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_cqring (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>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 struct bio_vec *bvec;
199 unsigned int nr_bvecs;
200 unsigned long acct_pages;
206 struct list_head list;
213 struct fixed_rsrc_table {
217 struct fixed_rsrc_ref_node {
218 struct percpu_ref refs;
219 struct list_head node;
220 struct list_head rsrc_list;
221 struct fixed_rsrc_data *rsrc_data;
222 void (*rsrc_put)(struct io_ring_ctx *ctx,
223 struct io_rsrc_put *prsrc);
224 struct llist_node llist;
228 struct fixed_rsrc_data {
229 struct fixed_rsrc_table *table;
230 struct io_ring_ctx *ctx;
232 struct fixed_rsrc_ref_node *node;
233 struct percpu_ref refs;
234 struct completion done;
239 struct list_head list;
245 struct io_restriction {
246 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
247 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
248 u8 sqe_flags_allowed;
249 u8 sqe_flags_required;
254 IO_SQ_THREAD_SHOULD_STOP = 0,
255 IO_SQ_THREAD_SHOULD_PARK,
260 atomic_t park_pending;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
266 struct task_struct *thread;
267 struct wait_queue_head wait;
269 unsigned sq_thread_idle;
275 struct completion exited;
276 struct callback_head *park_task_work;
279 #define IO_IOPOLL_BATCH 8
280 #define IO_COMPL_BATCH 32
281 #define IO_REQ_CACHE_SIZE 32
282 #define IO_REQ_ALLOC_BATCH 8
284 struct io_comp_state {
285 struct io_kiocb *reqs[IO_COMPL_BATCH];
287 unsigned int locked_free_nr;
288 /* inline/task_work completion list, under ->uring_lock */
289 struct list_head free_list;
290 /* IRQ completion list, under ->completion_lock */
291 struct list_head locked_free_list;
294 struct io_submit_link {
295 struct io_kiocb *head;
296 struct io_kiocb *last;
299 struct io_submit_state {
300 struct blk_plug plug;
301 struct io_submit_link link;
304 * io_kiocb alloc cache
306 void *reqs[IO_REQ_CACHE_SIZE];
307 unsigned int free_reqs;
312 * Batch completion logic
314 struct io_comp_state comp;
317 * File reference cache
321 unsigned int file_refs;
322 unsigned int ios_left;
327 struct percpu_ref refs;
328 } ____cacheline_aligned_in_smp;
332 unsigned int compat: 1;
333 unsigned int cq_overflow_flushed: 1;
334 unsigned int drain_next: 1;
335 unsigned int eventfd_async: 1;
336 unsigned int restricted: 1;
339 * Ring buffer of indices into array of io_uring_sqe, which is
340 * mmapped by the application using the IORING_OFF_SQES offset.
342 * This indirection could e.g. be used to assign fixed
343 * io_uring_sqe entries to operations and only submit them to
344 * the queue when needed.
346 * The kernel modifies neither the indices array nor the entries
350 unsigned cached_sq_head;
353 unsigned sq_thread_idle;
354 unsigned cached_sq_dropped;
355 unsigned cached_cq_overflow;
356 unsigned long sq_check_overflow;
358 /* hashed buffered write serialization */
359 struct io_wq_hash *hash_map;
361 struct list_head defer_list;
362 struct list_head timeout_list;
363 struct list_head cq_overflow_list;
365 struct io_uring_sqe *sq_sqes;
366 } ____cacheline_aligned_in_smp;
369 struct mutex uring_lock;
370 wait_queue_head_t wait;
371 } ____cacheline_aligned_in_smp;
373 struct io_submit_state submit_state;
375 struct io_rings *rings;
377 /* Only used for accounting purposes */
378 struct mm_struct *mm_account;
380 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
381 struct io_sq_data *sq_data; /* if using sq thread polling */
383 struct wait_queue_head sqo_sq_wait;
384 struct list_head sqd_list;
387 * If used, fixed file set. Writers must ensure that ->refs is dead,
388 * readers must ensure that ->refs is alive as long as the file* is
389 * used. Only updated through io_uring_register(2).
391 struct fixed_rsrc_data *file_data;
392 unsigned nr_user_files;
394 /* if used, fixed mapped user buffers */
395 unsigned nr_user_bufs;
396 struct io_mapped_ubuf *user_bufs;
398 struct user_struct *user;
400 struct completion ref_comp;
402 #if defined(CONFIG_UNIX)
403 struct socket *ring_sock;
406 struct xarray io_buffers;
408 struct xarray personalities;
412 unsigned cached_cq_tail;
415 atomic_t cq_timeouts;
416 unsigned cq_last_tm_flush;
417 unsigned long cq_check_overflow;
418 struct wait_queue_head cq_wait;
419 struct fasync_struct *cq_fasync;
420 struct eventfd_ctx *cq_ev_fd;
421 } ____cacheline_aligned_in_smp;
424 spinlock_t completion_lock;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list;
433 struct hlist_head *cancel_hash;
434 unsigned cancel_hash_bits;
435 bool poll_multi_file;
437 spinlock_t inflight_lock;
438 struct list_head inflight_list;
439 } ____cacheline_aligned_in_smp;
441 struct delayed_work rsrc_put_work;
442 struct llist_head rsrc_put_llist;
443 struct list_head rsrc_ref_list;
444 spinlock_t rsrc_ref_lock;
445 struct fixed_rsrc_ref_node *rsrc_backup_node;
447 struct io_restriction restrictions;
450 struct callback_head *exit_task_work;
452 struct wait_queue_head hash_wait;
454 /* Keep this last, we don't need it for the fast path */
455 struct work_struct exit_work;
456 struct list_head tctx_list;
459 struct io_uring_task {
460 /* submission side */
462 struct wait_queue_head wait;
463 const struct io_ring_ctx *last;
465 struct percpu_counter inflight;
469 spinlock_t task_lock;
470 struct io_wq_work_list task_list;
471 unsigned long task_state;
472 struct callback_head task_work;
476 * First field must be the file pointer in all the
477 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
479 struct io_poll_iocb {
481 struct wait_queue_head *head;
485 struct wait_queue_entry wait;
488 struct io_poll_remove {
498 struct io_timeout_data {
499 struct io_kiocb *req;
500 struct hrtimer timer;
501 struct timespec64 ts;
502 enum hrtimer_mode mode;
507 struct sockaddr __user *addr;
508 int __user *addr_len;
510 unsigned long nofile;
530 struct list_head list;
531 /* head of the link, used by linked timeouts only */
532 struct io_kiocb *head;
535 struct io_timeout_rem {
540 struct timespec64 ts;
545 /* NOTE: kiocb has the file as the first member, so don't do it here */
553 struct sockaddr __user *addr;
560 struct user_msghdr __user *umsg;
566 struct io_buffer *kbuf;
572 struct filename *filename;
574 unsigned long nofile;
577 struct io_rsrc_update {
603 struct epoll_event event;
607 struct file *file_out;
608 struct file *file_in;
615 struct io_provide_buf {
629 const char __user *filename;
630 struct statx __user *buffer;
642 struct filename *oldpath;
643 struct filename *newpath;
651 struct filename *filename;
654 struct io_completion {
656 struct list_head list;
660 struct io_async_connect {
661 struct sockaddr_storage address;
664 struct io_async_msghdr {
665 struct iovec fast_iov[UIO_FASTIOV];
666 /* points to an allocated iov, if NULL we use fast_iov instead */
667 struct iovec *free_iov;
668 struct sockaddr __user *uaddr;
670 struct sockaddr_storage addr;
674 struct iovec fast_iov[UIO_FASTIOV];
675 const struct iovec *free_iovec;
676 struct iov_iter iter;
678 struct wait_page_queue wpq;
682 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
683 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
684 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
685 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
686 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
687 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
693 REQ_F_LINK_TIMEOUT_BIT,
694 REQ_F_NEED_CLEANUP_BIT,
696 REQ_F_BUFFER_SELECTED_BIT,
697 REQ_F_NO_FILE_TABLE_BIT,
698 REQ_F_LTIMEOUT_ACTIVE_BIT,
699 REQ_F_COMPLETE_INLINE_BIT,
701 /* keep async read/write and isreg together and in order */
702 REQ_F_ASYNC_READ_BIT,
703 REQ_F_ASYNC_WRITE_BIT,
706 /* not a real bit, just to check we're not overflowing the space */
712 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
713 /* drain existing IO first */
714 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
716 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
717 /* doesn't sever on completion < 0 */
718 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
720 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
721 /* IOSQE_BUFFER_SELECT */
722 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
724 /* fail rest of links */
725 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
726 /* on inflight list, should be cancelled and waited on exit reliably */
727 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
728 /* read/write uses file position */
729 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
730 /* must not punt to workers */
731 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
732 /* has or had linked timeout */
733 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
735 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
736 /* already went through poll handler */
737 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
738 /* buffer already selected */
739 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
740 /* doesn't need file table for this request */
741 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
742 /* linked timeout is active, i.e. prepared by link's head */
743 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
744 /* completion is deferred through io_comp_state */
745 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
746 /* caller should reissue async */
747 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
748 /* supports async reads */
749 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
750 /* supports async writes */
751 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
753 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
757 struct io_poll_iocb poll;
758 struct io_poll_iocb *double_poll;
761 struct io_task_work {
762 struct io_wq_work_node node;
763 task_work_func_t func;
767 * NOTE! Each of the iocb union members has the file pointer
768 * as the first entry in their struct definition. So you can
769 * access the file pointer through any of the sub-structs,
770 * or directly as just 'ki_filp' in this struct.
776 struct io_poll_iocb poll;
777 struct io_poll_remove poll_remove;
778 struct io_accept accept;
780 struct io_cancel cancel;
781 struct io_timeout timeout;
782 struct io_timeout_rem timeout_rem;
783 struct io_connect connect;
784 struct io_sr_msg sr_msg;
786 struct io_close close;
787 struct io_rsrc_update rsrc_update;
788 struct io_fadvise fadvise;
789 struct io_madvise madvise;
790 struct io_epoll epoll;
791 struct io_splice splice;
792 struct io_provide_buf pbuf;
793 struct io_statx statx;
794 struct io_shutdown shutdown;
795 struct io_rename rename;
796 struct io_unlink unlink;
797 /* use only after cleaning per-op data, see io_clean_op() */
798 struct io_completion compl;
801 /* opcode allocated if it needs to store data for async defer */
804 /* polled IO has completed */
810 struct io_ring_ctx *ctx;
813 struct task_struct *task;
816 struct io_kiocb *link;
817 struct percpu_ref *fixed_rsrc_refs;
820 * 1. used with ctx->iopoll_list with reads/writes
821 * 2. to track reqs with ->files (see io_op_def::file_table)
823 struct list_head inflight_entry;
825 struct io_task_work io_task_work;
826 struct callback_head task_work;
828 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
829 struct hlist_node hash_node;
830 struct async_poll *apoll;
831 struct io_wq_work work;
834 struct io_tctx_node {
835 struct list_head ctx_node;
836 struct task_struct *task;
837 struct io_ring_ctx *ctx;
840 struct io_defer_entry {
841 struct list_head list;
842 struct io_kiocb *req;
847 /* needs req->file assigned */
848 unsigned needs_file : 1;
849 /* hash wq insertion if file is a regular file */
850 unsigned hash_reg_file : 1;
851 /* unbound wq insertion if file is a non-regular file */
852 unsigned unbound_nonreg_file : 1;
853 /* opcode is not supported by this kernel */
854 unsigned not_supported : 1;
855 /* set if opcode supports polled "wait" */
857 unsigned pollout : 1;
858 /* op supports buffer selection */
859 unsigned buffer_select : 1;
860 /* do prep async if is going to be punted */
861 unsigned needs_async_setup : 1;
862 /* should block plug */
864 /* size of async data needed, if any */
865 unsigned short async_size;
868 static const struct io_op_def io_op_defs[] = {
869 [IORING_OP_NOP] = {},
870 [IORING_OP_READV] = {
872 .unbound_nonreg_file = 1,
875 .needs_async_setup = 1,
877 .async_size = sizeof(struct io_async_rw),
879 [IORING_OP_WRITEV] = {
882 .unbound_nonreg_file = 1,
884 .needs_async_setup = 1,
886 .async_size = sizeof(struct io_async_rw),
888 [IORING_OP_FSYNC] = {
891 [IORING_OP_READ_FIXED] = {
893 .unbound_nonreg_file = 1,
896 .async_size = sizeof(struct io_async_rw),
898 [IORING_OP_WRITE_FIXED] = {
901 .unbound_nonreg_file = 1,
904 .async_size = sizeof(struct io_async_rw),
906 [IORING_OP_POLL_ADD] = {
908 .unbound_nonreg_file = 1,
910 [IORING_OP_POLL_REMOVE] = {},
911 [IORING_OP_SYNC_FILE_RANGE] = {
914 [IORING_OP_SENDMSG] = {
916 .unbound_nonreg_file = 1,
918 .needs_async_setup = 1,
919 .async_size = sizeof(struct io_async_msghdr),
921 [IORING_OP_RECVMSG] = {
923 .unbound_nonreg_file = 1,
926 .needs_async_setup = 1,
927 .async_size = sizeof(struct io_async_msghdr),
929 [IORING_OP_TIMEOUT] = {
930 .async_size = sizeof(struct io_timeout_data),
932 [IORING_OP_TIMEOUT_REMOVE] = {
933 /* used by timeout updates' prep() */
935 [IORING_OP_ACCEPT] = {
937 .unbound_nonreg_file = 1,
940 [IORING_OP_ASYNC_CANCEL] = {},
941 [IORING_OP_LINK_TIMEOUT] = {
942 .async_size = sizeof(struct io_timeout_data),
944 [IORING_OP_CONNECT] = {
946 .unbound_nonreg_file = 1,
948 .needs_async_setup = 1,
949 .async_size = sizeof(struct io_async_connect),
951 [IORING_OP_FALLOCATE] = {
954 [IORING_OP_OPENAT] = {},
955 [IORING_OP_CLOSE] = {},
956 [IORING_OP_FILES_UPDATE] = {},
957 [IORING_OP_STATX] = {},
960 .unbound_nonreg_file = 1,
964 .async_size = sizeof(struct io_async_rw),
966 [IORING_OP_WRITE] = {
968 .unbound_nonreg_file = 1,
971 .async_size = sizeof(struct io_async_rw),
973 [IORING_OP_FADVISE] = {
976 [IORING_OP_MADVISE] = {},
979 .unbound_nonreg_file = 1,
984 .unbound_nonreg_file = 1,
988 [IORING_OP_OPENAT2] = {
990 [IORING_OP_EPOLL_CTL] = {
991 .unbound_nonreg_file = 1,
993 [IORING_OP_SPLICE] = {
996 .unbound_nonreg_file = 1,
998 [IORING_OP_PROVIDE_BUFFERS] = {},
999 [IORING_OP_REMOVE_BUFFERS] = {},
1003 .unbound_nonreg_file = 1,
1005 [IORING_OP_SHUTDOWN] = {
1008 [IORING_OP_RENAMEAT] = {},
1009 [IORING_OP_UNLINKAT] = {},
1012 static bool io_disarm_next(struct io_kiocb *req);
1013 static void io_uring_del_task_file(unsigned long index);
1014 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1015 struct task_struct *task,
1016 struct files_struct *files);
1017 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1018 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1019 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1020 struct io_ring_ctx *ctx);
1021 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1023 static bool io_rw_reissue(struct io_kiocb *req);
1024 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1025 static void io_put_req(struct io_kiocb *req);
1026 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1027 static void io_dismantle_req(struct io_kiocb *req);
1028 static void io_put_task(struct task_struct *task, int nr);
1029 static void io_queue_next(struct io_kiocb *req);
1030 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1031 static void io_queue_linked_timeout(struct io_kiocb *req);
1032 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1033 struct io_uring_rsrc_update *ip,
1035 static void io_clean_op(struct io_kiocb *req);
1036 static struct file *io_file_get(struct io_submit_state *state,
1037 struct io_kiocb *req, int fd, bool fixed);
1038 static void __io_queue_sqe(struct io_kiocb *req);
1039 static void io_rsrc_put_work(struct work_struct *work);
1041 static void io_req_task_queue(struct io_kiocb *req);
1042 static void io_submit_flush_completions(struct io_comp_state *cs,
1043 struct io_ring_ctx *ctx);
1044 static int io_req_prep_async(struct io_kiocb *req);
1046 static struct kmem_cache *req_cachep;
1048 static const struct file_operations io_uring_fops;
1050 struct sock *io_uring_get_socket(struct file *file)
1052 #if defined(CONFIG_UNIX)
1053 if (file->f_op == &io_uring_fops) {
1054 struct io_ring_ctx *ctx = file->private_data;
1056 return ctx->ring_sock->sk;
1061 EXPORT_SYMBOL(io_uring_get_socket);
1063 #define io_for_each_link(pos, head) \
1064 for (pos = (head); pos; pos = pos->link)
1066 static inline void io_set_resource_node(struct io_kiocb *req)
1068 struct io_ring_ctx *ctx = req->ctx;
1070 if (!req->fixed_rsrc_refs) {
1071 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1072 percpu_ref_get(req->fixed_rsrc_refs);
1076 static bool io_match_task(struct io_kiocb *head,
1077 struct task_struct *task,
1078 struct files_struct *files)
1080 struct io_kiocb *req;
1082 if (task && head->task != task) {
1083 /* in terms of cancelation, always match if req task is dead */
1084 if (head->task->flags & PF_EXITING)
1091 io_for_each_link(req, head) {
1092 if (req->flags & REQ_F_INFLIGHT)
1098 static inline void req_set_fail_links(struct io_kiocb *req)
1100 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1101 req->flags |= REQ_F_FAIL_LINK;
1104 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1106 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1108 complete(&ctx->ref_comp);
1111 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1113 return !req->timeout.off;
1116 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1118 struct io_ring_ctx *ctx;
1121 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1126 * Use 5 bits less than the max cq entries, that should give us around
1127 * 32 entries per hash list if totally full and uniformly spread.
1129 hash_bits = ilog2(p->cq_entries);
1133 ctx->cancel_hash_bits = hash_bits;
1134 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1136 if (!ctx->cancel_hash)
1138 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1140 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1141 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1144 ctx->flags = p->flags;
1145 init_waitqueue_head(&ctx->sqo_sq_wait);
1146 INIT_LIST_HEAD(&ctx->sqd_list);
1147 init_waitqueue_head(&ctx->cq_wait);
1148 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1149 init_completion(&ctx->ref_comp);
1150 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1151 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1152 mutex_init(&ctx->uring_lock);
1153 init_waitqueue_head(&ctx->wait);
1154 spin_lock_init(&ctx->completion_lock);
1155 INIT_LIST_HEAD(&ctx->iopoll_list);
1156 INIT_LIST_HEAD(&ctx->defer_list);
1157 INIT_LIST_HEAD(&ctx->timeout_list);
1158 spin_lock_init(&ctx->inflight_lock);
1159 INIT_LIST_HEAD(&ctx->inflight_list);
1160 spin_lock_init(&ctx->rsrc_ref_lock);
1161 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1162 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1163 init_llist_head(&ctx->rsrc_put_llist);
1164 INIT_LIST_HEAD(&ctx->tctx_list);
1165 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1166 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1169 kfree(ctx->cancel_hash);
1174 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1176 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1177 struct io_ring_ctx *ctx = req->ctx;
1179 return seq != ctx->cached_cq_tail
1180 + READ_ONCE(ctx->cached_cq_overflow);
1186 static void io_req_track_inflight(struct io_kiocb *req)
1188 struct io_ring_ctx *ctx = req->ctx;
1190 if (!(req->flags & REQ_F_INFLIGHT)) {
1191 req->flags |= REQ_F_INFLIGHT;
1193 spin_lock_irq(&ctx->inflight_lock);
1194 list_add(&req->inflight_entry, &ctx->inflight_list);
1195 spin_unlock_irq(&ctx->inflight_lock);
1199 static void io_prep_async_work(struct io_kiocb *req)
1201 const struct io_op_def *def = &io_op_defs[req->opcode];
1202 struct io_ring_ctx *ctx = req->ctx;
1204 if (!req->work.creds)
1205 req->work.creds = get_current_cred();
1207 if (req->flags & REQ_F_FORCE_ASYNC)
1208 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1210 if (req->flags & REQ_F_ISREG) {
1211 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1212 io_wq_hash_work(&req->work, file_inode(req->file));
1213 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1214 if (def->unbound_nonreg_file)
1215 req->work.flags |= IO_WQ_WORK_UNBOUND;
1219 static void io_prep_async_link(struct io_kiocb *req)
1221 struct io_kiocb *cur;
1223 io_for_each_link(cur, req)
1224 io_prep_async_work(cur);
1227 static void io_queue_async_work(struct io_kiocb *req)
1229 struct io_ring_ctx *ctx = req->ctx;
1230 struct io_kiocb *link = io_prep_linked_timeout(req);
1231 struct io_uring_task *tctx = req->task->io_uring;
1234 BUG_ON(!tctx->io_wq);
1236 /* init ->work of the whole link before punting */
1237 io_prep_async_link(req);
1238 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1239 &req->work, req->flags);
1240 io_wq_enqueue(tctx->io_wq, &req->work);
1242 io_queue_linked_timeout(link);
1245 static void io_kill_timeout(struct io_kiocb *req, int status)
1247 struct io_timeout_data *io = req->async_data;
1250 ret = hrtimer_try_to_cancel(&io->timer);
1252 atomic_set(&req->ctx->cq_timeouts,
1253 atomic_read(&req->ctx->cq_timeouts) + 1);
1254 list_del_init(&req->timeout.list);
1255 io_cqring_fill_event(req, status);
1256 io_put_req_deferred(req, 1);
1260 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1263 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1264 struct io_defer_entry, list);
1266 if (req_need_defer(de->req, de->seq))
1268 list_del_init(&de->list);
1269 io_req_task_queue(de->req);
1271 } while (!list_empty(&ctx->defer_list));
1274 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1278 if (list_empty(&ctx->timeout_list))
1281 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1284 u32 events_needed, events_got;
1285 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1286 struct io_kiocb, timeout.list);
1288 if (io_is_timeout_noseq(req))
1292 * Since seq can easily wrap around over time, subtract
1293 * the last seq at which timeouts were flushed before comparing.
1294 * Assuming not more than 2^31-1 events have happened since,
1295 * these subtractions won't have wrapped, so we can check if
1296 * target is in [last_seq, current_seq] by comparing the two.
1298 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1299 events_got = seq - ctx->cq_last_tm_flush;
1300 if (events_got < events_needed)
1303 list_del_init(&req->timeout.list);
1304 io_kill_timeout(req, 0);
1305 } while (!list_empty(&ctx->timeout_list));
1307 ctx->cq_last_tm_flush = seq;
1310 static void io_commit_cqring(struct io_ring_ctx *ctx)
1312 io_flush_timeouts(ctx);
1314 /* order cqe stores with ring update */
1315 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1317 if (unlikely(!list_empty(&ctx->defer_list)))
1318 __io_queue_deferred(ctx);
1321 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1323 struct io_rings *r = ctx->rings;
1325 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1328 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1330 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1333 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1335 struct io_rings *rings = ctx->rings;
1339 * writes to the cq entry need to come after reading head; the
1340 * control dependency is enough as we're using WRITE_ONCE to
1343 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1346 tail = ctx->cached_cq_tail++;
1347 return &rings->cqes[tail & ctx->cq_mask];
1350 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1354 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1356 if (!ctx->eventfd_async)
1358 return io_wq_current_is_worker();
1361 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1363 /* see waitqueue_active() comment */
1366 if (waitqueue_active(&ctx->wait))
1367 wake_up(&ctx->wait);
1368 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1369 wake_up(&ctx->sq_data->wait);
1370 if (io_should_trigger_evfd(ctx))
1371 eventfd_signal(ctx->cq_ev_fd, 1);
1372 if (waitqueue_active(&ctx->cq_wait)) {
1373 wake_up_interruptible(&ctx->cq_wait);
1374 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1378 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1380 /* see waitqueue_active() comment */
1383 if (ctx->flags & IORING_SETUP_SQPOLL) {
1384 if (waitqueue_active(&ctx->wait))
1385 wake_up(&ctx->wait);
1387 if (io_should_trigger_evfd(ctx))
1388 eventfd_signal(ctx->cq_ev_fd, 1);
1389 if (waitqueue_active(&ctx->cq_wait)) {
1390 wake_up_interruptible(&ctx->cq_wait);
1391 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1395 /* Returns true if there are no backlogged entries after the flush */
1396 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1397 struct task_struct *tsk,
1398 struct files_struct *files)
1400 struct io_rings *rings = ctx->rings;
1401 struct io_kiocb *req, *tmp;
1402 struct io_uring_cqe *cqe;
1403 unsigned long flags;
1404 bool all_flushed, posted;
1407 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1411 spin_lock_irqsave(&ctx->completion_lock, flags);
1412 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1413 if (!io_match_task(req, tsk, files))
1416 cqe = io_get_cqring(ctx);
1420 list_move(&req->compl.list, &list);
1422 WRITE_ONCE(cqe->user_data, req->user_data);
1423 WRITE_ONCE(cqe->res, req->result);
1424 WRITE_ONCE(cqe->flags, req->compl.cflags);
1426 ctx->cached_cq_overflow++;
1427 WRITE_ONCE(ctx->rings->cq_overflow,
1428 ctx->cached_cq_overflow);
1433 all_flushed = list_empty(&ctx->cq_overflow_list);
1435 clear_bit(0, &ctx->sq_check_overflow);
1436 clear_bit(0, &ctx->cq_check_overflow);
1437 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1441 io_commit_cqring(ctx);
1442 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1444 io_cqring_ev_posted(ctx);
1446 while (!list_empty(&list)) {
1447 req = list_first_entry(&list, struct io_kiocb, compl.list);
1448 list_del(&req->compl.list);
1455 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1456 struct task_struct *tsk,
1457 struct files_struct *files)
1461 if (test_bit(0, &ctx->cq_check_overflow)) {
1462 /* iopoll syncs against uring_lock, not completion_lock */
1463 if (ctx->flags & IORING_SETUP_IOPOLL)
1464 mutex_lock(&ctx->uring_lock);
1465 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1466 if (ctx->flags & IORING_SETUP_IOPOLL)
1467 mutex_unlock(&ctx->uring_lock);
1474 * Shamelessly stolen from the mm implementation of page reference checking,
1475 * see commit f958d7b528b1 for details.
1477 #define req_ref_zero_or_close_to_overflow(req) \
1478 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1480 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1482 return atomic_inc_not_zero(&req->refs);
1485 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1487 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1488 return atomic_sub_and_test(refs, &req->refs);
1491 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1493 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1494 return atomic_dec_and_test(&req->refs);
1497 static inline void req_ref_put(struct io_kiocb *req)
1499 WARN_ON_ONCE(req_ref_put_and_test(req));
1502 static inline void req_ref_get(struct io_kiocb *req)
1504 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1505 atomic_inc(&req->refs);
1508 static void __io_cqring_fill_event(struct io_kiocb *req, long res,
1509 unsigned int cflags)
1511 struct io_ring_ctx *ctx = req->ctx;
1512 struct io_uring_cqe *cqe;
1514 trace_io_uring_complete(ctx, req->user_data, res);
1517 * If we can't get a cq entry, userspace overflowed the
1518 * submission (by quite a lot). Increment the overflow count in
1521 cqe = io_get_cqring(ctx);
1523 WRITE_ONCE(cqe->user_data, req->user_data);
1524 WRITE_ONCE(cqe->res, res);
1525 WRITE_ONCE(cqe->flags, cflags);
1526 } else if (ctx->cq_overflow_flushed ||
1527 atomic_read(&req->task->io_uring->in_idle)) {
1529 * If we're in ring overflow flush mode, or in task cancel mode,
1530 * then we cannot store the request for later flushing, we need
1531 * to drop it on the floor.
1533 ctx->cached_cq_overflow++;
1534 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1536 if (list_empty(&ctx->cq_overflow_list)) {
1537 set_bit(0, &ctx->sq_check_overflow);
1538 set_bit(0, &ctx->cq_check_overflow);
1539 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1541 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1545 req->compl.cflags = cflags;
1547 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1551 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1553 __io_cqring_fill_event(req, res, 0);
1556 static void io_req_complete_post(struct io_kiocb *req, long res,
1557 unsigned int cflags)
1559 struct io_ring_ctx *ctx = req->ctx;
1560 unsigned long flags;
1562 spin_lock_irqsave(&ctx->completion_lock, flags);
1563 __io_cqring_fill_event(req, res, cflags);
1565 * If we're the last reference to this request, add to our locked
1568 if (req_ref_put_and_test(req)) {
1569 struct io_comp_state *cs = &ctx->submit_state.comp;
1571 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1572 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1573 io_disarm_next(req);
1575 io_req_task_queue(req->link);
1579 io_dismantle_req(req);
1580 io_put_task(req->task, 1);
1581 list_add(&req->compl.list, &cs->locked_free_list);
1582 cs->locked_free_nr++;
1584 if (!percpu_ref_tryget(&ctx->refs))
1587 io_commit_cqring(ctx);
1588 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1591 io_cqring_ev_posted(ctx);
1592 percpu_ref_put(&ctx->refs);
1596 static void io_req_complete_state(struct io_kiocb *req, long res,
1597 unsigned int cflags)
1599 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1602 req->compl.cflags = cflags;
1603 req->flags |= REQ_F_COMPLETE_INLINE;
1606 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1607 long res, unsigned cflags)
1609 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1610 io_req_complete_state(req, res, cflags);
1612 io_req_complete_post(req, res, cflags);
1615 static inline void io_req_complete(struct io_kiocb *req, long res)
1617 __io_req_complete(req, 0, res, 0);
1620 static void io_req_complete_failed(struct io_kiocb *req, long res)
1622 req_set_fail_links(req);
1624 io_req_complete_post(req, res, 0);
1627 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1628 struct io_comp_state *cs)
1630 spin_lock_irq(&ctx->completion_lock);
1631 list_splice_init(&cs->locked_free_list, &cs->free_list);
1632 cs->locked_free_nr = 0;
1633 spin_unlock_irq(&ctx->completion_lock);
1636 /* Returns true IFF there are requests in the cache */
1637 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1639 struct io_submit_state *state = &ctx->submit_state;
1640 struct io_comp_state *cs = &state->comp;
1644 * If we have more than a batch's worth of requests in our IRQ side
1645 * locked cache, grab the lock and move them over to our submission
1648 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1649 io_flush_cached_locked_reqs(ctx, cs);
1651 nr = state->free_reqs;
1652 while (!list_empty(&cs->free_list)) {
1653 struct io_kiocb *req = list_first_entry(&cs->free_list,
1654 struct io_kiocb, compl.list);
1656 list_del(&req->compl.list);
1657 state->reqs[nr++] = req;
1658 if (nr == ARRAY_SIZE(state->reqs))
1662 state->free_reqs = nr;
1666 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1668 struct io_submit_state *state = &ctx->submit_state;
1670 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1672 if (!state->free_reqs) {
1673 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1676 if (io_flush_cached_reqs(ctx))
1679 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1683 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1684 * retry single alloc to be on the safe side.
1686 if (unlikely(ret <= 0)) {
1687 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1688 if (!state->reqs[0])
1692 state->free_reqs = ret;
1696 return state->reqs[state->free_reqs];
1699 static inline void io_put_file(struct file *file)
1705 static void io_dismantle_req(struct io_kiocb *req)
1707 unsigned int flags = req->flags;
1709 if (!(flags & REQ_F_FIXED_FILE))
1710 io_put_file(req->file);
1711 if (flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1715 if (req->flags & REQ_F_INFLIGHT) {
1716 struct io_ring_ctx *ctx = req->ctx;
1717 unsigned long flags;
1719 spin_lock_irqsave(&ctx->inflight_lock, flags);
1720 list_del(&req->inflight_entry);
1721 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1722 req->flags &= ~REQ_F_INFLIGHT;
1725 if (req->fixed_rsrc_refs)
1726 percpu_ref_put(req->fixed_rsrc_refs);
1727 if (req->async_data)
1728 kfree(req->async_data);
1729 if (req->work.creds) {
1730 put_cred(req->work.creds);
1731 req->work.creds = NULL;
1735 /* must to be called somewhat shortly after putting a request */
1736 static inline void io_put_task(struct task_struct *task, int nr)
1738 struct io_uring_task *tctx = task->io_uring;
1740 percpu_counter_sub(&tctx->inflight, nr);
1741 if (unlikely(atomic_read(&tctx->in_idle)))
1742 wake_up(&tctx->wait);
1743 put_task_struct_many(task, nr);
1746 static void __io_free_req(struct io_kiocb *req)
1748 struct io_ring_ctx *ctx = req->ctx;
1750 io_dismantle_req(req);
1751 io_put_task(req->task, 1);
1753 kmem_cache_free(req_cachep, req);
1754 percpu_ref_put(&ctx->refs);
1757 static inline void io_remove_next_linked(struct io_kiocb *req)
1759 struct io_kiocb *nxt = req->link;
1761 req->link = nxt->link;
1765 static bool io_kill_linked_timeout(struct io_kiocb *req)
1766 __must_hold(&req->ctx->completion_lock)
1768 struct io_kiocb *link = req->link;
1771 * Can happen if a linked timeout fired and link had been like
1772 * req -> link t-out -> link t-out [-> ...]
1774 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1775 struct io_timeout_data *io = link->async_data;
1778 io_remove_next_linked(req);
1779 link->timeout.head = NULL;
1780 ret = hrtimer_try_to_cancel(&io->timer);
1782 io_cqring_fill_event(link, -ECANCELED);
1783 io_put_req_deferred(link, 1);
1790 static void io_fail_links(struct io_kiocb *req)
1791 __must_hold(&req->ctx->completion_lock)
1793 struct io_kiocb *nxt, *link = req->link;
1800 trace_io_uring_fail_link(req, link);
1801 io_cqring_fill_event(link, -ECANCELED);
1802 io_put_req_deferred(link, 2);
1807 static bool io_disarm_next(struct io_kiocb *req)
1808 __must_hold(&req->ctx->completion_lock)
1810 bool posted = false;
1812 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1813 posted = io_kill_linked_timeout(req);
1814 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1815 posted |= (req->link != NULL);
1821 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1823 struct io_kiocb *nxt;
1826 * If LINK is set, we have dependent requests in this chain. If we
1827 * didn't fail this request, queue the first one up, moving any other
1828 * dependencies to the next request. In case of failure, fail the rest
1831 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1832 struct io_ring_ctx *ctx = req->ctx;
1833 unsigned long flags;
1836 spin_lock_irqsave(&ctx->completion_lock, flags);
1837 posted = io_disarm_next(req);
1839 io_commit_cqring(req->ctx);
1840 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1842 io_cqring_ev_posted(ctx);
1849 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1851 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1853 return __io_req_find_next(req);
1856 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1860 if (ctx->submit_state.comp.nr) {
1861 mutex_lock(&ctx->uring_lock);
1862 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1863 mutex_unlock(&ctx->uring_lock);
1865 percpu_ref_put(&ctx->refs);
1868 static bool __tctx_task_work(struct io_uring_task *tctx)
1870 struct io_ring_ctx *ctx = NULL;
1871 struct io_wq_work_list list;
1872 struct io_wq_work_node *node;
1874 if (wq_list_empty(&tctx->task_list))
1877 spin_lock_irq(&tctx->task_lock);
1878 list = tctx->task_list;
1879 INIT_WQ_LIST(&tctx->task_list);
1880 spin_unlock_irq(&tctx->task_lock);
1884 struct io_wq_work_node *next = node->next;
1885 struct io_kiocb *req;
1887 req = container_of(node, struct io_kiocb, io_task_work.node);
1888 if (req->ctx != ctx) {
1889 ctx_flush_and_put(ctx);
1891 percpu_ref_get(&ctx->refs);
1894 req->task_work.func(&req->task_work);
1898 ctx_flush_and_put(ctx);
1899 return list.first != NULL;
1902 static void tctx_task_work(struct callback_head *cb)
1904 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1906 clear_bit(0, &tctx->task_state);
1908 while (__tctx_task_work(tctx))
1912 static int io_req_task_work_add(struct io_kiocb *req)
1914 struct task_struct *tsk = req->task;
1915 struct io_uring_task *tctx = tsk->io_uring;
1916 enum task_work_notify_mode notify;
1917 struct io_wq_work_node *node, *prev;
1918 unsigned long flags;
1921 if (unlikely(tsk->flags & PF_EXITING))
1924 WARN_ON_ONCE(!tctx);
1926 spin_lock_irqsave(&tctx->task_lock, flags);
1927 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1928 spin_unlock_irqrestore(&tctx->task_lock, flags);
1930 /* task_work already pending, we're done */
1931 if (test_bit(0, &tctx->task_state) ||
1932 test_and_set_bit(0, &tctx->task_state))
1936 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1937 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1938 * processing task_work. There's no reliable way to tell if TWA_RESUME
1941 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1943 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1944 wake_up_process(tsk);
1949 * Slow path - we failed, find and delete work. if the work is not
1950 * in the list, it got run and we're fine.
1952 spin_lock_irqsave(&tctx->task_lock, flags);
1953 wq_list_for_each(node, prev, &tctx->task_list) {
1954 if (&req->io_task_work.node == node) {
1955 wq_list_del(&tctx->task_list, node, prev);
1960 spin_unlock_irqrestore(&tctx->task_lock, flags);
1961 clear_bit(0, &tctx->task_state);
1965 static bool io_run_task_work_head(struct callback_head **work_head)
1967 struct callback_head *work, *next;
1968 bool executed = false;
1971 work = xchg(work_head, NULL);
1987 static void io_task_work_add_head(struct callback_head **work_head,
1988 struct callback_head *task_work)
1990 struct callback_head *head;
1993 head = READ_ONCE(*work_head);
1994 task_work->next = head;
1995 } while (cmpxchg(work_head, head, task_work) != head);
1998 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1999 task_work_func_t cb)
2001 init_task_work(&req->task_work, cb);
2002 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2005 static void io_req_task_cancel(struct callback_head *cb)
2007 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2008 struct io_ring_ctx *ctx = req->ctx;
2010 /* ctx is guaranteed to stay alive while we hold uring_lock */
2011 mutex_lock(&ctx->uring_lock);
2012 io_req_complete_failed(req, req->result);
2013 mutex_unlock(&ctx->uring_lock);
2016 static void __io_req_task_submit(struct io_kiocb *req)
2018 struct io_ring_ctx *ctx = req->ctx;
2020 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2021 mutex_lock(&ctx->uring_lock);
2022 if (!(current->flags & PF_EXITING) && !current->in_execve)
2023 __io_queue_sqe(req);
2025 io_req_complete_failed(req, -EFAULT);
2026 mutex_unlock(&ctx->uring_lock);
2029 static void io_req_task_submit(struct callback_head *cb)
2031 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2033 __io_req_task_submit(req);
2036 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2039 req->task_work.func = io_req_task_cancel;
2041 if (unlikely(io_req_task_work_add(req)))
2042 io_req_task_work_add_fallback(req, io_req_task_cancel);
2045 static void io_req_task_queue(struct io_kiocb *req)
2047 req->task_work.func = io_req_task_submit;
2049 if (unlikely(io_req_task_work_add(req)))
2050 io_req_task_queue_fail(req, -ECANCELED);
2053 static inline void io_queue_next(struct io_kiocb *req)
2055 struct io_kiocb *nxt = io_req_find_next(req);
2058 io_req_task_queue(nxt);
2061 static void io_free_req(struct io_kiocb *req)
2068 struct task_struct *task;
2073 static inline void io_init_req_batch(struct req_batch *rb)
2080 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2081 struct req_batch *rb)
2084 io_put_task(rb->task, rb->task_refs);
2086 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2089 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2090 struct io_submit_state *state)
2093 io_dismantle_req(req);
2095 if (req->task != rb->task) {
2097 io_put_task(rb->task, rb->task_refs);
2098 rb->task = req->task;
2104 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2105 state->reqs[state->free_reqs++] = req;
2107 list_add(&req->compl.list, &state->comp.free_list);
2110 static void io_submit_flush_completions(struct io_comp_state *cs,
2111 struct io_ring_ctx *ctx)
2114 struct io_kiocb *req;
2115 struct req_batch rb;
2117 io_init_req_batch(&rb);
2118 spin_lock_irq(&ctx->completion_lock);
2119 for (i = 0; i < nr; i++) {
2121 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2123 io_commit_cqring(ctx);
2124 spin_unlock_irq(&ctx->completion_lock);
2126 io_cqring_ev_posted(ctx);
2127 for (i = 0; i < nr; i++) {
2130 /* submission and completion refs */
2131 if (req_ref_sub_and_test(req, 2))
2132 io_req_free_batch(&rb, req, &ctx->submit_state);
2135 io_req_free_batch_finish(ctx, &rb);
2140 * Drop reference to request, return next in chain (if there is one) if this
2141 * was the last reference to this request.
2143 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2145 struct io_kiocb *nxt = NULL;
2147 if (req_ref_put_and_test(req)) {
2148 nxt = io_req_find_next(req);
2154 static inline void io_put_req(struct io_kiocb *req)
2156 if (req_ref_put_and_test(req))
2160 static void io_put_req_deferred_cb(struct callback_head *cb)
2162 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2167 static void io_free_req_deferred(struct io_kiocb *req)
2169 req->task_work.func = io_put_req_deferred_cb;
2170 if (unlikely(io_req_task_work_add(req)))
2171 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2174 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2176 if (req_ref_sub_and_test(req, refs))
2177 io_free_req_deferred(req);
2180 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2182 /* See comment at the top of this file */
2184 return __io_cqring_events(ctx);
2187 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2189 struct io_rings *rings = ctx->rings;
2191 /* make sure SQ entry isn't read before tail */
2192 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2195 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2197 unsigned int cflags;
2199 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2200 cflags |= IORING_CQE_F_BUFFER;
2201 req->flags &= ~REQ_F_BUFFER_SELECTED;
2206 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2208 struct io_buffer *kbuf;
2210 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2211 return io_put_kbuf(req, kbuf);
2214 static inline bool io_run_task_work(void)
2217 * Not safe to run on exiting task, and the task_work handling will
2218 * not add work to such a task.
2220 if (unlikely(current->flags & PF_EXITING))
2222 if (current->task_works) {
2223 __set_current_state(TASK_RUNNING);
2232 * Find and free completed poll iocbs
2234 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2235 struct list_head *done)
2237 struct req_batch rb;
2238 struct io_kiocb *req;
2240 /* order with ->result store in io_complete_rw_iopoll() */
2243 io_init_req_batch(&rb);
2244 while (!list_empty(done)) {
2247 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2248 list_del(&req->inflight_entry);
2250 if (READ_ONCE(req->result) == -EAGAIN) {
2251 req->iopoll_completed = 0;
2252 if (io_rw_reissue(req))
2256 if (req->flags & REQ_F_BUFFER_SELECTED)
2257 cflags = io_put_rw_kbuf(req);
2259 __io_cqring_fill_event(req, req->result, cflags);
2262 if (req_ref_put_and_test(req))
2263 io_req_free_batch(&rb, req, &ctx->submit_state);
2266 io_commit_cqring(ctx);
2267 io_cqring_ev_posted_iopoll(ctx);
2268 io_req_free_batch_finish(ctx, &rb);
2271 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2274 struct io_kiocb *req, *tmp;
2280 * Only spin for completions if we don't have multiple devices hanging
2281 * off our complete list, and we're under the requested amount.
2283 spin = !ctx->poll_multi_file && *nr_events < min;
2286 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2287 struct kiocb *kiocb = &req->rw.kiocb;
2290 * Move completed and retryable entries to our local lists.
2291 * If we find a request that requires polling, break out
2292 * and complete those lists first, if we have entries there.
2294 if (READ_ONCE(req->iopoll_completed)) {
2295 list_move_tail(&req->inflight_entry, &done);
2298 if (!list_empty(&done))
2301 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2305 /* iopoll may have completed current req */
2306 if (READ_ONCE(req->iopoll_completed))
2307 list_move_tail(&req->inflight_entry, &done);
2314 if (!list_empty(&done))
2315 io_iopoll_complete(ctx, nr_events, &done);
2321 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2322 * non-spinning poll check - we'll still enter the driver poll loop, but only
2323 * as a non-spinning completion check.
2325 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2328 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2331 ret = io_do_iopoll(ctx, nr_events, min);
2334 if (*nr_events >= min)
2342 * We can't just wait for polled events to come to us, we have to actively
2343 * find and complete them.
2345 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2347 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2350 mutex_lock(&ctx->uring_lock);
2351 while (!list_empty(&ctx->iopoll_list)) {
2352 unsigned int nr_events = 0;
2354 io_do_iopoll(ctx, &nr_events, 0);
2356 /* let it sleep and repeat later if can't complete a request */
2360 * Ensure we allow local-to-the-cpu processing to take place,
2361 * in this case we need to ensure that we reap all events.
2362 * Also let task_work, etc. to progress by releasing the mutex
2364 if (need_resched()) {
2365 mutex_unlock(&ctx->uring_lock);
2367 mutex_lock(&ctx->uring_lock);
2370 mutex_unlock(&ctx->uring_lock);
2373 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2375 unsigned int nr_events = 0;
2376 int iters = 0, ret = 0;
2379 * We disallow the app entering submit/complete with polling, but we
2380 * still need to lock the ring to prevent racing with polled issue
2381 * that got punted to a workqueue.
2383 mutex_lock(&ctx->uring_lock);
2386 * Don't enter poll loop if we already have events pending.
2387 * If we do, we can potentially be spinning for commands that
2388 * already triggered a CQE (eg in error).
2390 if (test_bit(0, &ctx->cq_check_overflow))
2391 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2392 if (io_cqring_events(ctx))
2396 * If a submit got punted to a workqueue, we can have the
2397 * application entering polling for a command before it gets
2398 * issued. That app will hold the uring_lock for the duration
2399 * of the poll right here, so we need to take a breather every
2400 * now and then to ensure that the issue has a chance to add
2401 * the poll to the issued list. Otherwise we can spin here
2402 * forever, while the workqueue is stuck trying to acquire the
2405 if (!(++iters & 7)) {
2406 mutex_unlock(&ctx->uring_lock);
2408 mutex_lock(&ctx->uring_lock);
2411 ret = io_iopoll_getevents(ctx, &nr_events, min);
2415 } while (min && !nr_events && !need_resched());
2417 mutex_unlock(&ctx->uring_lock);
2421 static void kiocb_end_write(struct io_kiocb *req)
2424 * Tell lockdep we inherited freeze protection from submission
2427 if (req->flags & REQ_F_ISREG) {
2428 struct inode *inode = file_inode(req->file);
2430 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2432 file_end_write(req->file);
2436 static bool io_resubmit_prep(struct io_kiocb *req)
2438 /* either already prepared or successfully done */
2439 return req->async_data || !io_req_prep_async(req);
2442 static bool io_rw_should_reissue(struct io_kiocb *req)
2444 umode_t mode = file_inode(req->file)->i_mode;
2445 struct io_ring_ctx *ctx = req->ctx;
2447 if (!S_ISBLK(mode) && !S_ISREG(mode))
2449 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2450 !(ctx->flags & IORING_SETUP_IOPOLL)))
2453 * If ref is dying, we might be running poll reap from the exit work.
2454 * Don't attempt to reissue from that path, just let it fail with
2457 if (percpu_ref_is_dying(&ctx->refs))
2462 static bool io_rw_should_reissue(struct io_kiocb *req)
2468 static bool io_rw_reissue(struct io_kiocb *req)
2471 if (!io_rw_should_reissue(req))
2474 lockdep_assert_held(&req->ctx->uring_lock);
2476 if (io_resubmit_prep(req)) {
2478 io_queue_async_work(req);
2481 req_set_fail_links(req);
2486 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2487 unsigned int issue_flags)
2491 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2492 kiocb_end_write(req);
2493 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_should_reissue(req)) {
2494 req->flags |= REQ_F_REISSUE;
2497 if (res != req->result)
2498 req_set_fail_links(req);
2499 if (req->flags & REQ_F_BUFFER_SELECTED)
2500 cflags = io_put_rw_kbuf(req);
2501 __io_req_complete(req, issue_flags, res, cflags);
2504 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2506 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2508 __io_complete_rw(req, res, res2, 0);
2511 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2513 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2516 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2517 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2518 struct io_async_rw *rw = req->async_data;
2521 iov_iter_revert(&rw->iter,
2522 req->result - iov_iter_count(&rw->iter));
2523 else if (!io_resubmit_prep(req))
2528 if (kiocb->ki_flags & IOCB_WRITE)
2529 kiocb_end_write(req);
2531 if (res != -EAGAIN && res != req->result)
2532 req_set_fail_links(req);
2534 WRITE_ONCE(req->result, res);
2535 /* order with io_iopoll_complete() checking ->result */
2537 WRITE_ONCE(req->iopoll_completed, 1);
2541 * After the iocb has been issued, it's safe to be found on the poll list.
2542 * Adding the kiocb to the list AFTER submission ensures that we don't
2543 * find it from a io_iopoll_getevents() thread before the issuer is done
2544 * accessing the kiocb cookie.
2546 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2548 struct io_ring_ctx *ctx = req->ctx;
2551 * Track whether we have multiple files in our lists. This will impact
2552 * how we do polling eventually, not spinning if we're on potentially
2553 * different devices.
2555 if (list_empty(&ctx->iopoll_list)) {
2556 ctx->poll_multi_file = false;
2557 } else if (!ctx->poll_multi_file) {
2558 struct io_kiocb *list_req;
2560 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2562 if (list_req->file != req->file)
2563 ctx->poll_multi_file = true;
2567 * For fast devices, IO may have already completed. If it has, add
2568 * it to the front so we find it first.
2570 if (READ_ONCE(req->iopoll_completed))
2571 list_add(&req->inflight_entry, &ctx->iopoll_list);
2573 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2576 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2577 * task context or in io worker task context. If current task context is
2578 * sq thread, we don't need to check whether should wake up sq thread.
2580 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2581 wq_has_sleeper(&ctx->sq_data->wait))
2582 wake_up(&ctx->sq_data->wait);
2585 static inline void io_state_file_put(struct io_submit_state *state)
2587 if (state->file_refs) {
2588 fput_many(state->file, state->file_refs);
2589 state->file_refs = 0;
2594 * Get as many references to a file as we have IOs left in this submission,
2595 * assuming most submissions are for one file, or at least that each file
2596 * has more than one submission.
2598 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2603 if (state->file_refs) {
2604 if (state->fd == fd) {
2608 io_state_file_put(state);
2610 state->file = fget_many(fd, state->ios_left);
2611 if (unlikely(!state->file))
2615 state->file_refs = state->ios_left - 1;
2619 static bool io_bdev_nowait(struct block_device *bdev)
2621 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2625 * If we tracked the file through the SCM inflight mechanism, we could support
2626 * any file. For now, just ensure that anything potentially problematic is done
2629 static bool __io_file_supports_async(struct file *file, int rw)
2631 umode_t mode = file_inode(file)->i_mode;
2633 if (S_ISBLK(mode)) {
2634 if (IS_ENABLED(CONFIG_BLOCK) &&
2635 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2639 if (S_ISCHR(mode) || S_ISSOCK(mode))
2641 if (S_ISREG(mode)) {
2642 if (IS_ENABLED(CONFIG_BLOCK) &&
2643 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2644 file->f_op != &io_uring_fops)
2649 /* any ->read/write should understand O_NONBLOCK */
2650 if (file->f_flags & O_NONBLOCK)
2653 if (!(file->f_mode & FMODE_NOWAIT))
2657 return file->f_op->read_iter != NULL;
2659 return file->f_op->write_iter != NULL;
2662 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2664 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2666 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2669 return __io_file_supports_async(req->file, rw);
2672 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2674 struct io_ring_ctx *ctx = req->ctx;
2675 struct kiocb *kiocb = &req->rw.kiocb;
2676 struct file *file = req->file;
2680 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2681 req->flags |= REQ_F_ISREG;
2683 kiocb->ki_pos = READ_ONCE(sqe->off);
2684 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2685 req->flags |= REQ_F_CUR_POS;
2686 kiocb->ki_pos = file->f_pos;
2688 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2689 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2690 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2694 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2695 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2696 req->flags |= REQ_F_NOWAIT;
2698 ioprio = READ_ONCE(sqe->ioprio);
2700 ret = ioprio_check_cap(ioprio);
2704 kiocb->ki_ioprio = ioprio;
2706 kiocb->ki_ioprio = get_current_ioprio();
2708 if (ctx->flags & IORING_SETUP_IOPOLL) {
2709 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2710 !kiocb->ki_filp->f_op->iopoll)
2713 kiocb->ki_flags |= IOCB_HIPRI;
2714 kiocb->ki_complete = io_complete_rw_iopoll;
2715 req->iopoll_completed = 0;
2717 if (kiocb->ki_flags & IOCB_HIPRI)
2719 kiocb->ki_complete = io_complete_rw;
2722 req->rw.addr = READ_ONCE(sqe->addr);
2723 req->rw.len = READ_ONCE(sqe->len);
2724 req->buf_index = READ_ONCE(sqe->buf_index);
2728 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2734 case -ERESTARTNOINTR:
2735 case -ERESTARTNOHAND:
2736 case -ERESTART_RESTARTBLOCK:
2738 * We can't just restart the syscall, since previously
2739 * submitted sqes may already be in progress. Just fail this
2745 kiocb->ki_complete(kiocb, ret, 0);
2749 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2750 unsigned int issue_flags)
2752 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2753 struct io_async_rw *io = req->async_data;
2754 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2756 /* add previously done IO, if any */
2757 if (io && io->bytes_done > 0) {
2759 ret = io->bytes_done;
2761 ret += io->bytes_done;
2764 if (req->flags & REQ_F_CUR_POS)
2765 req->file->f_pos = kiocb->ki_pos;
2766 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2767 __io_complete_rw(req, ret, 0, issue_flags);
2769 io_rw_done(kiocb, ret);
2771 if (check_reissue && req->flags & REQ_F_REISSUE) {
2772 req->flags &= ~REQ_F_REISSUE;
2773 if (!io_rw_reissue(req)) {
2776 req_set_fail_links(req);
2777 if (req->flags & REQ_F_BUFFER_SELECTED)
2778 cflags = io_put_rw_kbuf(req);
2779 __io_req_complete(req, issue_flags, ret, cflags);
2784 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2786 struct io_ring_ctx *ctx = req->ctx;
2787 size_t len = req->rw.len;
2788 struct io_mapped_ubuf *imu;
2789 u16 index, buf_index = req->buf_index;
2793 if (unlikely(buf_index >= ctx->nr_user_bufs))
2795 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2796 imu = &ctx->user_bufs[index];
2797 buf_addr = req->rw.addr;
2800 if (buf_addr + len < buf_addr)
2802 /* not inside the mapped region */
2803 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2807 * May not be a start of buffer, set size appropriately
2808 * and advance us to the beginning.
2810 offset = buf_addr - imu->ubuf;
2811 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2815 * Don't use iov_iter_advance() here, as it's really slow for
2816 * using the latter parts of a big fixed buffer - it iterates
2817 * over each segment manually. We can cheat a bit here, because
2820 * 1) it's a BVEC iter, we set it up
2821 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2822 * first and last bvec
2824 * So just find our index, and adjust the iterator afterwards.
2825 * If the offset is within the first bvec (or the whole first
2826 * bvec, just use iov_iter_advance(). This makes it easier
2827 * since we can just skip the first segment, which may not
2828 * be PAGE_SIZE aligned.
2830 const struct bio_vec *bvec = imu->bvec;
2832 if (offset <= bvec->bv_len) {
2833 iov_iter_advance(iter, offset);
2835 unsigned long seg_skip;
2837 /* skip first vec */
2838 offset -= bvec->bv_len;
2839 seg_skip = 1 + (offset >> PAGE_SHIFT);
2841 iter->bvec = bvec + seg_skip;
2842 iter->nr_segs -= seg_skip;
2843 iter->count -= bvec->bv_len + offset;
2844 iter->iov_offset = offset & ~PAGE_MASK;
2851 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2854 mutex_unlock(&ctx->uring_lock);
2857 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2860 * "Normal" inline submissions always hold the uring_lock, since we
2861 * grab it from the system call. Same is true for the SQPOLL offload.
2862 * The only exception is when we've detached the request and issue it
2863 * from an async worker thread, grab the lock for that case.
2866 mutex_lock(&ctx->uring_lock);
2869 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2870 int bgid, struct io_buffer *kbuf,
2873 struct io_buffer *head;
2875 if (req->flags & REQ_F_BUFFER_SELECTED)
2878 io_ring_submit_lock(req->ctx, needs_lock);
2880 lockdep_assert_held(&req->ctx->uring_lock);
2882 head = xa_load(&req->ctx->io_buffers, bgid);
2884 if (!list_empty(&head->list)) {
2885 kbuf = list_last_entry(&head->list, struct io_buffer,
2887 list_del(&kbuf->list);
2890 xa_erase(&req->ctx->io_buffers, bgid);
2892 if (*len > kbuf->len)
2895 kbuf = ERR_PTR(-ENOBUFS);
2898 io_ring_submit_unlock(req->ctx, needs_lock);
2903 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2906 struct io_buffer *kbuf;
2909 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2910 bgid = req->buf_index;
2911 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2914 req->rw.addr = (u64) (unsigned long) kbuf;
2915 req->flags |= REQ_F_BUFFER_SELECTED;
2916 return u64_to_user_ptr(kbuf->addr);
2919 #ifdef CONFIG_COMPAT
2920 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2923 struct compat_iovec __user *uiov;
2924 compat_ssize_t clen;
2928 uiov = u64_to_user_ptr(req->rw.addr);
2929 if (!access_ok(uiov, sizeof(*uiov)))
2931 if (__get_user(clen, &uiov->iov_len))
2937 buf = io_rw_buffer_select(req, &len, needs_lock);
2939 return PTR_ERR(buf);
2940 iov[0].iov_base = buf;
2941 iov[0].iov_len = (compat_size_t) len;
2946 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2949 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2953 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2956 len = iov[0].iov_len;
2959 buf = io_rw_buffer_select(req, &len, needs_lock);
2961 return PTR_ERR(buf);
2962 iov[0].iov_base = buf;
2963 iov[0].iov_len = len;
2967 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2970 if (req->flags & REQ_F_BUFFER_SELECTED) {
2971 struct io_buffer *kbuf;
2973 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2974 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2975 iov[0].iov_len = kbuf->len;
2978 if (req->rw.len != 1)
2981 #ifdef CONFIG_COMPAT
2982 if (req->ctx->compat)
2983 return io_compat_import(req, iov, needs_lock);
2986 return __io_iov_buffer_select(req, iov, needs_lock);
2989 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2990 struct iov_iter *iter, bool needs_lock)
2992 void __user *buf = u64_to_user_ptr(req->rw.addr);
2993 size_t sqe_len = req->rw.len;
2994 u8 opcode = req->opcode;
2997 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2999 return io_import_fixed(req, rw, iter);
3002 /* buffer index only valid with fixed read/write, or buffer select */
3003 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3006 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3007 if (req->flags & REQ_F_BUFFER_SELECT) {
3008 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3010 return PTR_ERR(buf);
3011 req->rw.len = sqe_len;
3014 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3019 if (req->flags & REQ_F_BUFFER_SELECT) {
3020 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3022 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3027 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3031 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3033 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3037 * For files that don't have ->read_iter() and ->write_iter(), handle them
3038 * by looping over ->read() or ->write() manually.
3040 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3042 struct kiocb *kiocb = &req->rw.kiocb;
3043 struct file *file = req->file;
3047 * Don't support polled IO through this interface, and we can't
3048 * support non-blocking either. For the latter, this just causes
3049 * the kiocb to be handled from an async context.
3051 if (kiocb->ki_flags & IOCB_HIPRI)
3053 if (kiocb->ki_flags & IOCB_NOWAIT)
3056 while (iov_iter_count(iter)) {
3060 if (!iov_iter_is_bvec(iter)) {
3061 iovec = iov_iter_iovec(iter);
3063 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3064 iovec.iov_len = req->rw.len;
3068 nr = file->f_op->read(file, iovec.iov_base,
3069 iovec.iov_len, io_kiocb_ppos(kiocb));
3071 nr = file->f_op->write(file, iovec.iov_base,
3072 iovec.iov_len, io_kiocb_ppos(kiocb));
3081 if (nr != iovec.iov_len)
3085 iov_iter_advance(iter, nr);
3091 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3092 const struct iovec *fast_iov, struct iov_iter *iter)
3094 struct io_async_rw *rw = req->async_data;
3096 memcpy(&rw->iter, iter, sizeof(*iter));
3097 rw->free_iovec = iovec;
3099 /* can only be fixed buffers, no need to do anything */
3100 if (iov_iter_is_bvec(iter))
3103 unsigned iov_off = 0;
3105 rw->iter.iov = rw->fast_iov;
3106 if (iter->iov != fast_iov) {
3107 iov_off = iter->iov - fast_iov;
3108 rw->iter.iov += iov_off;
3110 if (rw->fast_iov != fast_iov)
3111 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3112 sizeof(struct iovec) * iter->nr_segs);
3114 req->flags |= REQ_F_NEED_CLEANUP;
3118 static inline int io_alloc_async_data(struct io_kiocb *req)
3120 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3121 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3122 return req->async_data == NULL;
3125 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3126 const struct iovec *fast_iov,
3127 struct iov_iter *iter, bool force)
3129 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3131 if (!req->async_data) {
3132 if (io_alloc_async_data(req)) {
3137 io_req_map_rw(req, iovec, fast_iov, iter);
3142 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3144 struct io_async_rw *iorw = req->async_data;
3145 struct iovec *iov = iorw->fast_iov;
3148 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3149 if (unlikely(ret < 0))
3152 iorw->bytes_done = 0;
3153 iorw->free_iovec = iov;
3155 req->flags |= REQ_F_NEED_CLEANUP;
3159 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3161 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3163 return io_prep_rw(req, sqe);
3167 * This is our waitqueue callback handler, registered through lock_page_async()
3168 * when we initially tried to do the IO with the iocb armed our waitqueue.
3169 * This gets called when the page is unlocked, and we generally expect that to
3170 * happen when the page IO is completed and the page is now uptodate. This will
3171 * queue a task_work based retry of the operation, attempting to copy the data
3172 * again. If the latter fails because the page was NOT uptodate, then we will
3173 * do a thread based blocking retry of the operation. That's the unexpected
3176 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3177 int sync, void *arg)
3179 struct wait_page_queue *wpq;
3180 struct io_kiocb *req = wait->private;
3181 struct wait_page_key *key = arg;
3183 wpq = container_of(wait, struct wait_page_queue, wait);
3185 if (!wake_page_match(wpq, key))
3188 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3189 list_del_init(&wait->entry);
3191 /* submit ref gets dropped, acquire a new one */
3193 io_req_task_queue(req);
3198 * This controls whether a given IO request should be armed for async page
3199 * based retry. If we return false here, the request is handed to the async
3200 * worker threads for retry. If we're doing buffered reads on a regular file,
3201 * we prepare a private wait_page_queue entry and retry the operation. This
3202 * will either succeed because the page is now uptodate and unlocked, or it
3203 * will register a callback when the page is unlocked at IO completion. Through
3204 * that callback, io_uring uses task_work to setup a retry of the operation.
3205 * That retry will attempt the buffered read again. The retry will generally
3206 * succeed, or in rare cases where it fails, we then fall back to using the
3207 * async worker threads for a blocking retry.
3209 static bool io_rw_should_retry(struct io_kiocb *req)
3211 struct io_async_rw *rw = req->async_data;
3212 struct wait_page_queue *wait = &rw->wpq;
3213 struct kiocb *kiocb = &req->rw.kiocb;
3215 /* never retry for NOWAIT, we just complete with -EAGAIN */
3216 if (req->flags & REQ_F_NOWAIT)
3219 /* Only for buffered IO */
3220 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3224 * just use poll if we can, and don't attempt if the fs doesn't
3225 * support callback based unlocks
3227 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3230 wait->wait.func = io_async_buf_func;
3231 wait->wait.private = req;
3232 wait->wait.flags = 0;
3233 INIT_LIST_HEAD(&wait->wait.entry);
3234 kiocb->ki_flags |= IOCB_WAITQ;
3235 kiocb->ki_flags &= ~IOCB_NOWAIT;
3236 kiocb->ki_waitq = wait;
3240 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3242 if (req->file->f_op->read_iter)
3243 return call_read_iter(req->file, &req->rw.kiocb, iter);
3244 else if (req->file->f_op->read)
3245 return loop_rw_iter(READ, req, iter);
3250 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3252 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3253 struct kiocb *kiocb = &req->rw.kiocb;
3254 struct iov_iter __iter, *iter = &__iter;
3255 struct io_async_rw *rw = req->async_data;
3256 ssize_t io_size, ret, ret2;
3257 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3263 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3267 io_size = iov_iter_count(iter);
3268 req->result = io_size;
3270 /* Ensure we clear previously set non-block flag */
3271 if (!force_nonblock)
3272 kiocb->ki_flags &= ~IOCB_NOWAIT;
3274 kiocb->ki_flags |= IOCB_NOWAIT;
3276 /* If the file doesn't support async, just async punt */
3277 if (force_nonblock && !io_file_supports_async(req, READ)) {
3278 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3279 return ret ?: -EAGAIN;
3282 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3283 if (unlikely(ret)) {
3288 ret = io_iter_do_read(req, iter);
3290 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3291 req->flags &= ~REQ_F_REISSUE;
3292 /* IOPOLL retry should happen for io-wq threads */
3293 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3295 /* no retry on NONBLOCK nor RWF_NOWAIT */
3296 if (req->flags & REQ_F_NOWAIT)
3298 /* some cases will consume bytes even on error returns */
3299 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3301 } else if (ret == -EIOCBQUEUED) {
3303 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3304 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3305 /* read all, failed, already did sync or don't want to retry */
3309 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3314 rw = req->async_data;
3315 /* now use our persistent iterator, if we aren't already */
3320 rw->bytes_done += ret;
3321 /* if we can retry, do so with the callbacks armed */
3322 if (!io_rw_should_retry(req)) {
3323 kiocb->ki_flags &= ~IOCB_WAITQ;
3328 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3329 * we get -EIOCBQUEUED, then we'll get a notification when the
3330 * desired page gets unlocked. We can also get a partial read
3331 * here, and if we do, then just retry at the new offset.
3333 ret = io_iter_do_read(req, iter);
3334 if (ret == -EIOCBQUEUED)
3336 /* we got some bytes, but not all. retry. */
3337 kiocb->ki_flags &= ~IOCB_WAITQ;
3338 } while (ret > 0 && ret < io_size);
3340 kiocb_done(kiocb, ret, issue_flags);
3342 /* it's faster to check here then delegate to kfree */
3348 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3350 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3352 return io_prep_rw(req, sqe);
3355 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3357 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3358 struct kiocb *kiocb = &req->rw.kiocb;
3359 struct iov_iter __iter, *iter = &__iter;
3360 struct io_async_rw *rw = req->async_data;
3361 ssize_t ret, ret2, io_size;
3362 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3368 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3372 io_size = iov_iter_count(iter);
3373 req->result = io_size;
3375 /* Ensure we clear previously set non-block flag */
3376 if (!force_nonblock)
3377 kiocb->ki_flags &= ~IOCB_NOWAIT;
3379 kiocb->ki_flags |= IOCB_NOWAIT;
3381 /* If the file doesn't support async, just async punt */
3382 if (force_nonblock && !io_file_supports_async(req, WRITE))
3385 /* file path doesn't support NOWAIT for non-direct_IO */
3386 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3387 (req->flags & REQ_F_ISREG))
3390 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3395 * Open-code file_start_write here to grab freeze protection,
3396 * which will be released by another thread in
3397 * io_complete_rw(). Fool lockdep by telling it the lock got
3398 * released so that it doesn't complain about the held lock when
3399 * we return to userspace.
3401 if (req->flags & REQ_F_ISREG) {
3402 sb_start_write(file_inode(req->file)->i_sb);
3403 __sb_writers_release(file_inode(req->file)->i_sb,
3406 kiocb->ki_flags |= IOCB_WRITE;
3408 if (req->file->f_op->write_iter)
3409 ret2 = call_write_iter(req->file, kiocb, iter);
3410 else if (req->file->f_op->write)
3411 ret2 = loop_rw_iter(WRITE, req, iter);
3415 if (req->flags & REQ_F_REISSUE) {
3416 req->flags &= ~REQ_F_REISSUE;
3421 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3422 * retry them without IOCB_NOWAIT.
3424 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3426 /* no retry on NONBLOCK nor RWF_NOWAIT */
3427 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3429 if (!force_nonblock || ret2 != -EAGAIN) {
3430 /* IOPOLL retry should happen for io-wq threads */
3431 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3434 kiocb_done(kiocb, ret2, issue_flags);
3437 /* some cases will consume bytes even on error returns */
3438 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3439 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3440 return ret ?: -EAGAIN;
3443 /* it's reportedly faster than delegating the null check to kfree() */
3449 static int io_renameat_prep(struct io_kiocb *req,
3450 const struct io_uring_sqe *sqe)
3452 struct io_rename *ren = &req->rename;
3453 const char __user *oldf, *newf;
3455 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3458 ren->old_dfd = READ_ONCE(sqe->fd);
3459 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3460 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3461 ren->new_dfd = READ_ONCE(sqe->len);
3462 ren->flags = READ_ONCE(sqe->rename_flags);
3464 ren->oldpath = getname(oldf);
3465 if (IS_ERR(ren->oldpath))
3466 return PTR_ERR(ren->oldpath);
3468 ren->newpath = getname(newf);
3469 if (IS_ERR(ren->newpath)) {
3470 putname(ren->oldpath);
3471 return PTR_ERR(ren->newpath);
3474 req->flags |= REQ_F_NEED_CLEANUP;
3478 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3480 struct io_rename *ren = &req->rename;
3483 if (issue_flags & IO_URING_F_NONBLOCK)
3486 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3487 ren->newpath, ren->flags);
3489 req->flags &= ~REQ_F_NEED_CLEANUP;
3491 req_set_fail_links(req);
3492 io_req_complete(req, ret);
3496 static int io_unlinkat_prep(struct io_kiocb *req,
3497 const struct io_uring_sqe *sqe)
3499 struct io_unlink *un = &req->unlink;
3500 const char __user *fname;
3502 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3505 un->dfd = READ_ONCE(sqe->fd);
3507 un->flags = READ_ONCE(sqe->unlink_flags);
3508 if (un->flags & ~AT_REMOVEDIR)
3511 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3512 un->filename = getname(fname);
3513 if (IS_ERR(un->filename))
3514 return PTR_ERR(un->filename);
3516 req->flags |= REQ_F_NEED_CLEANUP;
3520 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3522 struct io_unlink *un = &req->unlink;
3525 if (issue_flags & IO_URING_F_NONBLOCK)
3528 if (un->flags & AT_REMOVEDIR)
3529 ret = do_rmdir(un->dfd, un->filename);
3531 ret = do_unlinkat(un->dfd, un->filename);
3533 req->flags &= ~REQ_F_NEED_CLEANUP;
3535 req_set_fail_links(req);
3536 io_req_complete(req, ret);
3540 static int io_shutdown_prep(struct io_kiocb *req,
3541 const struct io_uring_sqe *sqe)
3543 #if defined(CONFIG_NET)
3544 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3546 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3550 req->shutdown.how = READ_ONCE(sqe->len);
3557 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3559 #if defined(CONFIG_NET)
3560 struct socket *sock;
3563 if (issue_flags & IO_URING_F_NONBLOCK)
3566 sock = sock_from_file(req->file);
3567 if (unlikely(!sock))
3570 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3572 req_set_fail_links(req);
3573 io_req_complete(req, ret);
3580 static int __io_splice_prep(struct io_kiocb *req,
3581 const struct io_uring_sqe *sqe)
3583 struct io_splice* sp = &req->splice;
3584 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3586 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3590 sp->len = READ_ONCE(sqe->len);
3591 sp->flags = READ_ONCE(sqe->splice_flags);
3593 if (unlikely(sp->flags & ~valid_flags))
3596 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3597 (sp->flags & SPLICE_F_FD_IN_FIXED));
3600 req->flags |= REQ_F_NEED_CLEANUP;
3602 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3604 * Splice operation will be punted aync, and here need to
3605 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3607 req->work.flags |= IO_WQ_WORK_UNBOUND;
3613 static int io_tee_prep(struct io_kiocb *req,
3614 const struct io_uring_sqe *sqe)
3616 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3618 return __io_splice_prep(req, sqe);
3621 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3623 struct io_splice *sp = &req->splice;
3624 struct file *in = sp->file_in;
3625 struct file *out = sp->file_out;
3626 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3629 if (issue_flags & IO_URING_F_NONBLOCK)
3632 ret = do_tee(in, out, sp->len, flags);
3634 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3636 req->flags &= ~REQ_F_NEED_CLEANUP;
3639 req_set_fail_links(req);
3640 io_req_complete(req, ret);
3644 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3646 struct io_splice* sp = &req->splice;
3648 sp->off_in = READ_ONCE(sqe->splice_off_in);
3649 sp->off_out = READ_ONCE(sqe->off);
3650 return __io_splice_prep(req, sqe);
3653 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3655 struct io_splice *sp = &req->splice;
3656 struct file *in = sp->file_in;
3657 struct file *out = sp->file_out;
3658 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3659 loff_t *poff_in, *poff_out;
3662 if (issue_flags & IO_URING_F_NONBLOCK)
3665 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3666 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3669 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3671 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3673 req->flags &= ~REQ_F_NEED_CLEANUP;
3676 req_set_fail_links(req);
3677 io_req_complete(req, ret);
3682 * IORING_OP_NOP just posts a completion event, nothing else.
3684 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3686 struct io_ring_ctx *ctx = req->ctx;
3688 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3691 __io_req_complete(req, issue_flags, 0, 0);
3695 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3697 struct io_ring_ctx *ctx = req->ctx;
3702 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3704 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3707 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3708 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3711 req->sync.off = READ_ONCE(sqe->off);
3712 req->sync.len = READ_ONCE(sqe->len);
3716 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3718 loff_t end = req->sync.off + req->sync.len;
3721 /* fsync always requires a blocking context */
3722 if (issue_flags & IO_URING_F_NONBLOCK)
3725 ret = vfs_fsync_range(req->file, req->sync.off,
3726 end > 0 ? end : LLONG_MAX,
3727 req->sync.flags & IORING_FSYNC_DATASYNC);
3729 req_set_fail_links(req);
3730 io_req_complete(req, ret);
3734 static int io_fallocate_prep(struct io_kiocb *req,
3735 const struct io_uring_sqe *sqe)
3737 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3742 req->sync.off = READ_ONCE(sqe->off);
3743 req->sync.len = READ_ONCE(sqe->addr);
3744 req->sync.mode = READ_ONCE(sqe->len);
3748 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3752 /* fallocate always requiring blocking context */
3753 if (issue_flags & IO_URING_F_NONBLOCK)
3755 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3758 req_set_fail_links(req);
3759 io_req_complete(req, ret);
3763 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3765 const char __user *fname;
3768 if (unlikely(sqe->ioprio || sqe->buf_index))
3770 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3773 /* open.how should be already initialised */
3774 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3775 req->open.how.flags |= O_LARGEFILE;
3777 req->open.dfd = READ_ONCE(sqe->fd);
3778 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3779 req->open.filename = getname(fname);
3780 if (IS_ERR(req->open.filename)) {
3781 ret = PTR_ERR(req->open.filename);
3782 req->open.filename = NULL;
3785 req->open.nofile = rlimit(RLIMIT_NOFILE);
3786 req->flags |= REQ_F_NEED_CLEANUP;
3790 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3794 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3796 mode = READ_ONCE(sqe->len);
3797 flags = READ_ONCE(sqe->open_flags);
3798 req->open.how = build_open_how(flags, mode);
3799 return __io_openat_prep(req, sqe);
3802 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3804 struct open_how __user *how;
3808 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3810 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3811 len = READ_ONCE(sqe->len);
3812 if (len < OPEN_HOW_SIZE_VER0)
3815 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3820 return __io_openat_prep(req, sqe);
3823 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3825 struct open_flags op;
3828 bool resolve_nonblock;
3831 ret = build_open_flags(&req->open.how, &op);
3834 nonblock_set = op.open_flag & O_NONBLOCK;
3835 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3836 if (issue_flags & IO_URING_F_NONBLOCK) {
3838 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3839 * it'll always -EAGAIN
3841 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3843 op.lookup_flags |= LOOKUP_CACHED;
3844 op.open_flag |= O_NONBLOCK;
3847 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3851 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3852 /* only retry if RESOLVE_CACHED wasn't already set by application */
3853 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3854 file == ERR_PTR(-EAGAIN)) {
3856 * We could hang on to this 'fd', but seems like marginal
3857 * gain for something that is now known to be a slower path.
3858 * So just put it, and we'll get a new one when we retry.
3866 ret = PTR_ERR(file);
3868 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3869 file->f_flags &= ~O_NONBLOCK;
3870 fsnotify_open(file);
3871 fd_install(ret, file);
3874 putname(req->open.filename);
3875 req->flags &= ~REQ_F_NEED_CLEANUP;
3877 req_set_fail_links(req);
3878 io_req_complete(req, ret);
3882 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3884 return io_openat2(req, issue_flags);
3887 static int io_remove_buffers_prep(struct io_kiocb *req,
3888 const struct io_uring_sqe *sqe)
3890 struct io_provide_buf *p = &req->pbuf;
3893 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3896 tmp = READ_ONCE(sqe->fd);
3897 if (!tmp || tmp > USHRT_MAX)
3900 memset(p, 0, sizeof(*p));
3902 p->bgid = READ_ONCE(sqe->buf_group);
3906 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3907 int bgid, unsigned nbufs)
3911 /* shouldn't happen */
3915 /* the head kbuf is the list itself */
3916 while (!list_empty(&buf->list)) {
3917 struct io_buffer *nxt;
3919 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3920 list_del(&nxt->list);
3927 xa_erase(&ctx->io_buffers, bgid);
3932 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3934 struct io_provide_buf *p = &req->pbuf;
3935 struct io_ring_ctx *ctx = req->ctx;
3936 struct io_buffer *head;
3938 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3940 io_ring_submit_lock(ctx, !force_nonblock);
3942 lockdep_assert_held(&ctx->uring_lock);
3945 head = xa_load(&ctx->io_buffers, p->bgid);
3947 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3949 req_set_fail_links(req);
3951 /* complete before unlock, IOPOLL may need the lock */
3952 __io_req_complete(req, issue_flags, ret, 0);
3953 io_ring_submit_unlock(ctx, !force_nonblock);
3957 static int io_provide_buffers_prep(struct io_kiocb *req,
3958 const struct io_uring_sqe *sqe)
3961 struct io_provide_buf *p = &req->pbuf;
3964 if (sqe->ioprio || sqe->rw_flags)
3967 tmp = READ_ONCE(sqe->fd);
3968 if (!tmp || tmp > USHRT_MAX)
3971 p->addr = READ_ONCE(sqe->addr);
3972 p->len = READ_ONCE(sqe->len);
3974 size = (unsigned long)p->len * p->nbufs;
3975 if (!access_ok(u64_to_user_ptr(p->addr), size))
3978 p->bgid = READ_ONCE(sqe->buf_group);
3979 tmp = READ_ONCE(sqe->off);
3980 if (tmp > USHRT_MAX)
3986 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3988 struct io_buffer *buf;
3989 u64 addr = pbuf->addr;
3990 int i, bid = pbuf->bid;
3992 for (i = 0; i < pbuf->nbufs; i++) {
3993 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3998 buf->len = pbuf->len;
4003 INIT_LIST_HEAD(&buf->list);
4006 list_add_tail(&buf->list, &(*head)->list);
4010 return i ? i : -ENOMEM;
4013 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4015 struct io_provide_buf *p = &req->pbuf;
4016 struct io_ring_ctx *ctx = req->ctx;
4017 struct io_buffer *head, *list;
4019 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4021 io_ring_submit_lock(ctx, !force_nonblock);
4023 lockdep_assert_held(&ctx->uring_lock);
4025 list = head = xa_load(&ctx->io_buffers, p->bgid);
4027 ret = io_add_buffers(p, &head);
4028 if (ret >= 0 && !list) {
4029 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4031 __io_remove_buffers(ctx, head, p->bgid, -1U);
4034 req_set_fail_links(req);
4035 /* complete before unlock, IOPOLL may need the lock */
4036 __io_req_complete(req, issue_flags, ret, 0);
4037 io_ring_submit_unlock(ctx, !force_nonblock);
4041 static int io_epoll_ctl_prep(struct io_kiocb *req,
4042 const struct io_uring_sqe *sqe)
4044 #if defined(CONFIG_EPOLL)
4045 if (sqe->ioprio || sqe->buf_index)
4047 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4050 req->epoll.epfd = READ_ONCE(sqe->fd);
4051 req->epoll.op = READ_ONCE(sqe->len);
4052 req->epoll.fd = READ_ONCE(sqe->off);
4054 if (ep_op_has_event(req->epoll.op)) {
4055 struct epoll_event __user *ev;
4057 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4058 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4068 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4070 #if defined(CONFIG_EPOLL)
4071 struct io_epoll *ie = &req->epoll;
4073 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4075 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4076 if (force_nonblock && ret == -EAGAIN)
4080 req_set_fail_links(req);
4081 __io_req_complete(req, issue_flags, ret, 0);
4088 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4090 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4091 if (sqe->ioprio || sqe->buf_index || sqe->off)
4093 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4096 req->madvise.addr = READ_ONCE(sqe->addr);
4097 req->madvise.len = READ_ONCE(sqe->len);
4098 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4105 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4107 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4108 struct io_madvise *ma = &req->madvise;
4111 if (issue_flags & IO_URING_F_NONBLOCK)
4114 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4116 req_set_fail_links(req);
4117 io_req_complete(req, ret);
4124 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4126 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4128 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4131 req->fadvise.offset = READ_ONCE(sqe->off);
4132 req->fadvise.len = READ_ONCE(sqe->len);
4133 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4137 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4139 struct io_fadvise *fa = &req->fadvise;
4142 if (issue_flags & IO_URING_F_NONBLOCK) {
4143 switch (fa->advice) {
4144 case POSIX_FADV_NORMAL:
4145 case POSIX_FADV_RANDOM:
4146 case POSIX_FADV_SEQUENTIAL:
4153 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4155 req_set_fail_links(req);
4156 io_req_complete(req, ret);
4160 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4162 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4164 if (sqe->ioprio || sqe->buf_index)
4166 if (req->flags & REQ_F_FIXED_FILE)
4169 req->statx.dfd = READ_ONCE(sqe->fd);
4170 req->statx.mask = READ_ONCE(sqe->len);
4171 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4172 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4173 req->statx.flags = READ_ONCE(sqe->statx_flags);
4178 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4180 struct io_statx *ctx = &req->statx;
4183 if (issue_flags & IO_URING_F_NONBLOCK) {
4184 /* only need file table for an actual valid fd */
4185 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4186 req->flags |= REQ_F_NO_FILE_TABLE;
4190 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4194 req_set_fail_links(req);
4195 io_req_complete(req, ret);
4199 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4201 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4203 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4204 sqe->rw_flags || sqe->buf_index)
4206 if (req->flags & REQ_F_FIXED_FILE)
4209 req->close.fd = READ_ONCE(sqe->fd);
4213 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4215 struct files_struct *files = current->files;
4216 struct io_close *close = &req->close;
4217 struct fdtable *fdt;
4223 spin_lock(&files->file_lock);
4224 fdt = files_fdtable(files);
4225 if (close->fd >= fdt->max_fds) {
4226 spin_unlock(&files->file_lock);
4229 file = fdt->fd[close->fd];
4231 spin_unlock(&files->file_lock);
4235 if (file->f_op == &io_uring_fops) {
4236 spin_unlock(&files->file_lock);
4241 /* if the file has a flush method, be safe and punt to async */
4242 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4243 spin_unlock(&files->file_lock);
4247 ret = __close_fd_get_file(close->fd, &file);
4248 spin_unlock(&files->file_lock);
4255 /* No ->flush() or already async, safely close from here */
4256 ret = filp_close(file, current->files);
4259 req_set_fail_links(req);
4262 __io_req_complete(req, issue_flags, ret, 0);
4266 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4268 struct io_ring_ctx *ctx = req->ctx;
4270 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4272 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4275 req->sync.off = READ_ONCE(sqe->off);
4276 req->sync.len = READ_ONCE(sqe->len);
4277 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4281 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4285 /* sync_file_range always requires a blocking context */
4286 if (issue_flags & IO_URING_F_NONBLOCK)
4289 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4292 req_set_fail_links(req);
4293 io_req_complete(req, ret);
4297 #if defined(CONFIG_NET)
4298 static int io_setup_async_msg(struct io_kiocb *req,
4299 struct io_async_msghdr *kmsg)
4301 struct io_async_msghdr *async_msg = req->async_data;
4305 if (io_alloc_async_data(req)) {
4306 kfree(kmsg->free_iov);
4309 async_msg = req->async_data;
4310 req->flags |= REQ_F_NEED_CLEANUP;
4311 memcpy(async_msg, kmsg, sizeof(*kmsg));
4312 async_msg->msg.msg_name = &async_msg->addr;
4313 /* if were using fast_iov, set it to the new one */
4314 if (!async_msg->free_iov)
4315 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4320 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4321 struct io_async_msghdr *iomsg)
4323 iomsg->msg.msg_name = &iomsg->addr;
4324 iomsg->free_iov = iomsg->fast_iov;
4325 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4326 req->sr_msg.msg_flags, &iomsg->free_iov);
4329 static int io_sendmsg_prep_async(struct io_kiocb *req)
4333 ret = io_sendmsg_copy_hdr(req, req->async_data);
4335 req->flags |= REQ_F_NEED_CLEANUP;
4339 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4341 struct io_sr_msg *sr = &req->sr_msg;
4343 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4346 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4347 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4348 sr->len = READ_ONCE(sqe->len);
4350 #ifdef CONFIG_COMPAT
4351 if (req->ctx->compat)
4352 sr->msg_flags |= MSG_CMSG_COMPAT;
4357 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4359 struct io_async_msghdr iomsg, *kmsg;
4360 struct socket *sock;
4365 sock = sock_from_file(req->file);
4366 if (unlikely(!sock))
4369 kmsg = req->async_data;
4371 ret = io_sendmsg_copy_hdr(req, &iomsg);
4377 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4378 if (flags & MSG_DONTWAIT)
4379 req->flags |= REQ_F_NOWAIT;
4380 else if (issue_flags & IO_URING_F_NONBLOCK)
4381 flags |= MSG_DONTWAIT;
4383 if (flags & MSG_WAITALL)
4384 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4386 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4387 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4388 return io_setup_async_msg(req, kmsg);
4389 if (ret == -ERESTARTSYS)
4392 /* fast path, check for non-NULL to avoid function call */
4394 kfree(kmsg->free_iov);
4395 req->flags &= ~REQ_F_NEED_CLEANUP;
4397 req_set_fail_links(req);
4398 __io_req_complete(req, issue_flags, ret, 0);
4402 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4404 struct io_sr_msg *sr = &req->sr_msg;
4407 struct socket *sock;
4412 sock = sock_from_file(req->file);
4413 if (unlikely(!sock))
4416 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4420 msg.msg_name = NULL;
4421 msg.msg_control = NULL;
4422 msg.msg_controllen = 0;
4423 msg.msg_namelen = 0;
4425 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4426 if (flags & MSG_DONTWAIT)
4427 req->flags |= REQ_F_NOWAIT;
4428 else if (issue_flags & IO_URING_F_NONBLOCK)
4429 flags |= MSG_DONTWAIT;
4431 if (flags & MSG_WAITALL)
4432 min_ret = iov_iter_count(&msg.msg_iter);
4434 msg.msg_flags = flags;
4435 ret = sock_sendmsg(sock, &msg);
4436 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4438 if (ret == -ERESTARTSYS)
4442 req_set_fail_links(req);
4443 __io_req_complete(req, issue_flags, ret, 0);
4447 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4448 struct io_async_msghdr *iomsg)
4450 struct io_sr_msg *sr = &req->sr_msg;
4451 struct iovec __user *uiov;
4455 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4456 &iomsg->uaddr, &uiov, &iov_len);
4460 if (req->flags & REQ_F_BUFFER_SELECT) {
4463 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4465 sr->len = iomsg->fast_iov[0].iov_len;
4466 iomsg->free_iov = NULL;
4468 iomsg->free_iov = iomsg->fast_iov;
4469 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4470 &iomsg->free_iov, &iomsg->msg.msg_iter,
4479 #ifdef CONFIG_COMPAT
4480 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4481 struct io_async_msghdr *iomsg)
4483 struct compat_msghdr __user *msg_compat;
4484 struct io_sr_msg *sr = &req->sr_msg;
4485 struct compat_iovec __user *uiov;
4490 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4491 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4496 uiov = compat_ptr(ptr);
4497 if (req->flags & REQ_F_BUFFER_SELECT) {
4498 compat_ssize_t clen;
4502 if (!access_ok(uiov, sizeof(*uiov)))
4504 if (__get_user(clen, &uiov->iov_len))
4509 iomsg->free_iov = NULL;
4511 iomsg->free_iov = iomsg->fast_iov;
4512 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4513 UIO_FASTIOV, &iomsg->free_iov,
4514 &iomsg->msg.msg_iter, true);
4523 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4524 struct io_async_msghdr *iomsg)
4526 iomsg->msg.msg_name = &iomsg->addr;
4528 #ifdef CONFIG_COMPAT
4529 if (req->ctx->compat)
4530 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4533 return __io_recvmsg_copy_hdr(req, iomsg);
4536 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4539 struct io_sr_msg *sr = &req->sr_msg;
4540 struct io_buffer *kbuf;
4542 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4547 req->flags |= REQ_F_BUFFER_SELECTED;
4551 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4553 return io_put_kbuf(req, req->sr_msg.kbuf);
4556 static int io_recvmsg_prep_async(struct io_kiocb *req)
4560 ret = io_recvmsg_copy_hdr(req, req->async_data);
4562 req->flags |= REQ_F_NEED_CLEANUP;
4566 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4568 struct io_sr_msg *sr = &req->sr_msg;
4570 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4573 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4574 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4575 sr->len = READ_ONCE(sqe->len);
4576 sr->bgid = READ_ONCE(sqe->buf_group);
4578 #ifdef CONFIG_COMPAT
4579 if (req->ctx->compat)
4580 sr->msg_flags |= MSG_CMSG_COMPAT;
4585 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4587 struct io_async_msghdr iomsg, *kmsg;
4588 struct socket *sock;
4589 struct io_buffer *kbuf;
4592 int ret, cflags = 0;
4593 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4595 sock = sock_from_file(req->file);
4596 if (unlikely(!sock))
4599 kmsg = req->async_data;
4601 ret = io_recvmsg_copy_hdr(req, &iomsg);
4607 if (req->flags & REQ_F_BUFFER_SELECT) {
4608 kbuf = io_recv_buffer_select(req, !force_nonblock);
4610 return PTR_ERR(kbuf);
4611 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4612 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4613 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4614 1, req->sr_msg.len);
4617 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4618 if (flags & MSG_DONTWAIT)
4619 req->flags |= REQ_F_NOWAIT;
4620 else if (force_nonblock)
4621 flags |= MSG_DONTWAIT;
4623 if (flags & MSG_WAITALL)
4624 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4626 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4627 kmsg->uaddr, flags);
4628 if (force_nonblock && ret == -EAGAIN)
4629 return io_setup_async_msg(req, kmsg);
4630 if (ret == -ERESTARTSYS)
4633 if (req->flags & REQ_F_BUFFER_SELECTED)
4634 cflags = io_put_recv_kbuf(req);
4635 /* fast path, check for non-NULL to avoid function call */
4637 kfree(kmsg->free_iov);
4638 req->flags &= ~REQ_F_NEED_CLEANUP;
4639 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4640 req_set_fail_links(req);
4641 __io_req_complete(req, issue_flags, ret, cflags);
4645 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4647 struct io_buffer *kbuf;
4648 struct io_sr_msg *sr = &req->sr_msg;
4650 void __user *buf = sr->buf;
4651 struct socket *sock;
4655 int ret, cflags = 0;
4656 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4658 sock = sock_from_file(req->file);
4659 if (unlikely(!sock))
4662 if (req->flags & REQ_F_BUFFER_SELECT) {
4663 kbuf = io_recv_buffer_select(req, !force_nonblock);
4665 return PTR_ERR(kbuf);
4666 buf = u64_to_user_ptr(kbuf->addr);
4669 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4673 msg.msg_name = NULL;
4674 msg.msg_control = NULL;
4675 msg.msg_controllen = 0;
4676 msg.msg_namelen = 0;
4677 msg.msg_iocb = NULL;
4680 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4681 if (flags & MSG_DONTWAIT)
4682 req->flags |= REQ_F_NOWAIT;
4683 else if (force_nonblock)
4684 flags |= MSG_DONTWAIT;
4686 if (flags & MSG_WAITALL)
4687 min_ret = iov_iter_count(&msg.msg_iter);
4689 ret = sock_recvmsg(sock, &msg, flags);
4690 if (force_nonblock && ret == -EAGAIN)
4692 if (ret == -ERESTARTSYS)
4695 if (req->flags & REQ_F_BUFFER_SELECTED)
4696 cflags = io_put_recv_kbuf(req);
4697 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4698 req_set_fail_links(req);
4699 __io_req_complete(req, issue_flags, ret, cflags);
4703 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4705 struct io_accept *accept = &req->accept;
4707 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4709 if (sqe->ioprio || sqe->len || sqe->buf_index)
4712 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4713 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4714 accept->flags = READ_ONCE(sqe->accept_flags);
4715 accept->nofile = rlimit(RLIMIT_NOFILE);
4719 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4721 struct io_accept *accept = &req->accept;
4722 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4723 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4726 if (req->file->f_flags & O_NONBLOCK)
4727 req->flags |= REQ_F_NOWAIT;
4729 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4730 accept->addr_len, accept->flags,
4732 if (ret == -EAGAIN && force_nonblock)
4735 if (ret == -ERESTARTSYS)
4737 req_set_fail_links(req);
4739 __io_req_complete(req, issue_flags, ret, 0);
4743 static int io_connect_prep_async(struct io_kiocb *req)
4745 struct io_async_connect *io = req->async_data;
4746 struct io_connect *conn = &req->connect;
4748 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4751 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4753 struct io_connect *conn = &req->connect;
4755 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4757 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4760 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4761 conn->addr_len = READ_ONCE(sqe->addr2);
4765 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4767 struct io_async_connect __io, *io;
4768 unsigned file_flags;
4770 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4772 if (req->async_data) {
4773 io = req->async_data;
4775 ret = move_addr_to_kernel(req->connect.addr,
4776 req->connect.addr_len,
4783 file_flags = force_nonblock ? O_NONBLOCK : 0;
4785 ret = __sys_connect_file(req->file, &io->address,
4786 req->connect.addr_len, file_flags);
4787 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4788 if (req->async_data)
4790 if (io_alloc_async_data(req)) {
4794 memcpy(req->async_data, &__io, sizeof(__io));
4797 if (ret == -ERESTARTSYS)
4801 req_set_fail_links(req);
4802 __io_req_complete(req, issue_flags, ret, 0);
4805 #else /* !CONFIG_NET */
4806 #define IO_NETOP_FN(op) \
4807 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4809 return -EOPNOTSUPP; \
4812 #define IO_NETOP_PREP(op) \
4814 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4816 return -EOPNOTSUPP; \
4819 #define IO_NETOP_PREP_ASYNC(op) \
4821 static int io_##op##_prep_async(struct io_kiocb *req) \
4823 return -EOPNOTSUPP; \
4826 IO_NETOP_PREP_ASYNC(sendmsg);
4827 IO_NETOP_PREP_ASYNC(recvmsg);
4828 IO_NETOP_PREP_ASYNC(connect);
4829 IO_NETOP_PREP(accept);
4832 #endif /* CONFIG_NET */
4834 struct io_poll_table {
4835 struct poll_table_struct pt;
4836 struct io_kiocb *req;
4840 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4841 __poll_t mask, task_work_func_t func)
4845 /* for instances that support it check for an event match first: */
4846 if (mask && !(mask & poll->events))
4849 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4851 list_del_init(&poll->wait.entry);
4854 req->task_work.func = func;
4857 * If this fails, then the task is exiting. When a task exits, the
4858 * work gets canceled, so just cancel this request as well instead
4859 * of executing it. We can't safely execute it anyway, as we may not
4860 * have the needed state needed for it anyway.
4862 ret = io_req_task_work_add(req);
4863 if (unlikely(ret)) {
4864 WRITE_ONCE(poll->canceled, true);
4865 io_req_task_work_add_fallback(req, func);
4870 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4871 __acquires(&req->ctx->completion_lock)
4873 struct io_ring_ctx *ctx = req->ctx;
4875 if (!req->result && !READ_ONCE(poll->canceled)) {
4876 struct poll_table_struct pt = { ._key = poll->events };
4878 req->result = vfs_poll(req->file, &pt) & poll->events;
4881 spin_lock_irq(&ctx->completion_lock);
4882 if (!req->result && !READ_ONCE(poll->canceled)) {
4883 add_wait_queue(poll->head, &poll->wait);
4890 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4892 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4893 if (req->opcode == IORING_OP_POLL_ADD)
4894 return req->async_data;
4895 return req->apoll->double_poll;
4898 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4900 if (req->opcode == IORING_OP_POLL_ADD)
4902 return &req->apoll->poll;
4905 static void io_poll_remove_double(struct io_kiocb *req)
4907 struct io_poll_iocb *poll = io_poll_get_double(req);
4909 lockdep_assert_held(&req->ctx->completion_lock);
4911 if (poll && poll->head) {
4912 struct wait_queue_head *head = poll->head;
4914 spin_lock(&head->lock);
4915 list_del_init(&poll->wait.entry);
4916 if (poll->wait.private)
4919 spin_unlock(&head->lock);
4923 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4925 struct io_ring_ctx *ctx = req->ctx;
4927 if (!error && req->poll.canceled)
4930 io_poll_remove_double(req);
4931 req->poll.done = true;
4932 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4933 io_commit_cqring(ctx);
4936 static void io_poll_task_func(struct callback_head *cb)
4938 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4939 struct io_ring_ctx *ctx = req->ctx;
4940 struct io_kiocb *nxt;
4942 if (io_poll_rewait(req, &req->poll)) {
4943 spin_unlock_irq(&ctx->completion_lock);
4945 hash_del(&req->hash_node);
4946 io_poll_complete(req, req->result, 0);
4947 spin_unlock_irq(&ctx->completion_lock);
4949 nxt = io_put_req_find_next(req);
4950 io_cqring_ev_posted(ctx);
4952 __io_req_task_submit(nxt);
4956 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4957 int sync, void *key)
4959 struct io_kiocb *req = wait->private;
4960 struct io_poll_iocb *poll = io_poll_get_single(req);
4961 __poll_t mask = key_to_poll(key);
4963 /* for instances that support it check for an event match first: */
4964 if (mask && !(mask & poll->events))
4967 list_del_init(&wait->entry);
4969 if (poll && poll->head) {
4972 spin_lock(&poll->head->lock);
4973 done = list_empty(&poll->wait.entry);
4975 list_del_init(&poll->wait.entry);
4976 /* make sure double remove sees this as being gone */
4977 wait->private = NULL;
4978 spin_unlock(&poll->head->lock);
4980 /* use wait func handler, so it matches the rq type */
4981 poll->wait.func(&poll->wait, mode, sync, key);
4988 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4989 wait_queue_func_t wake_func)
4993 poll->canceled = false;
4994 poll->events = events;
4995 INIT_LIST_HEAD(&poll->wait.entry);
4996 init_waitqueue_func_entry(&poll->wait, wake_func);
4999 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5000 struct wait_queue_head *head,
5001 struct io_poll_iocb **poll_ptr)
5003 struct io_kiocb *req = pt->req;
5006 * If poll->head is already set, it's because the file being polled
5007 * uses multiple waitqueues for poll handling (eg one for read, one
5008 * for write). Setup a separate io_poll_iocb if this happens.
5010 if (unlikely(poll->head)) {
5011 struct io_poll_iocb *poll_one = poll;
5013 /* already have a 2nd entry, fail a third attempt */
5015 pt->error = -EINVAL;
5018 /* double add on the same waitqueue head, ignore */
5019 if (poll->head == head)
5021 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5023 pt->error = -ENOMEM;
5026 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5028 poll->wait.private = req;
5035 if (poll->events & EPOLLEXCLUSIVE)
5036 add_wait_queue_exclusive(head, &poll->wait);
5038 add_wait_queue(head, &poll->wait);
5041 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5042 struct poll_table_struct *p)
5044 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5045 struct async_poll *apoll = pt->req->apoll;
5047 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5050 static void io_async_task_func(struct callback_head *cb)
5052 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5053 struct async_poll *apoll = req->apoll;
5054 struct io_ring_ctx *ctx = req->ctx;
5056 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5058 if (io_poll_rewait(req, &apoll->poll)) {
5059 spin_unlock_irq(&ctx->completion_lock);
5063 /* If req is still hashed, it cannot have been canceled. Don't check. */
5064 if (hash_hashed(&req->hash_node))
5065 hash_del(&req->hash_node);
5067 io_poll_remove_double(req);
5068 spin_unlock_irq(&ctx->completion_lock);
5070 if (!READ_ONCE(apoll->poll.canceled))
5071 __io_req_task_submit(req);
5073 io_req_complete_failed(req, -ECANCELED);
5075 kfree(apoll->double_poll);
5079 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5082 struct io_kiocb *req = wait->private;
5083 struct io_poll_iocb *poll = &req->apoll->poll;
5085 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5088 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5091 static void io_poll_req_insert(struct io_kiocb *req)
5093 struct io_ring_ctx *ctx = req->ctx;
5094 struct hlist_head *list;
5096 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5097 hlist_add_head(&req->hash_node, list);
5100 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5101 struct io_poll_iocb *poll,
5102 struct io_poll_table *ipt, __poll_t mask,
5103 wait_queue_func_t wake_func)
5104 __acquires(&ctx->completion_lock)
5106 struct io_ring_ctx *ctx = req->ctx;
5107 bool cancel = false;
5109 INIT_HLIST_NODE(&req->hash_node);
5110 io_init_poll_iocb(poll, mask, wake_func);
5111 poll->file = req->file;
5112 poll->wait.private = req;
5114 ipt->pt._key = mask;
5116 ipt->error = -EINVAL;
5118 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5120 spin_lock_irq(&ctx->completion_lock);
5121 if (likely(poll->head)) {
5122 spin_lock(&poll->head->lock);
5123 if (unlikely(list_empty(&poll->wait.entry))) {
5129 if (mask || ipt->error)
5130 list_del_init(&poll->wait.entry);
5132 WRITE_ONCE(poll->canceled, true);
5133 else if (!poll->done) /* actually waiting for an event */
5134 io_poll_req_insert(req);
5135 spin_unlock(&poll->head->lock);
5141 static bool io_arm_poll_handler(struct io_kiocb *req)
5143 const struct io_op_def *def = &io_op_defs[req->opcode];
5144 struct io_ring_ctx *ctx = req->ctx;
5145 struct async_poll *apoll;
5146 struct io_poll_table ipt;
5150 if (!req->file || !file_can_poll(req->file))
5152 if (req->flags & REQ_F_POLLED)
5156 else if (def->pollout)
5160 /* if we can't nonblock try, then no point in arming a poll handler */
5161 if (!io_file_supports_async(req, rw))
5164 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5165 if (unlikely(!apoll))
5167 apoll->double_poll = NULL;
5169 req->flags |= REQ_F_POLLED;
5174 mask |= POLLIN | POLLRDNORM;
5176 mask |= POLLOUT | POLLWRNORM;
5178 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5179 if ((req->opcode == IORING_OP_RECVMSG) &&
5180 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5183 mask |= POLLERR | POLLPRI;
5185 ipt.pt._qproc = io_async_queue_proc;
5187 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5189 if (ret || ipt.error) {
5190 io_poll_remove_double(req);
5191 spin_unlock_irq(&ctx->completion_lock);
5192 kfree(apoll->double_poll);
5196 spin_unlock_irq(&ctx->completion_lock);
5197 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5198 apoll->poll.events);
5202 static bool __io_poll_remove_one(struct io_kiocb *req,
5203 struct io_poll_iocb *poll)
5205 bool do_complete = false;
5207 spin_lock(&poll->head->lock);
5208 WRITE_ONCE(poll->canceled, true);
5209 if (!list_empty(&poll->wait.entry)) {
5210 list_del_init(&poll->wait.entry);
5213 spin_unlock(&poll->head->lock);
5214 hash_del(&req->hash_node);
5218 static bool io_poll_remove_one(struct io_kiocb *req)
5222 io_poll_remove_double(req);
5224 if (req->opcode == IORING_OP_POLL_ADD) {
5225 do_complete = __io_poll_remove_one(req, &req->poll);
5227 struct async_poll *apoll = req->apoll;
5229 /* non-poll requests have submit ref still */
5230 do_complete = __io_poll_remove_one(req, &apoll->poll);
5233 kfree(apoll->double_poll);
5239 io_cqring_fill_event(req, -ECANCELED);
5240 io_commit_cqring(req->ctx);
5241 req_set_fail_links(req);
5242 io_put_req_deferred(req, 1);
5249 * Returns true if we found and killed one or more poll requests
5251 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5252 struct files_struct *files)
5254 struct hlist_node *tmp;
5255 struct io_kiocb *req;
5258 spin_lock_irq(&ctx->completion_lock);
5259 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5260 struct hlist_head *list;
5262 list = &ctx->cancel_hash[i];
5263 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5264 if (io_match_task(req, tsk, files))
5265 posted += io_poll_remove_one(req);
5268 spin_unlock_irq(&ctx->completion_lock);
5271 io_cqring_ev_posted(ctx);
5276 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5278 struct hlist_head *list;
5279 struct io_kiocb *req;
5281 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5282 hlist_for_each_entry(req, list, hash_node) {
5283 if (sqe_addr != req->user_data)
5285 if (io_poll_remove_one(req))
5293 static int io_poll_remove_prep(struct io_kiocb *req,
5294 const struct io_uring_sqe *sqe)
5296 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5298 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5302 req->poll_remove.addr = READ_ONCE(sqe->addr);
5307 * Find a running poll command that matches one specified in sqe->addr,
5308 * and remove it if found.
5310 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5312 struct io_ring_ctx *ctx = req->ctx;
5315 spin_lock_irq(&ctx->completion_lock);
5316 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5317 spin_unlock_irq(&ctx->completion_lock);
5320 req_set_fail_links(req);
5321 io_req_complete(req, ret);
5325 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5328 struct io_kiocb *req = wait->private;
5329 struct io_poll_iocb *poll = &req->poll;
5331 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5334 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5335 struct poll_table_struct *p)
5337 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5339 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5342 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5344 struct io_poll_iocb *poll = &req->poll;
5347 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5349 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5352 events = READ_ONCE(sqe->poll32_events);
5354 events = swahw32(events);
5356 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5357 (events & EPOLLEXCLUSIVE);
5361 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5363 struct io_poll_iocb *poll = &req->poll;
5364 struct io_ring_ctx *ctx = req->ctx;
5365 struct io_poll_table ipt;
5368 ipt.pt._qproc = io_poll_queue_proc;
5370 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5373 if (mask) { /* no async, we'd stolen it */
5375 io_poll_complete(req, mask, 0);
5377 spin_unlock_irq(&ctx->completion_lock);
5380 io_cqring_ev_posted(ctx);
5386 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5388 struct io_timeout_data *data = container_of(timer,
5389 struct io_timeout_data, timer);
5390 struct io_kiocb *req = data->req;
5391 struct io_ring_ctx *ctx = req->ctx;
5392 unsigned long flags;
5394 spin_lock_irqsave(&ctx->completion_lock, flags);
5395 list_del_init(&req->timeout.list);
5396 atomic_set(&req->ctx->cq_timeouts,
5397 atomic_read(&req->ctx->cq_timeouts) + 1);
5399 io_cqring_fill_event(req, -ETIME);
5400 io_commit_cqring(ctx);
5401 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5403 io_cqring_ev_posted(ctx);
5404 req_set_fail_links(req);
5406 return HRTIMER_NORESTART;
5409 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5412 struct io_timeout_data *io;
5413 struct io_kiocb *req;
5416 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5417 if (user_data == req->user_data) {
5424 return ERR_PTR(ret);
5426 io = req->async_data;
5427 ret = hrtimer_try_to_cancel(&io->timer);
5429 return ERR_PTR(-EALREADY);
5430 list_del_init(&req->timeout.list);
5434 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5436 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5439 return PTR_ERR(req);
5441 req_set_fail_links(req);
5442 io_cqring_fill_event(req, -ECANCELED);
5443 io_put_req_deferred(req, 1);
5447 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5448 struct timespec64 *ts, enum hrtimer_mode mode)
5450 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5451 struct io_timeout_data *data;
5454 return PTR_ERR(req);
5456 req->timeout.off = 0; /* noseq */
5457 data = req->async_data;
5458 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5459 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5460 data->timer.function = io_timeout_fn;
5461 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5465 static int io_timeout_remove_prep(struct io_kiocb *req,
5466 const struct io_uring_sqe *sqe)
5468 struct io_timeout_rem *tr = &req->timeout_rem;
5470 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5472 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5474 if (sqe->ioprio || sqe->buf_index || sqe->len)
5477 tr->addr = READ_ONCE(sqe->addr);
5478 tr->flags = READ_ONCE(sqe->timeout_flags);
5479 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5480 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5482 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5484 } else if (tr->flags) {
5485 /* timeout removal doesn't support flags */
5492 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5494 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5499 * Remove or update an existing timeout command
5501 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5503 struct io_timeout_rem *tr = &req->timeout_rem;
5504 struct io_ring_ctx *ctx = req->ctx;
5507 spin_lock_irq(&ctx->completion_lock);
5508 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5509 ret = io_timeout_cancel(ctx, tr->addr);
5511 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5512 io_translate_timeout_mode(tr->flags));
5514 io_cqring_fill_event(req, ret);
5515 io_commit_cqring(ctx);
5516 spin_unlock_irq(&ctx->completion_lock);
5517 io_cqring_ev_posted(ctx);
5519 req_set_fail_links(req);
5524 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5525 bool is_timeout_link)
5527 struct io_timeout_data *data;
5529 u32 off = READ_ONCE(sqe->off);
5531 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5533 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5535 if (off && is_timeout_link)
5537 flags = READ_ONCE(sqe->timeout_flags);
5538 if (flags & ~IORING_TIMEOUT_ABS)
5541 req->timeout.off = off;
5543 if (!req->async_data && io_alloc_async_data(req))
5546 data = req->async_data;
5549 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5552 data->mode = io_translate_timeout_mode(flags);
5553 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5554 if (is_timeout_link)
5555 io_req_track_inflight(req);
5559 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5561 struct io_ring_ctx *ctx = req->ctx;
5562 struct io_timeout_data *data = req->async_data;
5563 struct list_head *entry;
5564 u32 tail, off = req->timeout.off;
5566 spin_lock_irq(&ctx->completion_lock);
5569 * sqe->off holds how many events that need to occur for this
5570 * timeout event to be satisfied. If it isn't set, then this is
5571 * a pure timeout request, sequence isn't used.
5573 if (io_is_timeout_noseq(req)) {
5574 entry = ctx->timeout_list.prev;
5578 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5579 req->timeout.target_seq = tail + off;
5581 /* Update the last seq here in case io_flush_timeouts() hasn't.
5582 * This is safe because ->completion_lock is held, and submissions
5583 * and completions are never mixed in the same ->completion_lock section.
5585 ctx->cq_last_tm_flush = tail;
5588 * Insertion sort, ensuring the first entry in the list is always
5589 * the one we need first.
5591 list_for_each_prev(entry, &ctx->timeout_list) {
5592 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5595 if (io_is_timeout_noseq(nxt))
5597 /* nxt.seq is behind @tail, otherwise would've been completed */
5598 if (off >= nxt->timeout.target_seq - tail)
5602 list_add(&req->timeout.list, entry);
5603 data->timer.function = io_timeout_fn;
5604 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5605 spin_unlock_irq(&ctx->completion_lock);
5609 struct io_cancel_data {
5610 struct io_ring_ctx *ctx;
5614 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5616 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5617 struct io_cancel_data *cd = data;
5619 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5622 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5623 struct io_ring_ctx *ctx)
5625 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5626 enum io_wq_cancel cancel_ret;
5629 if (!tctx || !tctx->io_wq)
5632 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5633 switch (cancel_ret) {
5634 case IO_WQ_CANCEL_OK:
5637 case IO_WQ_CANCEL_RUNNING:
5640 case IO_WQ_CANCEL_NOTFOUND:
5648 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5649 struct io_kiocb *req, __u64 sqe_addr,
5652 unsigned long flags;
5655 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5656 if (ret != -ENOENT) {
5657 spin_lock_irqsave(&ctx->completion_lock, flags);
5661 spin_lock_irqsave(&ctx->completion_lock, flags);
5662 ret = io_timeout_cancel(ctx, sqe_addr);
5665 ret = io_poll_cancel(ctx, sqe_addr);
5669 io_cqring_fill_event(req, ret);
5670 io_commit_cqring(ctx);
5671 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5672 io_cqring_ev_posted(ctx);
5675 req_set_fail_links(req);
5679 static int io_async_cancel_prep(struct io_kiocb *req,
5680 const struct io_uring_sqe *sqe)
5682 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5684 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5686 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5689 req->cancel.addr = READ_ONCE(sqe->addr);
5693 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5695 struct io_ring_ctx *ctx = req->ctx;
5696 u64 sqe_addr = req->cancel.addr;
5697 struct io_tctx_node *node;
5700 /* tasks should wait for their io-wq threads, so safe w/o sync */
5701 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5702 spin_lock_irq(&ctx->completion_lock);
5705 ret = io_timeout_cancel(ctx, sqe_addr);
5708 ret = io_poll_cancel(ctx, sqe_addr);
5711 spin_unlock_irq(&ctx->completion_lock);
5713 /* slow path, try all io-wq's */
5714 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5716 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5717 struct io_uring_task *tctx = node->task->io_uring;
5719 if (!tctx || !tctx->io_wq)
5721 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5725 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5727 spin_lock_irq(&ctx->completion_lock);
5729 io_cqring_fill_event(req, ret);
5730 io_commit_cqring(ctx);
5731 spin_unlock_irq(&ctx->completion_lock);
5732 io_cqring_ev_posted(ctx);
5735 req_set_fail_links(req);
5740 static int io_rsrc_update_prep(struct io_kiocb *req,
5741 const struct io_uring_sqe *sqe)
5743 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5745 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5747 if (sqe->ioprio || sqe->rw_flags)
5750 req->rsrc_update.offset = READ_ONCE(sqe->off);
5751 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5752 if (!req->rsrc_update.nr_args)
5754 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5758 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5760 struct io_ring_ctx *ctx = req->ctx;
5761 struct io_uring_rsrc_update up;
5764 if (issue_flags & IO_URING_F_NONBLOCK)
5767 up.offset = req->rsrc_update.offset;
5768 up.data = req->rsrc_update.arg;
5770 mutex_lock(&ctx->uring_lock);
5771 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5772 mutex_unlock(&ctx->uring_lock);
5775 req_set_fail_links(req);
5776 __io_req_complete(req, issue_flags, ret, 0);
5780 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5782 switch (req->opcode) {
5785 case IORING_OP_READV:
5786 case IORING_OP_READ_FIXED:
5787 case IORING_OP_READ:
5788 return io_read_prep(req, sqe);
5789 case IORING_OP_WRITEV:
5790 case IORING_OP_WRITE_FIXED:
5791 case IORING_OP_WRITE:
5792 return io_write_prep(req, sqe);
5793 case IORING_OP_POLL_ADD:
5794 return io_poll_add_prep(req, sqe);
5795 case IORING_OP_POLL_REMOVE:
5796 return io_poll_remove_prep(req, sqe);
5797 case IORING_OP_FSYNC:
5798 return io_fsync_prep(req, sqe);
5799 case IORING_OP_SYNC_FILE_RANGE:
5800 return io_sfr_prep(req, sqe);
5801 case IORING_OP_SENDMSG:
5802 case IORING_OP_SEND:
5803 return io_sendmsg_prep(req, sqe);
5804 case IORING_OP_RECVMSG:
5805 case IORING_OP_RECV:
5806 return io_recvmsg_prep(req, sqe);
5807 case IORING_OP_CONNECT:
5808 return io_connect_prep(req, sqe);
5809 case IORING_OP_TIMEOUT:
5810 return io_timeout_prep(req, sqe, false);
5811 case IORING_OP_TIMEOUT_REMOVE:
5812 return io_timeout_remove_prep(req, sqe);
5813 case IORING_OP_ASYNC_CANCEL:
5814 return io_async_cancel_prep(req, sqe);
5815 case IORING_OP_LINK_TIMEOUT:
5816 return io_timeout_prep(req, sqe, true);
5817 case IORING_OP_ACCEPT:
5818 return io_accept_prep(req, sqe);
5819 case IORING_OP_FALLOCATE:
5820 return io_fallocate_prep(req, sqe);
5821 case IORING_OP_OPENAT:
5822 return io_openat_prep(req, sqe);
5823 case IORING_OP_CLOSE:
5824 return io_close_prep(req, sqe);
5825 case IORING_OP_FILES_UPDATE:
5826 return io_rsrc_update_prep(req, sqe);
5827 case IORING_OP_STATX:
5828 return io_statx_prep(req, sqe);
5829 case IORING_OP_FADVISE:
5830 return io_fadvise_prep(req, sqe);
5831 case IORING_OP_MADVISE:
5832 return io_madvise_prep(req, sqe);
5833 case IORING_OP_OPENAT2:
5834 return io_openat2_prep(req, sqe);
5835 case IORING_OP_EPOLL_CTL:
5836 return io_epoll_ctl_prep(req, sqe);
5837 case IORING_OP_SPLICE:
5838 return io_splice_prep(req, sqe);
5839 case IORING_OP_PROVIDE_BUFFERS:
5840 return io_provide_buffers_prep(req, sqe);
5841 case IORING_OP_REMOVE_BUFFERS:
5842 return io_remove_buffers_prep(req, sqe);
5844 return io_tee_prep(req, sqe);
5845 case IORING_OP_SHUTDOWN:
5846 return io_shutdown_prep(req, sqe);
5847 case IORING_OP_RENAMEAT:
5848 return io_renameat_prep(req, sqe);
5849 case IORING_OP_UNLINKAT:
5850 return io_unlinkat_prep(req, sqe);
5853 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5858 static int io_req_prep_async(struct io_kiocb *req)
5860 if (!io_op_defs[req->opcode].needs_async_setup)
5862 if (WARN_ON_ONCE(req->async_data))
5864 if (io_alloc_async_data(req))
5867 switch (req->opcode) {
5868 case IORING_OP_READV:
5869 return io_rw_prep_async(req, READ);
5870 case IORING_OP_WRITEV:
5871 return io_rw_prep_async(req, WRITE);
5872 case IORING_OP_SENDMSG:
5873 return io_sendmsg_prep_async(req);
5874 case IORING_OP_RECVMSG:
5875 return io_recvmsg_prep_async(req);
5876 case IORING_OP_CONNECT:
5877 return io_connect_prep_async(req);
5879 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5884 static u32 io_get_sequence(struct io_kiocb *req)
5886 struct io_kiocb *pos;
5887 struct io_ring_ctx *ctx = req->ctx;
5888 u32 total_submitted, nr_reqs = 0;
5890 io_for_each_link(pos, req)
5893 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5894 return total_submitted - nr_reqs;
5897 static int io_req_defer(struct io_kiocb *req)
5899 struct io_ring_ctx *ctx = req->ctx;
5900 struct io_defer_entry *de;
5904 /* Still need defer if there is pending req in defer list. */
5905 if (likely(list_empty_careful(&ctx->defer_list) &&
5906 !(req->flags & REQ_F_IO_DRAIN)))
5909 seq = io_get_sequence(req);
5910 /* Still a chance to pass the sequence check */
5911 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5914 ret = io_req_prep_async(req);
5917 io_prep_async_link(req);
5918 de = kmalloc(sizeof(*de), GFP_KERNEL);
5922 spin_lock_irq(&ctx->completion_lock);
5923 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5924 spin_unlock_irq(&ctx->completion_lock);
5926 io_queue_async_work(req);
5927 return -EIOCBQUEUED;
5930 trace_io_uring_defer(ctx, req, req->user_data);
5933 list_add_tail(&de->list, &ctx->defer_list);
5934 spin_unlock_irq(&ctx->completion_lock);
5935 return -EIOCBQUEUED;
5938 static void io_clean_op(struct io_kiocb *req)
5940 if (req->flags & REQ_F_BUFFER_SELECTED) {
5941 switch (req->opcode) {
5942 case IORING_OP_READV:
5943 case IORING_OP_READ_FIXED:
5944 case IORING_OP_READ:
5945 kfree((void *)(unsigned long)req->rw.addr);
5947 case IORING_OP_RECVMSG:
5948 case IORING_OP_RECV:
5949 kfree(req->sr_msg.kbuf);
5952 req->flags &= ~REQ_F_BUFFER_SELECTED;
5955 if (req->flags & REQ_F_NEED_CLEANUP) {
5956 switch (req->opcode) {
5957 case IORING_OP_READV:
5958 case IORING_OP_READ_FIXED:
5959 case IORING_OP_READ:
5960 case IORING_OP_WRITEV:
5961 case IORING_OP_WRITE_FIXED:
5962 case IORING_OP_WRITE: {
5963 struct io_async_rw *io = req->async_data;
5965 kfree(io->free_iovec);
5968 case IORING_OP_RECVMSG:
5969 case IORING_OP_SENDMSG: {
5970 struct io_async_msghdr *io = req->async_data;
5972 kfree(io->free_iov);
5975 case IORING_OP_SPLICE:
5977 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
5978 io_put_file(req->splice.file_in);
5980 case IORING_OP_OPENAT:
5981 case IORING_OP_OPENAT2:
5982 if (req->open.filename)
5983 putname(req->open.filename);
5985 case IORING_OP_RENAMEAT:
5986 putname(req->rename.oldpath);
5987 putname(req->rename.newpath);
5989 case IORING_OP_UNLINKAT:
5990 putname(req->unlink.filename);
5993 req->flags &= ~REQ_F_NEED_CLEANUP;
5997 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5999 struct io_ring_ctx *ctx = req->ctx;
6000 const struct cred *creds = NULL;
6003 if (req->work.creds && req->work.creds != current_cred())
6004 creds = override_creds(req->work.creds);
6006 switch (req->opcode) {
6008 ret = io_nop(req, issue_flags);
6010 case IORING_OP_READV:
6011 case IORING_OP_READ_FIXED:
6012 case IORING_OP_READ:
6013 ret = io_read(req, issue_flags);
6015 case IORING_OP_WRITEV:
6016 case IORING_OP_WRITE_FIXED:
6017 case IORING_OP_WRITE:
6018 ret = io_write(req, issue_flags);
6020 case IORING_OP_FSYNC:
6021 ret = io_fsync(req, issue_flags);
6023 case IORING_OP_POLL_ADD:
6024 ret = io_poll_add(req, issue_flags);
6026 case IORING_OP_POLL_REMOVE:
6027 ret = io_poll_remove(req, issue_flags);
6029 case IORING_OP_SYNC_FILE_RANGE:
6030 ret = io_sync_file_range(req, issue_flags);
6032 case IORING_OP_SENDMSG:
6033 ret = io_sendmsg(req, issue_flags);
6035 case IORING_OP_SEND:
6036 ret = io_send(req, issue_flags);
6038 case IORING_OP_RECVMSG:
6039 ret = io_recvmsg(req, issue_flags);
6041 case IORING_OP_RECV:
6042 ret = io_recv(req, issue_flags);
6044 case IORING_OP_TIMEOUT:
6045 ret = io_timeout(req, issue_flags);
6047 case IORING_OP_TIMEOUT_REMOVE:
6048 ret = io_timeout_remove(req, issue_flags);
6050 case IORING_OP_ACCEPT:
6051 ret = io_accept(req, issue_flags);
6053 case IORING_OP_CONNECT:
6054 ret = io_connect(req, issue_flags);
6056 case IORING_OP_ASYNC_CANCEL:
6057 ret = io_async_cancel(req, issue_flags);
6059 case IORING_OP_FALLOCATE:
6060 ret = io_fallocate(req, issue_flags);
6062 case IORING_OP_OPENAT:
6063 ret = io_openat(req, issue_flags);
6065 case IORING_OP_CLOSE:
6066 ret = io_close(req, issue_flags);
6068 case IORING_OP_FILES_UPDATE:
6069 ret = io_files_update(req, issue_flags);
6071 case IORING_OP_STATX:
6072 ret = io_statx(req, issue_flags);
6074 case IORING_OP_FADVISE:
6075 ret = io_fadvise(req, issue_flags);
6077 case IORING_OP_MADVISE:
6078 ret = io_madvise(req, issue_flags);
6080 case IORING_OP_OPENAT2:
6081 ret = io_openat2(req, issue_flags);
6083 case IORING_OP_EPOLL_CTL:
6084 ret = io_epoll_ctl(req, issue_flags);
6086 case IORING_OP_SPLICE:
6087 ret = io_splice(req, issue_flags);
6089 case IORING_OP_PROVIDE_BUFFERS:
6090 ret = io_provide_buffers(req, issue_flags);
6092 case IORING_OP_REMOVE_BUFFERS:
6093 ret = io_remove_buffers(req, issue_flags);
6096 ret = io_tee(req, issue_flags);
6098 case IORING_OP_SHUTDOWN:
6099 ret = io_shutdown(req, issue_flags);
6101 case IORING_OP_RENAMEAT:
6102 ret = io_renameat(req, issue_flags);
6104 case IORING_OP_UNLINKAT:
6105 ret = io_unlinkat(req, issue_flags);
6113 revert_creds(creds);
6118 /* If the op doesn't have a file, we're not polling for it */
6119 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6120 const bool in_async = io_wq_current_is_worker();
6122 /* workqueue context doesn't hold uring_lock, grab it now */
6124 mutex_lock(&ctx->uring_lock);
6126 io_iopoll_req_issued(req, in_async);
6129 mutex_unlock(&ctx->uring_lock);
6135 static void io_wq_submit_work(struct io_wq_work *work)
6137 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6138 struct io_kiocb *timeout;
6141 timeout = io_prep_linked_timeout(req);
6143 io_queue_linked_timeout(timeout);
6145 if (work->flags & IO_WQ_WORK_CANCEL)
6150 ret = io_issue_sqe(req, 0);
6152 * We can get EAGAIN for polled IO even though we're
6153 * forcing a sync submission from here, since we can't
6154 * wait for request slots on the block side.
6162 /* avoid locking problems by failing it from a clean context */
6164 /* io-wq is going to take one down */
6166 io_req_task_queue_fail(req, ret);
6170 #define FFS_ASYNC_READ 0x1UL
6171 #define FFS_ASYNC_WRITE 0x2UL
6173 #define FFS_ISREG 0x4UL
6175 #define FFS_ISREG 0x0UL
6177 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6179 static inline struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
6182 struct fixed_rsrc_table *table;
6184 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6185 return &table->files[i & IORING_FILE_TABLE_MASK];
6188 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6191 struct file **file_slot = io_fixed_file_slot(ctx->file_data, index);
6193 return (struct file *) ((unsigned long) *file_slot & FFS_MASK);
6196 static struct file *io_file_get(struct io_submit_state *state,
6197 struct io_kiocb *req, int fd, bool fixed)
6199 struct io_ring_ctx *ctx = req->ctx;
6203 unsigned long file_ptr;
6205 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6207 fd = array_index_nospec(fd, ctx->nr_user_files);
6208 file_ptr = (unsigned long) *io_fixed_file_slot(ctx->file_data, fd);
6209 file = (struct file *) (file_ptr & FFS_MASK);
6210 file_ptr &= ~FFS_MASK;
6211 /* mask in overlapping REQ_F and FFS bits */
6212 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6213 io_set_resource_node(req);
6215 trace_io_uring_file_get(ctx, fd);
6216 file = __io_file_get(state, fd);
6218 /* we don't allow fixed io_uring files */
6219 if (file && unlikely(file->f_op == &io_uring_fops))
6220 io_req_track_inflight(req);
6226 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6228 struct io_timeout_data *data = container_of(timer,
6229 struct io_timeout_data, timer);
6230 struct io_kiocb *prev, *req = data->req;
6231 struct io_ring_ctx *ctx = req->ctx;
6232 unsigned long flags;
6234 spin_lock_irqsave(&ctx->completion_lock, flags);
6235 prev = req->timeout.head;
6236 req->timeout.head = NULL;
6239 * We don't expect the list to be empty, that will only happen if we
6240 * race with the completion of the linked work.
6242 if (prev && req_ref_inc_not_zero(prev))
6243 io_remove_next_linked(prev);
6246 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6249 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6250 io_put_req_deferred(prev, 1);
6252 io_req_complete_post(req, -ETIME, 0);
6253 io_put_req_deferred(req, 1);
6255 return HRTIMER_NORESTART;
6258 static void io_queue_linked_timeout(struct io_kiocb *req)
6260 struct io_ring_ctx *ctx = req->ctx;
6262 spin_lock_irq(&ctx->completion_lock);
6264 * If the back reference is NULL, then our linked request finished
6265 * before we got a chance to setup the timer
6267 if (req->timeout.head) {
6268 struct io_timeout_data *data = req->async_data;
6270 data->timer.function = io_link_timeout_fn;
6271 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6274 spin_unlock_irq(&ctx->completion_lock);
6275 /* drop submission reference */
6279 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6281 struct io_kiocb *nxt = req->link;
6283 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6284 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6287 nxt->timeout.head = req;
6288 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6289 req->flags |= REQ_F_LINK_TIMEOUT;
6293 static void __io_queue_sqe(struct io_kiocb *req)
6295 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6298 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6301 * We async punt it if the file wasn't marked NOWAIT, or if the file
6302 * doesn't support non-blocking read/write attempts
6305 /* drop submission reference */
6306 if (req->flags & REQ_F_COMPLETE_INLINE) {
6307 struct io_ring_ctx *ctx = req->ctx;
6308 struct io_comp_state *cs = &ctx->submit_state.comp;
6310 cs->reqs[cs->nr++] = req;
6311 if (cs->nr == ARRAY_SIZE(cs->reqs))
6312 io_submit_flush_completions(cs, ctx);
6316 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6317 if (!io_arm_poll_handler(req)) {
6319 * Queued up for async execution, worker will release
6320 * submit reference when the iocb is actually submitted.
6322 io_queue_async_work(req);
6325 io_req_complete_failed(req, ret);
6328 io_queue_linked_timeout(linked_timeout);
6331 static void io_queue_sqe(struct io_kiocb *req)
6335 ret = io_req_defer(req);
6337 if (ret != -EIOCBQUEUED) {
6339 io_req_complete_failed(req, ret);
6341 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6342 ret = io_req_prep_async(req);
6345 io_queue_async_work(req);
6347 __io_queue_sqe(req);
6352 * Check SQE restrictions (opcode and flags).
6354 * Returns 'true' if SQE is allowed, 'false' otherwise.
6356 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6357 struct io_kiocb *req,
6358 unsigned int sqe_flags)
6360 if (!ctx->restricted)
6363 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6366 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6367 ctx->restrictions.sqe_flags_required)
6370 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6371 ctx->restrictions.sqe_flags_required))
6377 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6378 const struct io_uring_sqe *sqe)
6380 struct io_submit_state *state;
6381 unsigned int sqe_flags;
6382 int personality, ret = 0;
6384 req->opcode = READ_ONCE(sqe->opcode);
6385 /* same numerical values with corresponding REQ_F_*, safe to copy */
6386 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6387 req->user_data = READ_ONCE(sqe->user_data);
6388 req->async_data = NULL;
6392 req->fixed_rsrc_refs = NULL;
6393 /* one is dropped after submission, the other at completion */
6394 atomic_set(&req->refs, 2);
6395 req->task = current;
6397 req->work.list.next = NULL;
6398 req->work.creds = NULL;
6399 req->work.flags = 0;
6401 /* enforce forwards compatibility on users */
6402 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6407 if (unlikely(req->opcode >= IORING_OP_LAST))
6410 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6413 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6414 !io_op_defs[req->opcode].buffer_select)
6417 personality = READ_ONCE(sqe->personality);
6419 req->work.creds = xa_load(&ctx->personalities, personality);
6420 if (!req->work.creds)
6422 get_cred(req->work.creds);
6424 state = &ctx->submit_state;
6427 * Plug now if we have more than 1 IO left after this, and the target
6428 * is potentially a read/write to block based storage.
6430 if (!state->plug_started && state->ios_left > 1 &&
6431 io_op_defs[req->opcode].plug) {
6432 blk_start_plug(&state->plug);
6433 state->plug_started = true;
6436 if (io_op_defs[req->opcode].needs_file) {
6437 bool fixed = req->flags & REQ_F_FIXED_FILE;
6439 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6440 if (unlikely(!req->file))
6448 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6449 const struct io_uring_sqe *sqe)
6451 struct io_submit_link *link = &ctx->submit_state.link;
6454 ret = io_init_req(ctx, req, sqe);
6455 if (unlikely(ret)) {
6458 /* fail even hard links since we don't submit */
6459 link->head->flags |= REQ_F_FAIL_LINK;
6460 io_req_complete_failed(link->head, -ECANCELED);
6463 io_req_complete_failed(req, ret);
6466 ret = io_req_prep(req, sqe);
6470 /* don't need @sqe from now on */
6471 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6472 true, ctx->flags & IORING_SETUP_SQPOLL);
6475 * If we already have a head request, queue this one for async
6476 * submittal once the head completes. If we don't have a head but
6477 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6478 * submitted sync once the chain is complete. If none of those
6479 * conditions are true (normal request), then just queue it.
6482 struct io_kiocb *head = link->head;
6485 * Taking sequential execution of a link, draining both sides
6486 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6487 * requests in the link. So, it drains the head and the
6488 * next after the link request. The last one is done via
6489 * drain_next flag to persist the effect across calls.
6491 if (req->flags & REQ_F_IO_DRAIN) {
6492 head->flags |= REQ_F_IO_DRAIN;
6493 ctx->drain_next = 1;
6495 ret = io_req_prep_async(req);
6498 trace_io_uring_link(ctx, req, head);
6499 link->last->link = req;
6502 /* last request of a link, enqueue the link */
6503 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6508 if (unlikely(ctx->drain_next)) {
6509 req->flags |= REQ_F_IO_DRAIN;
6510 ctx->drain_next = 0;
6512 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6524 * Batched submission is done, ensure local IO is flushed out.
6526 static void io_submit_state_end(struct io_submit_state *state,
6527 struct io_ring_ctx *ctx)
6529 if (state->link.head)
6530 io_queue_sqe(state->link.head);
6532 io_submit_flush_completions(&state->comp, ctx);
6533 if (state->plug_started)
6534 blk_finish_plug(&state->plug);
6535 io_state_file_put(state);
6539 * Start submission side cache.
6541 static void io_submit_state_start(struct io_submit_state *state,
6542 unsigned int max_ios)
6544 state->plug_started = false;
6545 state->ios_left = max_ios;
6546 /* set only head, no need to init link_last in advance */
6547 state->link.head = NULL;
6550 static void io_commit_sqring(struct io_ring_ctx *ctx)
6552 struct io_rings *rings = ctx->rings;
6555 * Ensure any loads from the SQEs are done at this point,
6556 * since once we write the new head, the application could
6557 * write new data to them.
6559 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6563 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6564 * that is mapped by userspace. This means that care needs to be taken to
6565 * ensure that reads are stable, as we cannot rely on userspace always
6566 * being a good citizen. If members of the sqe are validated and then later
6567 * used, it's important that those reads are done through READ_ONCE() to
6568 * prevent a re-load down the line.
6570 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6572 u32 *sq_array = ctx->sq_array;
6576 * The cached sq head (or cq tail) serves two purposes:
6578 * 1) allows us to batch the cost of updating the user visible
6580 * 2) allows the kernel side to track the head on its own, even
6581 * though the application is the one updating it.
6583 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6584 if (likely(head < ctx->sq_entries))
6585 return &ctx->sq_sqes[head];
6587 /* drop invalid entries */
6588 ctx->cached_sq_dropped++;
6589 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6593 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6597 /* if we have a backlog and couldn't flush it all, return BUSY */
6598 if (test_bit(0, &ctx->sq_check_overflow)) {
6599 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6603 /* make sure SQ entry isn't read before tail */
6604 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6606 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6609 percpu_counter_add(¤t->io_uring->inflight, nr);
6610 refcount_add(nr, ¤t->usage);
6611 io_submit_state_start(&ctx->submit_state, nr);
6613 while (submitted < nr) {
6614 const struct io_uring_sqe *sqe;
6615 struct io_kiocb *req;
6617 req = io_alloc_req(ctx);
6618 if (unlikely(!req)) {
6620 submitted = -EAGAIN;
6623 sqe = io_get_sqe(ctx);
6624 if (unlikely(!sqe)) {
6625 kmem_cache_free(req_cachep, req);
6628 /* will complete beyond this point, count as submitted */
6630 if (io_submit_sqe(ctx, req, sqe))
6634 if (unlikely(submitted != nr)) {
6635 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6636 struct io_uring_task *tctx = current->io_uring;
6637 int unused = nr - ref_used;
6639 percpu_ref_put_many(&ctx->refs, unused);
6640 percpu_counter_sub(&tctx->inflight, unused);
6641 put_task_struct_many(current, unused);
6644 io_submit_state_end(&ctx->submit_state, ctx);
6645 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6646 io_commit_sqring(ctx);
6651 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6653 /* Tell userspace we may need a wakeup call */
6654 spin_lock_irq(&ctx->completion_lock);
6655 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6656 spin_unlock_irq(&ctx->completion_lock);
6659 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6661 spin_lock_irq(&ctx->completion_lock);
6662 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6663 spin_unlock_irq(&ctx->completion_lock);
6666 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6668 unsigned int to_submit;
6671 to_submit = io_sqring_entries(ctx);
6672 /* if we're handling multiple rings, cap submit size for fairness */
6673 if (cap_entries && to_submit > 8)
6676 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6677 unsigned nr_events = 0;
6679 mutex_lock(&ctx->uring_lock);
6680 if (!list_empty(&ctx->iopoll_list))
6681 io_do_iopoll(ctx, &nr_events, 0);
6683 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6684 !(ctx->flags & IORING_SETUP_R_DISABLED))
6685 ret = io_submit_sqes(ctx, to_submit);
6686 mutex_unlock(&ctx->uring_lock);
6689 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6690 wake_up(&ctx->sqo_sq_wait);
6695 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6697 struct io_ring_ctx *ctx;
6698 unsigned sq_thread_idle = 0;
6700 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6701 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6702 sqd->sq_thread_idle = sq_thread_idle;
6705 static int io_sq_thread(void *data)
6707 struct io_sq_data *sqd = data;
6708 struct io_ring_ctx *ctx;
6709 unsigned long timeout = 0;
6710 char buf[TASK_COMM_LEN];
6713 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6714 set_task_comm(current, buf);
6715 current->pf_io_worker = NULL;
6717 if (sqd->sq_cpu != -1)
6718 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6720 set_cpus_allowed_ptr(current, cpu_online_mask);
6721 current->flags |= PF_NO_SETAFFINITY;
6723 mutex_lock(&sqd->lock);
6724 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6726 bool cap_entries, sqt_spin, needs_sched;
6728 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6729 signal_pending(current)) {
6730 bool did_sig = false;
6732 mutex_unlock(&sqd->lock);
6733 if (signal_pending(current)) {
6734 struct ksignal ksig;
6736 did_sig = get_signal(&ksig);
6739 mutex_lock(&sqd->lock);
6743 io_run_task_work_head(&sqd->park_task_work);
6744 timeout = jiffies + sqd->sq_thread_idle;
6748 cap_entries = !list_is_singular(&sqd->ctx_list);
6749 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6750 const struct cred *creds = NULL;
6752 if (ctx->sq_creds != current_cred())
6753 creds = override_creds(ctx->sq_creds);
6754 ret = __io_sq_thread(ctx, cap_entries);
6756 revert_creds(creds);
6757 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6761 if (sqt_spin || !time_after(jiffies, timeout)) {
6765 timeout = jiffies + sqd->sq_thread_idle;
6770 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6771 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6772 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6773 !list_empty_careful(&ctx->iopoll_list)) {
6774 needs_sched = false;
6777 if (io_sqring_entries(ctx)) {
6778 needs_sched = false;
6783 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6784 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6785 io_ring_set_wakeup_flag(ctx);
6787 mutex_unlock(&sqd->lock);
6789 mutex_lock(&sqd->lock);
6790 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6791 io_ring_clear_wakeup_flag(ctx);
6794 finish_wait(&sqd->wait, &wait);
6795 io_run_task_work_head(&sqd->park_task_work);
6796 timeout = jiffies + sqd->sq_thread_idle;
6799 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6800 io_uring_cancel_sqpoll(ctx);
6802 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6803 io_ring_set_wakeup_flag(ctx);
6804 mutex_unlock(&sqd->lock);
6807 io_run_task_work_head(&sqd->park_task_work);
6808 complete(&sqd->exited);
6812 struct io_wait_queue {
6813 struct wait_queue_entry wq;
6814 struct io_ring_ctx *ctx;
6816 unsigned nr_timeouts;
6819 static inline bool io_should_wake(struct io_wait_queue *iowq)
6821 struct io_ring_ctx *ctx = iowq->ctx;
6824 * Wake up if we have enough events, or if a timeout occurred since we
6825 * started waiting. For timeouts, we always want to return to userspace,
6826 * regardless of event count.
6828 return io_cqring_events(ctx) >= iowq->to_wait ||
6829 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6832 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6833 int wake_flags, void *key)
6835 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6839 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6840 * the task, and the next invocation will do it.
6842 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6843 return autoremove_wake_function(curr, mode, wake_flags, key);
6847 static int io_run_task_work_sig(void)
6849 if (io_run_task_work())
6851 if (!signal_pending(current))
6853 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6854 return -ERESTARTSYS;
6858 /* when returns >0, the caller should retry */
6859 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6860 struct io_wait_queue *iowq,
6861 signed long *timeout)
6865 /* make sure we run task_work before checking for signals */
6866 ret = io_run_task_work_sig();
6867 if (ret || io_should_wake(iowq))
6869 /* let the caller flush overflows, retry */
6870 if (test_bit(0, &ctx->cq_check_overflow))
6873 *timeout = schedule_timeout(*timeout);
6874 return !*timeout ? -ETIME : 1;
6878 * Wait until events become available, if we don't already have some. The
6879 * application must reap them itself, as they reside on the shared cq ring.
6881 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6882 const sigset_t __user *sig, size_t sigsz,
6883 struct __kernel_timespec __user *uts)
6885 struct io_wait_queue iowq = {
6888 .func = io_wake_function,
6889 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6892 .to_wait = min_events,
6894 struct io_rings *rings = ctx->rings;
6895 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6899 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6900 if (io_cqring_events(ctx) >= min_events)
6902 if (!io_run_task_work())
6907 #ifdef CONFIG_COMPAT
6908 if (in_compat_syscall())
6909 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6913 ret = set_user_sigmask(sig, sigsz);
6920 struct timespec64 ts;
6922 if (get_timespec64(&ts, uts))
6924 timeout = timespec64_to_jiffies(&ts);
6927 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6928 trace_io_uring_cqring_wait(ctx, min_events);
6930 /* if we can't even flush overflow, don't wait for more */
6931 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6935 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6936 TASK_INTERRUPTIBLE);
6937 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6938 finish_wait(&ctx->wait, &iowq.wq);
6942 restore_saved_sigmask_unless(ret == -EINTR);
6944 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6947 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6949 #if defined(CONFIG_UNIX)
6950 if (ctx->ring_sock) {
6951 struct sock *sock = ctx->ring_sock->sk;
6952 struct sk_buff *skb;
6954 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6960 for (i = 0; i < ctx->nr_user_files; i++) {
6963 file = io_file_from_index(ctx, i);
6970 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6972 struct fixed_rsrc_data *data;
6974 data = container_of(ref, struct fixed_rsrc_data, refs);
6975 complete(&data->done);
6978 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6980 spin_lock_bh(&ctx->rsrc_ref_lock);
6983 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6985 spin_unlock_bh(&ctx->rsrc_ref_lock);
6988 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6989 struct fixed_rsrc_data *rsrc_data,
6990 struct fixed_rsrc_ref_node *ref_node)
6992 io_rsrc_ref_lock(ctx);
6993 rsrc_data->node = ref_node;
6994 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6995 io_rsrc_ref_unlock(ctx);
6996 percpu_ref_get(&rsrc_data->refs);
6999 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7001 struct fixed_rsrc_ref_node *ref_node = NULL;
7003 io_rsrc_ref_lock(ctx);
7004 ref_node = data->node;
7006 io_rsrc_ref_unlock(ctx);
7008 percpu_ref_kill(&ref_node->refs);
7011 static int io_rsrc_refnode_prealloc(struct io_ring_ctx *ctx)
7013 if (ctx->rsrc_backup_node)
7015 ctx->rsrc_backup_node = alloc_fixed_rsrc_ref_node(ctx);
7016 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7019 static struct fixed_rsrc_ref_node *
7020 io_rsrc_refnode_get(struct io_ring_ctx *ctx,
7021 struct fixed_rsrc_data *rsrc_data,
7022 void (*rsrc_put)(struct io_ring_ctx *ctx,
7023 struct io_rsrc_put *prsrc))
7025 struct fixed_rsrc_ref_node *node = ctx->rsrc_backup_node;
7027 WARN_ON_ONCE(!node);
7029 ctx->rsrc_backup_node = NULL;
7030 node->rsrc_data = rsrc_data;
7031 node->rsrc_put = rsrc_put;
7035 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7036 struct io_ring_ctx *ctx,
7037 void (*rsrc_put)(struct io_ring_ctx *ctx,
7038 struct io_rsrc_put *prsrc))
7040 struct fixed_rsrc_ref_node *node;
7046 data->quiesce = true;
7048 ret = io_rsrc_refnode_prealloc(ctx);
7051 io_sqe_rsrc_kill_node(ctx, data);
7052 percpu_ref_kill(&data->refs);
7053 flush_delayed_work(&ctx->rsrc_put_work);
7055 ret = wait_for_completion_interruptible(&data->done);
7059 percpu_ref_resurrect(&data->refs);
7060 node = io_rsrc_refnode_get(ctx, data, rsrc_put);
7061 io_sqe_rsrc_set_node(ctx, data, node);
7062 reinit_completion(&data->done);
7064 mutex_unlock(&ctx->uring_lock);
7065 ret = io_run_task_work_sig();
7066 mutex_lock(&ctx->uring_lock);
7068 data->quiesce = false;
7073 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7075 struct fixed_rsrc_data *data;
7077 data = kzalloc(sizeof(*data), GFP_KERNEL);
7081 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7082 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7087 init_completion(&data->done);
7091 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7093 percpu_ref_exit(&data->refs);
7098 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7100 struct fixed_rsrc_data *data = ctx->file_data;
7101 unsigned nr_tables, i;
7105 * percpu_ref_is_dying() is to stop parallel files unregister
7106 * Since we possibly drop uring lock later in this function to
7109 if (!data || percpu_ref_is_dying(&data->refs))
7111 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7115 __io_sqe_files_unregister(ctx);
7116 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7117 for (i = 0; i < nr_tables; i++)
7118 kfree(data->table[i].files);
7119 free_fixed_rsrc_data(data);
7120 ctx->file_data = NULL;
7121 ctx->nr_user_files = 0;
7125 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7126 __releases(&sqd->lock)
7128 WARN_ON_ONCE(sqd->thread == current);
7131 * Do the dance but not conditional clear_bit() because it'd race with
7132 * other threads incrementing park_pending and setting the bit.
7134 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7135 if (atomic_dec_return(&sqd->park_pending))
7136 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7137 mutex_unlock(&sqd->lock);
7140 static void io_sq_thread_park(struct io_sq_data *sqd)
7141 __acquires(&sqd->lock)
7143 WARN_ON_ONCE(sqd->thread == current);
7145 atomic_inc(&sqd->park_pending);
7146 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7147 mutex_lock(&sqd->lock);
7149 wake_up_process(sqd->thread);
7152 static void io_sq_thread_stop(struct io_sq_data *sqd)
7154 WARN_ON_ONCE(sqd->thread == current);
7156 mutex_lock(&sqd->lock);
7157 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7159 wake_up_process(sqd->thread);
7160 mutex_unlock(&sqd->lock);
7161 wait_for_completion(&sqd->exited);
7164 static void io_put_sq_data(struct io_sq_data *sqd)
7166 if (refcount_dec_and_test(&sqd->refs)) {
7167 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7169 io_sq_thread_stop(sqd);
7174 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7176 struct io_sq_data *sqd = ctx->sq_data;
7179 io_sq_thread_park(sqd);
7180 list_del_init(&ctx->sqd_list);
7181 io_sqd_update_thread_idle(sqd);
7182 io_sq_thread_unpark(sqd);
7184 io_put_sq_data(sqd);
7185 ctx->sq_data = NULL;
7187 put_cred(ctx->sq_creds);
7191 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7193 struct io_ring_ctx *ctx_attach;
7194 struct io_sq_data *sqd;
7197 f = fdget(p->wq_fd);
7199 return ERR_PTR(-ENXIO);
7200 if (f.file->f_op != &io_uring_fops) {
7202 return ERR_PTR(-EINVAL);
7205 ctx_attach = f.file->private_data;
7206 sqd = ctx_attach->sq_data;
7209 return ERR_PTR(-EINVAL);
7211 if (sqd->task_tgid != current->tgid) {
7213 return ERR_PTR(-EPERM);
7216 refcount_inc(&sqd->refs);
7221 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7224 struct io_sq_data *sqd;
7227 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7228 sqd = io_attach_sq_data(p);
7233 /* fall through for EPERM case, setup new sqd/task */
7234 if (PTR_ERR(sqd) != -EPERM)
7238 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7240 return ERR_PTR(-ENOMEM);
7242 atomic_set(&sqd->park_pending, 0);
7243 refcount_set(&sqd->refs, 1);
7244 INIT_LIST_HEAD(&sqd->ctx_list);
7245 mutex_init(&sqd->lock);
7246 init_waitqueue_head(&sqd->wait);
7247 init_completion(&sqd->exited);
7251 #if defined(CONFIG_UNIX)
7253 * Ensure the UNIX gc is aware of our file set, so we are certain that
7254 * the io_uring can be safely unregistered on process exit, even if we have
7255 * loops in the file referencing.
7257 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7259 struct sock *sk = ctx->ring_sock->sk;
7260 struct scm_fp_list *fpl;
7261 struct sk_buff *skb;
7264 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7268 skb = alloc_skb(0, GFP_KERNEL);
7277 fpl->user = get_uid(current_user());
7278 for (i = 0; i < nr; i++) {
7279 struct file *file = io_file_from_index(ctx, i + offset);
7283 fpl->fp[nr_files] = get_file(file);
7284 unix_inflight(fpl->user, fpl->fp[nr_files]);
7289 fpl->max = SCM_MAX_FD;
7290 fpl->count = nr_files;
7291 UNIXCB(skb).fp = fpl;
7292 skb->destructor = unix_destruct_scm;
7293 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7294 skb_queue_head(&sk->sk_receive_queue, skb);
7296 for (i = 0; i < nr_files; i++)
7307 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7308 * causes regular reference counting to break down. We rely on the UNIX
7309 * garbage collection to take care of this problem for us.
7311 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7313 unsigned left, total;
7317 left = ctx->nr_user_files;
7319 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7321 ret = __io_sqe_files_scm(ctx, this_files, total);
7325 total += this_files;
7331 while (total < ctx->nr_user_files) {
7332 struct file *file = io_file_from_index(ctx, total);
7342 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7348 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7349 unsigned nr_tables, unsigned nr_files)
7353 for (i = 0; i < nr_tables; i++) {
7354 struct fixed_rsrc_table *table = &file_data->table[i];
7355 unsigned this_files;
7357 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7358 table->files = kcalloc(this_files, sizeof(struct file *),
7362 nr_files -= this_files;
7368 for (i = 0; i < nr_tables; i++) {
7369 struct fixed_rsrc_table *table = &file_data->table[i];
7370 kfree(table->files);
7375 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7377 struct file *file = prsrc->file;
7378 #if defined(CONFIG_UNIX)
7379 struct sock *sock = ctx->ring_sock->sk;
7380 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7381 struct sk_buff *skb;
7384 __skb_queue_head_init(&list);
7387 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7388 * remove this entry and rearrange the file array.
7390 skb = skb_dequeue(head);
7392 struct scm_fp_list *fp;
7394 fp = UNIXCB(skb).fp;
7395 for (i = 0; i < fp->count; i++) {
7398 if (fp->fp[i] != file)
7401 unix_notinflight(fp->user, fp->fp[i]);
7402 left = fp->count - 1 - i;
7404 memmove(&fp->fp[i], &fp->fp[i + 1],
7405 left * sizeof(struct file *));
7412 __skb_queue_tail(&list, skb);
7422 __skb_queue_tail(&list, skb);
7424 skb = skb_dequeue(head);
7427 if (skb_peek(&list)) {
7428 spin_lock_irq(&head->lock);
7429 while ((skb = __skb_dequeue(&list)) != NULL)
7430 __skb_queue_tail(head, skb);
7431 spin_unlock_irq(&head->lock);
7438 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7440 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7441 struct io_ring_ctx *ctx = rsrc_data->ctx;
7442 struct io_rsrc_put *prsrc, *tmp;
7444 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7445 list_del(&prsrc->list);
7446 ref_node->rsrc_put(ctx, prsrc);
7450 percpu_ref_exit(&ref_node->refs);
7452 percpu_ref_put(&rsrc_data->refs);
7455 static void io_rsrc_put_work(struct work_struct *work)
7457 struct io_ring_ctx *ctx;
7458 struct llist_node *node;
7460 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7461 node = llist_del_all(&ctx->rsrc_put_llist);
7464 struct fixed_rsrc_ref_node *ref_node;
7465 struct llist_node *next = node->next;
7467 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7468 __io_rsrc_put_work(ref_node);
7473 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7475 struct fixed_rsrc_ref_node *ref_node;
7476 struct fixed_rsrc_data *data;
7477 struct io_ring_ctx *ctx;
7478 bool first_add = false;
7481 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7482 data = ref_node->rsrc_data;
7485 io_rsrc_ref_lock(ctx);
7486 ref_node->done = true;
7488 while (!list_empty(&ctx->rsrc_ref_list)) {
7489 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7490 struct fixed_rsrc_ref_node, node);
7491 /* recycle ref nodes in order */
7492 if (!ref_node->done)
7494 list_del(&ref_node->node);
7495 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7497 io_rsrc_ref_unlock(ctx);
7499 if (percpu_ref_is_dying(&data->refs))
7503 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7505 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7508 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7509 struct io_ring_ctx *ctx)
7511 struct fixed_rsrc_ref_node *ref_node;
7513 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7517 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7522 INIT_LIST_HEAD(&ref_node->node);
7523 INIT_LIST_HEAD(&ref_node->rsrc_list);
7524 ref_node->done = false;
7528 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7529 struct fixed_rsrc_ref_node *ref_node)
7531 ref_node->rsrc_data = ctx->file_data;
7532 ref_node->rsrc_put = io_ring_file_put;
7535 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7537 percpu_ref_exit(&ref_node->refs);
7542 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7545 __s32 __user *fds = (__s32 __user *) arg;
7546 unsigned nr_tables, i;
7548 int fd, ret = -ENOMEM;
7549 struct fixed_rsrc_ref_node *ref_node;
7550 struct fixed_rsrc_data *file_data;
7556 if (nr_args > IORING_MAX_FIXED_FILES)
7559 file_data = alloc_fixed_rsrc_data(ctx);
7562 ctx->file_data = file_data;
7564 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7565 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7567 if (!file_data->table)
7570 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7573 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7574 unsigned long file_ptr;
7576 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7580 /* allow sparse sets */
7590 * Don't allow io_uring instances to be registered. If UNIX
7591 * isn't enabled, then this causes a reference cycle and this
7592 * instance can never get freed. If UNIX is enabled we'll
7593 * handle it just fine, but there's still no point in allowing
7594 * a ring fd as it doesn't support regular read/write anyway.
7596 if (file->f_op == &io_uring_fops) {
7600 file_ptr = (unsigned long) file;
7601 if (__io_file_supports_async(file, READ))
7602 file_ptr |= FFS_ASYNC_READ;
7603 if (__io_file_supports_async(file, WRITE))
7604 file_ptr |= FFS_ASYNC_WRITE;
7605 if (S_ISREG(file_inode(file)->i_mode))
7606 file_ptr |= FFS_ISREG;
7607 *io_fixed_file_slot(file_data, i) = (struct file *) file_ptr;
7610 ret = io_sqe_files_scm(ctx);
7612 io_sqe_files_unregister(ctx);
7616 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7618 io_sqe_files_unregister(ctx);
7621 init_fixed_file_ref_node(ctx, ref_node);
7623 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7626 for (i = 0; i < ctx->nr_user_files; i++) {
7627 file = io_file_from_index(ctx, i);
7631 for (i = 0; i < nr_tables; i++)
7632 kfree(file_data->table[i].files);
7633 ctx->nr_user_files = 0;
7635 free_fixed_rsrc_data(ctx->file_data);
7636 ctx->file_data = NULL;
7640 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7643 #if defined(CONFIG_UNIX)
7644 struct sock *sock = ctx->ring_sock->sk;
7645 struct sk_buff_head *head = &sock->sk_receive_queue;
7646 struct sk_buff *skb;
7649 * See if we can merge this file into an existing skb SCM_RIGHTS
7650 * file set. If there's no room, fall back to allocating a new skb
7651 * and filling it in.
7653 spin_lock_irq(&head->lock);
7654 skb = skb_peek(head);
7656 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7658 if (fpl->count < SCM_MAX_FD) {
7659 __skb_unlink(skb, head);
7660 spin_unlock_irq(&head->lock);
7661 fpl->fp[fpl->count] = get_file(file);
7662 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7664 spin_lock_irq(&head->lock);
7665 __skb_queue_head(head, skb);
7670 spin_unlock_irq(&head->lock);
7677 return __io_sqe_files_scm(ctx, 1, index);
7683 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7685 struct io_rsrc_put *prsrc;
7686 struct fixed_rsrc_ref_node *ref_node = data->node;
7688 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7693 list_add(&prsrc->list, &ref_node->rsrc_list);
7698 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7701 return io_queue_rsrc_removal(data, (void *)file);
7704 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7705 struct io_uring_rsrc_update *up,
7708 struct fixed_rsrc_data *data = ctx->file_data;
7709 struct fixed_rsrc_ref_node *ref_node;
7710 struct file *file, **file_slot;
7714 bool needs_switch = false;
7716 if (check_add_overflow(up->offset, nr_args, &done))
7718 if (done > ctx->nr_user_files)
7720 err = io_rsrc_refnode_prealloc(ctx);
7724 fds = u64_to_user_ptr(up->data);
7725 for (done = 0; done < nr_args; done++) {
7727 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7731 if (fd == IORING_REGISTER_FILES_SKIP)
7734 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7735 file_slot = io_fixed_file_slot(ctx->file_data, i);
7738 file = (struct file *) ((unsigned long) *file_slot & FFS_MASK);
7739 err = io_queue_file_removal(data, file);
7743 needs_switch = true;
7752 * Don't allow io_uring instances to be registered. If
7753 * UNIX isn't enabled, then this causes a reference
7754 * cycle and this instance can never get freed. If UNIX
7755 * is enabled we'll handle it just fine, but there's
7756 * still no point in allowing a ring fd as it doesn't
7757 * support regular read/write anyway.
7759 if (file->f_op == &io_uring_fops) {
7765 err = io_sqe_file_register(ctx, file, i);
7775 percpu_ref_kill(&data->node->refs);
7776 ref_node = io_rsrc_refnode_get(ctx, data, io_ring_file_put);
7777 io_sqe_rsrc_set_node(ctx, data, ref_node);
7779 return done ? done : err;
7782 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7785 struct io_uring_rsrc_update up;
7787 if (!ctx->file_data)
7791 if (copy_from_user(&up, arg, sizeof(up)))
7796 return __io_sqe_files_update(ctx, &up, nr_args);
7799 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7801 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7803 req = io_put_req_find_next(req);
7804 return req ? &req->work : NULL;
7807 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7809 struct io_wq_hash *hash;
7810 struct io_wq_data data;
7811 unsigned int concurrency;
7813 hash = ctx->hash_map;
7815 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7817 return ERR_PTR(-ENOMEM);
7818 refcount_set(&hash->refs, 1);
7819 init_waitqueue_head(&hash->wait);
7820 ctx->hash_map = hash;
7824 data.free_work = io_free_work;
7825 data.do_work = io_wq_submit_work;
7827 /* Do QD, or 4 * CPUS, whatever is smallest */
7828 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7830 return io_wq_create(concurrency, &data);
7833 static int io_uring_alloc_task_context(struct task_struct *task,
7834 struct io_ring_ctx *ctx)
7836 struct io_uring_task *tctx;
7839 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7840 if (unlikely(!tctx))
7843 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7844 if (unlikely(ret)) {
7849 tctx->io_wq = io_init_wq_offload(ctx);
7850 if (IS_ERR(tctx->io_wq)) {
7851 ret = PTR_ERR(tctx->io_wq);
7852 percpu_counter_destroy(&tctx->inflight);
7858 init_waitqueue_head(&tctx->wait);
7860 atomic_set(&tctx->in_idle, 0);
7861 task->io_uring = tctx;
7862 spin_lock_init(&tctx->task_lock);
7863 INIT_WQ_LIST(&tctx->task_list);
7864 tctx->task_state = 0;
7865 init_task_work(&tctx->task_work, tctx_task_work);
7869 void __io_uring_free(struct task_struct *tsk)
7871 struct io_uring_task *tctx = tsk->io_uring;
7873 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7874 WARN_ON_ONCE(tctx->io_wq);
7876 percpu_counter_destroy(&tctx->inflight);
7878 tsk->io_uring = NULL;
7881 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7882 struct io_uring_params *p)
7886 /* Retain compatibility with failing for an invalid attach attempt */
7887 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7888 IORING_SETUP_ATTACH_WQ) {
7891 f = fdget(p->wq_fd);
7894 if (f.file->f_op != &io_uring_fops) {
7900 if (ctx->flags & IORING_SETUP_SQPOLL) {
7901 struct task_struct *tsk;
7902 struct io_sq_data *sqd;
7906 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7909 sqd = io_get_sq_data(p, &attached);
7915 ctx->sq_creds = get_current_cred();
7917 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7918 if (!ctx->sq_thread_idle)
7919 ctx->sq_thread_idle = HZ;
7922 io_sq_thread_park(sqd);
7923 list_add(&ctx->sqd_list, &sqd->ctx_list);
7924 io_sqd_update_thread_idle(sqd);
7925 /* don't attach to a dying SQPOLL thread, would be racy */
7926 if (attached && !sqd->thread)
7928 io_sq_thread_unpark(sqd);
7935 if (p->flags & IORING_SETUP_SQ_AFF) {
7936 int cpu = p->sq_thread_cpu;
7939 if (cpu >= nr_cpu_ids)
7941 if (!cpu_online(cpu))
7949 sqd->task_pid = current->pid;
7950 sqd->task_tgid = current->tgid;
7951 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7958 ret = io_uring_alloc_task_context(tsk, ctx);
7959 wake_up_new_task(tsk);
7962 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7963 /* Can't have SQ_AFF without SQPOLL */
7970 io_sq_thread_finish(ctx);
7973 complete(&ctx->sq_data->exited);
7977 static inline void __io_unaccount_mem(struct user_struct *user,
7978 unsigned long nr_pages)
7980 atomic_long_sub(nr_pages, &user->locked_vm);
7983 static inline int __io_account_mem(struct user_struct *user,
7984 unsigned long nr_pages)
7986 unsigned long page_limit, cur_pages, new_pages;
7988 /* Don't allow more pages than we can safely lock */
7989 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7992 cur_pages = atomic_long_read(&user->locked_vm);
7993 new_pages = cur_pages + nr_pages;
7994 if (new_pages > page_limit)
7996 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7997 new_pages) != cur_pages);
8002 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8005 __io_unaccount_mem(ctx->user, nr_pages);
8007 if (ctx->mm_account)
8008 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8011 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8016 ret = __io_account_mem(ctx->user, nr_pages);
8021 if (ctx->mm_account)
8022 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8027 static void io_mem_free(void *ptr)
8034 page = virt_to_head_page(ptr);
8035 if (put_page_testzero(page))
8036 free_compound_page(page);
8039 static void *io_mem_alloc(size_t size)
8041 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8042 __GFP_NORETRY | __GFP_ACCOUNT;
8044 return (void *) __get_free_pages(gfp_flags, get_order(size));
8047 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8050 struct io_rings *rings;
8051 size_t off, sq_array_size;
8053 off = struct_size(rings, cqes, cq_entries);
8054 if (off == SIZE_MAX)
8058 off = ALIGN(off, SMP_CACHE_BYTES);
8066 sq_array_size = array_size(sizeof(u32), sq_entries);
8067 if (sq_array_size == SIZE_MAX)
8070 if (check_add_overflow(off, sq_array_size, &off))
8076 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8080 if (!ctx->user_bufs)
8083 for (i = 0; i < ctx->nr_user_bufs; i++) {
8084 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8086 for (j = 0; j < imu->nr_bvecs; j++)
8087 unpin_user_page(imu->bvec[j].bv_page);
8089 if (imu->acct_pages)
8090 io_unaccount_mem(ctx, imu->acct_pages);
8095 kfree(ctx->user_bufs);
8096 ctx->user_bufs = NULL;
8097 ctx->nr_user_bufs = 0;
8101 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8102 void __user *arg, unsigned index)
8104 struct iovec __user *src;
8106 #ifdef CONFIG_COMPAT
8108 struct compat_iovec __user *ciovs;
8109 struct compat_iovec ciov;
8111 ciovs = (struct compat_iovec __user *) arg;
8112 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8115 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8116 dst->iov_len = ciov.iov_len;
8120 src = (struct iovec __user *) arg;
8121 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8127 * Not super efficient, but this is just a registration time. And we do cache
8128 * the last compound head, so generally we'll only do a full search if we don't
8131 * We check if the given compound head page has already been accounted, to
8132 * avoid double accounting it. This allows us to account the full size of the
8133 * page, not just the constituent pages of a huge page.
8135 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8136 int nr_pages, struct page *hpage)
8140 /* check current page array */
8141 for (i = 0; i < nr_pages; i++) {
8142 if (!PageCompound(pages[i]))
8144 if (compound_head(pages[i]) == hpage)
8148 /* check previously registered pages */
8149 for (i = 0; i < ctx->nr_user_bufs; i++) {
8150 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8152 for (j = 0; j < imu->nr_bvecs; j++) {
8153 if (!PageCompound(imu->bvec[j].bv_page))
8155 if (compound_head(imu->bvec[j].bv_page) == hpage)
8163 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8164 int nr_pages, struct io_mapped_ubuf *imu,
8165 struct page **last_hpage)
8169 for (i = 0; i < nr_pages; i++) {
8170 if (!PageCompound(pages[i])) {
8175 hpage = compound_head(pages[i]);
8176 if (hpage == *last_hpage)
8178 *last_hpage = hpage;
8179 if (headpage_already_acct(ctx, pages, i, hpage))
8181 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8185 if (!imu->acct_pages)
8188 ret = io_account_mem(ctx, imu->acct_pages);
8190 imu->acct_pages = 0;
8194 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8195 struct io_mapped_ubuf *imu,
8196 struct page **last_hpage)
8198 struct vm_area_struct **vmas = NULL;
8199 struct page **pages = NULL;
8200 unsigned long off, start, end, ubuf;
8202 int ret, pret, nr_pages, i;
8204 ubuf = (unsigned long) iov->iov_base;
8205 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8206 start = ubuf >> PAGE_SHIFT;
8207 nr_pages = end - start;
8211 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8215 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8220 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8226 mmap_read_lock(current->mm);
8227 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8229 if (pret == nr_pages) {
8230 /* don't support file backed memory */
8231 for (i = 0; i < nr_pages; i++) {
8232 struct vm_area_struct *vma = vmas[i];
8235 !is_file_hugepages(vma->vm_file)) {
8241 ret = pret < 0 ? pret : -EFAULT;
8243 mmap_read_unlock(current->mm);
8246 * if we did partial map, or found file backed vmas,
8247 * release any pages we did get
8250 unpin_user_pages(pages, pret);
8255 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8257 unpin_user_pages(pages, pret);
8262 off = ubuf & ~PAGE_MASK;
8263 size = iov->iov_len;
8264 for (i = 0; i < nr_pages; i++) {
8267 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8268 imu->bvec[i].bv_page = pages[i];
8269 imu->bvec[i].bv_len = vec_len;
8270 imu->bvec[i].bv_offset = off;
8274 /* store original address for later verification */
8276 imu->len = iov->iov_len;
8277 imu->nr_bvecs = nr_pages;
8285 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8289 if (!nr_args || nr_args > UIO_MAXIOV)
8292 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8294 if (!ctx->user_bufs)
8300 static int io_buffer_validate(struct iovec *iov)
8303 * Don't impose further limits on the size and buffer
8304 * constraints here, we'll -EINVAL later when IO is
8305 * submitted if they are wrong.
8307 if (!iov->iov_base || !iov->iov_len)
8310 /* arbitrary limit, but we need something */
8311 if (iov->iov_len > SZ_1G)
8317 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8318 unsigned int nr_args)
8322 struct page *last_hpage = NULL;
8324 ret = io_buffers_map_alloc(ctx, nr_args);
8328 for (i = 0; i < nr_args; i++) {
8329 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8331 ret = io_copy_iov(ctx, &iov, arg, i);
8335 ret = io_buffer_validate(&iov);
8339 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8343 ctx->nr_user_bufs++;
8347 io_sqe_buffers_unregister(ctx);
8352 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8354 __s32 __user *fds = arg;
8360 if (copy_from_user(&fd, fds, sizeof(*fds)))
8363 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8364 if (IS_ERR(ctx->cq_ev_fd)) {
8365 int ret = PTR_ERR(ctx->cq_ev_fd);
8366 ctx->cq_ev_fd = NULL;
8373 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8375 if (ctx->cq_ev_fd) {
8376 eventfd_ctx_put(ctx->cq_ev_fd);
8377 ctx->cq_ev_fd = NULL;
8384 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8386 struct io_buffer *buf;
8387 unsigned long index;
8389 xa_for_each(&ctx->io_buffers, index, buf)
8390 __io_remove_buffers(ctx, buf, index, -1U);
8393 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8395 struct io_kiocb *req, *nxt;
8397 list_for_each_entry_safe(req, nxt, list, compl.list) {
8398 if (tsk && req->task != tsk)
8400 list_del(&req->compl.list);
8401 kmem_cache_free(req_cachep, req);
8405 static void io_req_caches_free(struct io_ring_ctx *ctx)
8407 struct io_submit_state *submit_state = &ctx->submit_state;
8408 struct io_comp_state *cs = &ctx->submit_state.comp;
8410 mutex_lock(&ctx->uring_lock);
8412 if (submit_state->free_reqs) {
8413 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8414 submit_state->reqs);
8415 submit_state->free_reqs = 0;
8418 io_flush_cached_locked_reqs(ctx, cs);
8419 io_req_cache_free(&cs->free_list, NULL);
8420 mutex_unlock(&ctx->uring_lock);
8423 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8426 * Some may use context even when all refs and requests have been put,
8427 * and they are free to do so while still holding uring_lock or
8428 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8430 mutex_lock(&ctx->uring_lock);
8431 mutex_unlock(&ctx->uring_lock);
8432 spin_lock_irq(&ctx->completion_lock);
8433 spin_unlock_irq(&ctx->completion_lock);
8435 io_sq_thread_finish(ctx);
8436 io_sqe_buffers_unregister(ctx);
8438 if (ctx->mm_account) {
8439 mmdrop(ctx->mm_account);
8440 ctx->mm_account = NULL;
8443 mutex_lock(&ctx->uring_lock);
8444 io_sqe_files_unregister(ctx);
8445 mutex_unlock(&ctx->uring_lock);
8446 io_eventfd_unregister(ctx);
8447 io_destroy_buffers(ctx);
8449 if (ctx->rsrc_backup_node)
8450 destroy_fixed_rsrc_ref_node(ctx->rsrc_backup_node);
8452 #if defined(CONFIG_UNIX)
8453 if (ctx->ring_sock) {
8454 ctx->ring_sock->file = NULL; /* so that iput() is called */
8455 sock_release(ctx->ring_sock);
8459 io_mem_free(ctx->rings);
8460 io_mem_free(ctx->sq_sqes);
8462 percpu_ref_exit(&ctx->refs);
8463 free_uid(ctx->user);
8464 io_req_caches_free(ctx);
8466 io_wq_put_hash(ctx->hash_map);
8467 kfree(ctx->cancel_hash);
8471 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8473 struct io_ring_ctx *ctx = file->private_data;
8476 poll_wait(file, &ctx->cq_wait, wait);
8478 * synchronizes with barrier from wq_has_sleeper call in
8482 if (!io_sqring_full(ctx))
8483 mask |= EPOLLOUT | EPOLLWRNORM;
8486 * Don't flush cqring overflow list here, just do a simple check.
8487 * Otherwise there could possible be ABBA deadlock:
8490 * lock(&ctx->uring_lock);
8492 * lock(&ctx->uring_lock);
8495 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8496 * pushs them to do the flush.
8498 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8499 mask |= EPOLLIN | EPOLLRDNORM;
8504 static int io_uring_fasync(int fd, struct file *file, int on)
8506 struct io_ring_ctx *ctx = file->private_data;
8508 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8511 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8513 const struct cred *creds;
8515 creds = xa_erase(&ctx->personalities, id);
8524 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8526 return io_run_task_work_head(&ctx->exit_task_work);
8529 struct io_tctx_exit {
8530 struct callback_head task_work;
8531 struct completion completion;
8532 struct io_ring_ctx *ctx;
8535 static void io_tctx_exit_cb(struct callback_head *cb)
8537 struct io_uring_task *tctx = current->io_uring;
8538 struct io_tctx_exit *work;
8540 work = container_of(cb, struct io_tctx_exit, task_work);
8542 * When @in_idle, we're in cancellation and it's racy to remove the
8543 * node. It'll be removed by the end of cancellation, just ignore it.
8545 if (!atomic_read(&tctx->in_idle))
8546 io_uring_del_task_file((unsigned long)work->ctx);
8547 complete(&work->completion);
8550 static void io_ring_exit_work(struct work_struct *work)
8552 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8553 unsigned long timeout = jiffies + HZ * 60 * 5;
8554 struct io_tctx_exit exit;
8555 struct io_tctx_node *node;
8558 /* prevent SQPOLL from submitting new requests */
8560 io_sq_thread_park(ctx->sq_data);
8561 list_del_init(&ctx->sqd_list);
8562 io_sqd_update_thread_idle(ctx->sq_data);
8563 io_sq_thread_unpark(ctx->sq_data);
8567 * If we're doing polled IO and end up having requests being
8568 * submitted async (out-of-line), then completions can come in while
8569 * we're waiting for refs to drop. We need to reap these manually,
8570 * as nobody else will be looking for them.
8573 io_uring_try_cancel_requests(ctx, NULL, NULL);
8575 WARN_ON_ONCE(time_after(jiffies, timeout));
8576 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8578 mutex_lock(&ctx->uring_lock);
8579 while (!list_empty(&ctx->tctx_list)) {
8580 WARN_ON_ONCE(time_after(jiffies, timeout));
8582 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8585 init_completion(&exit.completion);
8586 init_task_work(&exit.task_work, io_tctx_exit_cb);
8587 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8588 if (WARN_ON_ONCE(ret))
8590 wake_up_process(node->task);
8592 mutex_unlock(&ctx->uring_lock);
8593 wait_for_completion(&exit.completion);
8595 mutex_lock(&ctx->uring_lock);
8597 mutex_unlock(&ctx->uring_lock);
8599 io_ring_ctx_free(ctx);
8602 /* Returns true if we found and killed one or more timeouts */
8603 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8604 struct files_struct *files)
8606 struct io_kiocb *req, *tmp;
8609 spin_lock_irq(&ctx->completion_lock);
8610 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8611 if (io_match_task(req, tsk, files)) {
8612 io_kill_timeout(req, -ECANCELED);
8617 io_commit_cqring(ctx);
8618 spin_unlock_irq(&ctx->completion_lock);
8620 io_cqring_ev_posted(ctx);
8621 return canceled != 0;
8624 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8626 unsigned long index;
8627 struct creds *creds;
8629 mutex_lock(&ctx->uring_lock);
8630 percpu_ref_kill(&ctx->refs);
8631 /* if force is set, the ring is going away. always drop after that */
8632 ctx->cq_overflow_flushed = 1;
8634 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8635 xa_for_each(&ctx->personalities, index, creds)
8636 io_unregister_personality(ctx, index);
8637 mutex_unlock(&ctx->uring_lock);
8639 io_kill_timeouts(ctx, NULL, NULL);
8640 io_poll_remove_all(ctx, NULL, NULL);
8642 /* if we failed setting up the ctx, we might not have any rings */
8643 io_iopoll_try_reap_events(ctx);
8645 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8647 * Use system_unbound_wq to avoid spawning tons of event kworkers
8648 * if we're exiting a ton of rings at the same time. It just adds
8649 * noise and overhead, there's no discernable change in runtime
8650 * over using system_wq.
8652 queue_work(system_unbound_wq, &ctx->exit_work);
8655 static int io_uring_release(struct inode *inode, struct file *file)
8657 struct io_ring_ctx *ctx = file->private_data;
8659 file->private_data = NULL;
8660 io_ring_ctx_wait_and_kill(ctx);
8664 struct io_task_cancel {
8665 struct task_struct *task;
8666 struct files_struct *files;
8669 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8671 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8672 struct io_task_cancel *cancel = data;
8675 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8676 unsigned long flags;
8677 struct io_ring_ctx *ctx = req->ctx;
8679 /* protect against races with linked timeouts */
8680 spin_lock_irqsave(&ctx->completion_lock, flags);
8681 ret = io_match_task(req, cancel->task, cancel->files);
8682 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8684 ret = io_match_task(req, cancel->task, cancel->files);
8689 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8690 struct task_struct *task,
8691 struct files_struct *files)
8693 struct io_defer_entry *de;
8696 spin_lock_irq(&ctx->completion_lock);
8697 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8698 if (io_match_task(de->req, task, files)) {
8699 list_cut_position(&list, &ctx->defer_list, &de->list);
8703 spin_unlock_irq(&ctx->completion_lock);
8704 if (list_empty(&list))
8707 while (!list_empty(&list)) {
8708 de = list_first_entry(&list, struct io_defer_entry, list);
8709 list_del_init(&de->list);
8710 io_req_complete_failed(de->req, -ECANCELED);
8716 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8718 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8720 return req->ctx == data;
8723 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8725 struct io_tctx_node *node;
8726 enum io_wq_cancel cret;
8729 mutex_lock(&ctx->uring_lock);
8730 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8731 struct io_uring_task *tctx = node->task->io_uring;
8734 * io_wq will stay alive while we hold uring_lock, because it's
8735 * killed after ctx nodes, which requires to take the lock.
8737 if (!tctx || !tctx->io_wq)
8739 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8740 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8742 mutex_unlock(&ctx->uring_lock);
8747 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8748 struct task_struct *task,
8749 struct files_struct *files)
8751 struct io_task_cancel cancel = { .task = task, .files = files, };
8752 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8755 enum io_wq_cancel cret;
8759 ret |= io_uring_try_cancel_iowq(ctx);
8760 } else if (tctx && tctx->io_wq) {
8762 * Cancels requests of all rings, not only @ctx, but
8763 * it's fine as the task is in exit/exec.
8765 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8767 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8770 /* SQPOLL thread does its own polling */
8771 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8772 (ctx->sq_data && ctx->sq_data->thread == current)) {
8773 while (!list_empty_careful(&ctx->iopoll_list)) {
8774 io_iopoll_try_reap_events(ctx);
8779 ret |= io_cancel_defer_files(ctx, task, files);
8780 ret |= io_poll_remove_all(ctx, task, files);
8781 ret |= io_kill_timeouts(ctx, task, files);
8782 ret |= io_run_task_work();
8783 ret |= io_run_ctx_fallback(ctx);
8784 io_cqring_overflow_flush(ctx, true, task, files);
8791 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8792 struct task_struct *task,
8793 struct files_struct *files)
8795 struct io_kiocb *req;
8798 spin_lock_irq(&ctx->inflight_lock);
8799 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8800 cnt += io_match_task(req, task, files);
8801 spin_unlock_irq(&ctx->inflight_lock);
8805 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8806 struct task_struct *task,
8807 struct files_struct *files)
8809 while (!list_empty_careful(&ctx->inflight_list)) {
8813 inflight = io_uring_count_inflight(ctx, task, files);
8817 io_uring_try_cancel_requests(ctx, task, files);
8819 prepare_to_wait(&task->io_uring->wait, &wait,
8820 TASK_UNINTERRUPTIBLE);
8821 if (inflight == io_uring_count_inflight(ctx, task, files))
8823 finish_wait(&task->io_uring->wait, &wait);
8827 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8829 struct io_uring_task *tctx = current->io_uring;
8830 struct io_tctx_node *node;
8833 if (unlikely(!tctx)) {
8834 ret = io_uring_alloc_task_context(current, ctx);
8837 tctx = current->io_uring;
8839 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8840 node = kmalloc(sizeof(*node), GFP_KERNEL);
8844 node->task = current;
8846 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8853 mutex_lock(&ctx->uring_lock);
8854 list_add(&node->ctx_node, &ctx->tctx_list);
8855 mutex_unlock(&ctx->uring_lock);
8862 * Note that this task has used io_uring. We use it for cancelation purposes.
8864 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
8866 struct io_uring_task *tctx = current->io_uring;
8868 if (likely(tctx && tctx->last == ctx))
8870 return __io_uring_add_task_file(ctx);
8874 * Remove this io_uring_file -> task mapping.
8876 static void io_uring_del_task_file(unsigned long index)
8878 struct io_uring_task *tctx = current->io_uring;
8879 struct io_tctx_node *node;
8883 node = xa_erase(&tctx->xa, index);
8887 WARN_ON_ONCE(current != node->task);
8888 WARN_ON_ONCE(list_empty(&node->ctx_node));
8890 mutex_lock(&node->ctx->uring_lock);
8891 list_del(&node->ctx_node);
8892 mutex_unlock(&node->ctx->uring_lock);
8894 if (tctx->last == node->ctx)
8899 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8901 struct io_tctx_node *node;
8902 unsigned long index;
8904 xa_for_each(&tctx->xa, index, node)
8905 io_uring_del_task_file(index);
8907 io_wq_put_and_exit(tctx->io_wq);
8912 static s64 tctx_inflight(struct io_uring_task *tctx)
8914 return percpu_counter_sum(&tctx->inflight);
8917 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8919 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8920 struct io_ring_ctx *ctx = work->ctx;
8921 struct io_sq_data *sqd = ctx->sq_data;
8924 io_uring_cancel_sqpoll(ctx);
8925 complete(&work->completion);
8928 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8930 struct io_sq_data *sqd = ctx->sq_data;
8931 struct io_tctx_exit work = { .ctx = ctx, };
8932 struct task_struct *task;
8934 io_sq_thread_park(sqd);
8935 list_del_init(&ctx->sqd_list);
8936 io_sqd_update_thread_idle(sqd);
8939 init_completion(&work.completion);
8940 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8941 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8942 wake_up_process(task);
8944 io_sq_thread_unpark(sqd);
8947 wait_for_completion(&work.completion);
8950 void __io_uring_files_cancel(struct files_struct *files)
8952 struct io_uring_task *tctx = current->io_uring;
8953 struct io_tctx_node *node;
8954 unsigned long index;
8956 /* make sure overflow events are dropped */
8957 atomic_inc(&tctx->in_idle);
8958 xa_for_each(&tctx->xa, index, node) {
8959 struct io_ring_ctx *ctx = node->ctx;
8962 io_sqpoll_cancel_sync(ctx);
8965 io_uring_cancel_files(ctx, current, files);
8967 io_uring_try_cancel_requests(ctx, current, NULL);
8969 atomic_dec(&tctx->in_idle);
8972 io_uring_clean_tctx(tctx);
8975 /* should only be called by SQPOLL task */
8976 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8978 struct io_sq_data *sqd = ctx->sq_data;
8979 struct io_uring_task *tctx = current->io_uring;
8983 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8985 atomic_inc(&tctx->in_idle);
8987 /* read completions before cancelations */
8988 inflight = tctx_inflight(tctx);
8991 io_uring_try_cancel_requests(ctx, current, NULL);
8993 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8995 * If we've seen completions, retry without waiting. This
8996 * avoids a race where a completion comes in before we did
8997 * prepare_to_wait().
8999 if (inflight == tctx_inflight(tctx))
9001 finish_wait(&tctx->wait, &wait);
9003 atomic_dec(&tctx->in_idle);
9007 * Find any io_uring fd that this task has registered or done IO on, and cancel
9010 void __io_uring_task_cancel(void)
9012 struct io_uring_task *tctx = current->io_uring;
9016 /* make sure overflow events are dropped */
9017 atomic_inc(&tctx->in_idle);
9018 __io_uring_files_cancel(NULL);
9021 /* read completions before cancelations */
9022 inflight = tctx_inflight(tctx);
9025 __io_uring_files_cancel(NULL);
9027 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9030 * If we've seen completions, retry without waiting. This
9031 * avoids a race where a completion comes in before we did
9032 * prepare_to_wait().
9034 if (inflight == tctx_inflight(tctx))
9036 finish_wait(&tctx->wait, &wait);
9039 atomic_dec(&tctx->in_idle);
9041 io_uring_clean_tctx(tctx);
9042 /* all current's requests should be gone, we can kill tctx */
9043 __io_uring_free(current);
9046 static void *io_uring_validate_mmap_request(struct file *file,
9047 loff_t pgoff, size_t sz)
9049 struct io_ring_ctx *ctx = file->private_data;
9050 loff_t offset = pgoff << PAGE_SHIFT;
9055 case IORING_OFF_SQ_RING:
9056 case IORING_OFF_CQ_RING:
9059 case IORING_OFF_SQES:
9063 return ERR_PTR(-EINVAL);
9066 page = virt_to_head_page(ptr);
9067 if (sz > page_size(page))
9068 return ERR_PTR(-EINVAL);
9075 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9077 size_t sz = vma->vm_end - vma->vm_start;
9081 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9083 return PTR_ERR(ptr);
9085 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9086 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9089 #else /* !CONFIG_MMU */
9091 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9093 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9096 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9098 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9101 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9102 unsigned long addr, unsigned long len,
9103 unsigned long pgoff, unsigned long flags)
9107 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9109 return PTR_ERR(ptr);
9111 return (unsigned long) ptr;
9114 #endif /* !CONFIG_MMU */
9116 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9121 if (!io_sqring_full(ctx))
9123 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9125 if (!io_sqring_full(ctx))
9128 } while (!signal_pending(current));
9130 finish_wait(&ctx->sqo_sq_wait, &wait);
9134 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9135 struct __kernel_timespec __user **ts,
9136 const sigset_t __user **sig)
9138 struct io_uring_getevents_arg arg;
9141 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9142 * is just a pointer to the sigset_t.
9144 if (!(flags & IORING_ENTER_EXT_ARG)) {
9145 *sig = (const sigset_t __user *) argp;
9151 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9152 * timespec and sigset_t pointers if good.
9154 if (*argsz != sizeof(arg))
9156 if (copy_from_user(&arg, argp, sizeof(arg)))
9158 *sig = u64_to_user_ptr(arg.sigmask);
9159 *argsz = arg.sigmask_sz;
9160 *ts = u64_to_user_ptr(arg.ts);
9164 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9165 u32, min_complete, u32, flags, const void __user *, argp,
9168 struct io_ring_ctx *ctx;
9175 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9176 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9180 if (unlikely(!f.file))
9184 if (unlikely(f.file->f_op != &io_uring_fops))
9188 ctx = f.file->private_data;
9189 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9193 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9197 * For SQ polling, the thread will do all submissions and completions.
9198 * Just return the requested submit count, and wake the thread if
9202 if (ctx->flags & IORING_SETUP_SQPOLL) {
9203 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9206 if (unlikely(ctx->sq_data->thread == NULL)) {
9209 if (flags & IORING_ENTER_SQ_WAKEUP)
9210 wake_up(&ctx->sq_data->wait);
9211 if (flags & IORING_ENTER_SQ_WAIT) {
9212 ret = io_sqpoll_wait_sq(ctx);
9216 submitted = to_submit;
9217 } else if (to_submit) {
9218 ret = io_uring_add_task_file(ctx);
9221 mutex_lock(&ctx->uring_lock);
9222 submitted = io_submit_sqes(ctx, to_submit);
9223 mutex_unlock(&ctx->uring_lock);
9225 if (submitted != to_submit)
9228 if (flags & IORING_ENTER_GETEVENTS) {
9229 const sigset_t __user *sig;
9230 struct __kernel_timespec __user *ts;
9232 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9236 min_complete = min(min_complete, ctx->cq_entries);
9239 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9240 * space applications don't need to do io completion events
9241 * polling again, they can rely on io_sq_thread to do polling
9242 * work, which can reduce cpu usage and uring_lock contention.
9244 if (ctx->flags & IORING_SETUP_IOPOLL &&
9245 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9246 ret = io_iopoll_check(ctx, min_complete);
9248 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9253 percpu_ref_put(&ctx->refs);
9256 return submitted ? submitted : ret;
9259 #ifdef CONFIG_PROC_FS
9260 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9261 const struct cred *cred)
9263 struct user_namespace *uns = seq_user_ns(m);
9264 struct group_info *gi;
9269 seq_printf(m, "%5d\n", id);
9270 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9271 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9272 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9273 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9274 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9275 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9276 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9277 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9278 seq_puts(m, "\n\tGroups:\t");
9279 gi = cred->group_info;
9280 for (g = 0; g < gi->ngroups; g++) {
9281 seq_put_decimal_ull(m, g ? " " : "",
9282 from_kgid_munged(uns, gi->gid[g]));
9284 seq_puts(m, "\n\tCapEff:\t");
9285 cap = cred->cap_effective;
9286 CAP_FOR_EACH_U32(__capi)
9287 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9292 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9294 struct io_sq_data *sq = NULL;
9299 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9300 * since fdinfo case grabs it in the opposite direction of normal use
9301 * cases. If we fail to get the lock, we just don't iterate any
9302 * structures that could be going away outside the io_uring mutex.
9304 has_lock = mutex_trylock(&ctx->uring_lock);
9306 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9312 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9313 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9314 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9315 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9316 struct file *f = io_file_from_index(ctx, i);
9319 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9321 seq_printf(m, "%5u: <none>\n", i);
9323 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9324 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9325 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9327 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9328 (unsigned int) buf->len);
9330 if (has_lock && !xa_empty(&ctx->personalities)) {
9331 unsigned long index;
9332 const struct cred *cred;
9334 seq_printf(m, "Personalities:\n");
9335 xa_for_each(&ctx->personalities, index, cred)
9336 io_uring_show_cred(m, index, cred);
9338 seq_printf(m, "PollList:\n");
9339 spin_lock_irq(&ctx->completion_lock);
9340 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9341 struct hlist_head *list = &ctx->cancel_hash[i];
9342 struct io_kiocb *req;
9344 hlist_for_each_entry(req, list, hash_node)
9345 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9346 req->task->task_works != NULL);
9348 spin_unlock_irq(&ctx->completion_lock);
9350 mutex_unlock(&ctx->uring_lock);
9353 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9355 struct io_ring_ctx *ctx = f->private_data;
9357 if (percpu_ref_tryget(&ctx->refs)) {
9358 __io_uring_show_fdinfo(ctx, m);
9359 percpu_ref_put(&ctx->refs);
9364 static const struct file_operations io_uring_fops = {
9365 .release = io_uring_release,
9366 .mmap = io_uring_mmap,
9368 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9369 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9371 .poll = io_uring_poll,
9372 .fasync = io_uring_fasync,
9373 #ifdef CONFIG_PROC_FS
9374 .show_fdinfo = io_uring_show_fdinfo,
9378 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9379 struct io_uring_params *p)
9381 struct io_rings *rings;
9382 size_t size, sq_array_offset;
9384 /* make sure these are sane, as we already accounted them */
9385 ctx->sq_entries = p->sq_entries;
9386 ctx->cq_entries = p->cq_entries;
9388 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9389 if (size == SIZE_MAX)
9392 rings = io_mem_alloc(size);
9397 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9398 rings->sq_ring_mask = p->sq_entries - 1;
9399 rings->cq_ring_mask = p->cq_entries - 1;
9400 rings->sq_ring_entries = p->sq_entries;
9401 rings->cq_ring_entries = p->cq_entries;
9402 ctx->sq_mask = rings->sq_ring_mask;
9403 ctx->cq_mask = rings->cq_ring_mask;
9405 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9406 if (size == SIZE_MAX) {
9407 io_mem_free(ctx->rings);
9412 ctx->sq_sqes = io_mem_alloc(size);
9413 if (!ctx->sq_sqes) {
9414 io_mem_free(ctx->rings);
9422 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9426 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9430 ret = io_uring_add_task_file(ctx);
9435 fd_install(fd, file);
9440 * Allocate an anonymous fd, this is what constitutes the application
9441 * visible backing of an io_uring instance. The application mmaps this
9442 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9443 * we have to tie this fd to a socket for file garbage collection purposes.
9445 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9448 #if defined(CONFIG_UNIX)
9451 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9454 return ERR_PTR(ret);
9457 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9458 O_RDWR | O_CLOEXEC);
9459 #if defined(CONFIG_UNIX)
9461 sock_release(ctx->ring_sock);
9462 ctx->ring_sock = NULL;
9464 ctx->ring_sock->file = file;
9470 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9471 struct io_uring_params __user *params)
9473 struct io_ring_ctx *ctx;
9479 if (entries > IORING_MAX_ENTRIES) {
9480 if (!(p->flags & IORING_SETUP_CLAMP))
9482 entries = IORING_MAX_ENTRIES;
9486 * Use twice as many entries for the CQ ring. It's possible for the
9487 * application to drive a higher depth than the size of the SQ ring,
9488 * since the sqes are only used at submission time. This allows for
9489 * some flexibility in overcommitting a bit. If the application has
9490 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9491 * of CQ ring entries manually.
9493 p->sq_entries = roundup_pow_of_two(entries);
9494 if (p->flags & IORING_SETUP_CQSIZE) {
9496 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9497 * to a power-of-two, if it isn't already. We do NOT impose
9498 * any cq vs sq ring sizing.
9502 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9503 if (!(p->flags & IORING_SETUP_CLAMP))
9505 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9507 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9508 if (p->cq_entries < p->sq_entries)
9511 p->cq_entries = 2 * p->sq_entries;
9514 ctx = io_ring_ctx_alloc(p);
9517 ctx->compat = in_compat_syscall();
9518 if (!capable(CAP_IPC_LOCK))
9519 ctx->user = get_uid(current_user());
9522 * This is just grabbed for accounting purposes. When a process exits,
9523 * the mm is exited and dropped before the files, hence we need to hang
9524 * on to this mm purely for the purposes of being able to unaccount
9525 * memory (locked/pinned vm). It's not used for anything else.
9527 mmgrab(current->mm);
9528 ctx->mm_account = current->mm;
9530 ret = io_allocate_scq_urings(ctx, p);
9534 ret = io_sq_offload_create(ctx, p);
9538 memset(&p->sq_off, 0, sizeof(p->sq_off));
9539 p->sq_off.head = offsetof(struct io_rings, sq.head);
9540 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9541 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9542 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9543 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9544 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9545 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9547 memset(&p->cq_off, 0, sizeof(p->cq_off));
9548 p->cq_off.head = offsetof(struct io_rings, cq.head);
9549 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9550 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9551 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9552 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9553 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9554 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9556 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9557 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9558 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9559 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9560 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9562 if (copy_to_user(params, p, sizeof(*p))) {
9567 file = io_uring_get_file(ctx);
9569 ret = PTR_ERR(file);
9574 * Install ring fd as the very last thing, so we don't risk someone
9575 * having closed it before we finish setup
9577 ret = io_uring_install_fd(ctx, file);
9579 /* fput will clean it up */
9584 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9587 io_ring_ctx_wait_and_kill(ctx);
9592 * Sets up an aio uring context, and returns the fd. Applications asks for a
9593 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9594 * params structure passed in.
9596 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9598 struct io_uring_params p;
9601 if (copy_from_user(&p, params, sizeof(p)))
9603 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9608 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9609 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9610 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9611 IORING_SETUP_R_DISABLED))
9614 return io_uring_create(entries, &p, params);
9617 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9618 struct io_uring_params __user *, params)
9620 return io_uring_setup(entries, params);
9623 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9625 struct io_uring_probe *p;
9629 size = struct_size(p, ops, nr_args);
9630 if (size == SIZE_MAX)
9632 p = kzalloc(size, GFP_KERNEL);
9637 if (copy_from_user(p, arg, size))
9640 if (memchr_inv(p, 0, size))
9643 p->last_op = IORING_OP_LAST - 1;
9644 if (nr_args > IORING_OP_LAST)
9645 nr_args = IORING_OP_LAST;
9647 for (i = 0; i < nr_args; i++) {
9649 if (!io_op_defs[i].not_supported)
9650 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9655 if (copy_to_user(arg, p, size))
9662 static int io_register_personality(struct io_ring_ctx *ctx)
9664 const struct cred *creds;
9668 creds = get_current_cred();
9670 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9671 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9678 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9679 unsigned int nr_args)
9681 struct io_uring_restriction *res;
9685 /* Restrictions allowed only if rings started disabled */
9686 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9689 /* We allow only a single restrictions registration */
9690 if (ctx->restrictions.registered)
9693 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9696 size = array_size(nr_args, sizeof(*res));
9697 if (size == SIZE_MAX)
9700 res = memdup_user(arg, size);
9702 return PTR_ERR(res);
9706 for (i = 0; i < nr_args; i++) {
9707 switch (res[i].opcode) {
9708 case IORING_RESTRICTION_REGISTER_OP:
9709 if (res[i].register_op >= IORING_REGISTER_LAST) {
9714 __set_bit(res[i].register_op,
9715 ctx->restrictions.register_op);
9717 case IORING_RESTRICTION_SQE_OP:
9718 if (res[i].sqe_op >= IORING_OP_LAST) {
9723 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9725 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9726 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9728 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9729 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9738 /* Reset all restrictions if an error happened */
9740 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9742 ctx->restrictions.registered = true;
9748 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9750 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9753 if (ctx->restrictions.registered)
9754 ctx->restricted = 1;
9756 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9757 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9758 wake_up(&ctx->sq_data->wait);
9762 static bool io_register_op_must_quiesce(int op)
9765 case IORING_UNREGISTER_FILES:
9766 case IORING_REGISTER_FILES_UPDATE:
9767 case IORING_REGISTER_PROBE:
9768 case IORING_REGISTER_PERSONALITY:
9769 case IORING_UNREGISTER_PERSONALITY:
9776 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9777 void __user *arg, unsigned nr_args)
9778 __releases(ctx->uring_lock)
9779 __acquires(ctx->uring_lock)
9784 * We're inside the ring mutex, if the ref is already dying, then
9785 * someone else killed the ctx or is already going through
9786 * io_uring_register().
9788 if (percpu_ref_is_dying(&ctx->refs))
9791 if (io_register_op_must_quiesce(opcode)) {
9792 percpu_ref_kill(&ctx->refs);
9795 * Drop uring mutex before waiting for references to exit. If
9796 * another thread is currently inside io_uring_enter() it might
9797 * need to grab the uring_lock to make progress. If we hold it
9798 * here across the drain wait, then we can deadlock. It's safe
9799 * to drop the mutex here, since no new references will come in
9800 * after we've killed the percpu ref.
9802 mutex_unlock(&ctx->uring_lock);
9804 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9807 ret = io_run_task_work_sig();
9812 mutex_lock(&ctx->uring_lock);
9815 percpu_ref_resurrect(&ctx->refs);
9820 if (ctx->restricted) {
9821 if (opcode >= IORING_REGISTER_LAST) {
9826 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9833 case IORING_REGISTER_BUFFERS:
9834 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9836 case IORING_UNREGISTER_BUFFERS:
9840 ret = io_sqe_buffers_unregister(ctx);
9842 case IORING_REGISTER_FILES:
9843 ret = io_sqe_files_register(ctx, arg, nr_args);
9845 case IORING_UNREGISTER_FILES:
9849 ret = io_sqe_files_unregister(ctx);
9851 case IORING_REGISTER_FILES_UPDATE:
9852 ret = io_sqe_files_update(ctx, arg, nr_args);
9854 case IORING_REGISTER_EVENTFD:
9855 case IORING_REGISTER_EVENTFD_ASYNC:
9859 ret = io_eventfd_register(ctx, arg);
9862 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9863 ctx->eventfd_async = 1;
9865 ctx->eventfd_async = 0;
9867 case IORING_UNREGISTER_EVENTFD:
9871 ret = io_eventfd_unregister(ctx);
9873 case IORING_REGISTER_PROBE:
9875 if (!arg || nr_args > 256)
9877 ret = io_probe(ctx, arg, nr_args);
9879 case IORING_REGISTER_PERSONALITY:
9883 ret = io_register_personality(ctx);
9885 case IORING_UNREGISTER_PERSONALITY:
9889 ret = io_unregister_personality(ctx, nr_args);
9891 case IORING_REGISTER_ENABLE_RINGS:
9895 ret = io_register_enable_rings(ctx);
9897 case IORING_REGISTER_RESTRICTIONS:
9898 ret = io_register_restrictions(ctx, arg, nr_args);
9906 if (io_register_op_must_quiesce(opcode)) {
9907 /* bring the ctx back to life */
9908 percpu_ref_reinit(&ctx->refs);
9910 reinit_completion(&ctx->ref_comp);
9915 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9916 void __user *, arg, unsigned int, nr_args)
9918 struct io_ring_ctx *ctx;
9927 if (f.file->f_op != &io_uring_fops)
9930 ctx = f.file->private_data;
9934 mutex_lock(&ctx->uring_lock);
9935 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9936 mutex_unlock(&ctx->uring_lock);
9937 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9938 ctx->cq_ev_fd != NULL, ret);
9944 static int __init io_uring_init(void)
9946 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9947 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9948 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9951 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9952 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9953 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9954 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9955 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9956 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9957 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9958 BUILD_BUG_SQE_ELEM(8, __u64, off);
9959 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9960 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9961 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9962 BUILD_BUG_SQE_ELEM(24, __u32, len);
9963 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9964 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9965 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9966 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9967 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9968 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9969 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9970 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9971 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9972 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9973 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9974 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9975 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9976 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9977 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9978 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9979 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9980 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9981 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9983 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9984 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9985 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9989 __initcall(io_uring_init);