1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head ltimeout_list;
379 struct list_head cq_overflow_list;
380 struct xarray io_buffers;
381 struct xarray personalities;
383 unsigned sq_thread_idle;
384 } ____cacheline_aligned_in_smp;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list;
388 unsigned int locked_free_nr;
390 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
396 unsigned long check_cq_overflow;
399 unsigned cached_cq_tail;
401 struct eventfd_ctx *cq_ev_fd;
402 struct wait_queue_head poll_wait;
403 struct wait_queue_head cq_wait;
405 atomic_t cq_timeouts;
406 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
460 bool iowq_limits_set;
464 struct io_uring_task {
465 /* submission side */
468 struct wait_queue_head wait;
469 const struct io_ring_ctx *last;
471 struct percpu_counter inflight;
472 atomic_t inflight_tracked;
475 spinlock_t task_lock;
476 struct io_wq_work_list task_list;
477 struct callback_head task_work;
482 * First field must be the file pointer in all the
483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
485 struct io_poll_iocb {
487 struct wait_queue_head *head;
491 struct wait_queue_entry wait;
494 struct io_poll_update {
500 bool update_user_data;
509 struct io_timeout_data {
510 struct io_kiocb *req;
511 struct hrtimer timer;
512 struct timespec64 ts;
513 enum hrtimer_mode mode;
519 struct sockaddr __user *addr;
520 int __user *addr_len;
523 unsigned long nofile;
543 struct list_head list;
544 /* head of the link, used by linked timeouts only */
545 struct io_kiocb *head;
546 /* for linked completions */
547 struct io_kiocb *prev;
550 struct io_timeout_rem {
555 struct timespec64 ts;
561 /* NOTE: kiocb has the file as the first member, so don't do it here */
569 struct sockaddr __user *addr;
576 struct compat_msghdr __user *umsg_compat;
577 struct user_msghdr __user *umsg;
583 struct io_buffer *kbuf;
590 struct filename *filename;
592 unsigned long nofile;
595 struct io_rsrc_update {
621 struct epoll_event event;
625 struct file *file_out;
626 struct file *file_in;
633 struct io_provide_buf {
647 const char __user *filename;
648 struct statx __user *buffer;
660 struct filename *oldpath;
661 struct filename *newpath;
669 struct filename *filename;
676 struct filename *filename;
682 struct filename *oldpath;
683 struct filename *newpath;
690 struct filename *oldpath;
691 struct filename *newpath;
695 struct io_completion {
700 struct io_async_connect {
701 struct sockaddr_storage address;
704 struct io_async_msghdr {
705 struct iovec fast_iov[UIO_FASTIOV];
706 /* points to an allocated iov, if NULL we use fast_iov instead */
707 struct iovec *free_iov;
708 struct sockaddr __user *uaddr;
710 struct sockaddr_storage addr;
714 struct iovec fast_iov[UIO_FASTIOV];
715 const struct iovec *free_iovec;
716 struct iov_iter iter;
717 struct iov_iter_state iter_state;
719 struct wait_page_queue wpq;
723 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
724 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
725 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
726 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
727 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
728 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
730 /* first byte is taken by user flags, shift it to not overlap */
735 REQ_F_LINK_TIMEOUT_BIT,
736 REQ_F_NEED_CLEANUP_BIT,
738 REQ_F_BUFFER_SELECTED_BIT,
739 REQ_F_COMPLETE_INLINE_BIT,
743 REQ_F_ARM_LTIMEOUT_BIT,
744 /* keep async read/write and isreg together and in order */
745 REQ_F_NOWAIT_READ_BIT,
746 REQ_F_NOWAIT_WRITE_BIT,
749 /* not a real bit, just to check we're not overflowing the space */
755 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
756 /* drain existing IO first */
757 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
759 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
760 /* doesn't sever on completion < 0 */
761 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
763 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
764 /* IOSQE_BUFFER_SELECT */
765 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
767 /* fail rest of links */
768 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
769 /* on inflight list, should be cancelled and waited on exit reliably */
770 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
771 /* read/write uses file position */
772 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
773 /* must not punt to workers */
774 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
775 /* has or had linked timeout */
776 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
778 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
779 /* already went through poll handler */
780 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
781 /* buffer already selected */
782 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
783 /* completion is deferred through io_comp_state */
784 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
785 /* caller should reissue async */
786 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
787 /* supports async reads */
788 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
789 /* supports async writes */
790 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
792 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
793 /* has creds assigned */
794 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
795 /* skip refcounting if not set */
796 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
797 /* there is a linked timeout that has to be armed */
798 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
802 struct io_poll_iocb poll;
803 struct io_poll_iocb *double_poll;
806 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
808 struct io_task_work {
810 struct io_wq_work_node node;
811 struct llist_node fallback_node;
813 io_req_tw_func_t func;
817 IORING_RSRC_FILE = 0,
818 IORING_RSRC_BUFFER = 1,
822 * NOTE! Each of the iocb union members has the file pointer
823 * as the first entry in their struct definition. So you can
824 * access the file pointer through any of the sub-structs,
825 * or directly as just 'ki_filp' in this struct.
831 struct io_poll_iocb poll;
832 struct io_poll_update poll_update;
833 struct io_accept accept;
835 struct io_cancel cancel;
836 struct io_timeout timeout;
837 struct io_timeout_rem timeout_rem;
838 struct io_connect connect;
839 struct io_sr_msg sr_msg;
841 struct io_close close;
842 struct io_rsrc_update rsrc_update;
843 struct io_fadvise fadvise;
844 struct io_madvise madvise;
845 struct io_epoll epoll;
846 struct io_splice splice;
847 struct io_provide_buf pbuf;
848 struct io_statx statx;
849 struct io_shutdown shutdown;
850 struct io_rename rename;
851 struct io_unlink unlink;
852 struct io_mkdir mkdir;
853 struct io_symlink symlink;
854 struct io_hardlink hardlink;
855 /* use only after cleaning per-op data, see io_clean_op() */
856 struct io_completion compl;
859 /* opcode allocated if it needs to store data for async defer */
862 /* polled IO has completed */
868 struct io_ring_ctx *ctx;
871 struct task_struct *task;
874 struct io_kiocb *link;
875 struct percpu_ref *fixed_rsrc_refs;
877 /* used with ctx->iopoll_list with reads/writes */
878 struct list_head inflight_entry;
879 struct io_task_work io_task_work;
880 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
881 struct hlist_node hash_node;
882 struct async_poll *apoll;
883 struct io_wq_work work;
884 const struct cred *creds;
886 /* store used ubuf, so we can prevent reloading */
887 struct io_mapped_ubuf *imu;
890 struct io_tctx_node {
891 struct list_head ctx_node;
892 struct task_struct *task;
893 struct io_ring_ctx *ctx;
896 struct io_defer_entry {
897 struct list_head list;
898 struct io_kiocb *req;
903 /* needs req->file assigned */
904 unsigned needs_file : 1;
905 /* hash wq insertion if file is a regular file */
906 unsigned hash_reg_file : 1;
907 /* unbound wq insertion if file is a non-regular file */
908 unsigned unbound_nonreg_file : 1;
909 /* opcode is not supported by this kernel */
910 unsigned not_supported : 1;
911 /* set if opcode supports polled "wait" */
913 unsigned pollout : 1;
914 /* op supports buffer selection */
915 unsigned buffer_select : 1;
916 /* do prep async if is going to be punted */
917 unsigned needs_async_setup : 1;
918 /* should block plug */
920 /* size of async data needed, if any */
921 unsigned short async_size;
924 static const struct io_op_def io_op_defs[] = {
925 [IORING_OP_NOP] = {},
926 [IORING_OP_READV] = {
928 .unbound_nonreg_file = 1,
931 .needs_async_setup = 1,
933 .async_size = sizeof(struct io_async_rw),
935 [IORING_OP_WRITEV] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_setup = 1,
942 .async_size = sizeof(struct io_async_rw),
944 [IORING_OP_FSYNC] = {
947 [IORING_OP_READ_FIXED] = {
949 .unbound_nonreg_file = 1,
952 .async_size = sizeof(struct io_async_rw),
954 [IORING_OP_WRITE_FIXED] = {
957 .unbound_nonreg_file = 1,
960 .async_size = sizeof(struct io_async_rw),
962 [IORING_OP_POLL_ADD] = {
964 .unbound_nonreg_file = 1,
966 [IORING_OP_POLL_REMOVE] = {},
967 [IORING_OP_SYNC_FILE_RANGE] = {
970 [IORING_OP_SENDMSG] = {
972 .unbound_nonreg_file = 1,
974 .needs_async_setup = 1,
975 .async_size = sizeof(struct io_async_msghdr),
977 [IORING_OP_RECVMSG] = {
979 .unbound_nonreg_file = 1,
982 .needs_async_setup = 1,
983 .async_size = sizeof(struct io_async_msghdr),
985 [IORING_OP_TIMEOUT] = {
986 .async_size = sizeof(struct io_timeout_data),
988 [IORING_OP_TIMEOUT_REMOVE] = {
989 /* used by timeout updates' prep() */
991 [IORING_OP_ACCEPT] = {
993 .unbound_nonreg_file = 1,
996 [IORING_OP_ASYNC_CANCEL] = {},
997 [IORING_OP_LINK_TIMEOUT] = {
998 .async_size = sizeof(struct io_timeout_data),
1000 [IORING_OP_CONNECT] = {
1002 .unbound_nonreg_file = 1,
1004 .needs_async_setup = 1,
1005 .async_size = sizeof(struct io_async_connect),
1007 [IORING_OP_FALLOCATE] = {
1010 [IORING_OP_OPENAT] = {},
1011 [IORING_OP_CLOSE] = {},
1012 [IORING_OP_FILES_UPDATE] = {},
1013 [IORING_OP_STATX] = {},
1014 [IORING_OP_READ] = {
1016 .unbound_nonreg_file = 1,
1020 .async_size = sizeof(struct io_async_rw),
1022 [IORING_OP_WRITE] = {
1025 .unbound_nonreg_file = 1,
1028 .async_size = sizeof(struct io_async_rw),
1030 [IORING_OP_FADVISE] = {
1033 [IORING_OP_MADVISE] = {},
1034 [IORING_OP_SEND] = {
1036 .unbound_nonreg_file = 1,
1039 [IORING_OP_RECV] = {
1041 .unbound_nonreg_file = 1,
1045 [IORING_OP_OPENAT2] = {
1047 [IORING_OP_EPOLL_CTL] = {
1048 .unbound_nonreg_file = 1,
1050 [IORING_OP_SPLICE] = {
1053 .unbound_nonreg_file = 1,
1055 [IORING_OP_PROVIDE_BUFFERS] = {},
1056 [IORING_OP_REMOVE_BUFFERS] = {},
1060 .unbound_nonreg_file = 1,
1062 [IORING_OP_SHUTDOWN] = {
1065 [IORING_OP_RENAMEAT] = {},
1066 [IORING_OP_UNLINKAT] = {},
1067 [IORING_OP_MKDIRAT] = {},
1068 [IORING_OP_SYMLINKAT] = {},
1069 [IORING_OP_LINKAT] = {},
1072 /* requests with any of those set should undergo io_disarm_next() */
1073 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1075 static bool io_disarm_next(struct io_kiocb *req);
1076 static void io_uring_del_tctx_node(unsigned long index);
1077 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1078 struct task_struct *task,
1080 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1082 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1083 long res, unsigned int cflags);
1084 static void io_put_req(struct io_kiocb *req);
1085 static void io_put_req_deferred(struct io_kiocb *req);
1086 static void io_dismantle_req(struct io_kiocb *req);
1087 static void io_queue_linked_timeout(struct io_kiocb *req);
1088 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1089 struct io_uring_rsrc_update2 *up,
1091 static void io_clean_op(struct io_kiocb *req);
1092 static struct file *io_file_get(struct io_ring_ctx *ctx,
1093 struct io_kiocb *req, int fd, bool fixed);
1094 static void __io_queue_sqe(struct io_kiocb *req);
1095 static void io_rsrc_put_work(struct work_struct *work);
1097 static void io_req_task_queue(struct io_kiocb *req);
1098 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1099 static int io_req_prep_async(struct io_kiocb *req);
1101 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1102 unsigned int issue_flags, u32 slot_index);
1103 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1105 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1107 static struct kmem_cache *req_cachep;
1109 static const struct file_operations io_uring_fops;
1111 struct sock *io_uring_get_socket(struct file *file)
1113 #if defined(CONFIG_UNIX)
1114 if (file->f_op == &io_uring_fops) {
1115 struct io_ring_ctx *ctx = file->private_data;
1117 return ctx->ring_sock->sk;
1122 EXPORT_SYMBOL(io_uring_get_socket);
1124 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1127 mutex_lock(&ctx->uring_lock);
1132 #define io_for_each_link(pos, head) \
1133 for (pos = (head); pos; pos = pos->link)
1136 * Shamelessly stolen from the mm implementation of page reference checking,
1137 * see commit f958d7b528b1 for details.
1139 #define req_ref_zero_or_close_to_overflow(req) \
1140 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1142 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1144 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1145 return atomic_inc_not_zero(&req->refs);
1148 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1150 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1153 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1154 return atomic_dec_and_test(&req->refs);
1157 static inline void req_ref_put(struct io_kiocb *req)
1159 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1160 WARN_ON_ONCE(req_ref_put_and_test(req));
1163 static inline void req_ref_get(struct io_kiocb *req)
1165 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1166 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1167 atomic_inc(&req->refs);
1170 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1172 if (!(req->flags & REQ_F_REFCOUNT)) {
1173 req->flags |= REQ_F_REFCOUNT;
1174 atomic_set(&req->refs, nr);
1178 static inline void io_req_set_refcount(struct io_kiocb *req)
1180 __io_req_set_refcount(req, 1);
1183 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1185 struct io_ring_ctx *ctx = req->ctx;
1187 if (!req->fixed_rsrc_refs) {
1188 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1189 percpu_ref_get(req->fixed_rsrc_refs);
1193 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1195 bool got = percpu_ref_tryget(ref);
1197 /* already at zero, wait for ->release() */
1199 wait_for_completion(compl);
1200 percpu_ref_resurrect(ref);
1202 percpu_ref_put(ref);
1205 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1208 struct io_kiocb *req;
1210 if (task && head->task != task)
1215 io_for_each_link(req, head) {
1216 if (req->flags & REQ_F_INFLIGHT)
1222 static inline void req_set_fail(struct io_kiocb *req)
1224 req->flags |= REQ_F_FAIL;
1227 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1233 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1235 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1237 complete(&ctx->ref_comp);
1240 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1242 return !req->timeout.off;
1245 static void io_fallback_req_func(struct work_struct *work)
1247 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1248 fallback_work.work);
1249 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1250 struct io_kiocb *req, *tmp;
1251 bool locked = false;
1253 percpu_ref_get(&ctx->refs);
1254 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1255 req->io_task_work.func(req, &locked);
1258 if (ctx->submit_state.compl_nr)
1259 io_submit_flush_completions(ctx);
1260 mutex_unlock(&ctx->uring_lock);
1262 percpu_ref_put(&ctx->refs);
1266 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1268 struct io_ring_ctx *ctx;
1271 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1276 * Use 5 bits less than the max cq entries, that should give us around
1277 * 32 entries per hash list if totally full and uniformly spread.
1279 hash_bits = ilog2(p->cq_entries);
1283 ctx->cancel_hash_bits = hash_bits;
1284 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1286 if (!ctx->cancel_hash)
1288 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1290 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1291 if (!ctx->dummy_ubuf)
1293 /* set invalid range, so io_import_fixed() fails meeting it */
1294 ctx->dummy_ubuf->ubuf = -1UL;
1296 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1297 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1300 ctx->flags = p->flags;
1301 init_waitqueue_head(&ctx->sqo_sq_wait);
1302 INIT_LIST_HEAD(&ctx->sqd_list);
1303 init_waitqueue_head(&ctx->poll_wait);
1304 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1305 init_completion(&ctx->ref_comp);
1306 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1307 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1308 mutex_init(&ctx->uring_lock);
1309 init_waitqueue_head(&ctx->cq_wait);
1310 spin_lock_init(&ctx->completion_lock);
1311 spin_lock_init(&ctx->timeout_lock);
1312 INIT_LIST_HEAD(&ctx->iopoll_list);
1313 INIT_LIST_HEAD(&ctx->defer_list);
1314 INIT_LIST_HEAD(&ctx->timeout_list);
1315 INIT_LIST_HEAD(&ctx->ltimeout_list);
1316 spin_lock_init(&ctx->rsrc_ref_lock);
1317 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1318 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1319 init_llist_head(&ctx->rsrc_put_llist);
1320 INIT_LIST_HEAD(&ctx->tctx_list);
1321 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1322 INIT_LIST_HEAD(&ctx->locked_free_list);
1323 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1326 kfree(ctx->dummy_ubuf);
1327 kfree(ctx->cancel_hash);
1332 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1334 struct io_rings *r = ctx->rings;
1336 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1340 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1342 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1343 struct io_ring_ctx *ctx = req->ctx;
1345 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1351 #define FFS_ASYNC_READ 0x1UL
1352 #define FFS_ASYNC_WRITE 0x2UL
1354 #define FFS_ISREG 0x4UL
1356 #define FFS_ISREG 0x0UL
1358 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1360 static inline bool io_req_ffs_set(struct io_kiocb *req)
1362 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1365 static void io_req_track_inflight(struct io_kiocb *req)
1367 if (!(req->flags & REQ_F_INFLIGHT)) {
1368 req->flags |= REQ_F_INFLIGHT;
1369 atomic_inc(¤t->io_uring->inflight_tracked);
1373 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1375 req->flags &= ~REQ_F_LINK_TIMEOUT;
1378 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1380 if (WARN_ON_ONCE(!req->link))
1383 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1384 req->flags |= REQ_F_LINK_TIMEOUT;
1386 /* linked timeouts should have two refs once prep'ed */
1387 io_req_set_refcount(req);
1388 __io_req_set_refcount(req->link, 2);
1392 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1394 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1396 return __io_prep_linked_timeout(req);
1399 static void io_prep_async_work(struct io_kiocb *req)
1401 const struct io_op_def *def = &io_op_defs[req->opcode];
1402 struct io_ring_ctx *ctx = req->ctx;
1404 if (!(req->flags & REQ_F_CREDS)) {
1405 req->flags |= REQ_F_CREDS;
1406 req->creds = get_current_cred();
1409 req->work.list.next = NULL;
1410 req->work.flags = 0;
1411 if (req->flags & REQ_F_FORCE_ASYNC)
1412 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1414 if (req->flags & REQ_F_ISREG) {
1415 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1416 io_wq_hash_work(&req->work, file_inode(req->file));
1417 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1418 if (def->unbound_nonreg_file)
1419 req->work.flags |= IO_WQ_WORK_UNBOUND;
1422 switch (req->opcode) {
1423 case IORING_OP_SPLICE:
1425 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1426 req->work.flags |= IO_WQ_WORK_UNBOUND;
1431 static void io_prep_async_link(struct io_kiocb *req)
1433 struct io_kiocb *cur;
1435 if (req->flags & REQ_F_LINK_TIMEOUT) {
1436 struct io_ring_ctx *ctx = req->ctx;
1438 spin_lock(&ctx->completion_lock);
1439 io_for_each_link(cur, req)
1440 io_prep_async_work(cur);
1441 spin_unlock(&ctx->completion_lock);
1443 io_for_each_link(cur, req)
1444 io_prep_async_work(cur);
1448 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1450 struct io_ring_ctx *ctx = req->ctx;
1451 struct io_kiocb *link = io_prep_linked_timeout(req);
1452 struct io_uring_task *tctx = req->task->io_uring;
1454 /* must not take the lock, NULL it as a precaution */
1458 BUG_ON(!tctx->io_wq);
1460 /* init ->work of the whole link before punting */
1461 io_prep_async_link(req);
1464 * Not expected to happen, but if we do have a bug where this _can_
1465 * happen, catch it here and ensure the request is marked as
1466 * canceled. That will make io-wq go through the usual work cancel
1467 * procedure rather than attempt to run this request (or create a new
1470 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1471 req->work.flags |= IO_WQ_WORK_CANCEL;
1473 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1474 &req->work, req->flags);
1475 io_wq_enqueue(tctx->io_wq, &req->work);
1477 io_queue_linked_timeout(link);
1480 static void io_kill_timeout(struct io_kiocb *req, int status)
1481 __must_hold(&req->ctx->completion_lock)
1482 __must_hold(&req->ctx->timeout_lock)
1484 struct io_timeout_data *io = req->async_data;
1486 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1489 atomic_set(&req->ctx->cq_timeouts,
1490 atomic_read(&req->ctx->cq_timeouts) + 1);
1491 list_del_init(&req->timeout.list);
1492 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1493 io_put_req_deferred(req);
1497 static void io_queue_deferred(struct io_ring_ctx *ctx)
1499 while (!list_empty(&ctx->defer_list)) {
1500 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1501 struct io_defer_entry, list);
1503 if (req_need_defer(de->req, de->seq))
1505 list_del_init(&de->list);
1506 io_req_task_queue(de->req);
1511 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1512 __must_hold(&ctx->completion_lock)
1514 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1516 spin_lock_irq(&ctx->timeout_lock);
1517 while (!list_empty(&ctx->timeout_list)) {
1518 u32 events_needed, events_got;
1519 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1520 struct io_kiocb, timeout.list);
1522 if (io_is_timeout_noseq(req))
1526 * Since seq can easily wrap around over time, subtract
1527 * the last seq at which timeouts were flushed before comparing.
1528 * Assuming not more than 2^31-1 events have happened since,
1529 * these subtractions won't have wrapped, so we can check if
1530 * target is in [last_seq, current_seq] by comparing the two.
1532 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1533 events_got = seq - ctx->cq_last_tm_flush;
1534 if (events_got < events_needed)
1537 list_del_init(&req->timeout.list);
1538 io_kill_timeout(req, 0);
1540 ctx->cq_last_tm_flush = seq;
1541 spin_unlock_irq(&ctx->timeout_lock);
1544 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1546 if (ctx->off_timeout_used)
1547 io_flush_timeouts(ctx);
1548 if (ctx->drain_active)
1549 io_queue_deferred(ctx);
1552 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1554 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1555 __io_commit_cqring_flush(ctx);
1556 /* order cqe stores with ring update */
1557 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1560 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1562 struct io_rings *r = ctx->rings;
1564 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1567 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1569 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1572 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1574 struct io_rings *rings = ctx->rings;
1575 unsigned tail, mask = ctx->cq_entries - 1;
1578 * writes to the cq entry need to come after reading head; the
1579 * control dependency is enough as we're using WRITE_ONCE to
1582 if (__io_cqring_events(ctx) == ctx->cq_entries)
1585 tail = ctx->cached_cq_tail++;
1586 return &rings->cqes[tail & mask];
1589 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1591 if (likely(!ctx->cq_ev_fd))
1593 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1595 return !ctx->eventfd_async || io_wq_current_is_worker();
1599 * This should only get called when at least one event has been posted.
1600 * Some applications rely on the eventfd notification count only changing
1601 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1602 * 1:1 relationship between how many times this function is called (and
1603 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1605 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1608 * wake_up_all() may seem excessive, but io_wake_function() and
1609 * io_should_wake() handle the termination of the loop and only
1610 * wake as many waiters as we need to.
1612 if (wq_has_sleeper(&ctx->cq_wait))
1613 wake_up_all(&ctx->cq_wait);
1614 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1615 wake_up(&ctx->sq_data->wait);
1616 if (io_should_trigger_evfd(ctx))
1617 eventfd_signal(ctx->cq_ev_fd, 1);
1618 if (waitqueue_active(&ctx->poll_wait))
1619 wake_up_interruptible(&ctx->poll_wait);
1622 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1624 /* see waitqueue_active() comment */
1627 if (ctx->flags & IORING_SETUP_SQPOLL) {
1628 if (waitqueue_active(&ctx->cq_wait))
1629 wake_up_all(&ctx->cq_wait);
1631 if (io_should_trigger_evfd(ctx))
1632 eventfd_signal(ctx->cq_ev_fd, 1);
1633 if (waitqueue_active(&ctx->poll_wait))
1634 wake_up_interruptible(&ctx->poll_wait);
1637 /* Returns true if there are no backlogged entries after the flush */
1638 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1640 bool all_flushed, posted;
1642 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1646 spin_lock(&ctx->completion_lock);
1647 while (!list_empty(&ctx->cq_overflow_list)) {
1648 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1649 struct io_overflow_cqe *ocqe;
1653 ocqe = list_first_entry(&ctx->cq_overflow_list,
1654 struct io_overflow_cqe, list);
1656 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1658 io_account_cq_overflow(ctx);
1661 list_del(&ocqe->list);
1665 all_flushed = list_empty(&ctx->cq_overflow_list);
1667 clear_bit(0, &ctx->check_cq_overflow);
1668 WRITE_ONCE(ctx->rings->sq_flags,
1669 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1673 io_commit_cqring(ctx);
1674 spin_unlock(&ctx->completion_lock);
1676 io_cqring_ev_posted(ctx);
1680 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1684 if (test_bit(0, &ctx->check_cq_overflow)) {
1685 /* iopoll syncs against uring_lock, not completion_lock */
1686 if (ctx->flags & IORING_SETUP_IOPOLL)
1687 mutex_lock(&ctx->uring_lock);
1688 ret = __io_cqring_overflow_flush(ctx, false);
1689 if (ctx->flags & IORING_SETUP_IOPOLL)
1690 mutex_unlock(&ctx->uring_lock);
1696 /* must to be called somewhat shortly after putting a request */
1697 static inline void io_put_task(struct task_struct *task, int nr)
1699 struct io_uring_task *tctx = task->io_uring;
1701 if (likely(task == current)) {
1702 tctx->cached_refs += nr;
1704 percpu_counter_sub(&tctx->inflight, nr);
1705 if (unlikely(atomic_read(&tctx->in_idle)))
1706 wake_up(&tctx->wait);
1707 put_task_struct_many(task, nr);
1711 static void io_task_refs_refill(struct io_uring_task *tctx)
1713 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1715 percpu_counter_add(&tctx->inflight, refill);
1716 refcount_add(refill, ¤t->usage);
1717 tctx->cached_refs += refill;
1720 static inline void io_get_task_refs(int nr)
1722 struct io_uring_task *tctx = current->io_uring;
1724 tctx->cached_refs -= nr;
1725 if (unlikely(tctx->cached_refs < 0))
1726 io_task_refs_refill(tctx);
1729 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1730 long res, unsigned int cflags)
1732 struct io_overflow_cqe *ocqe;
1734 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1737 * If we're in ring overflow flush mode, or in task cancel mode,
1738 * or cannot allocate an overflow entry, then we need to drop it
1741 io_account_cq_overflow(ctx);
1744 if (list_empty(&ctx->cq_overflow_list)) {
1745 set_bit(0, &ctx->check_cq_overflow);
1746 WRITE_ONCE(ctx->rings->sq_flags,
1747 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1750 ocqe->cqe.user_data = user_data;
1751 ocqe->cqe.res = res;
1752 ocqe->cqe.flags = cflags;
1753 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1757 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1758 long res, unsigned int cflags)
1760 struct io_uring_cqe *cqe;
1762 trace_io_uring_complete(ctx, user_data, res, cflags);
1765 * If we can't get a cq entry, userspace overflowed the
1766 * submission (by quite a lot). Increment the overflow count in
1769 cqe = io_get_cqe(ctx);
1771 WRITE_ONCE(cqe->user_data, user_data);
1772 WRITE_ONCE(cqe->res, res);
1773 WRITE_ONCE(cqe->flags, cflags);
1776 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1779 /* not as hot to bloat with inlining */
1780 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1781 long res, unsigned int cflags)
1783 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1786 static void io_req_complete_post(struct io_kiocb *req, long res,
1787 unsigned int cflags)
1789 struct io_ring_ctx *ctx = req->ctx;
1791 spin_lock(&ctx->completion_lock);
1792 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1794 * If we're the last reference to this request, add to our locked
1797 if (req_ref_put_and_test(req)) {
1798 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1799 if (req->flags & IO_DISARM_MASK)
1800 io_disarm_next(req);
1802 io_req_task_queue(req->link);
1806 io_dismantle_req(req);
1807 io_put_task(req->task, 1);
1808 list_add(&req->inflight_entry, &ctx->locked_free_list);
1809 ctx->locked_free_nr++;
1811 if (!percpu_ref_tryget(&ctx->refs))
1814 io_commit_cqring(ctx);
1815 spin_unlock(&ctx->completion_lock);
1818 io_cqring_ev_posted(ctx);
1819 percpu_ref_put(&ctx->refs);
1823 static inline bool io_req_needs_clean(struct io_kiocb *req)
1825 return req->flags & IO_REQ_CLEAN_FLAGS;
1828 static void io_req_complete_state(struct io_kiocb *req, long res,
1829 unsigned int cflags)
1831 if (io_req_needs_clean(req))
1834 req->compl.cflags = cflags;
1835 req->flags |= REQ_F_COMPLETE_INLINE;
1838 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1839 long res, unsigned cflags)
1841 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1842 io_req_complete_state(req, res, cflags);
1844 io_req_complete_post(req, res, cflags);
1847 static inline void io_req_complete(struct io_kiocb *req, long res)
1849 __io_req_complete(req, 0, res, 0);
1852 static void io_req_complete_failed(struct io_kiocb *req, long res)
1855 io_req_complete_post(req, res, 0);
1858 static void io_req_complete_fail_submit(struct io_kiocb *req)
1861 * We don't submit, fail them all, for that replace hardlinks with
1862 * normal links. Extra REQ_F_LINK is tolerated.
1864 req->flags &= ~REQ_F_HARDLINK;
1865 req->flags |= REQ_F_LINK;
1866 io_req_complete_failed(req, req->result);
1870 * Don't initialise the fields below on every allocation, but do that in
1871 * advance and keep them valid across allocations.
1873 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1877 req->async_data = NULL;
1878 /* not necessary, but safer to zero */
1882 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1883 struct io_submit_state *state)
1885 spin_lock(&ctx->completion_lock);
1886 list_splice_init(&ctx->locked_free_list, &state->free_list);
1887 ctx->locked_free_nr = 0;
1888 spin_unlock(&ctx->completion_lock);
1891 /* Returns true IFF there are requests in the cache */
1892 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1894 struct io_submit_state *state = &ctx->submit_state;
1898 * If we have more than a batch's worth of requests in our IRQ side
1899 * locked cache, grab the lock and move them over to our submission
1902 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1903 io_flush_cached_locked_reqs(ctx, state);
1905 nr = state->free_reqs;
1906 while (!list_empty(&state->free_list)) {
1907 struct io_kiocb *req = list_first_entry(&state->free_list,
1908 struct io_kiocb, inflight_entry);
1910 list_del(&req->inflight_entry);
1911 state->reqs[nr++] = req;
1912 if (nr == ARRAY_SIZE(state->reqs))
1916 state->free_reqs = nr;
1921 * A request might get retired back into the request caches even before opcode
1922 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1923 * Because of that, io_alloc_req() should be called only under ->uring_lock
1924 * and with extra caution to not get a request that is still worked on.
1926 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1927 __must_hold(&ctx->uring_lock)
1929 struct io_submit_state *state = &ctx->submit_state;
1930 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1933 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1935 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1938 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1942 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1943 * retry single alloc to be on the safe side.
1945 if (unlikely(ret <= 0)) {
1946 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1947 if (!state->reqs[0])
1952 for (i = 0; i < ret; i++)
1953 io_preinit_req(state->reqs[i], ctx);
1954 state->free_reqs = ret;
1957 return state->reqs[state->free_reqs];
1960 static inline void io_put_file(struct file *file)
1966 static void io_dismantle_req(struct io_kiocb *req)
1968 unsigned int flags = req->flags;
1970 if (io_req_needs_clean(req))
1972 if (!(flags & REQ_F_FIXED_FILE))
1973 io_put_file(req->file);
1974 if (req->fixed_rsrc_refs)
1975 percpu_ref_put(req->fixed_rsrc_refs);
1976 if (req->async_data) {
1977 kfree(req->async_data);
1978 req->async_data = NULL;
1982 static void __io_free_req(struct io_kiocb *req)
1984 struct io_ring_ctx *ctx = req->ctx;
1986 io_dismantle_req(req);
1987 io_put_task(req->task, 1);
1989 spin_lock(&ctx->completion_lock);
1990 list_add(&req->inflight_entry, &ctx->locked_free_list);
1991 ctx->locked_free_nr++;
1992 spin_unlock(&ctx->completion_lock);
1994 percpu_ref_put(&ctx->refs);
1997 static inline void io_remove_next_linked(struct io_kiocb *req)
1999 struct io_kiocb *nxt = req->link;
2001 req->link = nxt->link;
2005 static bool io_kill_linked_timeout(struct io_kiocb *req)
2006 __must_hold(&req->ctx->completion_lock)
2007 __must_hold(&req->ctx->timeout_lock)
2009 struct io_kiocb *link = req->link;
2011 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2012 struct io_timeout_data *io = link->async_data;
2014 io_remove_next_linked(req);
2015 link->timeout.head = NULL;
2016 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2017 list_del(&link->timeout.list);
2018 io_cqring_fill_event(link->ctx, link->user_data,
2020 io_put_req_deferred(link);
2027 static void io_fail_links(struct io_kiocb *req)
2028 __must_hold(&req->ctx->completion_lock)
2030 struct io_kiocb *nxt, *link = req->link;
2034 long res = -ECANCELED;
2036 if (link->flags & REQ_F_FAIL)
2042 trace_io_uring_fail_link(req, link);
2043 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2044 io_put_req_deferred(link);
2049 static bool io_disarm_next(struct io_kiocb *req)
2050 __must_hold(&req->ctx->completion_lock)
2052 bool posted = false;
2054 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2055 struct io_kiocb *link = req->link;
2057 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2058 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2059 io_remove_next_linked(req);
2060 io_cqring_fill_event(link->ctx, link->user_data,
2062 io_put_req_deferred(link);
2065 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2066 struct io_ring_ctx *ctx = req->ctx;
2068 spin_lock_irq(&ctx->timeout_lock);
2069 posted = io_kill_linked_timeout(req);
2070 spin_unlock_irq(&ctx->timeout_lock);
2072 if (unlikely((req->flags & REQ_F_FAIL) &&
2073 !(req->flags & REQ_F_HARDLINK))) {
2074 posted |= (req->link != NULL);
2080 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2082 struct io_kiocb *nxt;
2085 * If LINK is set, we have dependent requests in this chain. If we
2086 * didn't fail this request, queue the first one up, moving any other
2087 * dependencies to the next request. In case of failure, fail the rest
2090 if (req->flags & IO_DISARM_MASK) {
2091 struct io_ring_ctx *ctx = req->ctx;
2094 spin_lock(&ctx->completion_lock);
2095 posted = io_disarm_next(req);
2097 io_commit_cqring(req->ctx);
2098 spin_unlock(&ctx->completion_lock);
2100 io_cqring_ev_posted(ctx);
2107 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2109 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2111 return __io_req_find_next(req);
2114 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2119 if (ctx->submit_state.compl_nr)
2120 io_submit_flush_completions(ctx);
2121 mutex_unlock(&ctx->uring_lock);
2124 percpu_ref_put(&ctx->refs);
2127 static void tctx_task_work(struct callback_head *cb)
2129 bool locked = false;
2130 struct io_ring_ctx *ctx = NULL;
2131 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2135 struct io_wq_work_node *node;
2137 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2138 io_submit_flush_completions(ctx);
2140 spin_lock_irq(&tctx->task_lock);
2141 node = tctx->task_list.first;
2142 INIT_WQ_LIST(&tctx->task_list);
2144 tctx->task_running = false;
2145 spin_unlock_irq(&tctx->task_lock);
2150 struct io_wq_work_node *next = node->next;
2151 struct io_kiocb *req = container_of(node, struct io_kiocb,
2154 if (req->ctx != ctx) {
2155 ctx_flush_and_put(ctx, &locked);
2157 /* if not contended, grab and improve batching */
2158 locked = mutex_trylock(&ctx->uring_lock);
2159 percpu_ref_get(&ctx->refs);
2161 req->io_task_work.func(req, &locked);
2168 ctx_flush_and_put(ctx, &locked);
2171 static void io_req_task_work_add(struct io_kiocb *req)
2173 struct task_struct *tsk = req->task;
2174 struct io_uring_task *tctx = tsk->io_uring;
2175 enum task_work_notify_mode notify;
2176 struct io_wq_work_node *node;
2177 unsigned long flags;
2180 WARN_ON_ONCE(!tctx);
2182 spin_lock_irqsave(&tctx->task_lock, flags);
2183 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2184 running = tctx->task_running;
2186 tctx->task_running = true;
2187 spin_unlock_irqrestore(&tctx->task_lock, flags);
2189 /* task_work already pending, we're done */
2194 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2195 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2196 * processing task_work. There's no reliable way to tell if TWA_RESUME
2199 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2200 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2201 wake_up_process(tsk);
2205 spin_lock_irqsave(&tctx->task_lock, flags);
2206 tctx->task_running = false;
2207 node = tctx->task_list.first;
2208 INIT_WQ_LIST(&tctx->task_list);
2209 spin_unlock_irqrestore(&tctx->task_lock, flags);
2212 req = container_of(node, struct io_kiocb, io_task_work.node);
2214 if (llist_add(&req->io_task_work.fallback_node,
2215 &req->ctx->fallback_llist))
2216 schedule_delayed_work(&req->ctx->fallback_work, 1);
2220 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2222 struct io_ring_ctx *ctx = req->ctx;
2224 /* not needed for normal modes, but SQPOLL depends on it */
2225 io_tw_lock(ctx, locked);
2226 io_req_complete_failed(req, req->result);
2229 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2231 struct io_ring_ctx *ctx = req->ctx;
2233 io_tw_lock(ctx, locked);
2234 /* req->task == current here, checking PF_EXITING is safe */
2235 if (likely(!(req->task->flags & PF_EXITING)))
2236 __io_queue_sqe(req);
2238 io_req_complete_failed(req, -EFAULT);
2241 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2244 req->io_task_work.func = io_req_task_cancel;
2245 io_req_task_work_add(req);
2248 static void io_req_task_queue(struct io_kiocb *req)
2250 req->io_task_work.func = io_req_task_submit;
2251 io_req_task_work_add(req);
2254 static void io_req_task_queue_reissue(struct io_kiocb *req)
2256 req->io_task_work.func = io_queue_async_work;
2257 io_req_task_work_add(req);
2260 static inline void io_queue_next(struct io_kiocb *req)
2262 struct io_kiocb *nxt = io_req_find_next(req);
2265 io_req_task_queue(nxt);
2268 static void io_free_req(struct io_kiocb *req)
2274 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2280 struct task_struct *task;
2285 static inline void io_init_req_batch(struct req_batch *rb)
2292 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2293 struct req_batch *rb)
2296 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2298 io_put_task(rb->task, rb->task_refs);
2301 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2302 struct io_submit_state *state)
2305 io_dismantle_req(req);
2307 if (req->task != rb->task) {
2309 io_put_task(rb->task, rb->task_refs);
2310 rb->task = req->task;
2316 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2317 state->reqs[state->free_reqs++] = req;
2319 list_add(&req->inflight_entry, &state->free_list);
2322 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2323 __must_hold(&ctx->uring_lock)
2325 struct io_submit_state *state = &ctx->submit_state;
2326 int i, nr = state->compl_nr;
2327 struct req_batch rb;
2329 spin_lock(&ctx->completion_lock);
2330 for (i = 0; i < nr; i++) {
2331 struct io_kiocb *req = state->compl_reqs[i];
2333 __io_cqring_fill_event(ctx, req->user_data, req->result,
2336 io_commit_cqring(ctx);
2337 spin_unlock(&ctx->completion_lock);
2338 io_cqring_ev_posted(ctx);
2340 io_init_req_batch(&rb);
2341 for (i = 0; i < nr; i++) {
2342 struct io_kiocb *req = state->compl_reqs[i];
2344 if (req_ref_put_and_test(req))
2345 io_req_free_batch(&rb, req, &ctx->submit_state);
2348 io_req_free_batch_finish(ctx, &rb);
2349 state->compl_nr = 0;
2353 * Drop reference to request, return next in chain (if there is one) if this
2354 * was the last reference to this request.
2356 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2358 struct io_kiocb *nxt = NULL;
2360 if (req_ref_put_and_test(req)) {
2361 nxt = io_req_find_next(req);
2367 static inline void io_put_req(struct io_kiocb *req)
2369 if (req_ref_put_and_test(req))
2373 static inline void io_put_req_deferred(struct io_kiocb *req)
2375 if (req_ref_put_and_test(req)) {
2376 req->io_task_work.func = io_free_req_work;
2377 io_req_task_work_add(req);
2381 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2383 /* See comment at the top of this file */
2385 return __io_cqring_events(ctx);
2388 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2390 struct io_rings *rings = ctx->rings;
2392 /* make sure SQ entry isn't read before tail */
2393 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2396 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2398 unsigned int cflags;
2400 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2401 cflags |= IORING_CQE_F_BUFFER;
2402 req->flags &= ~REQ_F_BUFFER_SELECTED;
2407 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2409 struct io_buffer *kbuf;
2411 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2413 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2414 return io_put_kbuf(req, kbuf);
2417 static inline bool io_run_task_work(void)
2419 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2420 __set_current_state(TASK_RUNNING);
2421 tracehook_notify_signal();
2429 * Find and free completed poll iocbs
2431 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2432 struct list_head *done)
2434 struct req_batch rb;
2435 struct io_kiocb *req;
2437 /* order with ->result store in io_complete_rw_iopoll() */
2440 io_init_req_batch(&rb);
2441 while (!list_empty(done)) {
2442 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2443 list_del(&req->inflight_entry);
2445 __io_cqring_fill_event(ctx, req->user_data, req->result,
2446 io_put_rw_kbuf(req));
2449 if (req_ref_put_and_test(req))
2450 io_req_free_batch(&rb, req, &ctx->submit_state);
2453 io_commit_cqring(ctx);
2454 io_cqring_ev_posted_iopoll(ctx);
2455 io_req_free_batch_finish(ctx, &rb);
2458 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2461 struct io_kiocb *req, *tmp;
2466 * Only spin for completions if we don't have multiple devices hanging
2467 * off our complete list, and we're under the requested amount.
2469 spin = !ctx->poll_multi_queue && *nr_events < min;
2471 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2472 struct kiocb *kiocb = &req->rw.kiocb;
2476 * Move completed and retryable entries to our local lists.
2477 * If we find a request that requires polling, break out
2478 * and complete those lists first, if we have entries there.
2480 if (READ_ONCE(req->iopoll_completed)) {
2481 list_move_tail(&req->inflight_entry, &done);
2484 if (!list_empty(&done))
2487 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2488 if (unlikely(ret < 0))
2493 /* iopoll may have completed current req */
2494 if (READ_ONCE(req->iopoll_completed))
2495 list_move_tail(&req->inflight_entry, &done);
2498 if (!list_empty(&done))
2499 io_iopoll_complete(ctx, nr_events, &done);
2505 * We can't just wait for polled events to come to us, we have to actively
2506 * find and complete them.
2508 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2510 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2513 mutex_lock(&ctx->uring_lock);
2514 while (!list_empty(&ctx->iopoll_list)) {
2515 unsigned int nr_events = 0;
2517 io_do_iopoll(ctx, &nr_events, 0);
2519 /* let it sleep and repeat later if can't complete a request */
2523 * Ensure we allow local-to-the-cpu processing to take place,
2524 * in this case we need to ensure that we reap all events.
2525 * Also let task_work, etc. to progress by releasing the mutex
2527 if (need_resched()) {
2528 mutex_unlock(&ctx->uring_lock);
2530 mutex_lock(&ctx->uring_lock);
2533 mutex_unlock(&ctx->uring_lock);
2536 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2538 unsigned int nr_events = 0;
2542 * We disallow the app entering submit/complete with polling, but we
2543 * still need to lock the ring to prevent racing with polled issue
2544 * that got punted to a workqueue.
2546 mutex_lock(&ctx->uring_lock);
2548 * Don't enter poll loop if we already have events pending.
2549 * If we do, we can potentially be spinning for commands that
2550 * already triggered a CQE (eg in error).
2552 if (test_bit(0, &ctx->check_cq_overflow))
2553 __io_cqring_overflow_flush(ctx, false);
2554 if (io_cqring_events(ctx))
2558 * If a submit got punted to a workqueue, we can have the
2559 * application entering polling for a command before it gets
2560 * issued. That app will hold the uring_lock for the duration
2561 * of the poll right here, so we need to take a breather every
2562 * now and then to ensure that the issue has a chance to add
2563 * the poll to the issued list. Otherwise we can spin here
2564 * forever, while the workqueue is stuck trying to acquire the
2567 if (list_empty(&ctx->iopoll_list)) {
2568 u32 tail = ctx->cached_cq_tail;
2570 mutex_unlock(&ctx->uring_lock);
2572 mutex_lock(&ctx->uring_lock);
2574 /* some requests don't go through iopoll_list */
2575 if (tail != ctx->cached_cq_tail ||
2576 list_empty(&ctx->iopoll_list))
2579 ret = io_do_iopoll(ctx, &nr_events, min);
2580 } while (!ret && nr_events < min && !need_resched());
2582 mutex_unlock(&ctx->uring_lock);
2586 static void kiocb_end_write(struct io_kiocb *req)
2589 * Tell lockdep we inherited freeze protection from submission
2592 if (req->flags & REQ_F_ISREG) {
2593 struct super_block *sb = file_inode(req->file)->i_sb;
2595 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2601 static bool io_resubmit_prep(struct io_kiocb *req)
2603 struct io_async_rw *rw = req->async_data;
2606 return !io_req_prep_async(req);
2607 iov_iter_restore(&rw->iter, &rw->iter_state);
2611 static bool io_rw_should_reissue(struct io_kiocb *req)
2613 umode_t mode = file_inode(req->file)->i_mode;
2614 struct io_ring_ctx *ctx = req->ctx;
2616 if (!S_ISBLK(mode) && !S_ISREG(mode))
2618 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2619 !(ctx->flags & IORING_SETUP_IOPOLL)))
2622 * If ref is dying, we might be running poll reap from the exit work.
2623 * Don't attempt to reissue from that path, just let it fail with
2626 if (percpu_ref_is_dying(&ctx->refs))
2629 * Play it safe and assume not safe to re-import and reissue if we're
2630 * not in the original thread group (or in task context).
2632 if (!same_thread_group(req->task, current) || !in_task())
2637 static bool io_resubmit_prep(struct io_kiocb *req)
2641 static bool io_rw_should_reissue(struct io_kiocb *req)
2647 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2649 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2650 kiocb_end_write(req);
2651 if (res != req->result) {
2652 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2653 io_rw_should_reissue(req)) {
2654 req->flags |= REQ_F_REISSUE;
2663 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2665 unsigned int cflags = io_put_rw_kbuf(req);
2666 long res = req->result;
2669 struct io_ring_ctx *ctx = req->ctx;
2670 struct io_submit_state *state = &ctx->submit_state;
2672 io_req_complete_state(req, res, cflags);
2673 state->compl_reqs[state->compl_nr++] = req;
2674 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2675 io_submit_flush_completions(ctx);
2677 io_req_complete_post(req, res, cflags);
2681 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2682 unsigned int issue_flags)
2684 if (__io_complete_rw_common(req, res))
2686 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2689 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2691 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2693 if (__io_complete_rw_common(req, res))
2696 req->io_task_work.func = io_req_task_complete;
2697 io_req_task_work_add(req);
2700 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2702 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2704 if (kiocb->ki_flags & IOCB_WRITE)
2705 kiocb_end_write(req);
2706 if (unlikely(res != req->result)) {
2707 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2708 req->flags |= REQ_F_REISSUE;
2713 WRITE_ONCE(req->result, res);
2714 /* order with io_iopoll_complete() checking ->result */
2716 WRITE_ONCE(req->iopoll_completed, 1);
2720 * After the iocb has been issued, it's safe to be found on the poll list.
2721 * Adding the kiocb to the list AFTER submission ensures that we don't
2722 * find it from a io_do_iopoll() thread before the issuer is done
2723 * accessing the kiocb cookie.
2725 static void io_iopoll_req_issued(struct io_kiocb *req)
2727 struct io_ring_ctx *ctx = req->ctx;
2728 const bool in_async = io_wq_current_is_worker();
2730 /* workqueue context doesn't hold uring_lock, grab it now */
2731 if (unlikely(in_async))
2732 mutex_lock(&ctx->uring_lock);
2735 * Track whether we have multiple files in our lists. This will impact
2736 * how we do polling eventually, not spinning if we're on potentially
2737 * different devices.
2739 if (list_empty(&ctx->iopoll_list)) {
2740 ctx->poll_multi_queue = false;
2741 } else if (!ctx->poll_multi_queue) {
2742 struct io_kiocb *list_req;
2743 unsigned int queue_num0, queue_num1;
2745 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2748 if (list_req->file != req->file) {
2749 ctx->poll_multi_queue = true;
2751 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2752 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2753 if (queue_num0 != queue_num1)
2754 ctx->poll_multi_queue = true;
2759 * For fast devices, IO may have already completed. If it has, add
2760 * it to the front so we find it first.
2762 if (READ_ONCE(req->iopoll_completed))
2763 list_add(&req->inflight_entry, &ctx->iopoll_list);
2765 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2767 if (unlikely(in_async)) {
2769 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2770 * in sq thread task context or in io worker task context. If
2771 * current task context is sq thread, we don't need to check
2772 * whether should wake up sq thread.
2774 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2775 wq_has_sleeper(&ctx->sq_data->wait))
2776 wake_up(&ctx->sq_data->wait);
2778 mutex_unlock(&ctx->uring_lock);
2782 static bool io_bdev_nowait(struct block_device *bdev)
2784 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2788 * If we tracked the file through the SCM inflight mechanism, we could support
2789 * any file. For now, just ensure that anything potentially problematic is done
2792 static bool __io_file_supports_nowait(struct file *file, int rw)
2794 umode_t mode = file_inode(file)->i_mode;
2796 if (S_ISBLK(mode)) {
2797 if (IS_ENABLED(CONFIG_BLOCK) &&
2798 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2804 if (S_ISREG(mode)) {
2805 if (IS_ENABLED(CONFIG_BLOCK) &&
2806 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2807 file->f_op != &io_uring_fops)
2812 /* any ->read/write should understand O_NONBLOCK */
2813 if (file->f_flags & O_NONBLOCK)
2816 if (!(file->f_mode & FMODE_NOWAIT))
2820 return file->f_op->read_iter != NULL;
2822 return file->f_op->write_iter != NULL;
2825 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2827 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2829 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2832 return __io_file_supports_nowait(req->file, rw);
2835 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2838 struct io_ring_ctx *ctx = req->ctx;
2839 struct kiocb *kiocb = &req->rw.kiocb;
2840 struct file *file = req->file;
2844 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2845 req->flags |= REQ_F_ISREG;
2847 kiocb->ki_pos = READ_ONCE(sqe->off);
2848 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2849 req->flags |= REQ_F_CUR_POS;
2850 kiocb->ki_pos = file->f_pos;
2852 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2853 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2854 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2859 * If the file is marked O_NONBLOCK, still allow retry for it if it
2860 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2861 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2863 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2864 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2865 req->flags |= REQ_F_NOWAIT;
2867 ioprio = READ_ONCE(sqe->ioprio);
2869 ret = ioprio_check_cap(ioprio);
2873 kiocb->ki_ioprio = ioprio;
2875 kiocb->ki_ioprio = get_current_ioprio();
2877 if (ctx->flags & IORING_SETUP_IOPOLL) {
2878 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2879 !kiocb->ki_filp->f_op->iopoll)
2882 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2883 kiocb->ki_complete = io_complete_rw_iopoll;
2884 req->iopoll_completed = 0;
2886 if (kiocb->ki_flags & IOCB_HIPRI)
2888 kiocb->ki_complete = io_complete_rw;
2891 if (req->opcode == IORING_OP_READ_FIXED ||
2892 req->opcode == IORING_OP_WRITE_FIXED) {
2894 io_req_set_rsrc_node(req);
2897 req->rw.addr = READ_ONCE(sqe->addr);
2898 req->rw.len = READ_ONCE(sqe->len);
2899 req->buf_index = READ_ONCE(sqe->buf_index);
2903 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2909 case -ERESTARTNOINTR:
2910 case -ERESTARTNOHAND:
2911 case -ERESTART_RESTARTBLOCK:
2913 * We can't just restart the syscall, since previously
2914 * submitted sqes may already be in progress. Just fail this
2920 kiocb->ki_complete(kiocb, ret, 0);
2924 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2925 unsigned int issue_flags)
2927 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2928 struct io_async_rw *io = req->async_data;
2930 /* add previously done IO, if any */
2931 if (io && io->bytes_done > 0) {
2933 ret = io->bytes_done;
2935 ret += io->bytes_done;
2938 if (req->flags & REQ_F_CUR_POS)
2939 req->file->f_pos = kiocb->ki_pos;
2940 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2941 __io_complete_rw(req, ret, 0, issue_flags);
2943 io_rw_done(kiocb, ret);
2945 if (req->flags & REQ_F_REISSUE) {
2946 req->flags &= ~REQ_F_REISSUE;
2947 if (io_resubmit_prep(req)) {
2948 io_req_task_queue_reissue(req);
2950 unsigned int cflags = io_put_rw_kbuf(req);
2951 struct io_ring_ctx *ctx = req->ctx;
2954 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
2955 mutex_lock(&ctx->uring_lock);
2956 __io_req_complete(req, issue_flags, ret, cflags);
2957 mutex_unlock(&ctx->uring_lock);
2959 __io_req_complete(req, issue_flags, ret, cflags);
2965 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2966 struct io_mapped_ubuf *imu)
2968 size_t len = req->rw.len;
2969 u64 buf_end, buf_addr = req->rw.addr;
2972 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2974 /* not inside the mapped region */
2975 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2979 * May not be a start of buffer, set size appropriately
2980 * and advance us to the beginning.
2982 offset = buf_addr - imu->ubuf;
2983 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2987 * Don't use iov_iter_advance() here, as it's really slow for
2988 * using the latter parts of a big fixed buffer - it iterates
2989 * over each segment manually. We can cheat a bit here, because
2992 * 1) it's a BVEC iter, we set it up
2993 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2994 * first and last bvec
2996 * So just find our index, and adjust the iterator afterwards.
2997 * If the offset is within the first bvec (or the whole first
2998 * bvec, just use iov_iter_advance(). This makes it easier
2999 * since we can just skip the first segment, which may not
3000 * be PAGE_SIZE aligned.
3002 const struct bio_vec *bvec = imu->bvec;
3004 if (offset <= bvec->bv_len) {
3005 iov_iter_advance(iter, offset);
3007 unsigned long seg_skip;
3009 /* skip first vec */
3010 offset -= bvec->bv_len;
3011 seg_skip = 1 + (offset >> PAGE_SHIFT);
3013 iter->bvec = bvec + seg_skip;
3014 iter->nr_segs -= seg_skip;
3015 iter->count -= bvec->bv_len + offset;
3016 iter->iov_offset = offset & ~PAGE_MASK;
3023 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3025 struct io_ring_ctx *ctx = req->ctx;
3026 struct io_mapped_ubuf *imu = req->imu;
3027 u16 index, buf_index = req->buf_index;
3030 if (unlikely(buf_index >= ctx->nr_user_bufs))
3032 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3033 imu = READ_ONCE(ctx->user_bufs[index]);
3036 return __io_import_fixed(req, rw, iter, imu);
3039 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3042 mutex_unlock(&ctx->uring_lock);
3045 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3048 * "Normal" inline submissions always hold the uring_lock, since we
3049 * grab it from the system call. Same is true for the SQPOLL offload.
3050 * The only exception is when we've detached the request and issue it
3051 * from an async worker thread, grab the lock for that case.
3054 mutex_lock(&ctx->uring_lock);
3057 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3058 int bgid, struct io_buffer *kbuf,
3061 struct io_buffer *head;
3063 if (req->flags & REQ_F_BUFFER_SELECTED)
3066 io_ring_submit_lock(req->ctx, needs_lock);
3068 lockdep_assert_held(&req->ctx->uring_lock);
3070 head = xa_load(&req->ctx->io_buffers, bgid);
3072 if (!list_empty(&head->list)) {
3073 kbuf = list_last_entry(&head->list, struct io_buffer,
3075 list_del(&kbuf->list);
3078 xa_erase(&req->ctx->io_buffers, bgid);
3080 if (*len > kbuf->len)
3083 kbuf = ERR_PTR(-ENOBUFS);
3086 io_ring_submit_unlock(req->ctx, needs_lock);
3091 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3094 struct io_buffer *kbuf;
3097 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3098 bgid = req->buf_index;
3099 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3102 req->rw.addr = (u64) (unsigned long) kbuf;
3103 req->flags |= REQ_F_BUFFER_SELECTED;
3104 return u64_to_user_ptr(kbuf->addr);
3107 #ifdef CONFIG_COMPAT
3108 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3111 struct compat_iovec __user *uiov;
3112 compat_ssize_t clen;
3116 uiov = u64_to_user_ptr(req->rw.addr);
3117 if (!access_ok(uiov, sizeof(*uiov)))
3119 if (__get_user(clen, &uiov->iov_len))
3125 buf = io_rw_buffer_select(req, &len, needs_lock);
3127 return PTR_ERR(buf);
3128 iov[0].iov_base = buf;
3129 iov[0].iov_len = (compat_size_t) len;
3134 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3137 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3141 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3144 len = iov[0].iov_len;
3147 buf = io_rw_buffer_select(req, &len, needs_lock);
3149 return PTR_ERR(buf);
3150 iov[0].iov_base = buf;
3151 iov[0].iov_len = len;
3155 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3158 if (req->flags & REQ_F_BUFFER_SELECTED) {
3159 struct io_buffer *kbuf;
3161 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3162 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3163 iov[0].iov_len = kbuf->len;
3166 if (req->rw.len != 1)
3169 #ifdef CONFIG_COMPAT
3170 if (req->ctx->compat)
3171 return io_compat_import(req, iov, needs_lock);
3174 return __io_iov_buffer_select(req, iov, needs_lock);
3177 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3178 struct iov_iter *iter, bool needs_lock)
3180 void __user *buf = u64_to_user_ptr(req->rw.addr);
3181 size_t sqe_len = req->rw.len;
3182 u8 opcode = req->opcode;
3185 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3187 return io_import_fixed(req, rw, iter);
3190 /* buffer index only valid with fixed read/write, or buffer select */
3191 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3194 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3195 if (req->flags & REQ_F_BUFFER_SELECT) {
3196 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3198 return PTR_ERR(buf);
3199 req->rw.len = sqe_len;
3202 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3207 if (req->flags & REQ_F_BUFFER_SELECT) {
3208 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3210 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3215 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3219 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3221 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3225 * For files that don't have ->read_iter() and ->write_iter(), handle them
3226 * by looping over ->read() or ->write() manually.
3228 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3230 struct kiocb *kiocb = &req->rw.kiocb;
3231 struct file *file = req->file;
3235 * Don't support polled IO through this interface, and we can't
3236 * support non-blocking either. For the latter, this just causes
3237 * the kiocb to be handled from an async context.
3239 if (kiocb->ki_flags & IOCB_HIPRI)
3241 if (kiocb->ki_flags & IOCB_NOWAIT)
3244 while (iov_iter_count(iter)) {
3248 if (!iov_iter_is_bvec(iter)) {
3249 iovec = iov_iter_iovec(iter);
3251 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3252 iovec.iov_len = req->rw.len;
3256 nr = file->f_op->read(file, iovec.iov_base,
3257 iovec.iov_len, io_kiocb_ppos(kiocb));
3259 nr = file->f_op->write(file, iovec.iov_base,
3260 iovec.iov_len, io_kiocb_ppos(kiocb));
3268 if (!iov_iter_is_bvec(iter)) {
3269 iov_iter_advance(iter, nr);
3275 if (nr != iovec.iov_len)
3282 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3283 const struct iovec *fast_iov, struct iov_iter *iter)
3285 struct io_async_rw *rw = req->async_data;
3287 memcpy(&rw->iter, iter, sizeof(*iter));
3288 rw->free_iovec = iovec;
3290 /* can only be fixed buffers, no need to do anything */
3291 if (iov_iter_is_bvec(iter))
3294 unsigned iov_off = 0;
3296 rw->iter.iov = rw->fast_iov;
3297 if (iter->iov != fast_iov) {
3298 iov_off = iter->iov - fast_iov;
3299 rw->iter.iov += iov_off;
3301 if (rw->fast_iov != fast_iov)
3302 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3303 sizeof(struct iovec) * iter->nr_segs);
3305 req->flags |= REQ_F_NEED_CLEANUP;
3309 static inline int io_alloc_async_data(struct io_kiocb *req)
3311 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3312 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3313 return req->async_data == NULL;
3316 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3317 const struct iovec *fast_iov,
3318 struct iov_iter *iter, bool force)
3320 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3322 if (!req->async_data) {
3323 struct io_async_rw *iorw;
3325 if (io_alloc_async_data(req)) {
3330 io_req_map_rw(req, iovec, fast_iov, iter);
3331 iorw = req->async_data;
3332 /* we've copied and mapped the iter, ensure state is saved */
3333 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3338 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3340 struct io_async_rw *iorw = req->async_data;
3341 struct iovec *iov = iorw->fast_iov;
3344 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3345 if (unlikely(ret < 0))
3348 iorw->bytes_done = 0;
3349 iorw->free_iovec = iov;
3351 req->flags |= REQ_F_NEED_CLEANUP;
3352 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3356 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3358 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3360 return io_prep_rw(req, sqe, READ);
3364 * This is our waitqueue callback handler, registered through lock_page_async()
3365 * when we initially tried to do the IO with the iocb armed our waitqueue.
3366 * This gets called when the page is unlocked, and we generally expect that to
3367 * happen when the page IO is completed and the page is now uptodate. This will
3368 * queue a task_work based retry of the operation, attempting to copy the data
3369 * again. If the latter fails because the page was NOT uptodate, then we will
3370 * do a thread based blocking retry of the operation. That's the unexpected
3373 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3374 int sync, void *arg)
3376 struct wait_page_queue *wpq;
3377 struct io_kiocb *req = wait->private;
3378 struct wait_page_key *key = arg;
3380 wpq = container_of(wait, struct wait_page_queue, wait);
3382 if (!wake_page_match(wpq, key))
3385 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3386 list_del_init(&wait->entry);
3387 io_req_task_queue(req);
3392 * This controls whether a given IO request should be armed for async page
3393 * based retry. If we return false here, the request is handed to the async
3394 * worker threads for retry. If we're doing buffered reads on a regular file,
3395 * we prepare a private wait_page_queue entry and retry the operation. This
3396 * will either succeed because the page is now uptodate and unlocked, or it
3397 * will register a callback when the page is unlocked at IO completion. Through
3398 * that callback, io_uring uses task_work to setup a retry of the operation.
3399 * That retry will attempt the buffered read again. The retry will generally
3400 * succeed, or in rare cases where it fails, we then fall back to using the
3401 * async worker threads for a blocking retry.
3403 static bool io_rw_should_retry(struct io_kiocb *req)
3405 struct io_async_rw *rw = req->async_data;
3406 struct wait_page_queue *wait = &rw->wpq;
3407 struct kiocb *kiocb = &req->rw.kiocb;
3409 /* never retry for NOWAIT, we just complete with -EAGAIN */
3410 if (req->flags & REQ_F_NOWAIT)
3413 /* Only for buffered IO */
3414 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3418 * just use poll if we can, and don't attempt if the fs doesn't
3419 * support callback based unlocks
3421 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3424 wait->wait.func = io_async_buf_func;
3425 wait->wait.private = req;
3426 wait->wait.flags = 0;
3427 INIT_LIST_HEAD(&wait->wait.entry);
3428 kiocb->ki_flags |= IOCB_WAITQ;
3429 kiocb->ki_flags &= ~IOCB_NOWAIT;
3430 kiocb->ki_waitq = wait;
3434 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3436 if (req->file->f_op->read_iter)
3437 return call_read_iter(req->file, &req->rw.kiocb, iter);
3438 else if (req->file->f_op->read)
3439 return loop_rw_iter(READ, req, iter);
3444 static bool need_read_all(struct io_kiocb *req)
3446 return req->flags & REQ_F_ISREG ||
3447 S_ISBLK(file_inode(req->file)->i_mode);
3450 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3452 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3453 struct kiocb *kiocb = &req->rw.kiocb;
3454 struct iov_iter __iter, *iter = &__iter;
3455 struct io_async_rw *rw = req->async_data;
3456 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3457 struct iov_iter_state __state, *state;
3462 state = &rw->iter_state;
3464 * We come here from an earlier attempt, restore our state to
3465 * match in case it doesn't. It's cheap enough that we don't
3466 * need to make this conditional.
3468 iov_iter_restore(iter, state);
3471 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3475 iov_iter_save_state(iter, state);
3477 req->result = iov_iter_count(iter);
3479 /* Ensure we clear previously set non-block flag */
3480 if (!force_nonblock)
3481 kiocb->ki_flags &= ~IOCB_NOWAIT;
3483 kiocb->ki_flags |= IOCB_NOWAIT;
3485 /* If the file doesn't support async, just async punt */
3486 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3487 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3488 return ret ?: -EAGAIN;
3491 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3492 if (unlikely(ret)) {
3497 ret = io_iter_do_read(req, iter);
3499 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3500 req->flags &= ~REQ_F_REISSUE;
3501 /* IOPOLL retry should happen for io-wq threads */
3502 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3504 /* no retry on NONBLOCK nor RWF_NOWAIT */
3505 if (req->flags & REQ_F_NOWAIT)
3508 } else if (ret == -EIOCBQUEUED) {
3510 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3511 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3512 /* read all, failed, already did sync or don't want to retry */
3517 * Don't depend on the iter state matching what was consumed, or being
3518 * untouched in case of error. Restore it and we'll advance it
3519 * manually if we need to.
3521 iov_iter_restore(iter, state);
3523 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3528 rw = req->async_data;
3530 * Now use our persistent iterator and state, if we aren't already.
3531 * We've restored and mapped the iter to match.
3533 if (iter != &rw->iter) {
3535 state = &rw->iter_state;
3540 * We end up here because of a partial read, either from
3541 * above or inside this loop. Advance the iter by the bytes
3542 * that were consumed.
3544 iov_iter_advance(iter, ret);
3545 if (!iov_iter_count(iter))
3547 rw->bytes_done += ret;
3548 iov_iter_save_state(iter, state);
3550 /* if we can retry, do so with the callbacks armed */
3551 if (!io_rw_should_retry(req)) {
3552 kiocb->ki_flags &= ~IOCB_WAITQ;
3557 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3558 * we get -EIOCBQUEUED, then we'll get a notification when the
3559 * desired page gets unlocked. We can also get a partial read
3560 * here, and if we do, then just retry at the new offset.
3562 ret = io_iter_do_read(req, iter);
3563 if (ret == -EIOCBQUEUED)
3565 /* we got some bytes, but not all. retry. */
3566 kiocb->ki_flags &= ~IOCB_WAITQ;
3567 iov_iter_restore(iter, state);
3570 kiocb_done(kiocb, ret, issue_flags);
3572 /* it's faster to check here then delegate to kfree */
3578 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3580 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3582 return io_prep_rw(req, sqe, WRITE);
3585 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3587 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3588 struct kiocb *kiocb = &req->rw.kiocb;
3589 struct iov_iter __iter, *iter = &__iter;
3590 struct io_async_rw *rw = req->async_data;
3591 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3592 struct iov_iter_state __state, *state;
3597 state = &rw->iter_state;
3598 iov_iter_restore(iter, state);
3601 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3605 iov_iter_save_state(iter, state);
3607 req->result = iov_iter_count(iter);
3609 /* Ensure we clear previously set non-block flag */
3610 if (!force_nonblock)
3611 kiocb->ki_flags &= ~IOCB_NOWAIT;
3613 kiocb->ki_flags |= IOCB_NOWAIT;
3615 /* If the file doesn't support async, just async punt */
3616 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3619 /* file path doesn't support NOWAIT for non-direct_IO */
3620 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3621 (req->flags & REQ_F_ISREG))
3624 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3629 * Open-code file_start_write here to grab freeze protection,
3630 * which will be released by another thread in
3631 * io_complete_rw(). Fool lockdep by telling it the lock got
3632 * released so that it doesn't complain about the held lock when
3633 * we return to userspace.
3635 if (req->flags & REQ_F_ISREG) {
3636 sb_start_write(file_inode(req->file)->i_sb);
3637 __sb_writers_release(file_inode(req->file)->i_sb,
3640 kiocb->ki_flags |= IOCB_WRITE;
3642 if (req->file->f_op->write_iter)
3643 ret2 = call_write_iter(req->file, kiocb, iter);
3644 else if (req->file->f_op->write)
3645 ret2 = loop_rw_iter(WRITE, req, iter);
3649 if (req->flags & REQ_F_REISSUE) {
3650 req->flags &= ~REQ_F_REISSUE;
3655 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3656 * retry them without IOCB_NOWAIT.
3658 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3660 /* no retry on NONBLOCK nor RWF_NOWAIT */
3661 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3663 if (!force_nonblock || ret2 != -EAGAIN) {
3664 /* IOPOLL retry should happen for io-wq threads */
3665 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3668 kiocb_done(kiocb, ret2, issue_flags);
3671 iov_iter_restore(iter, state);
3672 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3673 return ret ?: -EAGAIN;
3676 /* it's reportedly faster than delegating the null check to kfree() */
3682 static int io_renameat_prep(struct io_kiocb *req,
3683 const struct io_uring_sqe *sqe)
3685 struct io_rename *ren = &req->rename;
3686 const char __user *oldf, *newf;
3688 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3690 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3692 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3695 ren->old_dfd = READ_ONCE(sqe->fd);
3696 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3697 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3698 ren->new_dfd = READ_ONCE(sqe->len);
3699 ren->flags = READ_ONCE(sqe->rename_flags);
3701 ren->oldpath = getname(oldf);
3702 if (IS_ERR(ren->oldpath))
3703 return PTR_ERR(ren->oldpath);
3705 ren->newpath = getname(newf);
3706 if (IS_ERR(ren->newpath)) {
3707 putname(ren->oldpath);
3708 return PTR_ERR(ren->newpath);
3711 req->flags |= REQ_F_NEED_CLEANUP;
3715 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3717 struct io_rename *ren = &req->rename;
3720 if (issue_flags & IO_URING_F_NONBLOCK)
3723 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3724 ren->newpath, ren->flags);
3726 req->flags &= ~REQ_F_NEED_CLEANUP;
3729 io_req_complete(req, ret);
3733 static int io_unlinkat_prep(struct io_kiocb *req,
3734 const struct io_uring_sqe *sqe)
3736 struct io_unlink *un = &req->unlink;
3737 const char __user *fname;
3739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3741 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3744 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3747 un->dfd = READ_ONCE(sqe->fd);
3749 un->flags = READ_ONCE(sqe->unlink_flags);
3750 if (un->flags & ~AT_REMOVEDIR)
3753 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3754 un->filename = getname(fname);
3755 if (IS_ERR(un->filename))
3756 return PTR_ERR(un->filename);
3758 req->flags |= REQ_F_NEED_CLEANUP;
3762 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3764 struct io_unlink *un = &req->unlink;
3767 if (issue_flags & IO_URING_F_NONBLOCK)
3770 if (un->flags & AT_REMOVEDIR)
3771 ret = do_rmdir(un->dfd, un->filename);
3773 ret = do_unlinkat(un->dfd, un->filename);
3775 req->flags &= ~REQ_F_NEED_CLEANUP;
3778 io_req_complete(req, ret);
3782 static int io_mkdirat_prep(struct io_kiocb *req,
3783 const struct io_uring_sqe *sqe)
3785 struct io_mkdir *mkd = &req->mkdir;
3786 const char __user *fname;
3788 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3790 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3793 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3796 mkd->dfd = READ_ONCE(sqe->fd);
3797 mkd->mode = READ_ONCE(sqe->len);
3799 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3800 mkd->filename = getname(fname);
3801 if (IS_ERR(mkd->filename))
3802 return PTR_ERR(mkd->filename);
3804 req->flags |= REQ_F_NEED_CLEANUP;
3808 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3810 struct io_mkdir *mkd = &req->mkdir;
3813 if (issue_flags & IO_URING_F_NONBLOCK)
3816 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3818 req->flags &= ~REQ_F_NEED_CLEANUP;
3821 io_req_complete(req, ret);
3825 static int io_symlinkat_prep(struct io_kiocb *req,
3826 const struct io_uring_sqe *sqe)
3828 struct io_symlink *sl = &req->symlink;
3829 const char __user *oldpath, *newpath;
3831 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3833 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3836 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3839 sl->new_dfd = READ_ONCE(sqe->fd);
3840 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3841 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3843 sl->oldpath = getname(oldpath);
3844 if (IS_ERR(sl->oldpath))
3845 return PTR_ERR(sl->oldpath);
3847 sl->newpath = getname(newpath);
3848 if (IS_ERR(sl->newpath)) {
3849 putname(sl->oldpath);
3850 return PTR_ERR(sl->newpath);
3853 req->flags |= REQ_F_NEED_CLEANUP;
3857 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3859 struct io_symlink *sl = &req->symlink;
3862 if (issue_flags & IO_URING_F_NONBLOCK)
3865 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3867 req->flags &= ~REQ_F_NEED_CLEANUP;
3870 io_req_complete(req, ret);
3874 static int io_linkat_prep(struct io_kiocb *req,
3875 const struct io_uring_sqe *sqe)
3877 struct io_hardlink *lnk = &req->hardlink;
3878 const char __user *oldf, *newf;
3880 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3882 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3884 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3887 lnk->old_dfd = READ_ONCE(sqe->fd);
3888 lnk->new_dfd = READ_ONCE(sqe->len);
3889 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3890 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3891 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3893 lnk->oldpath = getname(oldf);
3894 if (IS_ERR(lnk->oldpath))
3895 return PTR_ERR(lnk->oldpath);
3897 lnk->newpath = getname(newf);
3898 if (IS_ERR(lnk->newpath)) {
3899 putname(lnk->oldpath);
3900 return PTR_ERR(lnk->newpath);
3903 req->flags |= REQ_F_NEED_CLEANUP;
3907 static int io_linkat(struct io_kiocb *req, int issue_flags)
3909 struct io_hardlink *lnk = &req->hardlink;
3912 if (issue_flags & IO_URING_F_NONBLOCK)
3915 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3916 lnk->newpath, lnk->flags);
3918 req->flags &= ~REQ_F_NEED_CLEANUP;
3921 io_req_complete(req, ret);
3925 static int io_shutdown_prep(struct io_kiocb *req,
3926 const struct io_uring_sqe *sqe)
3928 #if defined(CONFIG_NET)
3929 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3931 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3932 sqe->buf_index || sqe->splice_fd_in))
3935 req->shutdown.how = READ_ONCE(sqe->len);
3942 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3944 #if defined(CONFIG_NET)
3945 struct socket *sock;
3948 if (issue_flags & IO_URING_F_NONBLOCK)
3951 sock = sock_from_file(req->file);
3952 if (unlikely(!sock))
3955 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3958 io_req_complete(req, ret);
3965 static int __io_splice_prep(struct io_kiocb *req,
3966 const struct io_uring_sqe *sqe)
3968 struct io_splice *sp = &req->splice;
3969 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3971 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3975 sp->len = READ_ONCE(sqe->len);
3976 sp->flags = READ_ONCE(sqe->splice_flags);
3978 if (unlikely(sp->flags & ~valid_flags))
3981 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3982 (sp->flags & SPLICE_F_FD_IN_FIXED));
3985 req->flags |= REQ_F_NEED_CLEANUP;
3989 static int io_tee_prep(struct io_kiocb *req,
3990 const struct io_uring_sqe *sqe)
3992 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3994 return __io_splice_prep(req, sqe);
3997 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3999 struct io_splice *sp = &req->splice;
4000 struct file *in = sp->file_in;
4001 struct file *out = sp->file_out;
4002 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4005 if (issue_flags & IO_URING_F_NONBLOCK)
4008 ret = do_tee(in, out, sp->len, flags);
4010 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4012 req->flags &= ~REQ_F_NEED_CLEANUP;
4016 io_req_complete(req, ret);
4020 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4022 struct io_splice *sp = &req->splice;
4024 sp->off_in = READ_ONCE(sqe->splice_off_in);
4025 sp->off_out = READ_ONCE(sqe->off);
4026 return __io_splice_prep(req, sqe);
4029 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4031 struct io_splice *sp = &req->splice;
4032 struct file *in = sp->file_in;
4033 struct file *out = sp->file_out;
4034 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4035 loff_t *poff_in, *poff_out;
4038 if (issue_flags & IO_URING_F_NONBLOCK)
4041 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4042 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4045 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4047 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4049 req->flags &= ~REQ_F_NEED_CLEANUP;
4053 io_req_complete(req, ret);
4058 * IORING_OP_NOP just posts a completion event, nothing else.
4060 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4062 struct io_ring_ctx *ctx = req->ctx;
4064 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4067 __io_req_complete(req, issue_flags, 0, 0);
4071 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4073 struct io_ring_ctx *ctx = req->ctx;
4078 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4080 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4084 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4085 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4088 req->sync.off = READ_ONCE(sqe->off);
4089 req->sync.len = READ_ONCE(sqe->len);
4093 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4095 loff_t end = req->sync.off + req->sync.len;
4098 /* fsync always requires a blocking context */
4099 if (issue_flags & IO_URING_F_NONBLOCK)
4102 ret = vfs_fsync_range(req->file, req->sync.off,
4103 end > 0 ? end : LLONG_MAX,
4104 req->sync.flags & IORING_FSYNC_DATASYNC);
4107 io_req_complete(req, ret);
4111 static int io_fallocate_prep(struct io_kiocb *req,
4112 const struct io_uring_sqe *sqe)
4114 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4117 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4120 req->sync.off = READ_ONCE(sqe->off);
4121 req->sync.len = READ_ONCE(sqe->addr);
4122 req->sync.mode = READ_ONCE(sqe->len);
4126 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4130 /* fallocate always requiring blocking context */
4131 if (issue_flags & IO_URING_F_NONBLOCK)
4133 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4137 io_req_complete(req, ret);
4141 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4143 const char __user *fname;
4146 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4148 if (unlikely(sqe->ioprio || sqe->buf_index))
4150 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4153 /* open.how should be already initialised */
4154 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4155 req->open.how.flags |= O_LARGEFILE;
4157 req->open.dfd = READ_ONCE(sqe->fd);
4158 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4159 req->open.filename = getname(fname);
4160 if (IS_ERR(req->open.filename)) {
4161 ret = PTR_ERR(req->open.filename);
4162 req->open.filename = NULL;
4166 req->open.file_slot = READ_ONCE(sqe->file_index);
4167 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4170 req->open.nofile = rlimit(RLIMIT_NOFILE);
4171 req->flags |= REQ_F_NEED_CLEANUP;
4175 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4177 u64 mode = READ_ONCE(sqe->len);
4178 u64 flags = READ_ONCE(sqe->open_flags);
4180 req->open.how = build_open_how(flags, mode);
4181 return __io_openat_prep(req, sqe);
4184 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4186 struct open_how __user *how;
4190 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4191 len = READ_ONCE(sqe->len);
4192 if (len < OPEN_HOW_SIZE_VER0)
4195 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4200 return __io_openat_prep(req, sqe);
4203 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4205 struct open_flags op;
4207 bool resolve_nonblock, nonblock_set;
4208 bool fixed = !!req->open.file_slot;
4211 ret = build_open_flags(&req->open.how, &op);
4214 nonblock_set = op.open_flag & O_NONBLOCK;
4215 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4216 if (issue_flags & IO_URING_F_NONBLOCK) {
4218 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4219 * it'll always -EAGAIN
4221 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4223 op.lookup_flags |= LOOKUP_CACHED;
4224 op.open_flag |= O_NONBLOCK;
4228 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4233 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4236 * We could hang on to this 'fd' on retrying, but seems like
4237 * marginal gain for something that is now known to be a slower
4238 * path. So just put it, and we'll get a new one when we retry.
4243 ret = PTR_ERR(file);
4244 /* only retry if RESOLVE_CACHED wasn't already set by application */
4245 if (ret == -EAGAIN &&
4246 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4251 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4252 file->f_flags &= ~O_NONBLOCK;
4253 fsnotify_open(file);
4256 fd_install(ret, file);
4258 ret = io_install_fixed_file(req, file, issue_flags,
4259 req->open.file_slot - 1);
4261 putname(req->open.filename);
4262 req->flags &= ~REQ_F_NEED_CLEANUP;
4265 __io_req_complete(req, issue_flags, ret, 0);
4269 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4271 return io_openat2(req, issue_flags);
4274 static int io_remove_buffers_prep(struct io_kiocb *req,
4275 const struct io_uring_sqe *sqe)
4277 struct io_provide_buf *p = &req->pbuf;
4280 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4284 tmp = READ_ONCE(sqe->fd);
4285 if (!tmp || tmp > USHRT_MAX)
4288 memset(p, 0, sizeof(*p));
4290 p->bgid = READ_ONCE(sqe->buf_group);
4294 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4295 int bgid, unsigned nbufs)
4299 /* shouldn't happen */
4303 /* the head kbuf is the list itself */
4304 while (!list_empty(&buf->list)) {
4305 struct io_buffer *nxt;
4307 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4308 list_del(&nxt->list);
4315 xa_erase(&ctx->io_buffers, bgid);
4320 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4322 struct io_provide_buf *p = &req->pbuf;
4323 struct io_ring_ctx *ctx = req->ctx;
4324 struct io_buffer *head;
4326 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4328 io_ring_submit_lock(ctx, !force_nonblock);
4330 lockdep_assert_held(&ctx->uring_lock);
4333 head = xa_load(&ctx->io_buffers, p->bgid);
4335 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4339 /* complete before unlock, IOPOLL may need the lock */
4340 __io_req_complete(req, issue_flags, ret, 0);
4341 io_ring_submit_unlock(ctx, !force_nonblock);
4345 static int io_provide_buffers_prep(struct io_kiocb *req,
4346 const struct io_uring_sqe *sqe)
4348 unsigned long size, tmp_check;
4349 struct io_provide_buf *p = &req->pbuf;
4352 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4355 tmp = READ_ONCE(sqe->fd);
4356 if (!tmp || tmp > USHRT_MAX)
4359 p->addr = READ_ONCE(sqe->addr);
4360 p->len = READ_ONCE(sqe->len);
4362 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4365 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4368 size = (unsigned long)p->len * p->nbufs;
4369 if (!access_ok(u64_to_user_ptr(p->addr), size))
4372 p->bgid = READ_ONCE(sqe->buf_group);
4373 tmp = READ_ONCE(sqe->off);
4374 if (tmp > USHRT_MAX)
4380 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4382 struct io_buffer *buf;
4383 u64 addr = pbuf->addr;
4384 int i, bid = pbuf->bid;
4386 for (i = 0; i < pbuf->nbufs; i++) {
4387 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4392 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4397 INIT_LIST_HEAD(&buf->list);
4400 list_add_tail(&buf->list, &(*head)->list);
4404 return i ? i : -ENOMEM;
4407 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4409 struct io_provide_buf *p = &req->pbuf;
4410 struct io_ring_ctx *ctx = req->ctx;
4411 struct io_buffer *head, *list;
4413 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4415 io_ring_submit_lock(ctx, !force_nonblock);
4417 lockdep_assert_held(&ctx->uring_lock);
4419 list = head = xa_load(&ctx->io_buffers, p->bgid);
4421 ret = io_add_buffers(p, &head);
4422 if (ret >= 0 && !list) {
4423 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4425 __io_remove_buffers(ctx, head, p->bgid, -1U);
4429 /* complete before unlock, IOPOLL may need the lock */
4430 __io_req_complete(req, issue_flags, ret, 0);
4431 io_ring_submit_unlock(ctx, !force_nonblock);
4435 static int io_epoll_ctl_prep(struct io_kiocb *req,
4436 const struct io_uring_sqe *sqe)
4438 #if defined(CONFIG_EPOLL)
4439 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4441 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4444 req->epoll.epfd = READ_ONCE(sqe->fd);
4445 req->epoll.op = READ_ONCE(sqe->len);
4446 req->epoll.fd = READ_ONCE(sqe->off);
4448 if (ep_op_has_event(req->epoll.op)) {
4449 struct epoll_event __user *ev;
4451 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4452 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4462 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4464 #if defined(CONFIG_EPOLL)
4465 struct io_epoll *ie = &req->epoll;
4467 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4469 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4470 if (force_nonblock && ret == -EAGAIN)
4475 __io_req_complete(req, issue_flags, ret, 0);
4482 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4484 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4485 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4487 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4490 req->madvise.addr = READ_ONCE(sqe->addr);
4491 req->madvise.len = READ_ONCE(sqe->len);
4492 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4499 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4501 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4502 struct io_madvise *ma = &req->madvise;
4505 if (issue_flags & IO_URING_F_NONBLOCK)
4508 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4511 io_req_complete(req, ret);
4518 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4520 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4522 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4525 req->fadvise.offset = READ_ONCE(sqe->off);
4526 req->fadvise.len = READ_ONCE(sqe->len);
4527 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4531 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4533 struct io_fadvise *fa = &req->fadvise;
4536 if (issue_flags & IO_URING_F_NONBLOCK) {
4537 switch (fa->advice) {
4538 case POSIX_FADV_NORMAL:
4539 case POSIX_FADV_RANDOM:
4540 case POSIX_FADV_SEQUENTIAL:
4547 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4550 __io_req_complete(req, issue_flags, ret, 0);
4554 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4556 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4558 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4560 if (req->flags & REQ_F_FIXED_FILE)
4563 req->statx.dfd = READ_ONCE(sqe->fd);
4564 req->statx.mask = READ_ONCE(sqe->len);
4565 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4566 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4567 req->statx.flags = READ_ONCE(sqe->statx_flags);
4572 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4574 struct io_statx *ctx = &req->statx;
4577 if (issue_flags & IO_URING_F_NONBLOCK)
4580 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4585 io_req_complete(req, ret);
4589 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4591 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4593 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4594 sqe->rw_flags || sqe->buf_index)
4596 if (req->flags & REQ_F_FIXED_FILE)
4599 req->close.fd = READ_ONCE(sqe->fd);
4600 req->close.file_slot = READ_ONCE(sqe->file_index);
4601 if (req->close.file_slot && req->close.fd)
4607 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4609 struct files_struct *files = current->files;
4610 struct io_close *close = &req->close;
4611 struct fdtable *fdt;
4612 struct file *file = NULL;
4615 if (req->close.file_slot) {
4616 ret = io_close_fixed(req, issue_flags);
4620 spin_lock(&files->file_lock);
4621 fdt = files_fdtable(files);
4622 if (close->fd >= fdt->max_fds) {
4623 spin_unlock(&files->file_lock);
4626 file = fdt->fd[close->fd];
4627 if (!file || file->f_op == &io_uring_fops) {
4628 spin_unlock(&files->file_lock);
4633 /* if the file has a flush method, be safe and punt to async */
4634 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4635 spin_unlock(&files->file_lock);
4639 ret = __close_fd_get_file(close->fd, &file);
4640 spin_unlock(&files->file_lock);
4647 /* No ->flush() or already async, safely close from here */
4648 ret = filp_close(file, current->files);
4654 __io_req_complete(req, issue_flags, ret, 0);
4658 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4660 struct io_ring_ctx *ctx = req->ctx;
4662 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4664 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4668 req->sync.off = READ_ONCE(sqe->off);
4669 req->sync.len = READ_ONCE(sqe->len);
4670 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4674 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4678 /* sync_file_range always requires a blocking context */
4679 if (issue_flags & IO_URING_F_NONBLOCK)
4682 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4686 io_req_complete(req, ret);
4690 #if defined(CONFIG_NET)
4691 static int io_setup_async_msg(struct io_kiocb *req,
4692 struct io_async_msghdr *kmsg)
4694 struct io_async_msghdr *async_msg = req->async_data;
4698 if (io_alloc_async_data(req)) {
4699 kfree(kmsg->free_iov);
4702 async_msg = req->async_data;
4703 req->flags |= REQ_F_NEED_CLEANUP;
4704 memcpy(async_msg, kmsg, sizeof(*kmsg));
4705 async_msg->msg.msg_name = &async_msg->addr;
4706 /* if were using fast_iov, set it to the new one */
4707 if (!async_msg->free_iov)
4708 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4713 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4714 struct io_async_msghdr *iomsg)
4716 iomsg->msg.msg_name = &iomsg->addr;
4717 iomsg->free_iov = iomsg->fast_iov;
4718 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4719 req->sr_msg.msg_flags, &iomsg->free_iov);
4722 static int io_sendmsg_prep_async(struct io_kiocb *req)
4726 ret = io_sendmsg_copy_hdr(req, req->async_data);
4728 req->flags |= REQ_F_NEED_CLEANUP;
4732 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4734 struct io_sr_msg *sr = &req->sr_msg;
4736 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4739 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4740 sr->len = READ_ONCE(sqe->len);
4741 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4742 if (sr->msg_flags & MSG_DONTWAIT)
4743 req->flags |= REQ_F_NOWAIT;
4745 #ifdef CONFIG_COMPAT
4746 if (req->ctx->compat)
4747 sr->msg_flags |= MSG_CMSG_COMPAT;
4752 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4754 struct io_async_msghdr iomsg, *kmsg;
4755 struct socket *sock;
4760 sock = sock_from_file(req->file);
4761 if (unlikely(!sock))
4764 kmsg = req->async_data;
4766 ret = io_sendmsg_copy_hdr(req, &iomsg);
4772 flags = req->sr_msg.msg_flags;
4773 if (issue_flags & IO_URING_F_NONBLOCK)
4774 flags |= MSG_DONTWAIT;
4775 if (flags & MSG_WAITALL)
4776 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4778 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4779 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4780 return io_setup_async_msg(req, kmsg);
4781 if (ret == -ERESTARTSYS)
4784 /* fast path, check for non-NULL to avoid function call */
4786 kfree(kmsg->free_iov);
4787 req->flags &= ~REQ_F_NEED_CLEANUP;
4790 __io_req_complete(req, issue_flags, ret, 0);
4794 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4796 struct io_sr_msg *sr = &req->sr_msg;
4799 struct socket *sock;
4804 sock = sock_from_file(req->file);
4805 if (unlikely(!sock))
4808 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4812 msg.msg_name = NULL;
4813 msg.msg_control = NULL;
4814 msg.msg_controllen = 0;
4815 msg.msg_namelen = 0;
4817 flags = req->sr_msg.msg_flags;
4818 if (issue_flags & IO_URING_F_NONBLOCK)
4819 flags |= MSG_DONTWAIT;
4820 if (flags & MSG_WAITALL)
4821 min_ret = iov_iter_count(&msg.msg_iter);
4823 msg.msg_flags = flags;
4824 ret = sock_sendmsg(sock, &msg);
4825 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4827 if (ret == -ERESTARTSYS)
4832 __io_req_complete(req, issue_flags, ret, 0);
4836 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4837 struct io_async_msghdr *iomsg)
4839 struct io_sr_msg *sr = &req->sr_msg;
4840 struct iovec __user *uiov;
4844 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4845 &iomsg->uaddr, &uiov, &iov_len);
4849 if (req->flags & REQ_F_BUFFER_SELECT) {
4852 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4854 sr->len = iomsg->fast_iov[0].iov_len;
4855 iomsg->free_iov = NULL;
4857 iomsg->free_iov = iomsg->fast_iov;
4858 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4859 &iomsg->free_iov, &iomsg->msg.msg_iter,
4868 #ifdef CONFIG_COMPAT
4869 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4870 struct io_async_msghdr *iomsg)
4872 struct io_sr_msg *sr = &req->sr_msg;
4873 struct compat_iovec __user *uiov;
4878 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4883 uiov = compat_ptr(ptr);
4884 if (req->flags & REQ_F_BUFFER_SELECT) {
4885 compat_ssize_t clen;
4889 if (!access_ok(uiov, sizeof(*uiov)))
4891 if (__get_user(clen, &uiov->iov_len))
4896 iomsg->free_iov = NULL;
4898 iomsg->free_iov = iomsg->fast_iov;
4899 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4900 UIO_FASTIOV, &iomsg->free_iov,
4901 &iomsg->msg.msg_iter, true);
4910 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4911 struct io_async_msghdr *iomsg)
4913 iomsg->msg.msg_name = &iomsg->addr;
4915 #ifdef CONFIG_COMPAT
4916 if (req->ctx->compat)
4917 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4920 return __io_recvmsg_copy_hdr(req, iomsg);
4923 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4926 struct io_sr_msg *sr = &req->sr_msg;
4927 struct io_buffer *kbuf;
4929 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4934 req->flags |= REQ_F_BUFFER_SELECTED;
4938 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4940 return io_put_kbuf(req, req->sr_msg.kbuf);
4943 static int io_recvmsg_prep_async(struct io_kiocb *req)
4947 ret = io_recvmsg_copy_hdr(req, req->async_data);
4949 req->flags |= REQ_F_NEED_CLEANUP;
4953 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4955 struct io_sr_msg *sr = &req->sr_msg;
4957 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4960 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4961 sr->len = READ_ONCE(sqe->len);
4962 sr->bgid = READ_ONCE(sqe->buf_group);
4963 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4964 if (sr->msg_flags & MSG_DONTWAIT)
4965 req->flags |= REQ_F_NOWAIT;
4967 #ifdef CONFIG_COMPAT
4968 if (req->ctx->compat)
4969 sr->msg_flags |= MSG_CMSG_COMPAT;
4974 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4976 struct io_async_msghdr iomsg, *kmsg;
4977 struct socket *sock;
4978 struct io_buffer *kbuf;
4981 int ret, cflags = 0;
4982 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4984 sock = sock_from_file(req->file);
4985 if (unlikely(!sock))
4988 kmsg = req->async_data;
4990 ret = io_recvmsg_copy_hdr(req, &iomsg);
4996 if (req->flags & REQ_F_BUFFER_SELECT) {
4997 kbuf = io_recv_buffer_select(req, !force_nonblock);
4999 return PTR_ERR(kbuf);
5000 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5001 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5002 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5003 1, req->sr_msg.len);
5006 flags = req->sr_msg.msg_flags;
5008 flags |= MSG_DONTWAIT;
5009 if (flags & MSG_WAITALL)
5010 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5012 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5013 kmsg->uaddr, flags);
5014 if (force_nonblock && ret == -EAGAIN)
5015 return io_setup_async_msg(req, kmsg);
5016 if (ret == -ERESTARTSYS)
5019 if (req->flags & REQ_F_BUFFER_SELECTED)
5020 cflags = io_put_recv_kbuf(req);
5021 /* fast path, check for non-NULL to avoid function call */
5023 kfree(kmsg->free_iov);
5024 req->flags &= ~REQ_F_NEED_CLEANUP;
5025 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5027 __io_req_complete(req, issue_flags, ret, cflags);
5031 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5033 struct io_buffer *kbuf;
5034 struct io_sr_msg *sr = &req->sr_msg;
5036 void __user *buf = sr->buf;
5037 struct socket *sock;
5041 int ret, cflags = 0;
5042 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5044 sock = sock_from_file(req->file);
5045 if (unlikely(!sock))
5048 if (req->flags & REQ_F_BUFFER_SELECT) {
5049 kbuf = io_recv_buffer_select(req, !force_nonblock);
5051 return PTR_ERR(kbuf);
5052 buf = u64_to_user_ptr(kbuf->addr);
5055 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5059 msg.msg_name = NULL;
5060 msg.msg_control = NULL;
5061 msg.msg_controllen = 0;
5062 msg.msg_namelen = 0;
5063 msg.msg_iocb = NULL;
5066 flags = req->sr_msg.msg_flags;
5068 flags |= MSG_DONTWAIT;
5069 if (flags & MSG_WAITALL)
5070 min_ret = iov_iter_count(&msg.msg_iter);
5072 ret = sock_recvmsg(sock, &msg, flags);
5073 if (force_nonblock && ret == -EAGAIN)
5075 if (ret == -ERESTARTSYS)
5078 if (req->flags & REQ_F_BUFFER_SELECTED)
5079 cflags = io_put_recv_kbuf(req);
5080 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5082 __io_req_complete(req, issue_flags, ret, cflags);
5086 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5088 struct io_accept *accept = &req->accept;
5090 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5092 if (sqe->ioprio || sqe->len || sqe->buf_index)
5095 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5096 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5097 accept->flags = READ_ONCE(sqe->accept_flags);
5098 accept->nofile = rlimit(RLIMIT_NOFILE);
5100 accept->file_slot = READ_ONCE(sqe->file_index);
5101 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5102 (accept->flags & SOCK_CLOEXEC)))
5104 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5106 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5107 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5111 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5113 struct io_accept *accept = &req->accept;
5114 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5115 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5116 bool fixed = !!accept->file_slot;
5120 if (req->file->f_flags & O_NONBLOCK)
5121 req->flags |= REQ_F_NOWAIT;
5124 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5125 if (unlikely(fd < 0))
5128 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5133 ret = PTR_ERR(file);
5134 if (ret == -EAGAIN && force_nonblock)
5136 if (ret == -ERESTARTSYS)
5139 } else if (!fixed) {
5140 fd_install(fd, file);
5143 ret = io_install_fixed_file(req, file, issue_flags,
5144 accept->file_slot - 1);
5146 __io_req_complete(req, issue_flags, ret, 0);
5150 static int io_connect_prep_async(struct io_kiocb *req)
5152 struct io_async_connect *io = req->async_data;
5153 struct io_connect *conn = &req->connect;
5155 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5158 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5160 struct io_connect *conn = &req->connect;
5162 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5164 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5168 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5169 conn->addr_len = READ_ONCE(sqe->addr2);
5173 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5175 struct io_async_connect __io, *io;
5176 unsigned file_flags;
5178 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5180 if (req->async_data) {
5181 io = req->async_data;
5183 ret = move_addr_to_kernel(req->connect.addr,
5184 req->connect.addr_len,
5191 file_flags = force_nonblock ? O_NONBLOCK : 0;
5193 ret = __sys_connect_file(req->file, &io->address,
5194 req->connect.addr_len, file_flags);
5195 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5196 if (req->async_data)
5198 if (io_alloc_async_data(req)) {
5202 memcpy(req->async_data, &__io, sizeof(__io));
5205 if (ret == -ERESTARTSYS)
5210 __io_req_complete(req, issue_flags, ret, 0);
5213 #else /* !CONFIG_NET */
5214 #define IO_NETOP_FN(op) \
5215 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5217 return -EOPNOTSUPP; \
5220 #define IO_NETOP_PREP(op) \
5222 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5224 return -EOPNOTSUPP; \
5227 #define IO_NETOP_PREP_ASYNC(op) \
5229 static int io_##op##_prep_async(struct io_kiocb *req) \
5231 return -EOPNOTSUPP; \
5234 IO_NETOP_PREP_ASYNC(sendmsg);
5235 IO_NETOP_PREP_ASYNC(recvmsg);
5236 IO_NETOP_PREP_ASYNC(connect);
5237 IO_NETOP_PREP(accept);
5240 #endif /* CONFIG_NET */
5242 struct io_poll_table {
5243 struct poll_table_struct pt;
5244 struct io_kiocb *req;
5249 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5250 __poll_t mask, io_req_tw_func_t func)
5252 /* for instances that support it check for an event match first: */
5253 if (mask && !(mask & poll->events))
5256 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5258 list_del_init(&poll->wait.entry);
5261 req->io_task_work.func = func;
5264 * If this fails, then the task is exiting. When a task exits, the
5265 * work gets canceled, so just cancel this request as well instead
5266 * of executing it. We can't safely execute it anyway, as we may not
5267 * have the needed state needed for it anyway.
5269 io_req_task_work_add(req);
5273 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5274 __acquires(&req->ctx->completion_lock)
5276 struct io_ring_ctx *ctx = req->ctx;
5278 /* req->task == current here, checking PF_EXITING is safe */
5279 if (unlikely(req->task->flags & PF_EXITING))
5280 WRITE_ONCE(poll->canceled, true);
5282 if (!req->result && !READ_ONCE(poll->canceled)) {
5283 struct poll_table_struct pt = { ._key = poll->events };
5285 req->result = vfs_poll(req->file, &pt) & poll->events;
5288 spin_lock(&ctx->completion_lock);
5289 if (!req->result && !READ_ONCE(poll->canceled)) {
5290 add_wait_queue(poll->head, &poll->wait);
5297 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5299 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5300 if (req->opcode == IORING_OP_POLL_ADD)
5301 return req->async_data;
5302 return req->apoll->double_poll;
5305 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5307 if (req->opcode == IORING_OP_POLL_ADD)
5309 return &req->apoll->poll;
5312 static void io_poll_remove_double(struct io_kiocb *req)
5313 __must_hold(&req->ctx->completion_lock)
5315 struct io_poll_iocb *poll = io_poll_get_double(req);
5317 lockdep_assert_held(&req->ctx->completion_lock);
5319 if (poll && poll->head) {
5320 struct wait_queue_head *head = poll->head;
5322 spin_lock_irq(&head->lock);
5323 list_del_init(&poll->wait.entry);
5324 if (poll->wait.private)
5327 spin_unlock_irq(&head->lock);
5331 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5332 __must_hold(&req->ctx->completion_lock)
5334 struct io_ring_ctx *ctx = req->ctx;
5335 unsigned flags = IORING_CQE_F_MORE;
5338 if (READ_ONCE(req->poll.canceled)) {
5340 req->poll.events |= EPOLLONESHOT;
5342 error = mangle_poll(mask);
5344 if (req->poll.events & EPOLLONESHOT)
5346 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5347 req->poll.events |= EPOLLONESHOT;
5350 if (flags & IORING_CQE_F_MORE)
5353 return !(flags & IORING_CQE_F_MORE);
5356 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5357 __must_hold(&req->ctx->completion_lock)
5361 done = __io_poll_complete(req, mask);
5362 io_commit_cqring(req->ctx);
5366 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5368 struct io_ring_ctx *ctx = req->ctx;
5369 struct io_kiocb *nxt;
5371 if (io_poll_rewait(req, &req->poll)) {
5372 spin_unlock(&ctx->completion_lock);
5376 if (req->poll.done) {
5377 spin_unlock(&ctx->completion_lock);
5380 done = __io_poll_complete(req, req->result);
5382 io_poll_remove_double(req);
5383 hash_del(&req->hash_node);
5384 req->poll.done = true;
5387 add_wait_queue(req->poll.head, &req->poll.wait);
5389 io_commit_cqring(ctx);
5390 spin_unlock(&ctx->completion_lock);
5391 io_cqring_ev_posted(ctx);
5394 nxt = io_put_req_find_next(req);
5396 io_req_task_submit(nxt, locked);
5401 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5402 int sync, void *key)
5404 struct io_kiocb *req = wait->private;
5405 struct io_poll_iocb *poll = io_poll_get_single(req);
5406 __poll_t mask = key_to_poll(key);
5407 unsigned long flags;
5409 /* for instances that support it check for an event match first: */
5410 if (mask && !(mask & poll->events))
5412 if (!(poll->events & EPOLLONESHOT))
5413 return poll->wait.func(&poll->wait, mode, sync, key);
5415 list_del_init(&wait->entry);
5420 spin_lock_irqsave(&poll->head->lock, flags);
5421 done = list_empty(&poll->wait.entry);
5423 list_del_init(&poll->wait.entry);
5424 /* make sure double remove sees this as being gone */
5425 wait->private = NULL;
5426 spin_unlock_irqrestore(&poll->head->lock, flags);
5428 /* use wait func handler, so it matches the rq type */
5429 poll->wait.func(&poll->wait, mode, sync, key);
5436 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5437 wait_queue_func_t wake_func)
5441 poll->canceled = false;
5442 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5443 /* mask in events that we always want/need */
5444 poll->events = events | IO_POLL_UNMASK;
5445 INIT_LIST_HEAD(&poll->wait.entry);
5446 init_waitqueue_func_entry(&poll->wait, wake_func);
5449 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5450 struct wait_queue_head *head,
5451 struct io_poll_iocb **poll_ptr)
5453 struct io_kiocb *req = pt->req;
5456 * The file being polled uses multiple waitqueues for poll handling
5457 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5460 if (unlikely(pt->nr_entries)) {
5461 struct io_poll_iocb *poll_one = poll;
5463 /* double add on the same waitqueue head, ignore */
5464 if (poll_one->head == head)
5466 /* already have a 2nd entry, fail a third attempt */
5468 if ((*poll_ptr)->head == head)
5470 pt->error = -EINVAL;
5474 * Can't handle multishot for double wait for now, turn it
5475 * into one-shot mode.
5477 if (!(poll_one->events & EPOLLONESHOT))
5478 poll_one->events |= EPOLLONESHOT;
5479 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5481 pt->error = -ENOMEM;
5484 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5486 poll->wait.private = req;
5493 if (poll->events & EPOLLEXCLUSIVE)
5494 add_wait_queue_exclusive(head, &poll->wait);
5496 add_wait_queue(head, &poll->wait);
5499 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5500 struct poll_table_struct *p)
5502 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5503 struct async_poll *apoll = pt->req->apoll;
5505 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5508 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5510 struct async_poll *apoll = req->apoll;
5511 struct io_ring_ctx *ctx = req->ctx;
5513 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5515 if (io_poll_rewait(req, &apoll->poll)) {
5516 spin_unlock(&ctx->completion_lock);
5520 hash_del(&req->hash_node);
5521 io_poll_remove_double(req);
5522 apoll->poll.done = true;
5523 spin_unlock(&ctx->completion_lock);
5525 if (!READ_ONCE(apoll->poll.canceled))
5526 io_req_task_submit(req, locked);
5528 io_req_complete_failed(req, -ECANCELED);
5531 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5534 struct io_kiocb *req = wait->private;
5535 struct io_poll_iocb *poll = &req->apoll->poll;
5537 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5540 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5543 static void io_poll_req_insert(struct io_kiocb *req)
5545 struct io_ring_ctx *ctx = req->ctx;
5546 struct hlist_head *list;
5548 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5549 hlist_add_head(&req->hash_node, list);
5552 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5553 struct io_poll_iocb *poll,
5554 struct io_poll_table *ipt, __poll_t mask,
5555 wait_queue_func_t wake_func)
5556 __acquires(&ctx->completion_lock)
5558 struct io_ring_ctx *ctx = req->ctx;
5559 bool cancel = false;
5561 INIT_HLIST_NODE(&req->hash_node);
5562 io_init_poll_iocb(poll, mask, wake_func);
5563 poll->file = req->file;
5564 poll->wait.private = req;
5566 ipt->pt._key = mask;
5569 ipt->nr_entries = 0;
5571 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5572 if (unlikely(!ipt->nr_entries) && !ipt->error)
5573 ipt->error = -EINVAL;
5575 spin_lock(&ctx->completion_lock);
5576 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5577 io_poll_remove_double(req);
5578 if (likely(poll->head)) {
5579 spin_lock_irq(&poll->head->lock);
5580 if (unlikely(list_empty(&poll->wait.entry))) {
5586 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5587 list_del_init(&poll->wait.entry);
5589 WRITE_ONCE(poll->canceled, true);
5590 else if (!poll->done) /* actually waiting for an event */
5591 io_poll_req_insert(req);
5592 spin_unlock_irq(&poll->head->lock);
5604 static int io_arm_poll_handler(struct io_kiocb *req)
5606 const struct io_op_def *def = &io_op_defs[req->opcode];
5607 struct io_ring_ctx *ctx = req->ctx;
5608 struct async_poll *apoll;
5609 struct io_poll_table ipt;
5610 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5613 if (!req->file || !file_can_poll(req->file))
5614 return IO_APOLL_ABORTED;
5615 if (req->flags & REQ_F_POLLED)
5616 return IO_APOLL_ABORTED;
5617 if (!def->pollin && !def->pollout)
5618 return IO_APOLL_ABORTED;
5622 mask |= POLLIN | POLLRDNORM;
5624 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5625 if ((req->opcode == IORING_OP_RECVMSG) &&
5626 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5630 mask |= POLLOUT | POLLWRNORM;
5633 /* if we can't nonblock try, then no point in arming a poll handler */
5634 if (!io_file_supports_nowait(req, rw))
5635 return IO_APOLL_ABORTED;
5637 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5638 if (unlikely(!apoll))
5639 return IO_APOLL_ABORTED;
5640 apoll->double_poll = NULL;
5642 req->flags |= REQ_F_POLLED;
5643 ipt.pt._qproc = io_async_queue_proc;
5644 io_req_set_refcount(req);
5646 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5648 spin_unlock(&ctx->completion_lock);
5649 if (ret || ipt.error)
5650 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5652 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5653 mask, apoll->poll.events);
5657 static bool __io_poll_remove_one(struct io_kiocb *req,
5658 struct io_poll_iocb *poll, bool do_cancel)
5659 __must_hold(&req->ctx->completion_lock)
5661 bool do_complete = false;
5665 spin_lock_irq(&poll->head->lock);
5667 WRITE_ONCE(poll->canceled, true);
5668 if (!list_empty(&poll->wait.entry)) {
5669 list_del_init(&poll->wait.entry);
5672 spin_unlock_irq(&poll->head->lock);
5673 hash_del(&req->hash_node);
5677 static bool io_poll_remove_one(struct io_kiocb *req)
5678 __must_hold(&req->ctx->completion_lock)
5682 io_poll_remove_double(req);
5683 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5686 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5687 io_commit_cqring(req->ctx);
5689 io_put_req_deferred(req);
5695 * Returns true if we found and killed one or more poll requests
5697 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5700 struct hlist_node *tmp;
5701 struct io_kiocb *req;
5704 spin_lock(&ctx->completion_lock);
5705 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5706 struct hlist_head *list;
5708 list = &ctx->cancel_hash[i];
5709 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5710 if (io_match_task(req, tsk, cancel_all))
5711 posted += io_poll_remove_one(req);
5714 spin_unlock(&ctx->completion_lock);
5717 io_cqring_ev_posted(ctx);
5722 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5724 __must_hold(&ctx->completion_lock)
5726 struct hlist_head *list;
5727 struct io_kiocb *req;
5729 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5730 hlist_for_each_entry(req, list, hash_node) {
5731 if (sqe_addr != req->user_data)
5733 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5740 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5742 __must_hold(&ctx->completion_lock)
5744 struct io_kiocb *req;
5746 req = io_poll_find(ctx, sqe_addr, poll_only);
5749 if (io_poll_remove_one(req))
5755 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5760 events = READ_ONCE(sqe->poll32_events);
5762 events = swahw32(events);
5764 if (!(flags & IORING_POLL_ADD_MULTI))
5765 events |= EPOLLONESHOT;
5766 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5769 static int io_poll_update_prep(struct io_kiocb *req,
5770 const struct io_uring_sqe *sqe)
5772 struct io_poll_update *upd = &req->poll_update;
5775 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5777 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5779 flags = READ_ONCE(sqe->len);
5780 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5781 IORING_POLL_ADD_MULTI))
5783 /* meaningless without update */
5784 if (flags == IORING_POLL_ADD_MULTI)
5787 upd->old_user_data = READ_ONCE(sqe->addr);
5788 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5789 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5791 upd->new_user_data = READ_ONCE(sqe->off);
5792 if (!upd->update_user_data && upd->new_user_data)
5794 if (upd->update_events)
5795 upd->events = io_poll_parse_events(sqe, flags);
5796 else if (sqe->poll32_events)
5802 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5805 struct io_kiocb *req = wait->private;
5806 struct io_poll_iocb *poll = &req->poll;
5808 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5811 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5812 struct poll_table_struct *p)
5814 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5816 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5819 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5821 struct io_poll_iocb *poll = &req->poll;
5824 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5826 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5828 flags = READ_ONCE(sqe->len);
5829 if (flags & ~IORING_POLL_ADD_MULTI)
5832 io_req_set_refcount(req);
5833 poll->events = io_poll_parse_events(sqe, flags);
5837 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5839 struct io_poll_iocb *poll = &req->poll;
5840 struct io_ring_ctx *ctx = req->ctx;
5841 struct io_poll_table ipt;
5845 ipt.pt._qproc = io_poll_queue_proc;
5847 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5850 if (mask) { /* no async, we'd stolen it */
5852 done = io_poll_complete(req, mask);
5854 spin_unlock(&ctx->completion_lock);
5857 io_cqring_ev_posted(ctx);
5864 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5866 struct io_ring_ctx *ctx = req->ctx;
5867 struct io_kiocb *preq;
5871 spin_lock(&ctx->completion_lock);
5872 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5878 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5880 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5885 * Don't allow racy completion with singleshot, as we cannot safely
5886 * update those. For multishot, if we're racing with completion, just
5887 * let completion re-add it.
5889 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5890 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5894 /* we now have a detached poll request. reissue. */
5898 spin_unlock(&ctx->completion_lock);
5900 io_req_complete(req, ret);
5903 /* only mask one event flags, keep behavior flags */
5904 if (req->poll_update.update_events) {
5905 preq->poll.events &= ~0xffff;
5906 preq->poll.events |= req->poll_update.events & 0xffff;
5907 preq->poll.events |= IO_POLL_UNMASK;
5909 if (req->poll_update.update_user_data)
5910 preq->user_data = req->poll_update.new_user_data;
5911 spin_unlock(&ctx->completion_lock);
5913 /* complete update request, we're done with it */
5914 io_req_complete(req, ret);
5917 ret = io_poll_add(preq, issue_flags);
5920 io_req_complete(preq, ret);
5926 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5929 io_req_complete_post(req, -ETIME, 0);
5932 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5934 struct io_timeout_data *data = container_of(timer,
5935 struct io_timeout_data, timer);
5936 struct io_kiocb *req = data->req;
5937 struct io_ring_ctx *ctx = req->ctx;
5938 unsigned long flags;
5940 spin_lock_irqsave(&ctx->timeout_lock, flags);
5941 list_del_init(&req->timeout.list);
5942 atomic_set(&req->ctx->cq_timeouts,
5943 atomic_read(&req->ctx->cq_timeouts) + 1);
5944 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5946 req->io_task_work.func = io_req_task_timeout;
5947 io_req_task_work_add(req);
5948 return HRTIMER_NORESTART;
5951 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5953 __must_hold(&ctx->timeout_lock)
5955 struct io_timeout_data *io;
5956 struct io_kiocb *req;
5959 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5960 found = user_data == req->user_data;
5965 return ERR_PTR(-ENOENT);
5967 io = req->async_data;
5968 if (hrtimer_try_to_cancel(&io->timer) == -1)
5969 return ERR_PTR(-EALREADY);
5970 list_del_init(&req->timeout.list);
5974 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5975 __must_hold(&ctx->completion_lock)
5976 __must_hold(&ctx->timeout_lock)
5978 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5981 return PTR_ERR(req);
5984 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5985 io_put_req_deferred(req);
5989 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5991 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5992 case IORING_TIMEOUT_BOOTTIME:
5993 return CLOCK_BOOTTIME;
5994 case IORING_TIMEOUT_REALTIME:
5995 return CLOCK_REALTIME;
5997 /* can't happen, vetted at prep time */
6001 return CLOCK_MONOTONIC;
6005 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6006 struct timespec64 *ts, enum hrtimer_mode mode)
6007 __must_hold(&ctx->timeout_lock)
6009 struct io_timeout_data *io;
6010 struct io_kiocb *req;
6013 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6014 found = user_data == req->user_data;
6021 io = req->async_data;
6022 if (hrtimer_try_to_cancel(&io->timer) == -1)
6024 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6025 io->timer.function = io_link_timeout_fn;
6026 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6030 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6031 struct timespec64 *ts, enum hrtimer_mode mode)
6032 __must_hold(&ctx->timeout_lock)
6034 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6035 struct io_timeout_data *data;
6038 return PTR_ERR(req);
6040 req->timeout.off = 0; /* noseq */
6041 data = req->async_data;
6042 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6043 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6044 data->timer.function = io_timeout_fn;
6045 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6049 static int io_timeout_remove_prep(struct io_kiocb *req,
6050 const struct io_uring_sqe *sqe)
6052 struct io_timeout_rem *tr = &req->timeout_rem;
6054 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6056 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6058 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6061 tr->ltimeout = false;
6062 tr->addr = READ_ONCE(sqe->addr);
6063 tr->flags = READ_ONCE(sqe->timeout_flags);
6064 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6065 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6067 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6068 tr->ltimeout = true;
6069 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6071 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6073 } else if (tr->flags) {
6074 /* timeout removal doesn't support flags */
6081 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6083 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6088 * Remove or update an existing timeout command
6090 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6092 struct io_timeout_rem *tr = &req->timeout_rem;
6093 struct io_ring_ctx *ctx = req->ctx;
6096 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6097 spin_lock(&ctx->completion_lock);
6098 spin_lock_irq(&ctx->timeout_lock);
6099 ret = io_timeout_cancel(ctx, tr->addr);
6100 spin_unlock_irq(&ctx->timeout_lock);
6101 spin_unlock(&ctx->completion_lock);
6103 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6105 spin_lock_irq(&ctx->timeout_lock);
6107 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6109 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6110 spin_unlock_irq(&ctx->timeout_lock);
6115 io_req_complete_post(req, ret, 0);
6119 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6120 bool is_timeout_link)
6122 struct io_timeout_data *data;
6124 u32 off = READ_ONCE(sqe->off);
6126 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6128 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6131 if (off && is_timeout_link)
6133 flags = READ_ONCE(sqe->timeout_flags);
6134 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6136 /* more than one clock specified is invalid, obviously */
6137 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6140 INIT_LIST_HEAD(&req->timeout.list);
6141 req->timeout.off = off;
6142 if (unlikely(off && !req->ctx->off_timeout_used))
6143 req->ctx->off_timeout_used = true;
6145 if (!req->async_data && io_alloc_async_data(req))
6148 data = req->async_data;
6150 data->flags = flags;
6152 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6155 data->mode = io_translate_timeout_mode(flags);
6156 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6158 if (is_timeout_link) {
6159 struct io_submit_link *link = &req->ctx->submit_state.link;
6163 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6165 req->timeout.head = link->last;
6166 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6171 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6173 struct io_ring_ctx *ctx = req->ctx;
6174 struct io_timeout_data *data = req->async_data;
6175 struct list_head *entry;
6176 u32 tail, off = req->timeout.off;
6178 spin_lock_irq(&ctx->timeout_lock);
6181 * sqe->off holds how many events that need to occur for this
6182 * timeout event to be satisfied. If it isn't set, then this is
6183 * a pure timeout request, sequence isn't used.
6185 if (io_is_timeout_noseq(req)) {
6186 entry = ctx->timeout_list.prev;
6190 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6191 req->timeout.target_seq = tail + off;
6193 /* Update the last seq here in case io_flush_timeouts() hasn't.
6194 * This is safe because ->completion_lock is held, and submissions
6195 * and completions are never mixed in the same ->completion_lock section.
6197 ctx->cq_last_tm_flush = tail;
6200 * Insertion sort, ensuring the first entry in the list is always
6201 * the one we need first.
6203 list_for_each_prev(entry, &ctx->timeout_list) {
6204 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6207 if (io_is_timeout_noseq(nxt))
6209 /* nxt.seq is behind @tail, otherwise would've been completed */
6210 if (off >= nxt->timeout.target_seq - tail)
6214 list_add(&req->timeout.list, entry);
6215 data->timer.function = io_timeout_fn;
6216 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6217 spin_unlock_irq(&ctx->timeout_lock);
6221 struct io_cancel_data {
6222 struct io_ring_ctx *ctx;
6226 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6228 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6229 struct io_cancel_data *cd = data;
6231 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6234 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6235 struct io_ring_ctx *ctx)
6237 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6238 enum io_wq_cancel cancel_ret;
6241 if (!tctx || !tctx->io_wq)
6244 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6245 switch (cancel_ret) {
6246 case IO_WQ_CANCEL_OK:
6249 case IO_WQ_CANCEL_RUNNING:
6252 case IO_WQ_CANCEL_NOTFOUND:
6260 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6262 struct io_ring_ctx *ctx = req->ctx;
6265 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6267 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6271 spin_lock(&ctx->completion_lock);
6272 spin_lock_irq(&ctx->timeout_lock);
6273 ret = io_timeout_cancel(ctx, sqe_addr);
6274 spin_unlock_irq(&ctx->timeout_lock);
6277 ret = io_poll_cancel(ctx, sqe_addr, false);
6279 spin_unlock(&ctx->completion_lock);
6283 static int io_async_cancel_prep(struct io_kiocb *req,
6284 const struct io_uring_sqe *sqe)
6286 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6288 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6290 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6294 req->cancel.addr = READ_ONCE(sqe->addr);
6298 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6300 struct io_ring_ctx *ctx = req->ctx;
6301 u64 sqe_addr = req->cancel.addr;
6302 struct io_tctx_node *node;
6305 ret = io_try_cancel_userdata(req, sqe_addr);
6309 /* slow path, try all io-wq's */
6310 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6312 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6313 struct io_uring_task *tctx = node->task->io_uring;
6315 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6319 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6323 io_req_complete_post(req, ret, 0);
6327 static int io_rsrc_update_prep(struct io_kiocb *req,
6328 const struct io_uring_sqe *sqe)
6330 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6332 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6335 req->rsrc_update.offset = READ_ONCE(sqe->off);
6336 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6337 if (!req->rsrc_update.nr_args)
6339 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6343 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6345 struct io_ring_ctx *ctx = req->ctx;
6346 struct io_uring_rsrc_update2 up;
6349 up.offset = req->rsrc_update.offset;
6350 up.data = req->rsrc_update.arg;
6355 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6356 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6357 &up, req->rsrc_update.nr_args);
6358 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6362 __io_req_complete(req, issue_flags, ret, 0);
6366 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6368 switch (req->opcode) {
6371 case IORING_OP_READV:
6372 case IORING_OP_READ_FIXED:
6373 case IORING_OP_READ:
6374 return io_read_prep(req, sqe);
6375 case IORING_OP_WRITEV:
6376 case IORING_OP_WRITE_FIXED:
6377 case IORING_OP_WRITE:
6378 return io_write_prep(req, sqe);
6379 case IORING_OP_POLL_ADD:
6380 return io_poll_add_prep(req, sqe);
6381 case IORING_OP_POLL_REMOVE:
6382 return io_poll_update_prep(req, sqe);
6383 case IORING_OP_FSYNC:
6384 return io_fsync_prep(req, sqe);
6385 case IORING_OP_SYNC_FILE_RANGE:
6386 return io_sfr_prep(req, sqe);
6387 case IORING_OP_SENDMSG:
6388 case IORING_OP_SEND:
6389 return io_sendmsg_prep(req, sqe);
6390 case IORING_OP_RECVMSG:
6391 case IORING_OP_RECV:
6392 return io_recvmsg_prep(req, sqe);
6393 case IORING_OP_CONNECT:
6394 return io_connect_prep(req, sqe);
6395 case IORING_OP_TIMEOUT:
6396 return io_timeout_prep(req, sqe, false);
6397 case IORING_OP_TIMEOUT_REMOVE:
6398 return io_timeout_remove_prep(req, sqe);
6399 case IORING_OP_ASYNC_CANCEL:
6400 return io_async_cancel_prep(req, sqe);
6401 case IORING_OP_LINK_TIMEOUT:
6402 return io_timeout_prep(req, sqe, true);
6403 case IORING_OP_ACCEPT:
6404 return io_accept_prep(req, sqe);
6405 case IORING_OP_FALLOCATE:
6406 return io_fallocate_prep(req, sqe);
6407 case IORING_OP_OPENAT:
6408 return io_openat_prep(req, sqe);
6409 case IORING_OP_CLOSE:
6410 return io_close_prep(req, sqe);
6411 case IORING_OP_FILES_UPDATE:
6412 return io_rsrc_update_prep(req, sqe);
6413 case IORING_OP_STATX:
6414 return io_statx_prep(req, sqe);
6415 case IORING_OP_FADVISE:
6416 return io_fadvise_prep(req, sqe);
6417 case IORING_OP_MADVISE:
6418 return io_madvise_prep(req, sqe);
6419 case IORING_OP_OPENAT2:
6420 return io_openat2_prep(req, sqe);
6421 case IORING_OP_EPOLL_CTL:
6422 return io_epoll_ctl_prep(req, sqe);
6423 case IORING_OP_SPLICE:
6424 return io_splice_prep(req, sqe);
6425 case IORING_OP_PROVIDE_BUFFERS:
6426 return io_provide_buffers_prep(req, sqe);
6427 case IORING_OP_REMOVE_BUFFERS:
6428 return io_remove_buffers_prep(req, sqe);
6430 return io_tee_prep(req, sqe);
6431 case IORING_OP_SHUTDOWN:
6432 return io_shutdown_prep(req, sqe);
6433 case IORING_OP_RENAMEAT:
6434 return io_renameat_prep(req, sqe);
6435 case IORING_OP_UNLINKAT:
6436 return io_unlinkat_prep(req, sqe);
6437 case IORING_OP_MKDIRAT:
6438 return io_mkdirat_prep(req, sqe);
6439 case IORING_OP_SYMLINKAT:
6440 return io_symlinkat_prep(req, sqe);
6441 case IORING_OP_LINKAT:
6442 return io_linkat_prep(req, sqe);
6445 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6450 static int io_req_prep_async(struct io_kiocb *req)
6452 if (!io_op_defs[req->opcode].needs_async_setup)
6454 if (WARN_ON_ONCE(req->async_data))
6456 if (io_alloc_async_data(req))
6459 switch (req->opcode) {
6460 case IORING_OP_READV:
6461 return io_rw_prep_async(req, READ);
6462 case IORING_OP_WRITEV:
6463 return io_rw_prep_async(req, WRITE);
6464 case IORING_OP_SENDMSG:
6465 return io_sendmsg_prep_async(req);
6466 case IORING_OP_RECVMSG:
6467 return io_recvmsg_prep_async(req);
6468 case IORING_OP_CONNECT:
6469 return io_connect_prep_async(req);
6471 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6476 static u32 io_get_sequence(struct io_kiocb *req)
6478 u32 seq = req->ctx->cached_sq_head;
6480 /* need original cached_sq_head, but it was increased for each req */
6481 io_for_each_link(req, req)
6486 static bool io_drain_req(struct io_kiocb *req)
6488 struct io_kiocb *pos;
6489 struct io_ring_ctx *ctx = req->ctx;
6490 struct io_defer_entry *de;
6494 if (req->flags & REQ_F_FAIL) {
6495 io_req_complete_fail_submit(req);
6500 * If we need to drain a request in the middle of a link, drain the
6501 * head request and the next request/link after the current link.
6502 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6503 * maintained for every request of our link.
6505 if (ctx->drain_next) {
6506 req->flags |= REQ_F_IO_DRAIN;
6507 ctx->drain_next = false;
6509 /* not interested in head, start from the first linked */
6510 io_for_each_link(pos, req->link) {
6511 if (pos->flags & REQ_F_IO_DRAIN) {
6512 ctx->drain_next = true;
6513 req->flags |= REQ_F_IO_DRAIN;
6518 /* Still need defer if there is pending req in defer list. */
6519 if (likely(list_empty_careful(&ctx->defer_list) &&
6520 !(req->flags & REQ_F_IO_DRAIN))) {
6521 ctx->drain_active = false;
6525 seq = io_get_sequence(req);
6526 /* Still a chance to pass the sequence check */
6527 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6530 ret = io_req_prep_async(req);
6533 io_prep_async_link(req);
6534 de = kmalloc(sizeof(*de), GFP_KERNEL);
6538 io_req_complete_failed(req, ret);
6542 spin_lock(&ctx->completion_lock);
6543 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6544 spin_unlock(&ctx->completion_lock);
6546 io_queue_async_work(req, NULL);
6550 trace_io_uring_defer(ctx, req, req->user_data);
6553 list_add_tail(&de->list, &ctx->defer_list);
6554 spin_unlock(&ctx->completion_lock);
6558 static void io_clean_op(struct io_kiocb *req)
6560 if (req->flags & REQ_F_BUFFER_SELECTED) {
6561 switch (req->opcode) {
6562 case IORING_OP_READV:
6563 case IORING_OP_READ_FIXED:
6564 case IORING_OP_READ:
6565 kfree((void *)(unsigned long)req->rw.addr);
6567 case IORING_OP_RECVMSG:
6568 case IORING_OP_RECV:
6569 kfree(req->sr_msg.kbuf);
6574 if (req->flags & REQ_F_NEED_CLEANUP) {
6575 switch (req->opcode) {
6576 case IORING_OP_READV:
6577 case IORING_OP_READ_FIXED:
6578 case IORING_OP_READ:
6579 case IORING_OP_WRITEV:
6580 case IORING_OP_WRITE_FIXED:
6581 case IORING_OP_WRITE: {
6582 struct io_async_rw *io = req->async_data;
6584 kfree(io->free_iovec);
6587 case IORING_OP_RECVMSG:
6588 case IORING_OP_SENDMSG: {
6589 struct io_async_msghdr *io = req->async_data;
6591 kfree(io->free_iov);
6594 case IORING_OP_SPLICE:
6596 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6597 io_put_file(req->splice.file_in);
6599 case IORING_OP_OPENAT:
6600 case IORING_OP_OPENAT2:
6601 if (req->open.filename)
6602 putname(req->open.filename);
6604 case IORING_OP_RENAMEAT:
6605 putname(req->rename.oldpath);
6606 putname(req->rename.newpath);
6608 case IORING_OP_UNLINKAT:
6609 putname(req->unlink.filename);
6611 case IORING_OP_MKDIRAT:
6612 putname(req->mkdir.filename);
6614 case IORING_OP_SYMLINKAT:
6615 putname(req->symlink.oldpath);
6616 putname(req->symlink.newpath);
6618 case IORING_OP_LINKAT:
6619 putname(req->hardlink.oldpath);
6620 putname(req->hardlink.newpath);
6624 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6625 kfree(req->apoll->double_poll);
6629 if (req->flags & REQ_F_INFLIGHT) {
6630 struct io_uring_task *tctx = req->task->io_uring;
6632 atomic_dec(&tctx->inflight_tracked);
6634 if (req->flags & REQ_F_CREDS)
6635 put_cred(req->creds);
6637 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6640 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6642 struct io_ring_ctx *ctx = req->ctx;
6643 const struct cred *creds = NULL;
6646 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6647 creds = override_creds(req->creds);
6649 switch (req->opcode) {
6651 ret = io_nop(req, issue_flags);
6653 case IORING_OP_READV:
6654 case IORING_OP_READ_FIXED:
6655 case IORING_OP_READ:
6656 ret = io_read(req, issue_flags);
6658 case IORING_OP_WRITEV:
6659 case IORING_OP_WRITE_FIXED:
6660 case IORING_OP_WRITE:
6661 ret = io_write(req, issue_flags);
6663 case IORING_OP_FSYNC:
6664 ret = io_fsync(req, issue_flags);
6666 case IORING_OP_POLL_ADD:
6667 ret = io_poll_add(req, issue_flags);
6669 case IORING_OP_POLL_REMOVE:
6670 ret = io_poll_update(req, issue_flags);
6672 case IORING_OP_SYNC_FILE_RANGE:
6673 ret = io_sync_file_range(req, issue_flags);
6675 case IORING_OP_SENDMSG:
6676 ret = io_sendmsg(req, issue_flags);
6678 case IORING_OP_SEND:
6679 ret = io_send(req, issue_flags);
6681 case IORING_OP_RECVMSG:
6682 ret = io_recvmsg(req, issue_flags);
6684 case IORING_OP_RECV:
6685 ret = io_recv(req, issue_flags);
6687 case IORING_OP_TIMEOUT:
6688 ret = io_timeout(req, issue_flags);
6690 case IORING_OP_TIMEOUT_REMOVE:
6691 ret = io_timeout_remove(req, issue_flags);
6693 case IORING_OP_ACCEPT:
6694 ret = io_accept(req, issue_flags);
6696 case IORING_OP_CONNECT:
6697 ret = io_connect(req, issue_flags);
6699 case IORING_OP_ASYNC_CANCEL:
6700 ret = io_async_cancel(req, issue_flags);
6702 case IORING_OP_FALLOCATE:
6703 ret = io_fallocate(req, issue_flags);
6705 case IORING_OP_OPENAT:
6706 ret = io_openat(req, issue_flags);
6708 case IORING_OP_CLOSE:
6709 ret = io_close(req, issue_flags);
6711 case IORING_OP_FILES_UPDATE:
6712 ret = io_files_update(req, issue_flags);
6714 case IORING_OP_STATX:
6715 ret = io_statx(req, issue_flags);
6717 case IORING_OP_FADVISE:
6718 ret = io_fadvise(req, issue_flags);
6720 case IORING_OP_MADVISE:
6721 ret = io_madvise(req, issue_flags);
6723 case IORING_OP_OPENAT2:
6724 ret = io_openat2(req, issue_flags);
6726 case IORING_OP_EPOLL_CTL:
6727 ret = io_epoll_ctl(req, issue_flags);
6729 case IORING_OP_SPLICE:
6730 ret = io_splice(req, issue_flags);
6732 case IORING_OP_PROVIDE_BUFFERS:
6733 ret = io_provide_buffers(req, issue_flags);
6735 case IORING_OP_REMOVE_BUFFERS:
6736 ret = io_remove_buffers(req, issue_flags);
6739 ret = io_tee(req, issue_flags);
6741 case IORING_OP_SHUTDOWN:
6742 ret = io_shutdown(req, issue_flags);
6744 case IORING_OP_RENAMEAT:
6745 ret = io_renameat(req, issue_flags);
6747 case IORING_OP_UNLINKAT:
6748 ret = io_unlinkat(req, issue_flags);
6750 case IORING_OP_MKDIRAT:
6751 ret = io_mkdirat(req, issue_flags);
6753 case IORING_OP_SYMLINKAT:
6754 ret = io_symlinkat(req, issue_flags);
6756 case IORING_OP_LINKAT:
6757 ret = io_linkat(req, issue_flags);
6765 revert_creds(creds);
6768 /* If the op doesn't have a file, we're not polling for it */
6769 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6770 io_iopoll_req_issued(req);
6775 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6777 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6779 req = io_put_req_find_next(req);
6780 return req ? &req->work : NULL;
6783 static void io_wq_submit_work(struct io_wq_work *work)
6785 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6786 struct io_kiocb *timeout;
6789 /* one will be dropped by ->io_free_work() after returning to io-wq */
6790 if (!(req->flags & REQ_F_REFCOUNT))
6791 __io_req_set_refcount(req, 2);
6795 timeout = io_prep_linked_timeout(req);
6797 io_queue_linked_timeout(timeout);
6799 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6800 if (work->flags & IO_WQ_WORK_CANCEL)
6805 ret = io_issue_sqe(req, 0);
6807 * We can get EAGAIN for polled IO even though we're
6808 * forcing a sync submission from here, since we can't
6809 * wait for request slots on the block side.
6817 /* avoid locking problems by failing it from a clean context */
6819 io_req_task_queue_fail(req, ret);
6822 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6825 return &table->files[i];
6828 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6831 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6833 return (struct file *) (slot->file_ptr & FFS_MASK);
6836 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6838 unsigned long file_ptr = (unsigned long) file;
6840 if (__io_file_supports_nowait(file, READ))
6841 file_ptr |= FFS_ASYNC_READ;
6842 if (__io_file_supports_nowait(file, WRITE))
6843 file_ptr |= FFS_ASYNC_WRITE;
6844 if (S_ISREG(file_inode(file)->i_mode))
6845 file_ptr |= FFS_ISREG;
6846 file_slot->file_ptr = file_ptr;
6849 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6850 struct io_kiocb *req, int fd)
6853 unsigned long file_ptr;
6855 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6857 fd = array_index_nospec(fd, ctx->nr_user_files);
6858 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6859 file = (struct file *) (file_ptr & FFS_MASK);
6860 file_ptr &= ~FFS_MASK;
6861 /* mask in overlapping REQ_F and FFS bits */
6862 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6863 io_req_set_rsrc_node(req);
6867 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6868 struct io_kiocb *req, int fd)
6870 struct file *file = fget(fd);
6872 trace_io_uring_file_get(ctx, fd);
6874 /* we don't allow fixed io_uring files */
6875 if (file && unlikely(file->f_op == &io_uring_fops))
6876 io_req_track_inflight(req);
6880 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6881 struct io_kiocb *req, int fd, bool fixed)
6884 return io_file_get_fixed(ctx, req, fd);
6886 return io_file_get_normal(ctx, req, fd);
6889 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6891 struct io_kiocb *prev = req->timeout.prev;
6895 ret = io_try_cancel_userdata(req, prev->user_data);
6896 io_req_complete_post(req, ret ?: -ETIME, 0);
6899 io_req_complete_post(req, -ETIME, 0);
6903 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6905 struct io_timeout_data *data = container_of(timer,
6906 struct io_timeout_data, timer);
6907 struct io_kiocb *prev, *req = data->req;
6908 struct io_ring_ctx *ctx = req->ctx;
6909 unsigned long flags;
6911 spin_lock_irqsave(&ctx->timeout_lock, flags);
6912 prev = req->timeout.head;
6913 req->timeout.head = NULL;
6916 * We don't expect the list to be empty, that will only happen if we
6917 * race with the completion of the linked work.
6920 io_remove_next_linked(prev);
6921 if (!req_ref_inc_not_zero(prev))
6924 list_del(&req->timeout.list);
6925 req->timeout.prev = prev;
6926 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6928 req->io_task_work.func = io_req_task_link_timeout;
6929 io_req_task_work_add(req);
6930 return HRTIMER_NORESTART;
6933 static void io_queue_linked_timeout(struct io_kiocb *req)
6935 struct io_ring_ctx *ctx = req->ctx;
6937 spin_lock_irq(&ctx->timeout_lock);
6939 * If the back reference is NULL, then our linked request finished
6940 * before we got a chance to setup the timer
6942 if (req->timeout.head) {
6943 struct io_timeout_data *data = req->async_data;
6945 data->timer.function = io_link_timeout_fn;
6946 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6948 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6950 spin_unlock_irq(&ctx->timeout_lock);
6951 /* drop submission reference */
6955 static void __io_queue_sqe(struct io_kiocb *req)
6956 __must_hold(&req->ctx->uring_lock)
6958 struct io_kiocb *linked_timeout;
6962 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6965 * We async punt it if the file wasn't marked NOWAIT, or if the file
6966 * doesn't support non-blocking read/write attempts
6969 if (req->flags & REQ_F_COMPLETE_INLINE) {
6970 struct io_ring_ctx *ctx = req->ctx;
6971 struct io_submit_state *state = &ctx->submit_state;
6973 state->compl_reqs[state->compl_nr++] = req;
6974 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6975 io_submit_flush_completions(ctx);
6979 linked_timeout = io_prep_linked_timeout(req);
6981 io_queue_linked_timeout(linked_timeout);
6982 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6983 linked_timeout = io_prep_linked_timeout(req);
6985 switch (io_arm_poll_handler(req)) {
6986 case IO_APOLL_READY:
6988 io_unprep_linked_timeout(req);
6990 case IO_APOLL_ABORTED:
6992 * Queued up for async execution, worker will release
6993 * submit reference when the iocb is actually submitted.
6995 io_queue_async_work(req, NULL);
7000 io_queue_linked_timeout(linked_timeout);
7002 io_req_complete_failed(req, ret);
7006 static inline void io_queue_sqe(struct io_kiocb *req)
7007 __must_hold(&req->ctx->uring_lock)
7009 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7012 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7013 __io_queue_sqe(req);
7014 } else if (req->flags & REQ_F_FAIL) {
7015 io_req_complete_fail_submit(req);
7017 int ret = io_req_prep_async(req);
7020 io_req_complete_failed(req, ret);
7022 io_queue_async_work(req, NULL);
7027 * Check SQE restrictions (opcode and flags).
7029 * Returns 'true' if SQE is allowed, 'false' otherwise.
7031 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7032 struct io_kiocb *req,
7033 unsigned int sqe_flags)
7035 if (likely(!ctx->restricted))
7038 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7041 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7042 ctx->restrictions.sqe_flags_required)
7045 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7046 ctx->restrictions.sqe_flags_required))
7052 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7053 const struct io_uring_sqe *sqe)
7054 __must_hold(&ctx->uring_lock)
7056 struct io_submit_state *state;
7057 unsigned int sqe_flags;
7058 int personality, ret = 0;
7060 /* req is partially pre-initialised, see io_preinit_req() */
7061 req->opcode = READ_ONCE(sqe->opcode);
7062 /* same numerical values with corresponding REQ_F_*, safe to copy */
7063 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7064 req->user_data = READ_ONCE(sqe->user_data);
7066 req->fixed_rsrc_refs = NULL;
7067 req->task = current;
7069 /* enforce forwards compatibility on users */
7070 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7072 if (unlikely(req->opcode >= IORING_OP_LAST))
7074 if (!io_check_restriction(ctx, req, sqe_flags))
7077 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7078 !io_op_defs[req->opcode].buffer_select)
7080 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7081 ctx->drain_active = true;
7083 personality = READ_ONCE(sqe->personality);
7085 req->creds = xa_load(&ctx->personalities, personality);
7088 get_cred(req->creds);
7089 req->flags |= REQ_F_CREDS;
7091 state = &ctx->submit_state;
7094 * Plug now if we have more than 1 IO left after this, and the target
7095 * is potentially a read/write to block based storage.
7097 if (!state->plug_started && state->ios_left > 1 &&
7098 io_op_defs[req->opcode].plug) {
7099 blk_start_plug(&state->plug);
7100 state->plug_started = true;
7103 if (io_op_defs[req->opcode].needs_file) {
7104 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7105 (sqe_flags & IOSQE_FIXED_FILE));
7106 if (unlikely(!req->file))
7114 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7115 const struct io_uring_sqe *sqe)
7116 __must_hold(&ctx->uring_lock)
7118 struct io_submit_link *link = &ctx->submit_state.link;
7121 ret = io_init_req(ctx, req, sqe);
7122 if (unlikely(ret)) {
7124 /* fail even hard links since we don't submit */
7127 * we can judge a link req is failed or cancelled by if
7128 * REQ_F_FAIL is set, but the head is an exception since
7129 * it may be set REQ_F_FAIL because of other req's failure
7130 * so let's leverage req->result to distinguish if a head
7131 * is set REQ_F_FAIL because of its failure or other req's
7132 * failure so that we can set the correct ret code for it.
7133 * init result here to avoid affecting the normal path.
7135 if (!(link->head->flags & REQ_F_FAIL))
7136 req_fail_link_node(link->head, -ECANCELED);
7137 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7139 * the current req is a normal req, we should return
7140 * error and thus break the submittion loop.
7142 io_req_complete_failed(req, ret);
7145 req_fail_link_node(req, ret);
7147 ret = io_req_prep(req, sqe);
7152 /* don't need @sqe from now on */
7153 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7155 ctx->flags & IORING_SETUP_SQPOLL);
7158 * If we already have a head request, queue this one for async
7159 * submittal once the head completes. If we don't have a head but
7160 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7161 * submitted sync once the chain is complete. If none of those
7162 * conditions are true (normal request), then just queue it.
7165 struct io_kiocb *head = link->head;
7167 if (!(req->flags & REQ_F_FAIL)) {
7168 ret = io_req_prep_async(req);
7169 if (unlikely(ret)) {
7170 req_fail_link_node(req, ret);
7171 if (!(head->flags & REQ_F_FAIL))
7172 req_fail_link_node(head, -ECANCELED);
7175 trace_io_uring_link(ctx, req, head);
7176 link->last->link = req;
7179 /* last request of a link, enqueue the link */
7180 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7185 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7197 * Batched submission is done, ensure local IO is flushed out.
7199 static void io_submit_state_end(struct io_submit_state *state,
7200 struct io_ring_ctx *ctx)
7202 if (state->link.head)
7203 io_queue_sqe(state->link.head);
7204 if (state->compl_nr)
7205 io_submit_flush_completions(ctx);
7206 if (state->plug_started)
7207 blk_finish_plug(&state->plug);
7211 * Start submission side cache.
7213 static void io_submit_state_start(struct io_submit_state *state,
7214 unsigned int max_ios)
7216 state->plug_started = false;
7217 state->ios_left = max_ios;
7218 /* set only head, no need to init link_last in advance */
7219 state->link.head = NULL;
7222 static void io_commit_sqring(struct io_ring_ctx *ctx)
7224 struct io_rings *rings = ctx->rings;
7227 * Ensure any loads from the SQEs are done at this point,
7228 * since once we write the new head, the application could
7229 * write new data to them.
7231 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7235 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7236 * that is mapped by userspace. This means that care needs to be taken to
7237 * ensure that reads are stable, as we cannot rely on userspace always
7238 * being a good citizen. If members of the sqe are validated and then later
7239 * used, it's important that those reads are done through READ_ONCE() to
7240 * prevent a re-load down the line.
7242 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7244 unsigned head, mask = ctx->sq_entries - 1;
7245 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7248 * The cached sq head (or cq tail) serves two purposes:
7250 * 1) allows us to batch the cost of updating the user visible
7252 * 2) allows the kernel side to track the head on its own, even
7253 * though the application is the one updating it.
7255 head = READ_ONCE(ctx->sq_array[sq_idx]);
7256 if (likely(head < ctx->sq_entries))
7257 return &ctx->sq_sqes[head];
7259 /* drop invalid entries */
7261 WRITE_ONCE(ctx->rings->sq_dropped,
7262 READ_ONCE(ctx->rings->sq_dropped) + 1);
7266 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7267 __must_hold(&ctx->uring_lock)
7271 /* make sure SQ entry isn't read before tail */
7272 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7273 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7275 io_get_task_refs(nr);
7277 io_submit_state_start(&ctx->submit_state, nr);
7278 while (submitted < nr) {
7279 const struct io_uring_sqe *sqe;
7280 struct io_kiocb *req;
7282 req = io_alloc_req(ctx);
7283 if (unlikely(!req)) {
7285 submitted = -EAGAIN;
7288 sqe = io_get_sqe(ctx);
7289 if (unlikely(!sqe)) {
7290 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7293 /* will complete beyond this point, count as submitted */
7295 if (io_submit_sqe(ctx, req, sqe))
7299 if (unlikely(submitted != nr)) {
7300 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7301 int unused = nr - ref_used;
7303 current->io_uring->cached_refs += unused;
7304 percpu_ref_put_many(&ctx->refs, unused);
7307 io_submit_state_end(&ctx->submit_state, ctx);
7308 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7309 io_commit_sqring(ctx);
7314 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7316 return READ_ONCE(sqd->state);
7319 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7321 /* Tell userspace we may need a wakeup call */
7322 spin_lock(&ctx->completion_lock);
7323 WRITE_ONCE(ctx->rings->sq_flags,
7324 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7325 spin_unlock(&ctx->completion_lock);
7328 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7330 spin_lock(&ctx->completion_lock);
7331 WRITE_ONCE(ctx->rings->sq_flags,
7332 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7333 spin_unlock(&ctx->completion_lock);
7336 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7338 unsigned int to_submit;
7341 to_submit = io_sqring_entries(ctx);
7342 /* if we're handling multiple rings, cap submit size for fairness */
7343 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7344 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7346 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7347 unsigned nr_events = 0;
7348 const struct cred *creds = NULL;
7350 if (ctx->sq_creds != current_cred())
7351 creds = override_creds(ctx->sq_creds);
7353 mutex_lock(&ctx->uring_lock);
7354 if (!list_empty(&ctx->iopoll_list))
7355 io_do_iopoll(ctx, &nr_events, 0);
7358 * Don't submit if refs are dying, good for io_uring_register(),
7359 * but also it is relied upon by io_ring_exit_work()
7361 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7362 !(ctx->flags & IORING_SETUP_R_DISABLED))
7363 ret = io_submit_sqes(ctx, to_submit);
7364 mutex_unlock(&ctx->uring_lock);
7366 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7367 wake_up(&ctx->sqo_sq_wait);
7369 revert_creds(creds);
7375 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7377 struct io_ring_ctx *ctx;
7378 unsigned sq_thread_idle = 0;
7380 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7381 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7382 sqd->sq_thread_idle = sq_thread_idle;
7385 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7387 bool did_sig = false;
7388 struct ksignal ksig;
7390 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7391 signal_pending(current)) {
7392 mutex_unlock(&sqd->lock);
7393 if (signal_pending(current))
7394 did_sig = get_signal(&ksig);
7396 mutex_lock(&sqd->lock);
7398 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7401 static int io_sq_thread(void *data)
7403 struct io_sq_data *sqd = data;
7404 struct io_ring_ctx *ctx;
7405 unsigned long timeout = 0;
7406 char buf[TASK_COMM_LEN];
7409 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7410 set_task_comm(current, buf);
7412 if (sqd->sq_cpu != -1)
7413 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7415 set_cpus_allowed_ptr(current, cpu_online_mask);
7416 current->flags |= PF_NO_SETAFFINITY;
7418 mutex_lock(&sqd->lock);
7420 bool cap_entries, sqt_spin = false;
7422 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7423 if (io_sqd_handle_event(sqd))
7425 timeout = jiffies + sqd->sq_thread_idle;
7428 cap_entries = !list_is_singular(&sqd->ctx_list);
7429 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7430 int ret = __io_sq_thread(ctx, cap_entries);
7432 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7435 if (io_run_task_work())
7438 if (sqt_spin || !time_after(jiffies, timeout)) {
7441 timeout = jiffies + sqd->sq_thread_idle;
7445 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7446 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7447 bool needs_sched = true;
7449 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7450 io_ring_set_wakeup_flag(ctx);
7452 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7453 !list_empty_careful(&ctx->iopoll_list)) {
7454 needs_sched = false;
7457 if (io_sqring_entries(ctx)) {
7458 needs_sched = false;
7464 mutex_unlock(&sqd->lock);
7466 mutex_lock(&sqd->lock);
7468 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7469 io_ring_clear_wakeup_flag(ctx);
7472 finish_wait(&sqd->wait, &wait);
7473 timeout = jiffies + sqd->sq_thread_idle;
7476 io_uring_cancel_generic(true, sqd);
7478 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7479 io_ring_set_wakeup_flag(ctx);
7481 mutex_unlock(&sqd->lock);
7483 complete(&sqd->exited);
7487 struct io_wait_queue {
7488 struct wait_queue_entry wq;
7489 struct io_ring_ctx *ctx;
7491 unsigned nr_timeouts;
7494 static inline bool io_should_wake(struct io_wait_queue *iowq)
7496 struct io_ring_ctx *ctx = iowq->ctx;
7497 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7500 * Wake up if we have enough events, or if a timeout occurred since we
7501 * started waiting. For timeouts, we always want to return to userspace,
7502 * regardless of event count.
7504 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7507 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7508 int wake_flags, void *key)
7510 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7514 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7515 * the task, and the next invocation will do it.
7517 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7518 return autoremove_wake_function(curr, mode, wake_flags, key);
7522 static int io_run_task_work_sig(void)
7524 if (io_run_task_work())
7526 if (!signal_pending(current))
7528 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7529 return -ERESTARTSYS;
7533 /* when returns >0, the caller should retry */
7534 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7535 struct io_wait_queue *iowq,
7536 signed long *timeout)
7540 /* make sure we run task_work before checking for signals */
7541 ret = io_run_task_work_sig();
7542 if (ret || io_should_wake(iowq))
7544 /* let the caller flush overflows, retry */
7545 if (test_bit(0, &ctx->check_cq_overflow))
7548 *timeout = schedule_timeout(*timeout);
7549 return !*timeout ? -ETIME : 1;
7553 * Wait until events become available, if we don't already have some. The
7554 * application must reap them itself, as they reside on the shared cq ring.
7556 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7557 const sigset_t __user *sig, size_t sigsz,
7558 struct __kernel_timespec __user *uts)
7560 struct io_wait_queue iowq;
7561 struct io_rings *rings = ctx->rings;
7562 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7566 io_cqring_overflow_flush(ctx);
7567 if (io_cqring_events(ctx) >= min_events)
7569 if (!io_run_task_work())
7574 struct timespec64 ts;
7576 if (get_timespec64(&ts, uts))
7578 timeout = timespec64_to_jiffies(&ts);
7582 #ifdef CONFIG_COMPAT
7583 if (in_compat_syscall())
7584 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7588 ret = set_user_sigmask(sig, sigsz);
7594 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7595 iowq.wq.private = current;
7596 INIT_LIST_HEAD(&iowq.wq.entry);
7598 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7599 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7601 trace_io_uring_cqring_wait(ctx, min_events);
7603 /* if we can't even flush overflow, don't wait for more */
7604 if (!io_cqring_overflow_flush(ctx)) {
7608 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7609 TASK_INTERRUPTIBLE);
7610 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7611 finish_wait(&ctx->cq_wait, &iowq.wq);
7615 restore_saved_sigmask_unless(ret == -EINTR);
7617 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7620 static void io_free_page_table(void **table, size_t size)
7622 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7624 for (i = 0; i < nr_tables; i++)
7629 static void **io_alloc_page_table(size_t size)
7631 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7632 size_t init_size = size;
7635 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7639 for (i = 0; i < nr_tables; i++) {
7640 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7642 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7644 io_free_page_table(table, init_size);
7652 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7654 percpu_ref_exit(&ref_node->refs);
7658 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7660 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7661 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7662 unsigned long flags;
7663 bool first_add = false;
7665 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7668 while (!list_empty(&ctx->rsrc_ref_list)) {
7669 node = list_first_entry(&ctx->rsrc_ref_list,
7670 struct io_rsrc_node, node);
7671 /* recycle ref nodes in order */
7674 list_del(&node->node);
7675 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7677 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7680 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7683 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7685 struct io_rsrc_node *ref_node;
7687 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7691 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7696 INIT_LIST_HEAD(&ref_node->node);
7697 INIT_LIST_HEAD(&ref_node->rsrc_list);
7698 ref_node->done = false;
7702 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7703 struct io_rsrc_data *data_to_kill)
7705 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7706 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7709 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7711 rsrc_node->rsrc_data = data_to_kill;
7712 spin_lock_irq(&ctx->rsrc_ref_lock);
7713 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7714 spin_unlock_irq(&ctx->rsrc_ref_lock);
7716 atomic_inc(&data_to_kill->refs);
7717 percpu_ref_kill(&rsrc_node->refs);
7718 ctx->rsrc_node = NULL;
7721 if (!ctx->rsrc_node) {
7722 ctx->rsrc_node = ctx->rsrc_backup_node;
7723 ctx->rsrc_backup_node = NULL;
7727 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7729 if (ctx->rsrc_backup_node)
7731 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7732 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7735 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7739 /* As we may drop ->uring_lock, other task may have started quiesce */
7743 data->quiesce = true;
7745 ret = io_rsrc_node_switch_start(ctx);
7748 io_rsrc_node_switch(ctx, data);
7750 /* kill initial ref, already quiesced if zero */
7751 if (atomic_dec_and_test(&data->refs))
7753 mutex_unlock(&ctx->uring_lock);
7754 flush_delayed_work(&ctx->rsrc_put_work);
7755 ret = wait_for_completion_interruptible(&data->done);
7757 mutex_lock(&ctx->uring_lock);
7761 atomic_inc(&data->refs);
7762 /* wait for all works potentially completing data->done */
7763 flush_delayed_work(&ctx->rsrc_put_work);
7764 reinit_completion(&data->done);
7766 ret = io_run_task_work_sig();
7767 mutex_lock(&ctx->uring_lock);
7769 data->quiesce = false;
7774 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7776 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7777 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7779 return &data->tags[table_idx][off];
7782 static void io_rsrc_data_free(struct io_rsrc_data *data)
7784 size_t size = data->nr * sizeof(data->tags[0][0]);
7787 io_free_page_table((void **)data->tags, size);
7791 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7792 u64 __user *utags, unsigned nr,
7793 struct io_rsrc_data **pdata)
7795 struct io_rsrc_data *data;
7799 data = kzalloc(sizeof(*data), GFP_KERNEL);
7802 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7810 data->do_put = do_put;
7813 for (i = 0; i < nr; i++) {
7814 u64 *tag_slot = io_get_tag_slot(data, i);
7816 if (copy_from_user(tag_slot, &utags[i],
7822 atomic_set(&data->refs, 1);
7823 init_completion(&data->done);
7827 io_rsrc_data_free(data);
7831 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7833 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7834 GFP_KERNEL_ACCOUNT);
7835 return !!table->files;
7838 static void io_free_file_tables(struct io_file_table *table)
7840 kvfree(table->files);
7841 table->files = NULL;
7844 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7846 #if defined(CONFIG_UNIX)
7847 if (ctx->ring_sock) {
7848 struct sock *sock = ctx->ring_sock->sk;
7849 struct sk_buff *skb;
7851 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7857 for (i = 0; i < ctx->nr_user_files; i++) {
7860 file = io_file_from_index(ctx, i);
7865 io_free_file_tables(&ctx->file_table);
7866 io_rsrc_data_free(ctx->file_data);
7867 ctx->file_data = NULL;
7868 ctx->nr_user_files = 0;
7871 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7875 if (!ctx->file_data)
7877 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7879 __io_sqe_files_unregister(ctx);
7883 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7884 __releases(&sqd->lock)
7886 WARN_ON_ONCE(sqd->thread == current);
7889 * Do the dance but not conditional clear_bit() because it'd race with
7890 * other threads incrementing park_pending and setting the bit.
7892 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7893 if (atomic_dec_return(&sqd->park_pending))
7894 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7895 mutex_unlock(&sqd->lock);
7898 static void io_sq_thread_park(struct io_sq_data *sqd)
7899 __acquires(&sqd->lock)
7901 WARN_ON_ONCE(sqd->thread == current);
7903 atomic_inc(&sqd->park_pending);
7904 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7905 mutex_lock(&sqd->lock);
7907 wake_up_process(sqd->thread);
7910 static void io_sq_thread_stop(struct io_sq_data *sqd)
7912 WARN_ON_ONCE(sqd->thread == current);
7913 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7915 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7916 mutex_lock(&sqd->lock);
7918 wake_up_process(sqd->thread);
7919 mutex_unlock(&sqd->lock);
7920 wait_for_completion(&sqd->exited);
7923 static void io_put_sq_data(struct io_sq_data *sqd)
7925 if (refcount_dec_and_test(&sqd->refs)) {
7926 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7928 io_sq_thread_stop(sqd);
7933 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7935 struct io_sq_data *sqd = ctx->sq_data;
7938 io_sq_thread_park(sqd);
7939 list_del_init(&ctx->sqd_list);
7940 io_sqd_update_thread_idle(sqd);
7941 io_sq_thread_unpark(sqd);
7943 io_put_sq_data(sqd);
7944 ctx->sq_data = NULL;
7948 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7950 struct io_ring_ctx *ctx_attach;
7951 struct io_sq_data *sqd;
7954 f = fdget(p->wq_fd);
7956 return ERR_PTR(-ENXIO);
7957 if (f.file->f_op != &io_uring_fops) {
7959 return ERR_PTR(-EINVAL);
7962 ctx_attach = f.file->private_data;
7963 sqd = ctx_attach->sq_data;
7966 return ERR_PTR(-EINVAL);
7968 if (sqd->task_tgid != current->tgid) {
7970 return ERR_PTR(-EPERM);
7973 refcount_inc(&sqd->refs);
7978 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7981 struct io_sq_data *sqd;
7984 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7985 sqd = io_attach_sq_data(p);
7990 /* fall through for EPERM case, setup new sqd/task */
7991 if (PTR_ERR(sqd) != -EPERM)
7995 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7997 return ERR_PTR(-ENOMEM);
7999 atomic_set(&sqd->park_pending, 0);
8000 refcount_set(&sqd->refs, 1);
8001 INIT_LIST_HEAD(&sqd->ctx_list);
8002 mutex_init(&sqd->lock);
8003 init_waitqueue_head(&sqd->wait);
8004 init_completion(&sqd->exited);
8008 #if defined(CONFIG_UNIX)
8010 * Ensure the UNIX gc is aware of our file set, so we are certain that
8011 * the io_uring can be safely unregistered on process exit, even if we have
8012 * loops in the file referencing.
8014 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8016 struct sock *sk = ctx->ring_sock->sk;
8017 struct scm_fp_list *fpl;
8018 struct sk_buff *skb;
8021 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8025 skb = alloc_skb(0, GFP_KERNEL);
8034 fpl->user = get_uid(current_user());
8035 for (i = 0; i < nr; i++) {
8036 struct file *file = io_file_from_index(ctx, i + offset);
8040 fpl->fp[nr_files] = get_file(file);
8041 unix_inflight(fpl->user, fpl->fp[nr_files]);
8046 fpl->max = SCM_MAX_FD;
8047 fpl->count = nr_files;
8048 UNIXCB(skb).fp = fpl;
8049 skb->destructor = unix_destruct_scm;
8050 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8051 skb_queue_head(&sk->sk_receive_queue, skb);
8053 for (i = 0; i < nr_files; i++)
8064 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8065 * causes regular reference counting to break down. We rely on the UNIX
8066 * garbage collection to take care of this problem for us.
8068 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8070 unsigned left, total;
8074 left = ctx->nr_user_files;
8076 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8078 ret = __io_sqe_files_scm(ctx, this_files, total);
8082 total += this_files;
8088 while (total < ctx->nr_user_files) {
8089 struct file *file = io_file_from_index(ctx, total);
8099 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8105 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8107 struct file *file = prsrc->file;
8108 #if defined(CONFIG_UNIX)
8109 struct sock *sock = ctx->ring_sock->sk;
8110 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8111 struct sk_buff *skb;
8114 __skb_queue_head_init(&list);
8117 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8118 * remove this entry and rearrange the file array.
8120 skb = skb_dequeue(head);
8122 struct scm_fp_list *fp;
8124 fp = UNIXCB(skb).fp;
8125 for (i = 0; i < fp->count; i++) {
8128 if (fp->fp[i] != file)
8131 unix_notinflight(fp->user, fp->fp[i]);
8132 left = fp->count - 1 - i;
8134 memmove(&fp->fp[i], &fp->fp[i + 1],
8135 left * sizeof(struct file *));
8142 __skb_queue_tail(&list, skb);
8152 __skb_queue_tail(&list, skb);
8154 skb = skb_dequeue(head);
8157 if (skb_peek(&list)) {
8158 spin_lock_irq(&head->lock);
8159 while ((skb = __skb_dequeue(&list)) != NULL)
8160 __skb_queue_tail(head, skb);
8161 spin_unlock_irq(&head->lock);
8168 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8170 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8171 struct io_ring_ctx *ctx = rsrc_data->ctx;
8172 struct io_rsrc_put *prsrc, *tmp;
8174 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8175 list_del(&prsrc->list);
8178 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8180 io_ring_submit_lock(ctx, lock_ring);
8181 spin_lock(&ctx->completion_lock);
8182 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8184 io_commit_cqring(ctx);
8185 spin_unlock(&ctx->completion_lock);
8186 io_cqring_ev_posted(ctx);
8187 io_ring_submit_unlock(ctx, lock_ring);
8190 rsrc_data->do_put(ctx, prsrc);
8194 io_rsrc_node_destroy(ref_node);
8195 if (atomic_dec_and_test(&rsrc_data->refs))
8196 complete(&rsrc_data->done);
8199 static void io_rsrc_put_work(struct work_struct *work)
8201 struct io_ring_ctx *ctx;
8202 struct llist_node *node;
8204 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8205 node = llist_del_all(&ctx->rsrc_put_llist);
8208 struct io_rsrc_node *ref_node;
8209 struct llist_node *next = node->next;
8211 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8212 __io_rsrc_put_work(ref_node);
8217 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8218 unsigned nr_args, u64 __user *tags)
8220 __s32 __user *fds = (__s32 __user *) arg;
8229 if (nr_args > IORING_MAX_FIXED_FILES)
8231 if (nr_args > rlimit(RLIMIT_NOFILE))
8233 ret = io_rsrc_node_switch_start(ctx);
8236 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8242 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8245 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8246 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8250 /* allow sparse sets */
8253 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8260 if (unlikely(!file))
8264 * Don't allow io_uring instances to be registered. If UNIX
8265 * isn't enabled, then this causes a reference cycle and this
8266 * instance can never get freed. If UNIX is enabled we'll
8267 * handle it just fine, but there's still no point in allowing
8268 * a ring fd as it doesn't support regular read/write anyway.
8270 if (file->f_op == &io_uring_fops) {
8274 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8277 ret = io_sqe_files_scm(ctx);
8279 __io_sqe_files_unregister(ctx);
8283 io_rsrc_node_switch(ctx, NULL);
8286 for (i = 0; i < ctx->nr_user_files; i++) {
8287 file = io_file_from_index(ctx, i);
8291 io_free_file_tables(&ctx->file_table);
8292 ctx->nr_user_files = 0;
8294 io_rsrc_data_free(ctx->file_data);
8295 ctx->file_data = NULL;
8299 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8302 #if defined(CONFIG_UNIX)
8303 struct sock *sock = ctx->ring_sock->sk;
8304 struct sk_buff_head *head = &sock->sk_receive_queue;
8305 struct sk_buff *skb;
8308 * See if we can merge this file into an existing skb SCM_RIGHTS
8309 * file set. If there's no room, fall back to allocating a new skb
8310 * and filling it in.
8312 spin_lock_irq(&head->lock);
8313 skb = skb_peek(head);
8315 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8317 if (fpl->count < SCM_MAX_FD) {
8318 __skb_unlink(skb, head);
8319 spin_unlock_irq(&head->lock);
8320 fpl->fp[fpl->count] = get_file(file);
8321 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8323 spin_lock_irq(&head->lock);
8324 __skb_queue_head(head, skb);
8329 spin_unlock_irq(&head->lock);
8336 return __io_sqe_files_scm(ctx, 1, index);
8342 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8343 struct io_rsrc_node *node, void *rsrc)
8345 struct io_rsrc_put *prsrc;
8347 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8351 prsrc->tag = *io_get_tag_slot(data, idx);
8353 list_add(&prsrc->list, &node->rsrc_list);
8357 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8358 unsigned int issue_flags, u32 slot_index)
8360 struct io_ring_ctx *ctx = req->ctx;
8361 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8362 bool needs_switch = false;
8363 struct io_fixed_file *file_slot;
8366 io_ring_submit_lock(ctx, !force_nonblock);
8367 if (file->f_op == &io_uring_fops)
8370 if (!ctx->file_data)
8373 if (slot_index >= ctx->nr_user_files)
8376 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8377 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8379 if (file_slot->file_ptr) {
8380 struct file *old_file;
8382 ret = io_rsrc_node_switch_start(ctx);
8386 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8387 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8388 ctx->rsrc_node, old_file);
8391 file_slot->file_ptr = 0;
8392 needs_switch = true;
8395 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8396 io_fixed_file_set(file_slot, file);
8397 ret = io_sqe_file_register(ctx, file, slot_index);
8399 file_slot->file_ptr = 0;
8406 io_rsrc_node_switch(ctx, ctx->file_data);
8407 io_ring_submit_unlock(ctx, !force_nonblock);
8413 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8415 unsigned int offset = req->close.file_slot - 1;
8416 struct io_ring_ctx *ctx = req->ctx;
8417 struct io_fixed_file *file_slot;
8421 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8423 if (unlikely(!ctx->file_data))
8426 if (offset >= ctx->nr_user_files)
8428 ret = io_rsrc_node_switch_start(ctx);
8432 i = array_index_nospec(offset, ctx->nr_user_files);
8433 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8435 if (!file_slot->file_ptr)
8438 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8439 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8443 file_slot->file_ptr = 0;
8444 io_rsrc_node_switch(ctx, ctx->file_data);
8447 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8451 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8452 struct io_uring_rsrc_update2 *up,
8455 u64 __user *tags = u64_to_user_ptr(up->tags);
8456 __s32 __user *fds = u64_to_user_ptr(up->data);
8457 struct io_rsrc_data *data = ctx->file_data;
8458 struct io_fixed_file *file_slot;
8462 bool needs_switch = false;
8464 if (!ctx->file_data)
8466 if (up->offset + nr_args > ctx->nr_user_files)
8469 for (done = 0; done < nr_args; done++) {
8472 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8473 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8477 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8481 if (fd == IORING_REGISTER_FILES_SKIP)
8484 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8485 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8487 if (file_slot->file_ptr) {
8488 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8489 err = io_queue_rsrc_removal(data, up->offset + done,
8490 ctx->rsrc_node, file);
8493 file_slot->file_ptr = 0;
8494 needs_switch = true;
8503 * Don't allow io_uring instances to be registered. If
8504 * UNIX isn't enabled, then this causes a reference
8505 * cycle and this instance can never get freed. If UNIX
8506 * is enabled we'll handle it just fine, but there's
8507 * still no point in allowing a ring fd as it doesn't
8508 * support regular read/write anyway.
8510 if (file->f_op == &io_uring_fops) {
8515 *io_get_tag_slot(data, up->offset + done) = tag;
8516 io_fixed_file_set(file_slot, file);
8517 err = io_sqe_file_register(ctx, file, i);
8519 file_slot->file_ptr = 0;
8527 io_rsrc_node_switch(ctx, data);
8528 return done ? done : err;
8531 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8532 struct task_struct *task)
8534 struct io_wq_hash *hash;
8535 struct io_wq_data data;
8536 unsigned int concurrency;
8538 mutex_lock(&ctx->uring_lock);
8539 hash = ctx->hash_map;
8541 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8543 mutex_unlock(&ctx->uring_lock);
8544 return ERR_PTR(-ENOMEM);
8546 refcount_set(&hash->refs, 1);
8547 init_waitqueue_head(&hash->wait);
8548 ctx->hash_map = hash;
8550 mutex_unlock(&ctx->uring_lock);
8554 data.free_work = io_wq_free_work;
8555 data.do_work = io_wq_submit_work;
8557 /* Do QD, or 4 * CPUS, whatever is smallest */
8558 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8560 return io_wq_create(concurrency, &data);
8563 static int io_uring_alloc_task_context(struct task_struct *task,
8564 struct io_ring_ctx *ctx)
8566 struct io_uring_task *tctx;
8569 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8570 if (unlikely(!tctx))
8573 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8574 if (unlikely(ret)) {
8579 tctx->io_wq = io_init_wq_offload(ctx, task);
8580 if (IS_ERR(tctx->io_wq)) {
8581 ret = PTR_ERR(tctx->io_wq);
8582 percpu_counter_destroy(&tctx->inflight);
8588 init_waitqueue_head(&tctx->wait);
8589 atomic_set(&tctx->in_idle, 0);
8590 atomic_set(&tctx->inflight_tracked, 0);
8591 task->io_uring = tctx;
8592 spin_lock_init(&tctx->task_lock);
8593 INIT_WQ_LIST(&tctx->task_list);
8594 init_task_work(&tctx->task_work, tctx_task_work);
8598 void __io_uring_free(struct task_struct *tsk)
8600 struct io_uring_task *tctx = tsk->io_uring;
8602 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8603 WARN_ON_ONCE(tctx->io_wq);
8604 WARN_ON_ONCE(tctx->cached_refs);
8606 percpu_counter_destroy(&tctx->inflight);
8608 tsk->io_uring = NULL;
8611 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8612 struct io_uring_params *p)
8616 /* Retain compatibility with failing for an invalid attach attempt */
8617 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8618 IORING_SETUP_ATTACH_WQ) {
8621 f = fdget(p->wq_fd);
8624 if (f.file->f_op != &io_uring_fops) {
8630 if (ctx->flags & IORING_SETUP_SQPOLL) {
8631 struct task_struct *tsk;
8632 struct io_sq_data *sqd;
8635 sqd = io_get_sq_data(p, &attached);
8641 ctx->sq_creds = get_current_cred();
8643 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8644 if (!ctx->sq_thread_idle)
8645 ctx->sq_thread_idle = HZ;
8647 io_sq_thread_park(sqd);
8648 list_add(&ctx->sqd_list, &sqd->ctx_list);
8649 io_sqd_update_thread_idle(sqd);
8650 /* don't attach to a dying SQPOLL thread, would be racy */
8651 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8652 io_sq_thread_unpark(sqd);
8659 if (p->flags & IORING_SETUP_SQ_AFF) {
8660 int cpu = p->sq_thread_cpu;
8663 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8670 sqd->task_pid = current->pid;
8671 sqd->task_tgid = current->tgid;
8672 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8679 ret = io_uring_alloc_task_context(tsk, ctx);
8680 wake_up_new_task(tsk);
8683 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8684 /* Can't have SQ_AFF without SQPOLL */
8691 complete(&ctx->sq_data->exited);
8693 io_sq_thread_finish(ctx);
8697 static inline void __io_unaccount_mem(struct user_struct *user,
8698 unsigned long nr_pages)
8700 atomic_long_sub(nr_pages, &user->locked_vm);
8703 static inline int __io_account_mem(struct user_struct *user,
8704 unsigned long nr_pages)
8706 unsigned long page_limit, cur_pages, new_pages;
8708 /* Don't allow more pages than we can safely lock */
8709 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8712 cur_pages = atomic_long_read(&user->locked_vm);
8713 new_pages = cur_pages + nr_pages;
8714 if (new_pages > page_limit)
8716 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8717 new_pages) != cur_pages);
8722 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8725 __io_unaccount_mem(ctx->user, nr_pages);
8727 if (ctx->mm_account)
8728 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8731 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8736 ret = __io_account_mem(ctx->user, nr_pages);
8741 if (ctx->mm_account)
8742 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8747 static void io_mem_free(void *ptr)
8754 page = virt_to_head_page(ptr);
8755 if (put_page_testzero(page))
8756 free_compound_page(page);
8759 static void *io_mem_alloc(size_t size)
8761 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8762 __GFP_NORETRY | __GFP_ACCOUNT;
8764 return (void *) __get_free_pages(gfp_flags, get_order(size));
8767 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8770 struct io_rings *rings;
8771 size_t off, sq_array_size;
8773 off = struct_size(rings, cqes, cq_entries);
8774 if (off == SIZE_MAX)
8778 off = ALIGN(off, SMP_CACHE_BYTES);
8786 sq_array_size = array_size(sizeof(u32), sq_entries);
8787 if (sq_array_size == SIZE_MAX)
8790 if (check_add_overflow(off, sq_array_size, &off))
8796 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8798 struct io_mapped_ubuf *imu = *slot;
8801 if (imu != ctx->dummy_ubuf) {
8802 for (i = 0; i < imu->nr_bvecs; i++)
8803 unpin_user_page(imu->bvec[i].bv_page);
8804 if (imu->acct_pages)
8805 io_unaccount_mem(ctx, imu->acct_pages);
8811 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8813 io_buffer_unmap(ctx, &prsrc->buf);
8817 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8821 for (i = 0; i < ctx->nr_user_bufs; i++)
8822 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8823 kfree(ctx->user_bufs);
8824 io_rsrc_data_free(ctx->buf_data);
8825 ctx->user_bufs = NULL;
8826 ctx->buf_data = NULL;
8827 ctx->nr_user_bufs = 0;
8830 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8837 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8839 __io_sqe_buffers_unregister(ctx);
8843 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8844 void __user *arg, unsigned index)
8846 struct iovec __user *src;
8848 #ifdef CONFIG_COMPAT
8850 struct compat_iovec __user *ciovs;
8851 struct compat_iovec ciov;
8853 ciovs = (struct compat_iovec __user *) arg;
8854 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8857 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8858 dst->iov_len = ciov.iov_len;
8862 src = (struct iovec __user *) arg;
8863 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8869 * Not super efficient, but this is just a registration time. And we do cache
8870 * the last compound head, so generally we'll only do a full search if we don't
8873 * We check if the given compound head page has already been accounted, to
8874 * avoid double accounting it. This allows us to account the full size of the
8875 * page, not just the constituent pages of a huge page.
8877 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8878 int nr_pages, struct page *hpage)
8882 /* check current page array */
8883 for (i = 0; i < nr_pages; i++) {
8884 if (!PageCompound(pages[i]))
8886 if (compound_head(pages[i]) == hpage)
8890 /* check previously registered pages */
8891 for (i = 0; i < ctx->nr_user_bufs; i++) {
8892 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8894 for (j = 0; j < imu->nr_bvecs; j++) {
8895 if (!PageCompound(imu->bvec[j].bv_page))
8897 if (compound_head(imu->bvec[j].bv_page) == hpage)
8905 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8906 int nr_pages, struct io_mapped_ubuf *imu,
8907 struct page **last_hpage)
8911 imu->acct_pages = 0;
8912 for (i = 0; i < nr_pages; i++) {
8913 if (!PageCompound(pages[i])) {
8918 hpage = compound_head(pages[i]);
8919 if (hpage == *last_hpage)
8921 *last_hpage = hpage;
8922 if (headpage_already_acct(ctx, pages, i, hpage))
8924 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8928 if (!imu->acct_pages)
8931 ret = io_account_mem(ctx, imu->acct_pages);
8933 imu->acct_pages = 0;
8937 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8938 struct io_mapped_ubuf **pimu,
8939 struct page **last_hpage)
8941 struct io_mapped_ubuf *imu = NULL;
8942 struct vm_area_struct **vmas = NULL;
8943 struct page **pages = NULL;
8944 unsigned long off, start, end, ubuf;
8946 int ret, pret, nr_pages, i;
8948 if (!iov->iov_base) {
8949 *pimu = ctx->dummy_ubuf;
8953 ubuf = (unsigned long) iov->iov_base;
8954 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8955 start = ubuf >> PAGE_SHIFT;
8956 nr_pages = end - start;
8961 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8965 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8970 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8975 mmap_read_lock(current->mm);
8976 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8978 if (pret == nr_pages) {
8979 /* don't support file backed memory */
8980 for (i = 0; i < nr_pages; i++) {
8981 struct vm_area_struct *vma = vmas[i];
8983 if (vma_is_shmem(vma))
8986 !is_file_hugepages(vma->vm_file)) {
8992 ret = pret < 0 ? pret : -EFAULT;
8994 mmap_read_unlock(current->mm);
8997 * if we did partial map, or found file backed vmas,
8998 * release any pages we did get
9001 unpin_user_pages(pages, pret);
9005 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9007 unpin_user_pages(pages, pret);
9011 off = ubuf & ~PAGE_MASK;
9012 size = iov->iov_len;
9013 for (i = 0; i < nr_pages; i++) {
9016 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9017 imu->bvec[i].bv_page = pages[i];
9018 imu->bvec[i].bv_len = vec_len;
9019 imu->bvec[i].bv_offset = off;
9023 /* store original address for later verification */
9025 imu->ubuf_end = ubuf + iov->iov_len;
9026 imu->nr_bvecs = nr_pages;
9037 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9039 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9040 return ctx->user_bufs ? 0 : -ENOMEM;
9043 static int io_buffer_validate(struct iovec *iov)
9045 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9048 * Don't impose further limits on the size and buffer
9049 * constraints here, we'll -EINVAL later when IO is
9050 * submitted if they are wrong.
9053 return iov->iov_len ? -EFAULT : 0;
9057 /* arbitrary limit, but we need something */
9058 if (iov->iov_len > SZ_1G)
9061 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9067 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9068 unsigned int nr_args, u64 __user *tags)
9070 struct page *last_hpage = NULL;
9071 struct io_rsrc_data *data;
9077 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9079 ret = io_rsrc_node_switch_start(ctx);
9082 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9085 ret = io_buffers_map_alloc(ctx, nr_args);
9087 io_rsrc_data_free(data);
9091 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9092 ret = io_copy_iov(ctx, &iov, arg, i);
9095 ret = io_buffer_validate(&iov);
9098 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9103 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9109 WARN_ON_ONCE(ctx->buf_data);
9111 ctx->buf_data = data;
9113 __io_sqe_buffers_unregister(ctx);
9115 io_rsrc_node_switch(ctx, NULL);
9119 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9120 struct io_uring_rsrc_update2 *up,
9121 unsigned int nr_args)
9123 u64 __user *tags = u64_to_user_ptr(up->tags);
9124 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9125 struct page *last_hpage = NULL;
9126 bool needs_switch = false;
9132 if (up->offset + nr_args > ctx->nr_user_bufs)
9135 for (done = 0; done < nr_args; done++) {
9136 struct io_mapped_ubuf *imu;
9137 int offset = up->offset + done;
9140 err = io_copy_iov(ctx, &iov, iovs, done);
9143 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9147 err = io_buffer_validate(&iov);
9150 if (!iov.iov_base && tag) {
9154 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9158 i = array_index_nospec(offset, ctx->nr_user_bufs);
9159 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9160 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9161 ctx->rsrc_node, ctx->user_bufs[i]);
9162 if (unlikely(err)) {
9163 io_buffer_unmap(ctx, &imu);
9166 ctx->user_bufs[i] = NULL;
9167 needs_switch = true;
9170 ctx->user_bufs[i] = imu;
9171 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9175 io_rsrc_node_switch(ctx, ctx->buf_data);
9176 return done ? done : err;
9179 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9181 __s32 __user *fds = arg;
9187 if (copy_from_user(&fd, fds, sizeof(*fds)))
9190 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9191 if (IS_ERR(ctx->cq_ev_fd)) {
9192 int ret = PTR_ERR(ctx->cq_ev_fd);
9194 ctx->cq_ev_fd = NULL;
9201 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9203 if (ctx->cq_ev_fd) {
9204 eventfd_ctx_put(ctx->cq_ev_fd);
9205 ctx->cq_ev_fd = NULL;
9212 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9214 struct io_buffer *buf;
9215 unsigned long index;
9217 xa_for_each(&ctx->io_buffers, index, buf) {
9218 __io_remove_buffers(ctx, buf, index, -1U);
9223 static void io_req_cache_free(struct list_head *list)
9225 struct io_kiocb *req, *nxt;
9227 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9228 list_del(&req->inflight_entry);
9229 kmem_cache_free(req_cachep, req);
9233 static void io_req_caches_free(struct io_ring_ctx *ctx)
9235 struct io_submit_state *state = &ctx->submit_state;
9237 mutex_lock(&ctx->uring_lock);
9239 if (state->free_reqs) {
9240 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9241 state->free_reqs = 0;
9244 io_flush_cached_locked_reqs(ctx, state);
9245 io_req_cache_free(&state->free_list);
9246 mutex_unlock(&ctx->uring_lock);
9249 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9251 if (data && !atomic_dec_and_test(&data->refs))
9252 wait_for_completion(&data->done);
9255 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9257 io_sq_thread_finish(ctx);
9259 if (ctx->mm_account) {
9260 mmdrop(ctx->mm_account);
9261 ctx->mm_account = NULL;
9264 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9265 io_wait_rsrc_data(ctx->buf_data);
9266 io_wait_rsrc_data(ctx->file_data);
9268 mutex_lock(&ctx->uring_lock);
9270 __io_sqe_buffers_unregister(ctx);
9272 __io_sqe_files_unregister(ctx);
9274 __io_cqring_overflow_flush(ctx, true);
9275 mutex_unlock(&ctx->uring_lock);
9276 io_eventfd_unregister(ctx);
9277 io_destroy_buffers(ctx);
9279 put_cred(ctx->sq_creds);
9281 /* there are no registered resources left, nobody uses it */
9283 io_rsrc_node_destroy(ctx->rsrc_node);
9284 if (ctx->rsrc_backup_node)
9285 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9286 flush_delayed_work(&ctx->rsrc_put_work);
9288 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9289 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9291 #if defined(CONFIG_UNIX)
9292 if (ctx->ring_sock) {
9293 ctx->ring_sock->file = NULL; /* so that iput() is called */
9294 sock_release(ctx->ring_sock);
9297 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9299 io_mem_free(ctx->rings);
9300 io_mem_free(ctx->sq_sqes);
9302 percpu_ref_exit(&ctx->refs);
9303 free_uid(ctx->user);
9304 io_req_caches_free(ctx);
9306 io_wq_put_hash(ctx->hash_map);
9307 kfree(ctx->cancel_hash);
9308 kfree(ctx->dummy_ubuf);
9312 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9314 struct io_ring_ctx *ctx = file->private_data;
9317 poll_wait(file, &ctx->poll_wait, wait);
9319 * synchronizes with barrier from wq_has_sleeper call in
9323 if (!io_sqring_full(ctx))
9324 mask |= EPOLLOUT | EPOLLWRNORM;
9327 * Don't flush cqring overflow list here, just do a simple check.
9328 * Otherwise there could possible be ABBA deadlock:
9331 * lock(&ctx->uring_lock);
9333 * lock(&ctx->uring_lock);
9336 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9337 * pushs them to do the flush.
9339 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9340 mask |= EPOLLIN | EPOLLRDNORM;
9345 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9347 const struct cred *creds;
9349 creds = xa_erase(&ctx->personalities, id);
9358 struct io_tctx_exit {
9359 struct callback_head task_work;
9360 struct completion completion;
9361 struct io_ring_ctx *ctx;
9364 static void io_tctx_exit_cb(struct callback_head *cb)
9366 struct io_uring_task *tctx = current->io_uring;
9367 struct io_tctx_exit *work;
9369 work = container_of(cb, struct io_tctx_exit, task_work);
9371 * When @in_idle, we're in cancellation and it's racy to remove the
9372 * node. It'll be removed by the end of cancellation, just ignore it.
9374 if (!atomic_read(&tctx->in_idle))
9375 io_uring_del_tctx_node((unsigned long)work->ctx);
9376 complete(&work->completion);
9379 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9381 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9383 return req->ctx == data;
9386 static void io_ring_exit_work(struct work_struct *work)
9388 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9389 unsigned long timeout = jiffies + HZ * 60 * 5;
9390 unsigned long interval = HZ / 20;
9391 struct io_tctx_exit exit;
9392 struct io_tctx_node *node;
9396 * If we're doing polled IO and end up having requests being
9397 * submitted async (out-of-line), then completions can come in while
9398 * we're waiting for refs to drop. We need to reap these manually,
9399 * as nobody else will be looking for them.
9402 io_uring_try_cancel_requests(ctx, NULL, true);
9404 struct io_sq_data *sqd = ctx->sq_data;
9405 struct task_struct *tsk;
9407 io_sq_thread_park(sqd);
9409 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9410 io_wq_cancel_cb(tsk->io_uring->io_wq,
9411 io_cancel_ctx_cb, ctx, true);
9412 io_sq_thread_unpark(sqd);
9415 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9416 /* there is little hope left, don't run it too often */
9419 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9421 init_completion(&exit.completion);
9422 init_task_work(&exit.task_work, io_tctx_exit_cb);
9425 * Some may use context even when all refs and requests have been put,
9426 * and they are free to do so while still holding uring_lock or
9427 * completion_lock, see io_req_task_submit(). Apart from other work,
9428 * this lock/unlock section also waits them to finish.
9430 mutex_lock(&ctx->uring_lock);
9431 while (!list_empty(&ctx->tctx_list)) {
9432 WARN_ON_ONCE(time_after(jiffies, timeout));
9434 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9436 /* don't spin on a single task if cancellation failed */
9437 list_rotate_left(&ctx->tctx_list);
9438 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9439 if (WARN_ON_ONCE(ret))
9441 wake_up_process(node->task);
9443 mutex_unlock(&ctx->uring_lock);
9444 wait_for_completion(&exit.completion);
9445 mutex_lock(&ctx->uring_lock);
9447 mutex_unlock(&ctx->uring_lock);
9448 spin_lock(&ctx->completion_lock);
9449 spin_unlock(&ctx->completion_lock);
9451 io_ring_ctx_free(ctx);
9454 /* Returns true if we found and killed one or more timeouts */
9455 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9458 struct io_kiocb *req, *tmp;
9461 spin_lock(&ctx->completion_lock);
9462 spin_lock_irq(&ctx->timeout_lock);
9463 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9464 if (io_match_task(req, tsk, cancel_all)) {
9465 io_kill_timeout(req, -ECANCELED);
9469 spin_unlock_irq(&ctx->timeout_lock);
9471 io_commit_cqring(ctx);
9472 spin_unlock(&ctx->completion_lock);
9474 io_cqring_ev_posted(ctx);
9475 return canceled != 0;
9478 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9480 unsigned long index;
9481 struct creds *creds;
9483 mutex_lock(&ctx->uring_lock);
9484 percpu_ref_kill(&ctx->refs);
9486 __io_cqring_overflow_flush(ctx, true);
9487 xa_for_each(&ctx->personalities, index, creds)
9488 io_unregister_personality(ctx, index);
9489 mutex_unlock(&ctx->uring_lock);
9491 io_kill_timeouts(ctx, NULL, true);
9492 io_poll_remove_all(ctx, NULL, true);
9494 /* if we failed setting up the ctx, we might not have any rings */
9495 io_iopoll_try_reap_events(ctx);
9497 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9499 * Use system_unbound_wq to avoid spawning tons of event kworkers
9500 * if we're exiting a ton of rings at the same time. It just adds
9501 * noise and overhead, there's no discernable change in runtime
9502 * over using system_wq.
9504 queue_work(system_unbound_wq, &ctx->exit_work);
9507 static int io_uring_release(struct inode *inode, struct file *file)
9509 struct io_ring_ctx *ctx = file->private_data;
9511 file->private_data = NULL;
9512 io_ring_ctx_wait_and_kill(ctx);
9516 struct io_task_cancel {
9517 struct task_struct *task;
9521 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9523 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9524 struct io_task_cancel *cancel = data;
9527 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9528 struct io_ring_ctx *ctx = req->ctx;
9530 /* protect against races with linked timeouts */
9531 spin_lock(&ctx->completion_lock);
9532 ret = io_match_task(req, cancel->task, cancel->all);
9533 spin_unlock(&ctx->completion_lock);
9535 ret = io_match_task(req, cancel->task, cancel->all);
9540 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9541 struct task_struct *task, bool cancel_all)
9543 struct io_defer_entry *de;
9546 spin_lock(&ctx->completion_lock);
9547 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9548 if (io_match_task(de->req, task, cancel_all)) {
9549 list_cut_position(&list, &ctx->defer_list, &de->list);
9553 spin_unlock(&ctx->completion_lock);
9554 if (list_empty(&list))
9557 while (!list_empty(&list)) {
9558 de = list_first_entry(&list, struct io_defer_entry, list);
9559 list_del_init(&de->list);
9560 io_req_complete_failed(de->req, -ECANCELED);
9566 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9568 struct io_tctx_node *node;
9569 enum io_wq_cancel cret;
9572 mutex_lock(&ctx->uring_lock);
9573 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9574 struct io_uring_task *tctx = node->task->io_uring;
9577 * io_wq will stay alive while we hold uring_lock, because it's
9578 * killed after ctx nodes, which requires to take the lock.
9580 if (!tctx || !tctx->io_wq)
9582 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9583 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9585 mutex_unlock(&ctx->uring_lock);
9590 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9591 struct task_struct *task,
9594 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9595 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9598 enum io_wq_cancel cret;
9602 ret |= io_uring_try_cancel_iowq(ctx);
9603 } else if (tctx && tctx->io_wq) {
9605 * Cancels requests of all rings, not only @ctx, but
9606 * it's fine as the task is in exit/exec.
9608 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9610 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9613 /* SQPOLL thread does its own polling */
9614 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9615 (ctx->sq_data && ctx->sq_data->thread == current)) {
9616 while (!list_empty_careful(&ctx->iopoll_list)) {
9617 io_iopoll_try_reap_events(ctx);
9622 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9623 ret |= io_poll_remove_all(ctx, task, cancel_all);
9624 ret |= io_kill_timeouts(ctx, task, cancel_all);
9626 ret |= io_run_task_work();
9633 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9635 struct io_uring_task *tctx = current->io_uring;
9636 struct io_tctx_node *node;
9639 if (unlikely(!tctx)) {
9640 ret = io_uring_alloc_task_context(current, ctx);
9644 tctx = current->io_uring;
9645 if (ctx->iowq_limits_set) {
9646 unsigned int limits[2] = { ctx->iowq_limits[0],
9647 ctx->iowq_limits[1], };
9649 ret = io_wq_max_workers(tctx->io_wq, limits);
9654 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9655 node = kmalloc(sizeof(*node), GFP_KERNEL);
9659 node->task = current;
9661 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9668 mutex_lock(&ctx->uring_lock);
9669 list_add(&node->ctx_node, &ctx->tctx_list);
9670 mutex_unlock(&ctx->uring_lock);
9677 * Note that this task has used io_uring. We use it for cancelation purposes.
9679 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9681 struct io_uring_task *tctx = current->io_uring;
9683 if (likely(tctx && tctx->last == ctx))
9685 return __io_uring_add_tctx_node(ctx);
9689 * Remove this io_uring_file -> task mapping.
9691 static void io_uring_del_tctx_node(unsigned long index)
9693 struct io_uring_task *tctx = current->io_uring;
9694 struct io_tctx_node *node;
9698 node = xa_erase(&tctx->xa, index);
9702 WARN_ON_ONCE(current != node->task);
9703 WARN_ON_ONCE(list_empty(&node->ctx_node));
9705 mutex_lock(&node->ctx->uring_lock);
9706 list_del(&node->ctx_node);
9707 mutex_unlock(&node->ctx->uring_lock);
9709 if (tctx->last == node->ctx)
9714 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9716 struct io_wq *wq = tctx->io_wq;
9717 struct io_tctx_node *node;
9718 unsigned long index;
9720 xa_for_each(&tctx->xa, index, node) {
9721 io_uring_del_tctx_node(index);
9726 * Must be after io_uring_del_task_file() (removes nodes under
9727 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9729 io_wq_put_and_exit(wq);
9734 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9737 return atomic_read(&tctx->inflight_tracked);
9738 return percpu_counter_sum(&tctx->inflight);
9741 static void io_uring_drop_tctx_refs(struct task_struct *task)
9743 struct io_uring_task *tctx = task->io_uring;
9744 unsigned int refs = tctx->cached_refs;
9747 tctx->cached_refs = 0;
9748 percpu_counter_sub(&tctx->inflight, refs);
9749 put_task_struct_many(task, refs);
9754 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9755 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9757 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9759 struct io_uring_task *tctx = current->io_uring;
9760 struct io_ring_ctx *ctx;
9764 WARN_ON_ONCE(sqd && sqd->thread != current);
9766 if (!current->io_uring)
9769 io_wq_exit_start(tctx->io_wq);
9771 atomic_inc(&tctx->in_idle);
9773 io_uring_drop_tctx_refs(current);
9774 /* read completions before cancelations */
9775 inflight = tctx_inflight(tctx, !cancel_all);
9780 struct io_tctx_node *node;
9781 unsigned long index;
9783 xa_for_each(&tctx->xa, index, node) {
9784 /* sqpoll task will cancel all its requests */
9785 if (node->ctx->sq_data)
9787 io_uring_try_cancel_requests(node->ctx, current,
9791 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9792 io_uring_try_cancel_requests(ctx, current,
9796 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9797 io_uring_drop_tctx_refs(current);
9799 * If we've seen completions, retry without waiting. This
9800 * avoids a race where a completion comes in before we did
9801 * prepare_to_wait().
9803 if (inflight == tctx_inflight(tctx, !cancel_all))
9805 finish_wait(&tctx->wait, &wait);
9807 atomic_dec(&tctx->in_idle);
9809 io_uring_clean_tctx(tctx);
9811 /* for exec all current's requests should be gone, kill tctx */
9812 __io_uring_free(current);
9816 void __io_uring_cancel(bool cancel_all)
9818 io_uring_cancel_generic(cancel_all, NULL);
9821 static void *io_uring_validate_mmap_request(struct file *file,
9822 loff_t pgoff, size_t sz)
9824 struct io_ring_ctx *ctx = file->private_data;
9825 loff_t offset = pgoff << PAGE_SHIFT;
9830 case IORING_OFF_SQ_RING:
9831 case IORING_OFF_CQ_RING:
9834 case IORING_OFF_SQES:
9838 return ERR_PTR(-EINVAL);
9841 page = virt_to_head_page(ptr);
9842 if (sz > page_size(page))
9843 return ERR_PTR(-EINVAL);
9850 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9852 size_t sz = vma->vm_end - vma->vm_start;
9856 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9858 return PTR_ERR(ptr);
9860 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9861 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9864 #else /* !CONFIG_MMU */
9866 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9868 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9871 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9873 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9876 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9877 unsigned long addr, unsigned long len,
9878 unsigned long pgoff, unsigned long flags)
9882 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9884 return PTR_ERR(ptr);
9886 return (unsigned long) ptr;
9889 #endif /* !CONFIG_MMU */
9891 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9896 if (!io_sqring_full(ctx))
9898 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9900 if (!io_sqring_full(ctx))
9903 } while (!signal_pending(current));
9905 finish_wait(&ctx->sqo_sq_wait, &wait);
9909 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9910 struct __kernel_timespec __user **ts,
9911 const sigset_t __user **sig)
9913 struct io_uring_getevents_arg arg;
9916 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9917 * is just a pointer to the sigset_t.
9919 if (!(flags & IORING_ENTER_EXT_ARG)) {
9920 *sig = (const sigset_t __user *) argp;
9926 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9927 * timespec and sigset_t pointers if good.
9929 if (*argsz != sizeof(arg))
9931 if (copy_from_user(&arg, argp, sizeof(arg)))
9933 *sig = u64_to_user_ptr(arg.sigmask);
9934 *argsz = arg.sigmask_sz;
9935 *ts = u64_to_user_ptr(arg.ts);
9939 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9940 u32, min_complete, u32, flags, const void __user *, argp,
9943 struct io_ring_ctx *ctx;
9950 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9951 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9955 if (unlikely(!f.file))
9959 if (unlikely(f.file->f_op != &io_uring_fops))
9963 ctx = f.file->private_data;
9964 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9968 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9972 * For SQ polling, the thread will do all submissions and completions.
9973 * Just return the requested submit count, and wake the thread if
9977 if (ctx->flags & IORING_SETUP_SQPOLL) {
9978 io_cqring_overflow_flush(ctx);
9980 if (unlikely(ctx->sq_data->thread == NULL)) {
9984 if (flags & IORING_ENTER_SQ_WAKEUP)
9985 wake_up(&ctx->sq_data->wait);
9986 if (flags & IORING_ENTER_SQ_WAIT) {
9987 ret = io_sqpoll_wait_sq(ctx);
9991 submitted = to_submit;
9992 } else if (to_submit) {
9993 ret = io_uring_add_tctx_node(ctx);
9996 mutex_lock(&ctx->uring_lock);
9997 submitted = io_submit_sqes(ctx, to_submit);
9998 mutex_unlock(&ctx->uring_lock);
10000 if (submitted != to_submit)
10003 if (flags & IORING_ENTER_GETEVENTS) {
10004 const sigset_t __user *sig;
10005 struct __kernel_timespec __user *ts;
10007 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10011 min_complete = min(min_complete, ctx->cq_entries);
10014 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10015 * space applications don't need to do io completion events
10016 * polling again, they can rely on io_sq_thread to do polling
10017 * work, which can reduce cpu usage and uring_lock contention.
10019 if (ctx->flags & IORING_SETUP_IOPOLL &&
10020 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10021 ret = io_iopoll_check(ctx, min_complete);
10023 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10028 percpu_ref_put(&ctx->refs);
10031 return submitted ? submitted : ret;
10034 #ifdef CONFIG_PROC_FS
10035 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10036 const struct cred *cred)
10038 struct user_namespace *uns = seq_user_ns(m);
10039 struct group_info *gi;
10044 seq_printf(m, "%5d\n", id);
10045 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10046 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10047 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10048 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10049 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10050 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10051 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10052 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10053 seq_puts(m, "\n\tGroups:\t");
10054 gi = cred->group_info;
10055 for (g = 0; g < gi->ngroups; g++) {
10056 seq_put_decimal_ull(m, g ? " " : "",
10057 from_kgid_munged(uns, gi->gid[g]));
10059 seq_puts(m, "\n\tCapEff:\t");
10060 cap = cred->cap_effective;
10061 CAP_FOR_EACH_U32(__capi)
10062 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10067 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10069 struct io_sq_data *sq = NULL;
10074 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10075 * since fdinfo case grabs it in the opposite direction of normal use
10076 * cases. If we fail to get the lock, we just don't iterate any
10077 * structures that could be going away outside the io_uring mutex.
10079 has_lock = mutex_trylock(&ctx->uring_lock);
10081 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10087 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10088 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10089 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10090 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10091 struct file *f = io_file_from_index(ctx, i);
10094 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10096 seq_printf(m, "%5u: <none>\n", i);
10098 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10099 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10100 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10101 unsigned int len = buf->ubuf_end - buf->ubuf;
10103 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10105 if (has_lock && !xa_empty(&ctx->personalities)) {
10106 unsigned long index;
10107 const struct cred *cred;
10109 seq_printf(m, "Personalities:\n");
10110 xa_for_each(&ctx->personalities, index, cred)
10111 io_uring_show_cred(m, index, cred);
10113 seq_printf(m, "PollList:\n");
10114 spin_lock(&ctx->completion_lock);
10115 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10116 struct hlist_head *list = &ctx->cancel_hash[i];
10117 struct io_kiocb *req;
10119 hlist_for_each_entry(req, list, hash_node)
10120 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10121 req->task->task_works != NULL);
10123 spin_unlock(&ctx->completion_lock);
10125 mutex_unlock(&ctx->uring_lock);
10128 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10130 struct io_ring_ctx *ctx = f->private_data;
10132 if (percpu_ref_tryget(&ctx->refs)) {
10133 __io_uring_show_fdinfo(ctx, m);
10134 percpu_ref_put(&ctx->refs);
10139 static const struct file_operations io_uring_fops = {
10140 .release = io_uring_release,
10141 .mmap = io_uring_mmap,
10143 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10144 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10146 .poll = io_uring_poll,
10147 #ifdef CONFIG_PROC_FS
10148 .show_fdinfo = io_uring_show_fdinfo,
10152 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10153 struct io_uring_params *p)
10155 struct io_rings *rings;
10156 size_t size, sq_array_offset;
10158 /* make sure these are sane, as we already accounted them */
10159 ctx->sq_entries = p->sq_entries;
10160 ctx->cq_entries = p->cq_entries;
10162 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10163 if (size == SIZE_MAX)
10166 rings = io_mem_alloc(size);
10170 ctx->rings = rings;
10171 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10172 rings->sq_ring_mask = p->sq_entries - 1;
10173 rings->cq_ring_mask = p->cq_entries - 1;
10174 rings->sq_ring_entries = p->sq_entries;
10175 rings->cq_ring_entries = p->cq_entries;
10177 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10178 if (size == SIZE_MAX) {
10179 io_mem_free(ctx->rings);
10184 ctx->sq_sqes = io_mem_alloc(size);
10185 if (!ctx->sq_sqes) {
10186 io_mem_free(ctx->rings);
10194 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10198 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10202 ret = io_uring_add_tctx_node(ctx);
10207 fd_install(fd, file);
10212 * Allocate an anonymous fd, this is what constitutes the application
10213 * visible backing of an io_uring instance. The application mmaps this
10214 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10215 * we have to tie this fd to a socket for file garbage collection purposes.
10217 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10220 #if defined(CONFIG_UNIX)
10223 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10226 return ERR_PTR(ret);
10229 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10230 O_RDWR | O_CLOEXEC);
10231 #if defined(CONFIG_UNIX)
10232 if (IS_ERR(file)) {
10233 sock_release(ctx->ring_sock);
10234 ctx->ring_sock = NULL;
10236 ctx->ring_sock->file = file;
10242 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10243 struct io_uring_params __user *params)
10245 struct io_ring_ctx *ctx;
10251 if (entries > IORING_MAX_ENTRIES) {
10252 if (!(p->flags & IORING_SETUP_CLAMP))
10254 entries = IORING_MAX_ENTRIES;
10258 * Use twice as many entries for the CQ ring. It's possible for the
10259 * application to drive a higher depth than the size of the SQ ring,
10260 * since the sqes are only used at submission time. This allows for
10261 * some flexibility in overcommitting a bit. If the application has
10262 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10263 * of CQ ring entries manually.
10265 p->sq_entries = roundup_pow_of_two(entries);
10266 if (p->flags & IORING_SETUP_CQSIZE) {
10268 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10269 * to a power-of-two, if it isn't already. We do NOT impose
10270 * any cq vs sq ring sizing.
10272 if (!p->cq_entries)
10274 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10275 if (!(p->flags & IORING_SETUP_CLAMP))
10277 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10279 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10280 if (p->cq_entries < p->sq_entries)
10283 p->cq_entries = 2 * p->sq_entries;
10286 ctx = io_ring_ctx_alloc(p);
10289 ctx->compat = in_compat_syscall();
10290 if (!capable(CAP_IPC_LOCK))
10291 ctx->user = get_uid(current_user());
10294 * This is just grabbed for accounting purposes. When a process exits,
10295 * the mm is exited and dropped before the files, hence we need to hang
10296 * on to this mm purely for the purposes of being able to unaccount
10297 * memory (locked/pinned vm). It's not used for anything else.
10299 mmgrab(current->mm);
10300 ctx->mm_account = current->mm;
10302 ret = io_allocate_scq_urings(ctx, p);
10306 ret = io_sq_offload_create(ctx, p);
10309 /* always set a rsrc node */
10310 ret = io_rsrc_node_switch_start(ctx);
10313 io_rsrc_node_switch(ctx, NULL);
10315 memset(&p->sq_off, 0, sizeof(p->sq_off));
10316 p->sq_off.head = offsetof(struct io_rings, sq.head);
10317 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10318 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10319 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10320 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10321 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10322 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10324 memset(&p->cq_off, 0, sizeof(p->cq_off));
10325 p->cq_off.head = offsetof(struct io_rings, cq.head);
10326 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10327 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10328 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10329 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10330 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10331 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10333 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10334 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10335 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10336 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10337 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10338 IORING_FEAT_RSRC_TAGS;
10340 if (copy_to_user(params, p, sizeof(*p))) {
10345 file = io_uring_get_file(ctx);
10346 if (IS_ERR(file)) {
10347 ret = PTR_ERR(file);
10352 * Install ring fd as the very last thing, so we don't risk someone
10353 * having closed it before we finish setup
10355 ret = io_uring_install_fd(ctx, file);
10357 /* fput will clean it up */
10362 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10365 io_ring_ctx_wait_and_kill(ctx);
10370 * Sets up an aio uring context, and returns the fd. Applications asks for a
10371 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10372 * params structure passed in.
10374 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10376 struct io_uring_params p;
10379 if (copy_from_user(&p, params, sizeof(p)))
10381 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10386 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10387 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10388 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10389 IORING_SETUP_R_DISABLED))
10392 return io_uring_create(entries, &p, params);
10395 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10396 struct io_uring_params __user *, params)
10398 return io_uring_setup(entries, params);
10401 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10403 struct io_uring_probe *p;
10407 size = struct_size(p, ops, nr_args);
10408 if (size == SIZE_MAX)
10410 p = kzalloc(size, GFP_KERNEL);
10415 if (copy_from_user(p, arg, size))
10418 if (memchr_inv(p, 0, size))
10421 p->last_op = IORING_OP_LAST - 1;
10422 if (nr_args > IORING_OP_LAST)
10423 nr_args = IORING_OP_LAST;
10425 for (i = 0; i < nr_args; i++) {
10427 if (!io_op_defs[i].not_supported)
10428 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10433 if (copy_to_user(arg, p, size))
10440 static int io_register_personality(struct io_ring_ctx *ctx)
10442 const struct cred *creds;
10446 creds = get_current_cred();
10448 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10449 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10457 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10458 unsigned int nr_args)
10460 struct io_uring_restriction *res;
10464 /* Restrictions allowed only if rings started disabled */
10465 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10468 /* We allow only a single restrictions registration */
10469 if (ctx->restrictions.registered)
10472 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10475 size = array_size(nr_args, sizeof(*res));
10476 if (size == SIZE_MAX)
10479 res = memdup_user(arg, size);
10481 return PTR_ERR(res);
10485 for (i = 0; i < nr_args; i++) {
10486 switch (res[i].opcode) {
10487 case IORING_RESTRICTION_REGISTER_OP:
10488 if (res[i].register_op >= IORING_REGISTER_LAST) {
10493 __set_bit(res[i].register_op,
10494 ctx->restrictions.register_op);
10496 case IORING_RESTRICTION_SQE_OP:
10497 if (res[i].sqe_op >= IORING_OP_LAST) {
10502 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10504 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10505 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10507 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10508 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10517 /* Reset all restrictions if an error happened */
10519 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10521 ctx->restrictions.registered = true;
10527 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10529 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10532 if (ctx->restrictions.registered)
10533 ctx->restricted = 1;
10535 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10536 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10537 wake_up(&ctx->sq_data->wait);
10541 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10542 struct io_uring_rsrc_update2 *up,
10550 if (check_add_overflow(up->offset, nr_args, &tmp))
10552 err = io_rsrc_node_switch_start(ctx);
10557 case IORING_RSRC_FILE:
10558 return __io_sqe_files_update(ctx, up, nr_args);
10559 case IORING_RSRC_BUFFER:
10560 return __io_sqe_buffers_update(ctx, up, nr_args);
10565 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10568 struct io_uring_rsrc_update2 up;
10572 memset(&up, 0, sizeof(up));
10573 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10575 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10578 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10579 unsigned size, unsigned type)
10581 struct io_uring_rsrc_update2 up;
10583 if (size != sizeof(up))
10585 if (copy_from_user(&up, arg, sizeof(up)))
10587 if (!up.nr || up.resv)
10589 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10592 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10593 unsigned int size, unsigned int type)
10595 struct io_uring_rsrc_register rr;
10597 /* keep it extendible */
10598 if (size != sizeof(rr))
10601 memset(&rr, 0, sizeof(rr));
10602 if (copy_from_user(&rr, arg, size))
10604 if (!rr.nr || rr.resv || rr.resv2)
10608 case IORING_RSRC_FILE:
10609 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10610 rr.nr, u64_to_user_ptr(rr.tags));
10611 case IORING_RSRC_BUFFER:
10612 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10613 rr.nr, u64_to_user_ptr(rr.tags));
10618 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10621 struct io_uring_task *tctx = current->io_uring;
10622 cpumask_var_t new_mask;
10625 if (!tctx || !tctx->io_wq)
10628 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10631 cpumask_clear(new_mask);
10632 if (len > cpumask_size())
10633 len = cpumask_size();
10635 if (copy_from_user(new_mask, arg, len)) {
10636 free_cpumask_var(new_mask);
10640 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10641 free_cpumask_var(new_mask);
10645 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10647 struct io_uring_task *tctx = current->io_uring;
10649 if (!tctx || !tctx->io_wq)
10652 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10655 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10658 struct io_uring_task *tctx = NULL;
10659 struct io_sq_data *sqd = NULL;
10660 __u32 new_count[2];
10663 if (copy_from_user(new_count, arg, sizeof(new_count)))
10665 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10666 if (new_count[i] > INT_MAX)
10669 if (ctx->flags & IORING_SETUP_SQPOLL) {
10670 sqd = ctx->sq_data;
10673 * Observe the correct sqd->lock -> ctx->uring_lock
10674 * ordering. Fine to drop uring_lock here, we hold
10675 * a ref to the ctx.
10677 refcount_inc(&sqd->refs);
10678 mutex_unlock(&ctx->uring_lock);
10679 mutex_lock(&sqd->lock);
10680 mutex_lock(&ctx->uring_lock);
10682 tctx = sqd->thread->io_uring;
10685 tctx = current->io_uring;
10688 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10690 memcpy(ctx->iowq_limits, new_count, sizeof(new_count));
10691 ctx->iowq_limits_set = true;
10694 if (tctx && tctx->io_wq) {
10695 ret = io_wq_max_workers(tctx->io_wq, new_count);
10699 memset(new_count, 0, sizeof(new_count));
10703 mutex_unlock(&sqd->lock);
10704 io_put_sq_data(sqd);
10707 if (copy_to_user(arg, new_count, sizeof(new_count)))
10713 mutex_unlock(&sqd->lock);
10714 io_put_sq_data(sqd);
10719 static bool io_register_op_must_quiesce(int op)
10722 case IORING_REGISTER_BUFFERS:
10723 case IORING_UNREGISTER_BUFFERS:
10724 case IORING_REGISTER_FILES:
10725 case IORING_UNREGISTER_FILES:
10726 case IORING_REGISTER_FILES_UPDATE:
10727 case IORING_REGISTER_PROBE:
10728 case IORING_REGISTER_PERSONALITY:
10729 case IORING_UNREGISTER_PERSONALITY:
10730 case IORING_REGISTER_FILES2:
10731 case IORING_REGISTER_FILES_UPDATE2:
10732 case IORING_REGISTER_BUFFERS2:
10733 case IORING_REGISTER_BUFFERS_UPDATE:
10734 case IORING_REGISTER_IOWQ_AFF:
10735 case IORING_UNREGISTER_IOWQ_AFF:
10736 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10743 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10747 percpu_ref_kill(&ctx->refs);
10750 * Drop uring mutex before waiting for references to exit. If another
10751 * thread is currently inside io_uring_enter() it might need to grab the
10752 * uring_lock to make progress. If we hold it here across the drain
10753 * wait, then we can deadlock. It's safe to drop the mutex here, since
10754 * no new references will come in after we've killed the percpu ref.
10756 mutex_unlock(&ctx->uring_lock);
10758 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10761 ret = io_run_task_work_sig();
10762 } while (ret >= 0);
10763 mutex_lock(&ctx->uring_lock);
10766 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10770 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10771 void __user *arg, unsigned nr_args)
10772 __releases(ctx->uring_lock)
10773 __acquires(ctx->uring_lock)
10778 * We're inside the ring mutex, if the ref is already dying, then
10779 * someone else killed the ctx or is already going through
10780 * io_uring_register().
10782 if (percpu_ref_is_dying(&ctx->refs))
10785 if (ctx->restricted) {
10786 if (opcode >= IORING_REGISTER_LAST)
10788 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10789 if (!test_bit(opcode, ctx->restrictions.register_op))
10793 if (io_register_op_must_quiesce(opcode)) {
10794 ret = io_ctx_quiesce(ctx);
10800 case IORING_REGISTER_BUFFERS:
10801 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10803 case IORING_UNREGISTER_BUFFERS:
10805 if (arg || nr_args)
10807 ret = io_sqe_buffers_unregister(ctx);
10809 case IORING_REGISTER_FILES:
10810 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10812 case IORING_UNREGISTER_FILES:
10814 if (arg || nr_args)
10816 ret = io_sqe_files_unregister(ctx);
10818 case IORING_REGISTER_FILES_UPDATE:
10819 ret = io_register_files_update(ctx, arg, nr_args);
10821 case IORING_REGISTER_EVENTFD:
10822 case IORING_REGISTER_EVENTFD_ASYNC:
10826 ret = io_eventfd_register(ctx, arg);
10829 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10830 ctx->eventfd_async = 1;
10832 ctx->eventfd_async = 0;
10834 case IORING_UNREGISTER_EVENTFD:
10836 if (arg || nr_args)
10838 ret = io_eventfd_unregister(ctx);
10840 case IORING_REGISTER_PROBE:
10842 if (!arg || nr_args > 256)
10844 ret = io_probe(ctx, arg, nr_args);
10846 case IORING_REGISTER_PERSONALITY:
10848 if (arg || nr_args)
10850 ret = io_register_personality(ctx);
10852 case IORING_UNREGISTER_PERSONALITY:
10856 ret = io_unregister_personality(ctx, nr_args);
10858 case IORING_REGISTER_ENABLE_RINGS:
10860 if (arg || nr_args)
10862 ret = io_register_enable_rings(ctx);
10864 case IORING_REGISTER_RESTRICTIONS:
10865 ret = io_register_restrictions(ctx, arg, nr_args);
10867 case IORING_REGISTER_FILES2:
10868 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10870 case IORING_REGISTER_FILES_UPDATE2:
10871 ret = io_register_rsrc_update(ctx, arg, nr_args,
10874 case IORING_REGISTER_BUFFERS2:
10875 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10877 case IORING_REGISTER_BUFFERS_UPDATE:
10878 ret = io_register_rsrc_update(ctx, arg, nr_args,
10879 IORING_RSRC_BUFFER);
10881 case IORING_REGISTER_IOWQ_AFF:
10883 if (!arg || !nr_args)
10885 ret = io_register_iowq_aff(ctx, arg, nr_args);
10887 case IORING_UNREGISTER_IOWQ_AFF:
10889 if (arg || nr_args)
10891 ret = io_unregister_iowq_aff(ctx);
10893 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10895 if (!arg || nr_args != 2)
10897 ret = io_register_iowq_max_workers(ctx, arg);
10904 if (io_register_op_must_quiesce(opcode)) {
10905 /* bring the ctx back to life */
10906 percpu_ref_reinit(&ctx->refs);
10907 reinit_completion(&ctx->ref_comp);
10912 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10913 void __user *, arg, unsigned int, nr_args)
10915 struct io_ring_ctx *ctx;
10924 if (f.file->f_op != &io_uring_fops)
10927 ctx = f.file->private_data;
10929 io_run_task_work();
10931 mutex_lock(&ctx->uring_lock);
10932 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10933 mutex_unlock(&ctx->uring_lock);
10934 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10935 ctx->cq_ev_fd != NULL, ret);
10941 static int __init io_uring_init(void)
10943 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10944 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10945 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10948 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10949 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10950 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10951 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10952 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10953 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10954 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10955 BUILD_BUG_SQE_ELEM(8, __u64, off);
10956 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10957 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10958 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10959 BUILD_BUG_SQE_ELEM(24, __u32, len);
10960 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10961 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10962 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10963 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10964 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10965 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10966 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10967 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10968 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10969 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10970 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10971 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10972 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10973 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10974 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10975 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10976 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10977 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10978 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10979 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10980 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10982 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10983 sizeof(struct io_uring_rsrc_update));
10984 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10985 sizeof(struct io_uring_rsrc_update2));
10987 /* ->buf_index is u16 */
10988 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10990 /* should fit into one byte */
10991 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10993 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10994 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
10996 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11000 __initcall(io_uring_init);