remove_one(struct io_kiocb *req,
5719 struct io_poll_iocb *poll, bool do_cancel)
5720 __must_hold(&req->ctx->completion_lock)
5722 bool do_complete = false;
5726 spin_lock_irq(&poll->head->lock);
5728 WRITE_ONCE(poll->canceled, true);
5729 if (!list_empty(&poll->wait.entry)) {
5730 list_del_init(&poll->wait.entry);
5733 spin_unlock_irq(&poll->head->lock);
5734 hash_del(&req->hash_node);
5738 static bool io_poll_remove_one(struct io_kiocb *req)
5739 __must_hold(&req->ctx->completion_lock)
5743 io_poll_remove_double(req);
5744 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5748 io_fill_cqe_req(req, -ECANCELED, 0);
5749 io_commit_cqring(req->ctx);
5750 io_put_req_deferred(req);
5756 * Returns true if we found and killed one or more poll requests
5758 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5761 struct hlist_node *tmp;
5762 struct io_kiocb *req;
5765 spin_lock(&ctx->completion_lock);
5766 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5767 struct hlist_head *list;
5769 list = &ctx->cancel_hash[i];
5770 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5771 if (io_match_task_safe(req, tsk, cancel_all))
5772 posted += io_poll_remove_one(req);
5775 spin_unlock(&ctx->completion_lock);
5778 io_cqring_ev_posted(ctx);
5783 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5785 __must_hold(&ctx->completion_lock)
5787 struct hlist_head *list;
5788 struct io_kiocb *req;
5790 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5791 hlist_for_each_entry(req, list, hash_node) {
5792 if (sqe_addr != req->user_data)
5794 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5801 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5803 __must_hold(&ctx->completion_lock)
5805 struct io_kiocb *req;
5807 req = io_poll_find(ctx, sqe_addr, poll_only);
5810 if (io_poll_remove_one(req))
5816 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5821 events = READ_ONCE(sqe->poll32_events);
5823 events = swahw32(events);
5825 if (!(flags & IORING_POLL_ADD_MULTI))
5826 events |= EPOLLONESHOT;
5827 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5830 static int io_poll_update_prep(struct io_kiocb *req,
5831 const struct io_uring_sqe *sqe)
5833 struct io_poll_update *upd = &req->poll_update;
5836 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5838 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5840 flags = READ_ONCE(sqe->len);
5841 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5842 IORING_POLL_ADD_MULTI))
5844 /* meaningless without update */
5845 if (flags == IORING_POLL_ADD_MULTI)
5848 upd->old_user_data = READ_ONCE(sqe->addr);
5849 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5850 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5852 upd->new_user_data = READ_ONCE(sqe->off);
5853 if (!upd->update_user_data && upd->new_user_data)
5855 if (upd->update_events)
5856 upd->events = io_poll_parse_events(sqe, flags);
5857 else if (sqe->poll32_events)
5863 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5866 struct io_kiocb *req = wait->private;
5867 struct io_poll_iocb *poll = &req->poll;
5869 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5872 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5873 struct poll_table_struct *p)
5875 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5877 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5880 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5882 struct io_poll_iocb *poll = &req->poll;
5885 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5887 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5889 flags = READ_ONCE(sqe->len);
5890 if (flags & ~IORING_POLL_ADD_MULTI)
5893 io_req_set_refcount(req);
5894 poll->events = io_poll_parse_events(sqe, flags);
5898 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5900 struct io_poll_iocb *poll = &req->poll;
5901 struct io_ring_ctx *ctx = req->ctx;
5902 struct io_poll_table ipt;
5906 ipt.pt._qproc = io_poll_queue_proc;
5908 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5911 if (mask) { /* no async, we'd stolen it */
5913 done = io_poll_complete(req, mask);
5915 spin_unlock(&ctx->completion_lock);
5918 io_cqring_ev_posted(ctx);
5925 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5927 struct io_ring_ctx *ctx = req->ctx;
5928 struct io_kiocb *preq;
5932 spin_lock(&ctx->completion_lock);
5933 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5937 spin_unlock(&ctx->completion_lock);
5940 io_poll_remove_double(preq);
5942 * Don't allow racy completion with singleshot, as we cannot safely
5943 * update those. For multishot, if we're racing with completion, just
5944 * let completion re-add it.
5946 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5947 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5951 spin_unlock(&ctx->completion_lock);
5953 if (req->poll_update.update_events || req->poll_update.update_user_data) {
5954 /* only mask one event flags, keep behavior flags */
5955 if (req->poll_update.update_events) {
5956 preq->poll.events &= ~0xffff;
5957 preq->poll.events |= req->poll_update.events & 0xffff;
5958 preq->poll.events |= IO_POLL_UNMASK;
5960 if (req->poll_update.update_user_data)
5961 preq->user_data = req->poll_update.new_user_data;
5963 ret2 = io_poll_add(preq, issue_flags);
5964 /* successfully updated, don't complete poll request */
5969 io_req_complete(preq, -ECANCELED);
5973 /* complete update request, we're done with it */
5974 io_req_complete(req, 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_fill_cqe_req(req, -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.
6863 if (ret != -EAGAIN || !(req->ctx->flags & IORING_SETUP_IOPOLL))
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_fill_cqe_aux(ctx, prsrc->tag, 0, 0);
8263 io_commit_cqring(ctx);
8264 spin_unlock(&ctx->completion_lock);
8265 io_cqring_ev_posted(ctx);
8266 io_ring_submit_unlock(ctx, lock_ring);
8269 rsrc_data->do_put(ctx, prsrc);
8273 io_rsrc_node_destroy(ref_node);
8274 if (atomic_dec_and_test(&rsrc_data->refs))
8275 complete(&rsrc_data->done);
8278 static void io_rsrc_put_work(struct work_struct *work)
8280 struct io_ring_ctx *ctx;
8281 struct llist_node *node;
8283 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8284 node = llist_del_all(&ctx->rsrc_put_llist);
8287 struct io_rsrc_node *ref_node;
8288 struct llist_node *next = node->next;
8290 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8291 __io_rsrc_put_work(ref_node);
8296 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8297 unsigned nr_args, u64 __user *tags)
8299 __s32 __user *fds = (__s32 __user *) arg;
8308 if (nr_args > IORING_MAX_FIXED_FILES)
8310 if (nr_args > rlimit(RLIMIT_NOFILE))
8312 ret = io_rsrc_node_switch_start(ctx);
8315 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8321 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8324 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8325 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8329 /* allow sparse sets */
8332 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8339 if (unlikely(!file))
8343 * Don't allow io_uring instances to be registered. If UNIX
8344 * isn't enabled, then this causes a reference cycle and this
8345 * instance can never get freed. If UNIX is enabled we'll
8346 * handle it just fine, but there's still no point in allowing
8347 * a ring fd as it doesn't support regular read/write anyway.
8349 if (file->f_op == &io_uring_fops) {
8353 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8356 ret = io_sqe_files_scm(ctx);
8358 __io_sqe_files_unregister(ctx);
8362 io_rsrc_node_switch(ctx, NULL);
8365 for (i = 0; i < ctx->nr_user_files; i++) {
8366 file = io_file_from_index(ctx, i);
8370 io_free_file_tables(&ctx->file_table);
8371 ctx->nr_user_files = 0;
8373 io_rsrc_data_free(ctx->file_data);
8374 ctx->file_data = NULL;
8378 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8381 #if defined(CONFIG_UNIX)
8382 struct sock *sock = ctx->ring_sock->sk;
8383 struct sk_buff_head *head = &sock->sk_receive_queue;
8384 struct sk_buff *skb;
8387 * See if we can merge this file into an existing skb SCM_RIGHTS
8388 * file set. If there's no room, fall back to allocating a new skb
8389 * and filling it in.
8391 spin_lock_irq(&head->lock);
8392 skb = skb_peek(head);
8394 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8396 if (fpl->count < SCM_MAX_FD) {
8397 __skb_unlink(skb, head);
8398 spin_unlock_irq(&head->lock);
8399 fpl->fp[fpl->count] = get_file(file);
8400 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8402 spin_lock_irq(&head->lock);
8403 __skb_queue_head(head, skb);
8408 spin_unlock_irq(&head->lock);
8415 return __io_sqe_files_scm(ctx, 1, index);
8421 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8422 struct io_rsrc_node *node, void *rsrc)
8424 u64 *tag_slot = io_get_tag_slot(data, idx);
8425 struct io_rsrc_put *prsrc;
8427 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8431 prsrc->tag = *tag_slot;
8434 list_add(&prsrc->list, &node->rsrc_list);
8438 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8439 unsigned int issue_flags, u32 slot_index)
8441 struct io_ring_ctx *ctx = req->ctx;
8442 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8443 bool needs_switch = false;
8444 struct io_fixed_file *file_slot;
8447 io_ring_submit_lock(ctx, !force_nonblock);
8448 if (file->f_op == &io_uring_fops)
8451 if (!ctx->file_data)
8454 if (slot_index >= ctx->nr_user_files)
8457 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8458 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8460 if (file_slot->file_ptr) {
8461 struct file *old_file;
8463 ret = io_rsrc_node_switch_start(ctx);
8467 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8468 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8469 ctx->rsrc_node, old_file);
8472 file_slot->file_ptr = 0;
8473 needs_switch = true;
8476 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8477 io_fixed_file_set(file_slot, file);
8478 ret = io_sqe_file_register(ctx, file, slot_index);
8480 file_slot->file_ptr = 0;
8487 io_rsrc_node_switch(ctx, ctx->file_data);
8488 io_ring_submit_unlock(ctx, !force_nonblock);
8494 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8496 unsigned int offset = req->close.file_slot - 1;
8497 struct io_ring_ctx *ctx = req->ctx;
8498 struct io_fixed_file *file_slot;
8502 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8504 if (unlikely(!ctx->file_data))
8507 if (offset >= ctx->nr_user_files)
8509 ret = io_rsrc_node_switch_start(ctx);
8513 offset = array_index_nospec(offset, ctx->nr_user_files);
8514 file_slot = io_fixed_file_slot(&ctx->file_table, offset);
8516 if (!file_slot->file_ptr)
8519 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8520 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8524 file_slot->file_ptr = 0;
8525 io_rsrc_node_switch(ctx, ctx->file_data);
8528 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8532 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8533 struct io_uring_rsrc_update2 *up,
8536 u64 __user *tags = u64_to_user_ptr(up->tags);
8537 __s32 __user *fds = u64_to_user_ptr(up->data);
8538 struct io_rsrc_data *data = ctx->file_data;
8539 struct io_fixed_file *file_slot;
8543 bool needs_switch = false;
8545 if (!ctx->file_data)
8547 if (up->offset + nr_args > ctx->nr_user_files)
8550 for (done = 0; done < nr_args; done++) {
8553 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8554 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8558 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8562 if (fd == IORING_REGISTER_FILES_SKIP)
8565 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8566 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8568 if (file_slot->file_ptr) {
8569 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8570 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
8573 file_slot->file_ptr = 0;
8574 needs_switch = true;
8583 * Don't allow io_uring instances to be registered. If
8584 * UNIX isn't enabled, then this causes a reference
8585 * cycle and this instance can never get freed. If UNIX
8586 * is enabled we'll handle it just fine, but there's
8587 * still no point in allowing a ring fd as it doesn't
8588 * support regular read/write anyway.
8590 if (file->f_op == &io_uring_fops) {
8595 *io_get_tag_slot(data, i) = tag;
8596 io_fixed_file_set(file_slot, file);
8597 err = io_sqe_file_register(ctx, file, i);
8599 file_slot->file_ptr = 0;
8607 io_rsrc_node_switch(ctx, data);
8608 return done ? done : err;
8611 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8612 struct task_struct *task)
8614 struct io_wq_hash *hash;
8615 struct io_wq_data data;
8616 unsigned int concurrency;
8618 mutex_lock(&ctx->uring_lock);
8619 hash = ctx->hash_map;
8621 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8623 mutex_unlock(&ctx->uring_lock);
8624 return ERR_PTR(-ENOMEM);
8626 refcount_set(&hash->refs, 1);
8627 init_waitqueue_head(&hash->wait);
8628 ctx->hash_map = hash;
8630 mutex_unlock(&ctx->uring_lock);
8634 data.free_work = io_wq_free_work;
8635 data.do_work = io_wq_submit_work;
8637 /* Do QD, or 4 * CPUS, whatever is smallest */
8638 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8640 return io_wq_create(concurrency, &data);
8643 static int io_uring_alloc_task_context(struct task_struct *task,
8644 struct io_ring_ctx *ctx)
8646 struct io_uring_task *tctx;
8649 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8650 if (unlikely(!tctx))
8653 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8654 if (unlikely(ret)) {
8659 tctx->io_wq = io_init_wq_offload(ctx, task);
8660 if (IS_ERR(tctx->io_wq)) {
8661 ret = PTR_ERR(tctx->io_wq);
8662 percpu_counter_destroy(&tctx->inflight);
8668 init_waitqueue_head(&tctx->wait);
8669 atomic_set(&tctx->in_idle, 0);
8670 atomic_set(&tctx->inflight_tracked, 0);
8671 task->io_uring = tctx;
8672 spin_lock_init(&tctx->task_lock);
8673 INIT_WQ_LIST(&tctx->task_list);
8674 init_task_work(&tctx->task_work, tctx_task_work);
8678 void __io_uring_free(struct task_struct *tsk)
8680 struct io_uring_task *tctx = tsk->io_uring;
8682 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8683 WARN_ON_ONCE(tctx->io_wq);
8684 WARN_ON_ONCE(tctx->cached_refs);
8686 percpu_counter_destroy(&tctx->inflight);
8688 tsk->io_uring = NULL;
8691 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8692 struct io_uring_params *p)
8696 /* Retain compatibility with failing for an invalid attach attempt */
8697 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8698 IORING_SETUP_ATTACH_WQ) {
8701 f = fdget(p->wq_fd);
8704 if (f.file->f_op != &io_uring_fops) {
8710 if (ctx->flags & IORING_SETUP_SQPOLL) {
8711 struct task_struct *tsk;
8712 struct io_sq_data *sqd;
8715 sqd = io_get_sq_data(p, &attached);
8721 ctx->sq_creds = get_current_cred();
8723 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8724 if (!ctx->sq_thread_idle)
8725 ctx->sq_thread_idle = HZ;
8727 io_sq_thread_park(sqd);
8728 list_add(&ctx->sqd_list, &sqd->ctx_list);
8729 io_sqd_update_thread_idle(sqd);
8730 /* don't attach to a dying SQPOLL thread, would be racy */
8731 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8732 io_sq_thread_unpark(sqd);
8739 if (p->flags & IORING_SETUP_SQ_AFF) {
8740 int cpu = p->sq_thread_cpu;
8743 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8750 sqd->task_pid = current->pid;
8751 sqd->task_tgid = current->tgid;
8752 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8759 ret = io_uring_alloc_task_context(tsk, ctx);
8760 wake_up_new_task(tsk);
8763 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8764 /* Can't have SQ_AFF without SQPOLL */
8771 complete(&ctx->sq_data->exited);
8773 io_sq_thread_finish(ctx);
8777 static inline void __io_unaccount_mem(struct user_struct *user,
8778 unsigned long nr_pages)
8780 atomic_long_sub(nr_pages, &user->locked_vm);
8783 static inline int __io_account_mem(struct user_struct *user,
8784 unsigned long nr_pages)
8786 unsigned long page_limit, cur_pages, new_pages;
8788 /* Don't allow more pages than we can safely lock */
8789 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8792 cur_pages = atomic_long_read(&user->locked_vm);
8793 new_pages = cur_pages + nr_pages;
8794 if (new_pages > page_limit)
8796 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8797 new_pages) != cur_pages);
8802 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8805 __io_unaccount_mem(ctx->user, nr_pages);
8807 if (ctx->mm_account)
8808 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8811 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8816 ret = __io_account_mem(ctx->user, nr_pages);
8821 if (ctx->mm_account)
8822 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8827 static void io_mem_free(void *ptr)
8834 page = virt_to_head_page(ptr);
8835 if (put_page_testzero(page))
8836 free_compound_page(page);
8839 static void *io_mem_alloc(size_t size)
8841 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
8843 return (void *) __get_free_pages(gfp, get_order(size));
8846 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8849 struct io_rings *rings;
8850 size_t off, sq_array_size;
8852 off = struct_size(rings, cqes, cq_entries);
8853 if (off == SIZE_MAX)
8857 off = ALIGN(off, SMP_CACHE_BYTES);
8865 sq_array_size = array_size(sizeof(u32), sq_entries);
8866 if (sq_array_size == SIZE_MAX)
8869 if (check_add_overflow(off, sq_array_size, &off))
8875 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8877 struct io_mapped_ubuf *imu = *slot;
8880 if (imu != ctx->dummy_ubuf) {
8881 for (i = 0; i < imu->nr_bvecs; i++)
8882 unpin_user_page(imu->bvec[i].bv_page);
8883 if (imu->acct_pages)
8884 io_unaccount_mem(ctx, imu->acct_pages);
8890 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8892 io_buffer_unmap(ctx, &prsrc->buf);
8896 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8900 for (i = 0; i < ctx->nr_user_bufs; i++)
8901 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8902 kfree(ctx->user_bufs);
8903 io_rsrc_data_free(ctx->buf_data);
8904 ctx->user_bufs = NULL;
8905 ctx->buf_data = NULL;
8906 ctx->nr_user_bufs = 0;
8909 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8911 unsigned nr = ctx->nr_user_bufs;
8918 * Quiesce may unlock ->uring_lock, and while it's not held
8919 * prevent new requests using the table.
8921 ctx->nr_user_bufs = 0;
8922 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8923 ctx->nr_user_bufs = nr;
8925 __io_sqe_buffers_unregister(ctx);
8929 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8930 void __user *arg, unsigned index)
8932 struct iovec __user *src;
8934 #ifdef CONFIG_COMPAT
8936 struct compat_iovec __user *ciovs;
8937 struct compat_iovec ciov;
8939 ciovs = (struct compat_iovec __user *) arg;
8940 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8943 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8944 dst->iov_len = ciov.iov_len;
8948 src = (struct iovec __user *) arg;
8949 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8955 * Not super efficient, but this is just a registration time. And we do cache
8956 * the last compound head, so generally we'll only do a full search if we don't
8959 * We check if the given compound head page has already been accounted, to
8960 * avoid double accounting it. This allows us to account the full size of the
8961 * page, not just the constituent pages of a huge page.
8963 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8964 int nr_pages, struct page *hpage)
8968 /* check current page array */
8969 for (i = 0; i < nr_pages; i++) {
8970 if (!PageCompound(pages[i]))
8972 if (compound_head(pages[i]) == hpage)
8976 /* check previously registered pages */
8977 for (i = 0; i < ctx->nr_user_bufs; i++) {
8978 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8980 for (j = 0; j < imu->nr_bvecs; j++) {
8981 if (!PageCompound(imu->bvec[j].bv_page))
8983 if (compound_head(imu->bvec[j].bv_page) == hpage)
8991 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8992 int nr_pages, struct io_mapped_ubuf *imu,
8993 struct page **last_hpage)
8997 imu->acct_pages = 0;
8998 for (i = 0; i < nr_pages; i++) {
8999 if (!PageCompound(pages[i])) {
9004 hpage = compound_head(pages[i]);
9005 if (hpage == *last_hpage)
9007 *last_hpage = hpage;
9008 if (headpage_already_acct(ctx, pages, i, hpage))
9010 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
9014 if (!imu->acct_pages)
9017 ret = io_account_mem(ctx, imu->acct_pages);
9019 imu->acct_pages = 0;
9023 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
9024 struct io_mapped_ubuf **pimu,
9025 struct page **last_hpage)
9027 struct io_mapped_ubuf *imu = NULL;
9028 struct vm_area_struct **vmas = NULL;
9029 struct page **pages = NULL;
9030 unsigned long off, start, end, ubuf;
9032 int ret, pret, nr_pages, i;
9034 if (!iov->iov_base) {
9035 *pimu = ctx->dummy_ubuf;
9039 ubuf = (unsigned long) iov->iov_base;
9040 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9041 start = ubuf >> PAGE_SHIFT;
9042 nr_pages = end - start;
9047 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9051 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9056 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9061 mmap_read_lock(current->mm);
9062 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9064 if (pret == nr_pages) {
9065 /* don't support file backed memory */
9066 for (i = 0; i < nr_pages; i++) {
9067 struct vm_area_struct *vma = vmas[i];
9069 if (vma_is_shmem(vma))
9072 !is_file_hugepages(vma->vm_file)) {
9078 ret = pret < 0 ? pret : -EFAULT;
9080 mmap_read_unlock(current->mm);
9083 * if we did partial map, or found file backed vmas,
9084 * release any pages we did get
9087 unpin_user_pages(pages, pret);
9091 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9093 unpin_user_pages(pages, pret);
9097 off = ubuf & ~PAGE_MASK;
9098 size = iov->iov_len;
9099 for (i = 0; i < nr_pages; i++) {
9102 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9103 imu->bvec[i].bv_page = pages[i];
9104 imu->bvec[i].bv_len = vec_len;
9105 imu->bvec[i].bv_offset = off;
9109 /* store original address for later verification */
9111 imu->ubuf_end = ubuf + iov->iov_len;
9112 imu->nr_bvecs = nr_pages;
9123 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9125 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9126 return ctx->user_bufs ? 0 : -ENOMEM;
9129 static int io_buffer_validate(struct iovec *iov)
9131 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9134 * Don't impose further limits on the size and buffer
9135 * constraints here, we'll -EINVAL later when IO is
9136 * submitted if they are wrong.
9139 return iov->iov_len ? -EFAULT : 0;
9143 /* arbitrary limit, but we need something */
9144 if (iov->iov_len > SZ_1G)
9147 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9153 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9154 unsigned int nr_args, u64 __user *tags)
9156 struct page *last_hpage = NULL;
9157 struct io_rsrc_data *data;
9163 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9165 ret = io_rsrc_node_switch_start(ctx);
9168 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9171 ret = io_buffers_map_alloc(ctx, nr_args);
9173 io_rsrc_data_free(data);
9177 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9178 ret = io_copy_iov(ctx, &iov, arg, i);
9181 ret = io_buffer_validate(&iov);
9184 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9189 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9195 WARN_ON_ONCE(ctx->buf_data);
9197 ctx->buf_data = data;
9199 __io_sqe_buffers_unregister(ctx);
9201 io_rsrc_node_switch(ctx, NULL);
9205 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9206 struct io_uring_rsrc_update2 *up,
9207 unsigned int nr_args)
9209 u64 __user *tags = u64_to_user_ptr(up->tags);
9210 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9211 struct page *last_hpage = NULL;
9212 bool needs_switch = false;
9218 if (up->offset + nr_args > ctx->nr_user_bufs)
9221 for (done = 0; done < nr_args; done++) {
9222 struct io_mapped_ubuf *imu;
9223 int offset = up->offset + done;
9226 err = io_copy_iov(ctx, &iov, iovs, done);
9229 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9233 err = io_buffer_validate(&iov);
9236 if (!iov.iov_base && tag) {
9240 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9244 i = array_index_nospec(offset, ctx->nr_user_bufs);
9245 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9246 err = io_queue_rsrc_removal(ctx->buf_data, i,
9247 ctx->rsrc_node, ctx->user_bufs[i]);
9248 if (unlikely(err)) {
9249 io_buffer_unmap(ctx, &imu);
9252 ctx->user_bufs[i] = NULL;
9253 needs_switch = true;
9256 ctx->user_bufs[i] = imu;
9257 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9261 io_rsrc_node_switch(ctx, ctx->buf_data);
9262 return done ? done : err;
9265 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9267 __s32 __user *fds = arg;
9273 if (copy_from_user(&fd, fds, sizeof(*fds)))
9276 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9277 if (IS_ERR(ctx->cq_ev_fd)) {
9278 int ret = PTR_ERR(ctx->cq_ev_fd);
9280 ctx->cq_ev_fd = NULL;
9287 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9289 if (ctx->cq_ev_fd) {
9290 eventfd_ctx_put(ctx->cq_ev_fd);
9291 ctx->cq_ev_fd = NULL;
9298 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9300 struct io_buffer *buf;
9301 unsigned long index;
9303 xa_for_each(&ctx->io_buffers, index, buf)
9304 __io_remove_buffers(ctx, buf, index, -1U);
9307 static void io_req_cache_free(struct list_head *list)
9309 struct io_kiocb *req, *nxt;
9311 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9312 list_del(&req->inflight_entry);
9313 kmem_cache_free(req_cachep, req);
9317 static void io_req_caches_free(struct io_ring_ctx *ctx)
9319 struct io_submit_state *state = &ctx->submit_state;
9321 mutex_lock(&ctx->uring_lock);
9323 if (state->free_reqs) {
9324 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9325 state->free_reqs = 0;
9328 io_flush_cached_locked_reqs(ctx, state);
9329 io_req_cache_free(&state->free_list);
9330 mutex_unlock(&ctx->uring_lock);
9333 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9335 if (data && !atomic_dec_and_test(&data->refs))
9336 wait_for_completion(&data->done);
9339 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9341 io_sq_thread_finish(ctx);
9343 if (ctx->mm_account) {
9344 mmdrop(ctx->mm_account);
9345 ctx->mm_account = NULL;
9348 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9349 io_wait_rsrc_data(ctx->buf_data);
9350 io_wait_rsrc_data(ctx->file_data);
9352 mutex_lock(&ctx->uring_lock);
9354 __io_sqe_buffers_unregister(ctx);
9356 __io_sqe_files_unregister(ctx);
9358 __io_cqring_overflow_flush(ctx, true);
9359 mutex_unlock(&ctx->uring_lock);
9360 io_eventfd_unregister(ctx);
9361 io_destroy_buffers(ctx);
9363 put_cred(ctx->sq_creds);
9365 /* there are no registered resources left, nobody uses it */
9367 io_rsrc_node_destroy(ctx->rsrc_node);
9368 if (ctx->rsrc_backup_node)
9369 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9370 flush_delayed_work(&ctx->rsrc_put_work);
9372 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9373 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9375 #if defined(CONFIG_UNIX)
9376 if (ctx->ring_sock) {
9377 ctx->ring_sock->file = NULL; /* so that iput() is called */
9378 sock_release(ctx->ring_sock);
9381 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9383 io_mem_free(ctx->rings);
9384 io_mem_free(ctx->sq_sqes);
9386 percpu_ref_exit(&ctx->refs);
9387 free_uid(ctx->user);
9388 io_req_caches_free(ctx);
9390 io_wq_put_hash(ctx->hash_map);
9391 kfree(ctx->cancel_hash);
9392 kfree(ctx->dummy_ubuf);
9396 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9398 struct io_ring_ctx *ctx = file->private_data;
9401 poll_wait(file, &ctx->poll_wait, wait);
9403 * synchronizes with barrier from wq_has_sleeper call in
9407 if (!io_sqring_full(ctx))
9408 mask |= EPOLLOUT | EPOLLWRNORM;
9411 * Don't flush cqring overflow list here, just do a simple check.
9412 * Otherwise there could possible be ABBA deadlock:
9415 * lock(&ctx->uring_lock);
9417 * lock(&ctx->uring_lock);
9420 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9421 * pushs them to do the flush.
9423 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9424 mask |= EPOLLIN | EPOLLRDNORM;
9429 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9431 const struct cred *creds;
9433 creds = xa_erase(&ctx->personalities, id);
9442 struct io_tctx_exit {
9443 struct callback_head task_work;
9444 struct completion completion;
9445 struct io_ring_ctx *ctx;
9448 static void io_tctx_exit_cb(struct callback_head *cb)
9450 struct io_uring_task *tctx = current->io_uring;
9451 struct io_tctx_exit *work;
9453 work = container_of(cb, struct io_tctx_exit, task_work);
9455 * When @in_idle, we're in cancellation and it's racy to remove the
9456 * node. It'll be removed by the end of cancellation, just ignore it.
9458 if (!atomic_read(&tctx->in_idle))
9459 io_uring_del_tctx_node((unsigned long)work->ctx);
9460 complete(&work->completion);
9463 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9465 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9467 return req->ctx == data;
9470 static void io_ring_exit_work(struct work_struct *work)
9472 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9473 unsigned long timeout = jiffies + HZ * 60 * 5;
9474 unsigned long interval = HZ / 20;
9475 struct io_tctx_exit exit;
9476 struct io_tctx_node *node;
9480 * If we're doing polled IO and end up having requests being
9481 * submitted async (out-of-line), then completions can come in while
9482 * we're waiting for refs to drop. We need to reap these manually,
9483 * as nobody else will be looking for them.
9486 io_uring_try_cancel_requests(ctx, NULL, true);
9488 struct io_sq_data *sqd = ctx->sq_data;
9489 struct task_struct *tsk;
9491 io_sq_thread_park(sqd);
9493 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9494 io_wq_cancel_cb(tsk->io_uring->io_wq,
9495 io_cancel_ctx_cb, ctx, true);
9496 io_sq_thread_unpark(sqd);
9499 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9500 /* there is little hope left, don't run it too often */
9503 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9505 init_completion(&exit.completion);
9506 init_task_work(&exit.task_work, io_tctx_exit_cb);
9509 * Some may use context even when all refs and requests have been put,
9510 * and they are free to do so while still holding uring_lock or
9511 * completion_lock, see io_req_task_submit(). Apart from other work,
9512 * this lock/unlock section also waits them to finish.
9514 mutex_lock(&ctx->uring_lock);
9515 while (!list_empty(&ctx->tctx_list)) {
9516 WARN_ON_ONCE(time_after(jiffies, timeout));
9518 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9520 /* don't spin on a single task if cancellation failed */
9521 list_rotate_left(&ctx->tctx_list);
9522 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9523 if (WARN_ON_ONCE(ret))
9525 wake_up_process(node->task);
9527 mutex_unlock(&ctx->uring_lock);
9528 wait_for_completion(&exit.completion);
9529 mutex_lock(&ctx->uring_lock);
9531 mutex_unlock(&ctx->uring_lock);
9532 spin_lock(&ctx->completion_lock);
9533 spin_unlock(&ctx->completion_lock);
9535 io_ring_ctx_free(ctx);
9538 /* Returns true if we found and killed one or more timeouts */
9539 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9542 struct io_kiocb *req, *tmp;
9545 spin_lock(&ctx->completion_lock);
9546 spin_lock_irq(&ctx->timeout_lock);
9547 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9548 if (io_match_task(req, tsk, cancel_all)) {
9549 io_kill_timeout(req, -ECANCELED);
9553 spin_unlock_irq(&ctx->timeout_lock);
9555 io_commit_cqring(ctx);
9556 spin_unlock(&ctx->completion_lock);
9558 io_cqring_ev_posted(ctx);
9559 return canceled != 0;
9562 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9564 unsigned long index;
9565 struct creds *creds;
9567 mutex_lock(&ctx->uring_lock);
9568 percpu_ref_kill(&ctx->refs);
9570 __io_cqring_overflow_flush(ctx, true);
9571 xa_for_each(&ctx->personalities, index, creds)
9572 io_unregister_personality(ctx, index);
9573 mutex_unlock(&ctx->uring_lock);
9575 io_kill_timeouts(ctx, NULL, true);
9576 io_poll_remove_all(ctx, NULL, true);
9578 /* if we failed setting up the ctx, we might not have any rings */
9579 io_iopoll_try_reap_events(ctx);
9581 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9583 * Use system_unbound_wq to avoid spawning tons of event kworkers
9584 * if we're exiting a ton of rings at the same time. It just adds
9585 * noise and overhead, there's no discernable change in runtime
9586 * over using system_wq.
9588 queue_work(system_unbound_wq, &ctx->exit_work);
9591 static int io_uring_release(struct inode *inode, struct file *file)
9593 struct io_ring_ctx *ctx = file->private_data;
9595 file->private_data = NULL;
9596 io_ring_ctx_wait_and_kill(ctx);
9600 struct io_task_cancel {
9601 struct task_struct *task;
9605 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9607 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9608 struct io_task_cancel *cancel = data;
9610 return io_match_task_safe(req, cancel->task, cancel->all);
9613 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9614 struct task_struct *task, bool cancel_all)
9616 struct io_defer_entry *de;
9619 spin_lock(&ctx->completion_lock);
9620 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9621 if (io_match_task_safe(de->req, task, cancel_all)) {
9622 list_cut_position(&list, &ctx->defer_list, &de->list);
9626 spin_unlock(&ctx->completion_lock);
9627 if (list_empty(&list))
9630 while (!list_empty(&list)) {
9631 de = list_first_entry(&list, struct io_defer_entry, list);
9632 list_del_init(&de->list);
9633 io_req_complete_failed(de->req, -ECANCELED);
9639 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9641 struct io_tctx_node *node;
9642 enum io_wq_cancel cret;
9645 mutex_lock(&ctx->uring_lock);
9646 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9647 struct io_uring_task *tctx = node->task->io_uring;
9650 * io_wq will stay alive while we hold uring_lock, because it's
9651 * killed after ctx nodes, which requires to take the lock.
9653 if (!tctx || !tctx->io_wq)
9655 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9656 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9658 mutex_unlock(&ctx->uring_lock);
9663 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9664 struct task_struct *task,
9667 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9668 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9671 enum io_wq_cancel cret;
9675 ret |= io_uring_try_cancel_iowq(ctx);
9676 } else if (tctx && tctx->io_wq) {
9678 * Cancels requests of all rings, not only @ctx, but
9679 * it's fine as the task is in exit/exec.
9681 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9683 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9686 /* SQPOLL thread does its own polling */
9687 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9688 (ctx->sq_data && ctx->sq_data->thread == current)) {
9689 while (!list_empty_careful(&ctx->iopoll_list)) {
9690 io_iopoll_try_reap_events(ctx);
9695 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9696 ret |= io_poll_remove_all(ctx, task, cancel_all);
9697 ret |= io_kill_timeouts(ctx, task, cancel_all);
9699 ret |= io_run_task_work();
9706 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9708 struct io_uring_task *tctx = current->io_uring;
9709 struct io_tctx_node *node;
9712 if (unlikely(!tctx)) {
9713 ret = io_uring_alloc_task_context(current, ctx);
9717 tctx = current->io_uring;
9718 if (ctx->iowq_limits_set) {
9719 unsigned int limits[2] = { ctx->iowq_limits[0],
9720 ctx->iowq_limits[1], };
9722 ret = io_wq_max_workers(tctx->io_wq, limits);
9727 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9728 node = kmalloc(sizeof(*node), GFP_KERNEL);
9732 node->task = current;
9734 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9741 mutex_lock(&ctx->uring_lock);
9742 list_add(&node->ctx_node, &ctx->tctx_list);
9743 mutex_unlock(&ctx->uring_lock);
9750 * Note that this task has used io_uring. We use it for cancelation purposes.
9752 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9754 struct io_uring_task *tctx = current->io_uring;
9756 if (likely(tctx && tctx->last == ctx))
9758 return __io_uring_add_tctx_node(ctx);
9762 * Remove this io_uring_file -> task mapping.
9764 static void io_uring_del_tctx_node(unsigned long index)
9766 struct io_uring_task *tctx = current->io_uring;
9767 struct io_tctx_node *node;
9771 node = xa_erase(&tctx->xa, index);
9775 WARN_ON_ONCE(current != node->task);
9776 WARN_ON_ONCE(list_empty(&node->ctx_node));
9778 mutex_lock(&node->ctx->uring_lock);
9779 list_del(&node->ctx_node);
9780 mutex_unlock(&node->ctx->uring_lock);
9782 if (tctx->last == node->ctx)
9787 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9789 struct io_wq *wq = tctx->io_wq;
9790 struct io_tctx_node *node;
9791 unsigned long index;
9793 xa_for_each(&tctx->xa, index, node) {
9794 io_uring_del_tctx_node(index);
9799 * Must be after io_uring_del_task_file() (removes nodes under
9800 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9802 io_wq_put_and_exit(wq);
9807 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9810 return atomic_read(&tctx->inflight_tracked);
9811 return percpu_counter_sum(&tctx->inflight);
9815 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9816 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9818 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9820 struct io_uring_task *tctx = current->io_uring;
9821 struct io_ring_ctx *ctx;
9825 WARN_ON_ONCE(sqd && sqd->thread != current);
9827 if (!current->io_uring)
9830 io_wq_exit_start(tctx->io_wq);
9832 atomic_inc(&tctx->in_idle);
9834 io_uring_drop_tctx_refs(current);
9835 /* read completions before cancelations */
9836 inflight = tctx_inflight(tctx, !cancel_all);
9841 struct io_tctx_node *node;
9842 unsigned long index;
9844 xa_for_each(&tctx->xa, index, node) {
9845 /* sqpoll task will cancel all its requests */
9846 if (node->ctx->sq_data)
9848 io_uring_try_cancel_requests(node->ctx, current,
9852 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9853 io_uring_try_cancel_requests(ctx, current,
9857 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9859 io_uring_drop_tctx_refs(current);
9862 * If we've seen completions, retry without waiting. This
9863 * avoids a race where a completion comes in before we did
9864 * prepare_to_wait().
9866 if (inflight == tctx_inflight(tctx, !cancel_all))
9868 finish_wait(&tctx->wait, &wait);
9871 io_uring_clean_tctx(tctx);
9874 * We shouldn't run task_works after cancel, so just leave
9875 * ->in_idle set for normal exit.
9877 atomic_dec(&tctx->in_idle);
9878 /* for exec all current's requests should be gone, kill tctx */
9879 __io_uring_free(current);
9883 void __io_uring_cancel(bool cancel_all)
9885 io_uring_cancel_generic(cancel_all, NULL);
9888 static void *io_uring_validate_mmap_request(struct file *file,
9889 loff_t pgoff, size_t sz)
9891 struct io_ring_ctx *ctx = file->private_data;
9892 loff_t offset = pgoff << PAGE_SHIFT;
9897 case IORING_OFF_SQ_RING:
9898 case IORING_OFF_CQ_RING:
9901 case IORING_OFF_SQES:
9905 return ERR_PTR(-EINVAL);
9908 page = virt_to_head_page(ptr);
9909 if (sz > page_size(page))
9910 return ERR_PTR(-EINVAL);
9917 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9919 size_t sz = vma->vm_end - vma->vm_start;
9923 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9925 return PTR_ERR(ptr);
9927 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9928 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9931 #else /* !CONFIG_MMU */
9933 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9935 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9938 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9940 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9943 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9944 unsigned long addr, unsigned long len,
9945 unsigned long pgoff, unsigned long flags)
9949 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9951 return PTR_ERR(ptr);
9953 return (unsigned long) ptr;
9956 #endif /* !CONFIG_MMU */
9958 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9963 if (!io_sqring_full(ctx))
9965 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9967 if (!io_sqring_full(ctx))
9970 } while (!signal_pending(current));
9972 finish_wait(&ctx->sqo_sq_wait, &wait);
9976 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9977 struct __kernel_timespec __user **ts,
9978 const sigset_t __user **sig)
9980 struct io_uring_getevents_arg arg;
9983 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9984 * is just a pointer to the sigset_t.
9986 if (!(flags & IORING_ENTER_EXT_ARG)) {
9987 *sig = (const sigset_t __user *) argp;
9993 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9994 * timespec and sigset_t pointers if good.
9996 if (*argsz != sizeof(arg))
9998 if (copy_from_user(&arg, argp, sizeof(arg)))
10002 *sig = u64_to_user_ptr(arg.sigmask);
10003 *argsz = arg.sigmask_sz;
10004 *ts = u64_to_user_ptr(arg.ts);
10008 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
10009 u32, min_complete, u32, flags, const void __user *, argp,
10012 struct io_ring_ctx *ctx;
10017 io_run_task_work();
10019 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
10020 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
10024 if (unlikely(!f.file))
10028 if (unlikely(f.file->f_op != &io_uring_fops))
10032 ctx = f.file->private_data;
10033 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
10037 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10041 * For SQ polling, the thread will do all submissions and completions.
10042 * Just return the requested submit count, and wake the thread if
10043 * we were asked to.
10046 if (ctx->flags & IORING_SETUP_SQPOLL) {
10047 io_cqring_overflow_flush(ctx);
10049 if (unlikely(ctx->sq_data->thread == NULL)) {
10053 if (flags & IORING_ENTER_SQ_WAKEUP)
10054 wake_up(&ctx->sq_data->wait);
10055 if (flags & IORING_ENTER_SQ_WAIT) {
10056 ret = io_sqpoll_wait_sq(ctx);
10060 submitted = to_submit;
10061 } else if (to_submit) {
10062 ret = io_uring_add_tctx_node(ctx);
10065 mutex_lock(&ctx->uring_lock);
10066 submitted = io_submit_sqes(ctx, to_submit);
10067 mutex_unlock(&ctx->uring_lock);
10069 if (submitted != to_submit)
10072 if (flags & IORING_ENTER_GETEVENTS) {
10073 const sigset_t __user *sig;
10074 struct __kernel_timespec __user *ts;
10076 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10080 min_complete = min(min_complete, ctx->cq_entries);
10083 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10084 * space applications don't need to do io completion events
10085 * polling again, they can rely on io_sq_thread to do polling
10086 * work, which can reduce cpu usage and uring_lock contention.
10088 if (ctx->flags & IORING_SETUP_IOPOLL &&
10089 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10090 ret = io_iopoll_check(ctx, min_complete);
10092 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10097 percpu_ref_put(&ctx->refs);
10100 return submitted ? submitted : ret;
10103 #ifdef CONFIG_PROC_FS
10104 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10105 const struct cred *cred)
10107 struct user_namespace *uns = seq_user_ns(m);
10108 struct group_info *gi;
10113 seq_printf(m, "%5d\n", id);
10114 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10115 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10116 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10117 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10118 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10119 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10120 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10121 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10122 seq_puts(m, "\n\tGroups:\t");
10123 gi = cred->group_info;
10124 for (g = 0; g < gi->ngroups; g++) {
10125 seq_put_decimal_ull(m, g ? " " : "",
10126 from_kgid_munged(uns, gi->gid[g]));
10128 seq_puts(m, "\n\tCapEff:\t");
10129 cap = cred->cap_effective;
10130 CAP_FOR_EACH_U32(__capi)
10131 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10136 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10138 struct io_sq_data *sq = NULL;
10143 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10144 * since fdinfo case grabs it in the opposite direction of normal use
10145 * cases. If we fail to get the lock, we just don't iterate any
10146 * structures that could be going away outside the io_uring mutex.
10148 has_lock = mutex_trylock(&ctx->uring_lock);
10150 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10156 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10157 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10158 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10159 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10160 struct file *f = io_file_from_index(ctx, i);
10163 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10165 seq_printf(m, "%5u: <none>\n", i);
10167 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10168 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10169 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10170 unsigned int len = buf->ubuf_end - buf->ubuf;
10172 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10174 if (has_lock && !xa_empty(&ctx->personalities)) {
10175 unsigned long index;
10176 const struct cred *cred;
10178 seq_printf(m, "Personalities:\n");
10179 xa_for_each(&ctx->personalities, index, cred)
10180 io_uring_show_cred(m, index, cred);
10182 seq_printf(m, "PollList:\n");
10183 spin_lock(&ctx->completion_lock);
10184 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10185 struct hlist_head *list = &ctx->cancel_hash[i];
10186 struct io_kiocb *req;
10188 hlist_for_each_entry(req, list, hash_node)
10189 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10190 req->task->task_works != NULL);
10192 spin_unlock(&ctx->completion_lock);
10194 mutex_unlock(&ctx->uring_lock);
10197 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10199 struct io_ring_ctx *ctx = f->private_data;
10201 if (percpu_ref_tryget(&ctx->refs)) {
10202 __io_uring_show_fdinfo(ctx, m);
10203 percpu_ref_put(&ctx->refs);
10208 static const struct file_operations io_uring_fops = {
10209 .release = io_uring_release,
10210 .mmap = io_uring_mmap,
10212 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10213 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10215 .poll = io_uring_poll,
10216 #ifdef CONFIG_PROC_FS
10217 .show_fdinfo = io_uring_show_fdinfo,
10221 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10222 struct io_uring_params *p)
10224 struct io_rings *rings;
10225 size_t size, sq_array_offset;
10227 /* make sure these are sane, as we already accounted them */
10228 ctx->sq_entries = p->sq_entries;
10229 ctx->cq_entries = p->cq_entries;
10231 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10232 if (size == SIZE_MAX)
10235 rings = io_mem_alloc(size);
10239 ctx->rings = rings;
10240 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10241 rings->sq_ring_mask = p->sq_entries - 1;
10242 rings->cq_ring_mask = p->cq_entries - 1;
10243 rings->sq_ring_entries = p->sq_entries;
10244 rings->cq_ring_entries = p->cq_entries;
10246 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10247 if (size == SIZE_MAX) {
10248 io_mem_free(ctx->rings);
10253 ctx->sq_sqes = io_mem_alloc(size);
10254 if (!ctx->sq_sqes) {
10255 io_mem_free(ctx->rings);
10263 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10267 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10271 ret = io_uring_add_tctx_node(ctx);
10276 fd_install(fd, file);
10281 * Allocate an anonymous fd, this is what constitutes the application
10282 * visible backing of an io_uring instance. The application mmaps this
10283 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10284 * we have to tie this fd to a socket for file garbage collection purposes.
10286 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10289 #if defined(CONFIG_UNIX)
10292 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10295 return ERR_PTR(ret);
10298 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10299 O_RDWR | O_CLOEXEC);
10300 #if defined(CONFIG_UNIX)
10301 if (IS_ERR(file)) {
10302 sock_release(ctx->ring_sock);
10303 ctx->ring_sock = NULL;
10305 ctx->ring_sock->file = file;
10311 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10312 struct io_uring_params __user *params)
10314 struct io_ring_ctx *ctx;
10320 if (entries > IORING_MAX_ENTRIES) {
10321 if (!(p->flags & IORING_SETUP_CLAMP))
10323 entries = IORING_MAX_ENTRIES;
10327 * Use twice as many entries for the CQ ring. It's possible for the
10328 * application to drive a higher depth than the size of the SQ ring,
10329 * since the sqes are only used at submission time. This allows for
10330 * some flexibility in overcommitting a bit. If the application has
10331 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10332 * of CQ ring entries manually.
10334 p->sq_entries = roundup_pow_of_two(entries);
10335 if (p->flags & IORING_SETUP_CQSIZE) {
10337 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10338 * to a power-of-two, if it isn't already. We do NOT impose
10339 * any cq vs sq ring sizing.
10341 if (!p->cq_entries)
10343 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10344 if (!(p->flags & IORING_SETUP_CLAMP))
10346 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10348 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10349 if (p->cq_entries < p->sq_entries)
10352 p->cq_entries = 2 * p->sq_entries;
10355 ctx = io_ring_ctx_alloc(p);
10358 ctx->compat = in_compat_syscall();
10359 if (!capable(CAP_IPC_LOCK))
10360 ctx->user = get_uid(current_user());
10363 * This is just grabbed for accounting purposes. When a process exits,
10364 * the mm is exited and dropped before the files, hence we need to hang
10365 * on to this mm purely for the purposes of being able to unaccount
10366 * memory (locked/pinned vm). It's not used for anything else.
10368 mmgrab(current->mm);
10369 ctx->mm_account = current->mm;
10371 ret = io_allocate_scq_urings(ctx, p);
10375 ret = io_sq_offload_create(ctx, p);
10378 /* always set a rsrc node */
10379 ret = io_rsrc_node_switch_start(ctx);
10382 io_rsrc_node_switch(ctx, NULL);
10384 memset(&p->sq_off, 0, sizeof(p->sq_off));
10385 p->sq_off.head = offsetof(struct io_rings, sq.head);
10386 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10387 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10388 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10389 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10390 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10391 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10393 memset(&p->cq_off, 0, sizeof(p->cq_off));
10394 p->cq_off.head = offsetof(struct io_rings, cq.head);
10395 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10396 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10397 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10398 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10399 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10400 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10402 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10403 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10404 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10405 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10406 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10407 IORING_FEAT_RSRC_TAGS;
10409 if (copy_to_user(params, p, sizeof(*p))) {
10414 file = io_uring_get_file(ctx);
10415 if (IS_ERR(file)) {
10416 ret = PTR_ERR(file);
10421 * Install ring fd as the very last thing, so we don't risk someone
10422 * having closed it before we finish setup
10424 ret = io_uring_install_fd(ctx, file);
10426 /* fput will clean it up */
10431 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10434 io_ring_ctx_wait_and_kill(ctx);
10439 * Sets up an aio uring context, and returns the fd. Applications asks for a
10440 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10441 * params structure passed in.
10443 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10445 struct io_uring_params p;
10448 if (copy_from_user(&p, params, sizeof(p)))
10450 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10455 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10456 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10457 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10458 IORING_SETUP_R_DISABLED))
10461 return io_uring_create(entries, &p, params);
10464 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10465 struct io_uring_params __user *, params)
10467 return io_uring_setup(entries, params);
10470 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10472 struct io_uring_probe *p;
10476 size = struct_size(p, ops, nr_args);
10477 if (size == SIZE_MAX)
10479 p = kzalloc(size, GFP_KERNEL);
10484 if (copy_from_user(p, arg, size))
10487 if (memchr_inv(p, 0, size))
10490 p->last_op = IORING_OP_LAST - 1;
10491 if (nr_args > IORING_OP_LAST)
10492 nr_args = IORING_OP_LAST;
10494 for (i = 0; i < nr_args; i++) {
10496 if (!io_op_defs[i].not_supported)
10497 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10502 if (copy_to_user(arg, p, size))
10509 static int io_register_personality(struct io_ring_ctx *ctx)
10511 const struct cred *creds;
10515 creds = get_current_cred();
10517 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10518 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10526 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10527 unsigned int nr_args)
10529 struct io_uring_restriction *res;
10533 /* Restrictions allowed only if rings started disabled */
10534 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10537 /* We allow only a single restrictions registration */
10538 if (ctx->restrictions.registered)
10541 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10544 size = array_size(nr_args, sizeof(*res));
10545 if (size == SIZE_MAX)
10548 res = memdup_user(arg, size);
10550 return PTR_ERR(res);
10554 for (i = 0; i < nr_args; i++) {
10555 switch (res[i].opcode) {
10556 case IORING_RESTRICTION_REGISTER_OP:
10557 if (res[i].register_op >= IORING_REGISTER_LAST) {
10562 __set_bit(res[i].register_op,
10563 ctx->restrictions.register_op);
10565 case IORING_RESTRICTION_SQE_OP:
10566 if (res[i].sqe_op >= IORING_OP_LAST) {
10571 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10573 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10574 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10576 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10577 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10586 /* Reset all restrictions if an error happened */
10588 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10590 ctx->restrictions.registered = true;
10596 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10598 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10601 if (ctx->restrictions.registered)
10602 ctx->restricted = 1;
10604 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10605 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10606 wake_up(&ctx->sq_data->wait);
10610 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10611 struct io_uring_rsrc_update2 *up,
10617 if (check_add_overflow(up->offset, nr_args, &tmp))
10619 err = io_rsrc_node_switch_start(ctx);
10624 case IORING_RSRC_FILE:
10625 return __io_sqe_files_update(ctx, up, nr_args);
10626 case IORING_RSRC_BUFFER:
10627 return __io_sqe_buffers_update(ctx, up, nr_args);
10632 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10635 struct io_uring_rsrc_update2 up;
10639 memset(&up, 0, sizeof(up));
10640 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10642 if (up.resv || up.resv2)
10644 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10647 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10648 unsigned size, unsigned type)
10650 struct io_uring_rsrc_update2 up;
10652 if (size != sizeof(up))
10654 if (copy_from_user(&up, arg, sizeof(up)))
10656 if (!up.nr || up.resv || up.resv2)
10658 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10661 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10662 unsigned int size, unsigned int type)
10664 struct io_uring_rsrc_register rr;
10666 /* keep it extendible */
10667 if (size != sizeof(rr))
10670 memset(&rr, 0, sizeof(rr));
10671 if (copy_from_user(&rr, arg, size))
10673 if (!rr.nr || rr.resv || rr.resv2)
10677 case IORING_RSRC_FILE:
10678 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10679 rr.nr, u64_to_user_ptr(rr.tags));
10680 case IORING_RSRC_BUFFER:
10681 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10682 rr.nr, u64_to_user_ptr(rr.tags));
10687 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10690 struct io_uring_task *tctx = current->io_uring;
10691 cpumask_var_t new_mask;
10694 if (!tctx || !tctx->io_wq)
10697 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10700 cpumask_clear(new_mask);
10701 if (len > cpumask_size())
10702 len = cpumask_size();
10704 if (in_compat_syscall()) {
10705 ret = compat_get_bitmap(cpumask_bits(new_mask),
10706 (const compat_ulong_t __user *)arg,
10707 len * 8 /* CHAR_BIT */);
10709 ret = copy_from_user(new_mask, arg, len);
10713 free_cpumask_var(new_mask);
10717 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10718 free_cpumask_var(new_mask);
10722 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10724 struct io_uring_task *tctx = current->io_uring;
10726 if (!tctx || !tctx->io_wq)
10729 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10732 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10734 __must_hold(&ctx->uring_lock)
10736 struct io_tctx_node *node;
10737 struct io_uring_task *tctx = NULL;
10738 struct io_sq_data *sqd = NULL;
10739 __u32 new_count[2];
10742 if (copy_from_user(new_count, arg, sizeof(new_count)))
10744 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10745 if (new_count[i] > INT_MAX)
10748 if (ctx->flags & IORING_SETUP_SQPOLL) {
10749 sqd = ctx->sq_data;
10752 * Observe the correct sqd->lock -> ctx->uring_lock
10753 * ordering. Fine to drop uring_lock here, we hold
10754 * a ref to the ctx.
10756 refcount_inc(&sqd->refs);
10757 mutex_unlock(&ctx->uring_lock);
10758 mutex_lock(&sqd->lock);
10759 mutex_lock(&ctx->uring_lock);
10761 tctx = sqd->thread->io_uring;
10764 tctx = current->io_uring;
10767 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10769 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10771 ctx->iowq_limits[i] = new_count[i];
10772 ctx->iowq_limits_set = true;
10775 if (tctx && tctx->io_wq) {
10776 ret = io_wq_max_workers(tctx->io_wq, new_count);
10780 memset(new_count, 0, sizeof(new_count));
10784 mutex_unlock(&sqd->lock);
10785 io_put_sq_data(sqd);
10788 if (copy_to_user(arg, new_count, sizeof(new_count)))
10791 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10795 /* now propagate the restriction to all registered users */
10796 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10797 struct io_uring_task *tctx = node->task->io_uring;
10799 if (WARN_ON_ONCE(!tctx->io_wq))
10802 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10803 new_count[i] = ctx->iowq_limits[i];
10804 /* ignore errors, it always returns zero anyway */
10805 (void)io_wq_max_workers(tctx->io_wq, new_count);
10810 mutex_unlock(&sqd->lock);
10811 io_put_sq_data(sqd);
10816 static bool io_register_op_must_quiesce(int op)
10819 case IORING_REGISTER_BUFFERS:
10820 case IORING_UNREGISTER_BUFFERS:
10821 case IORING_REGISTER_FILES:
10822 case IORING_UNREGISTER_FILES:
10823 case IORING_REGISTER_FILES_UPDATE:
10824 case IORING_REGISTER_PROBE:
10825 case IORING_REGISTER_PERSONALITY:
10826 case IORING_UNREGISTER_PERSONALITY:
10827 case IORING_REGISTER_FILES2:
10828 case IORING_REGISTER_FILES_UPDATE2:
10829 case IORING_REGISTER_BUFFERS2:
10830 case IORING_REGISTER_BUFFERS_UPDATE:
10831 case IORING_REGISTER_IOWQ_AFF:
10832 case IORING_UNREGISTER_IOWQ_AFF:
10833 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10840 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10844 percpu_ref_kill(&ctx->refs);
10847 * Drop uring mutex before waiting for references to exit. If another
10848 * thread is currently inside io_uring_enter() it might need to grab the
10849 * uring_lock to make progress. If we hold it here across the drain
10850 * wait, then we can deadlock. It's safe to drop the mutex here, since
10851 * no new references will come in after we've killed the percpu ref.
10853 mutex_unlock(&ctx->uring_lock);
10855 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10858 ret = io_run_task_work_sig();
10859 } while (ret >= 0);
10860 mutex_lock(&ctx->uring_lock);
10863 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10867 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10868 void __user *arg, unsigned nr_args)
10869 __releases(ctx->uring_lock)
10870 __acquires(ctx->uring_lock)
10875 * We're inside the ring mutex, if the ref is already dying, then
10876 * someone else killed the ctx or is already going through
10877 * io_uring_register().
10879 if (percpu_ref_is_dying(&ctx->refs))
10882 if (ctx->restricted) {
10883 if (opcode >= IORING_REGISTER_LAST)
10885 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10886 if (!test_bit(opcode, ctx->restrictions.register_op))
10890 if (io_register_op_must_quiesce(opcode)) {
10891 ret = io_ctx_quiesce(ctx);
10897 case IORING_REGISTER_BUFFERS:
10898 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10900 case IORING_UNREGISTER_BUFFERS:
10902 if (arg || nr_args)
10904 ret = io_sqe_buffers_unregister(ctx);
10906 case IORING_REGISTER_FILES:
10907 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10909 case IORING_UNREGISTER_FILES:
10911 if (arg || nr_args)
10913 ret = io_sqe_files_unregister(ctx);
10915 case IORING_REGISTER_FILES_UPDATE:
10916 ret = io_register_files_update(ctx, arg, nr_args);
10918 case IORING_REGISTER_EVENTFD:
10919 case IORING_REGISTER_EVENTFD_ASYNC:
10923 ret = io_eventfd_register(ctx, arg);
10926 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10927 ctx->eventfd_async = 1;
10929 ctx->eventfd_async = 0;
10931 case IORING_UNREGISTER_EVENTFD:
10933 if (arg || nr_args)
10935 ret = io_eventfd_unregister(ctx);
10937 case IORING_REGISTER_PROBE:
10939 if (!arg || nr_args > 256)
10941 ret = io_probe(ctx, arg, nr_args);
10943 case IORING_REGISTER_PERSONALITY:
10945 if (arg || nr_args)
10947 ret = io_register_personality(ctx);
10949 case IORING_UNREGISTER_PERSONALITY:
10953 ret = io_unregister_personality(ctx, nr_args);
10955 case IORING_REGISTER_ENABLE_RINGS:
10957 if (arg || nr_args)
10959 ret = io_register_enable_rings(ctx);
10961 case IORING_REGISTER_RESTRICTIONS:
10962 ret = io_register_restrictions(ctx, arg, nr_args);
10964 case IORING_REGISTER_FILES2:
10965 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10967 case IORING_REGISTER_FILES_UPDATE2:
10968 ret = io_register_rsrc_update(ctx, arg, nr_args,
10971 case IORING_REGISTER_BUFFERS2:
10972 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10974 case IORING_REGISTER_BUFFERS_UPDATE:
10975 ret = io_register_rsrc_update(ctx, arg, nr_args,
10976 IORING_RSRC_BUFFER);
10978 case IORING_REGISTER_IOWQ_AFF:
10980 if (!arg || !nr_args)
10982 ret = io_register_iowq_aff(ctx, arg, nr_args);
10984 case IORING_UNREGISTER_IOWQ_AFF:
10986 if (arg || nr_args)
10988 ret = io_unregister_iowq_aff(ctx);
10990 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10992 if (!arg || nr_args != 2)
10994 ret = io_register_iowq_max_workers(ctx, arg);
11001 if (io_register_op_must_quiesce(opcode)) {
11002 /* bring the ctx back to life */
11003 percpu_ref_reinit(&ctx->refs);
11004 reinit_completion(&ctx->ref_comp);
11009 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
11010 void __user *, arg, unsigned int, nr_args)
11012 struct io_ring_ctx *ctx;
11021 if (f.file->f_op != &io_uring_fops)
11024 ctx = f.file->private_data;
11026 io_run_task_work();
11028 mutex_lock(&ctx->uring_lock);
11029 ret = __io_uring_register(ctx, opcode, arg, nr_args);
11030 mutex_unlock(&ctx->uring_lock);
11031 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
11032 ctx->cq_ev_fd != NULL, ret);
11038 static int __init io_uring_init(void)
11040 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11041 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11042 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11045 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11046 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11047 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11048 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11049 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11050 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11051 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11052 BUILD_BUG_SQE_ELEM(8, __u64, off);
11053 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11054 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11055 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11056 BUILD_BUG_SQE_ELEM(24, __u32, len);
11057 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11058 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11059 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11060 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11061 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11062 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11063 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11064 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11065 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11066 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11067 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11068 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11069 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11070 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11071 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11072 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11073 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11074 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11075 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11076 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11077 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11079 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11080 sizeof(struct io_uring_rsrc_update));
11081 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11082 sizeof(struct io_uring_rsrc_update2));
11084 /* ->buf_index is u16 */
11085 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11087 /* should fit into one byte */
11088 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11090 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11091 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11093 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11097 __initcall(io_uring_init);