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 io_kiocb *req, struct file *file,
1706 static void io_dismantle_req(struct io_kiocb *req)
1708 unsigned int flags = req->flags;
1711 io_put_file(req, req->file, (flags & REQ_F_FIXED_FILE));
1712 if (flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1716 if (req->flags & REQ_F_INFLIGHT) {
1717 struct io_ring_ctx *ctx = req->ctx;
1718 unsigned long flags;
1720 spin_lock_irqsave(&ctx->inflight_lock, flags);
1721 list_del(&req->inflight_entry);
1722 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1723 req->flags &= ~REQ_F_INFLIGHT;
1726 if (req->fixed_rsrc_refs)
1727 percpu_ref_put(req->fixed_rsrc_refs);
1728 if (req->async_data)
1729 kfree(req->async_data);
1730 if (req->work.creds) {
1731 put_cred(req->work.creds);
1732 req->work.creds = NULL;
1736 /* must to be called somewhat shortly after putting a request */
1737 static inline void io_put_task(struct task_struct *task, int nr)
1739 struct io_uring_task *tctx = task->io_uring;
1741 percpu_counter_sub(&tctx->inflight, nr);
1742 if (unlikely(atomic_read(&tctx->in_idle)))
1743 wake_up(&tctx->wait);
1744 put_task_struct_many(task, nr);
1747 static void __io_free_req(struct io_kiocb *req)
1749 struct io_ring_ctx *ctx = req->ctx;
1751 io_dismantle_req(req);
1752 io_put_task(req->task, 1);
1754 kmem_cache_free(req_cachep, req);
1755 percpu_ref_put(&ctx->refs);
1758 static inline void io_remove_next_linked(struct io_kiocb *req)
1760 struct io_kiocb *nxt = req->link;
1762 req->link = nxt->link;
1766 static bool io_kill_linked_timeout(struct io_kiocb *req)
1767 __must_hold(&req->ctx->completion_lock)
1769 struct io_kiocb *link = req->link;
1770 bool cancelled = false;
1773 * Can happen if a linked timeout fired and link had been like
1774 * req -> link t-out -> link t-out [-> ...]
1776 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1777 struct io_timeout_data *io = link->async_data;
1780 io_remove_next_linked(req);
1781 link->timeout.head = NULL;
1782 ret = hrtimer_try_to_cancel(&io->timer);
1784 io_cqring_fill_event(link, -ECANCELED);
1785 io_put_req_deferred(link, 1);
1789 req->flags &= ~REQ_F_LINK_TIMEOUT;
1793 static void io_fail_links(struct io_kiocb *req)
1794 __must_hold(&req->ctx->completion_lock)
1796 struct io_kiocb *nxt, *link = req->link;
1803 trace_io_uring_fail_link(req, link);
1804 io_cqring_fill_event(link, -ECANCELED);
1805 io_put_req_deferred(link, 2);
1810 static bool io_disarm_next(struct io_kiocb *req)
1811 __must_hold(&req->ctx->completion_lock)
1813 bool posted = false;
1815 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1816 posted = io_kill_linked_timeout(req);
1817 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1818 posted |= (req->link != NULL);
1824 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1826 struct io_kiocb *nxt;
1829 * If LINK is set, we have dependent requests in this chain. If we
1830 * didn't fail this request, queue the first one up, moving any other
1831 * dependencies to the next request. In case of failure, fail the rest
1834 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1835 struct io_ring_ctx *ctx = req->ctx;
1836 unsigned long flags;
1839 spin_lock_irqsave(&ctx->completion_lock, flags);
1840 posted = io_disarm_next(req);
1842 io_commit_cqring(req->ctx);
1843 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1845 io_cqring_ev_posted(ctx);
1852 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1854 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1856 return __io_req_find_next(req);
1859 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1863 if (ctx->submit_state.comp.nr) {
1864 mutex_lock(&ctx->uring_lock);
1865 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1866 mutex_unlock(&ctx->uring_lock);
1868 percpu_ref_put(&ctx->refs);
1871 static bool __tctx_task_work(struct io_uring_task *tctx)
1873 struct io_ring_ctx *ctx = NULL;
1874 struct io_wq_work_list list;
1875 struct io_wq_work_node *node;
1877 if (wq_list_empty(&tctx->task_list))
1880 spin_lock_irq(&tctx->task_lock);
1881 list = tctx->task_list;
1882 INIT_WQ_LIST(&tctx->task_list);
1883 spin_unlock_irq(&tctx->task_lock);
1887 struct io_wq_work_node *next = node->next;
1888 struct io_kiocb *req;
1890 req = container_of(node, struct io_kiocb, io_task_work.node);
1891 if (req->ctx != ctx) {
1892 ctx_flush_and_put(ctx);
1894 percpu_ref_get(&ctx->refs);
1897 req->task_work.func(&req->task_work);
1901 ctx_flush_and_put(ctx);
1902 return list.first != NULL;
1905 static void tctx_task_work(struct callback_head *cb)
1907 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1909 clear_bit(0, &tctx->task_state);
1911 while (__tctx_task_work(tctx))
1915 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1916 enum task_work_notify_mode notify)
1918 struct io_uring_task *tctx = tsk->io_uring;
1919 struct io_wq_work_node *node, *prev;
1920 unsigned long flags;
1923 WARN_ON_ONCE(!tctx);
1925 spin_lock_irqsave(&tctx->task_lock, flags);
1926 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1927 spin_unlock_irqrestore(&tctx->task_lock, flags);
1929 /* task_work already pending, we're done */
1930 if (test_bit(0, &tctx->task_state) ||
1931 test_and_set_bit(0, &tctx->task_state))
1934 if (!task_work_add(tsk, &tctx->task_work, notify))
1938 * Slow path - we failed, find and delete work. if the work is not
1939 * in the list, it got run and we're fine.
1942 spin_lock_irqsave(&tctx->task_lock, flags);
1943 wq_list_for_each(node, prev, &tctx->task_list) {
1944 if (&req->io_task_work.node == node) {
1945 wq_list_del(&tctx->task_list, node, prev);
1950 spin_unlock_irqrestore(&tctx->task_lock, flags);
1951 clear_bit(0, &tctx->task_state);
1955 static int io_req_task_work_add(struct io_kiocb *req)
1957 struct task_struct *tsk = req->task;
1958 struct io_ring_ctx *ctx = req->ctx;
1959 enum task_work_notify_mode notify;
1962 if (tsk->flags & PF_EXITING)
1966 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1967 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1968 * processing task_work. There's no reliable way to tell if TWA_RESUME
1972 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1973 notify = TWA_SIGNAL;
1975 ret = io_task_work_add(tsk, req, notify);
1977 wake_up_process(tsk);
1982 static bool io_run_task_work_head(struct callback_head **work_head)
1984 struct callback_head *work, *next;
1985 bool executed = false;
1988 work = xchg(work_head, NULL);
2004 static void io_task_work_add_head(struct callback_head **work_head,
2005 struct callback_head *task_work)
2007 struct callback_head *head;
2010 head = READ_ONCE(*work_head);
2011 task_work->next = head;
2012 } while (cmpxchg(work_head, head, task_work) != head);
2015 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2016 task_work_func_t cb)
2018 init_task_work(&req->task_work, cb);
2019 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2022 static void io_req_task_cancel(struct callback_head *cb)
2024 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2025 struct io_ring_ctx *ctx = req->ctx;
2027 /* ctx is guaranteed to stay alive while we hold uring_lock */
2028 mutex_lock(&ctx->uring_lock);
2029 io_req_complete_failed(req, req->result);
2030 mutex_unlock(&ctx->uring_lock);
2033 static void __io_req_task_submit(struct io_kiocb *req)
2035 struct io_ring_ctx *ctx = req->ctx;
2037 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2038 mutex_lock(&ctx->uring_lock);
2039 if (!(current->flags & PF_EXITING) && !current->in_execve)
2040 __io_queue_sqe(req);
2042 io_req_complete_failed(req, -EFAULT);
2043 mutex_unlock(&ctx->uring_lock);
2046 static void io_req_task_submit(struct callback_head *cb)
2048 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2050 __io_req_task_submit(req);
2053 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2056 req->task_work.func = io_req_task_cancel;
2058 if (unlikely(io_req_task_work_add(req)))
2059 io_req_task_work_add_fallback(req, io_req_task_cancel);
2062 static void io_req_task_queue(struct io_kiocb *req)
2064 req->task_work.func = io_req_task_submit;
2066 if (unlikely(io_req_task_work_add(req)))
2067 io_req_task_queue_fail(req, -ECANCELED);
2070 static inline void io_queue_next(struct io_kiocb *req)
2072 struct io_kiocb *nxt = io_req_find_next(req);
2075 io_req_task_queue(nxt);
2078 static void io_free_req(struct io_kiocb *req)
2085 struct task_struct *task;
2090 static inline void io_init_req_batch(struct req_batch *rb)
2097 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2098 struct req_batch *rb)
2101 io_put_task(rb->task, rb->task_refs);
2103 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2106 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2107 struct io_submit_state *state)
2110 io_dismantle_req(req);
2112 if (req->task != rb->task) {
2114 io_put_task(rb->task, rb->task_refs);
2115 rb->task = req->task;
2121 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2122 state->reqs[state->free_reqs++] = req;
2124 list_add(&req->compl.list, &state->comp.free_list);
2127 static void io_submit_flush_completions(struct io_comp_state *cs,
2128 struct io_ring_ctx *ctx)
2131 struct io_kiocb *req;
2132 struct req_batch rb;
2134 io_init_req_batch(&rb);
2135 spin_lock_irq(&ctx->completion_lock);
2136 for (i = 0; i < nr; i++) {
2138 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2140 io_commit_cqring(ctx);
2141 spin_unlock_irq(&ctx->completion_lock);
2143 io_cqring_ev_posted(ctx);
2144 for (i = 0; i < nr; i++) {
2147 /* submission and completion refs */
2148 if (req_ref_sub_and_test(req, 2))
2149 io_req_free_batch(&rb, req, &ctx->submit_state);
2152 io_req_free_batch_finish(ctx, &rb);
2157 * Drop reference to request, return next in chain (if there is one) if this
2158 * was the last reference to this request.
2160 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2162 struct io_kiocb *nxt = NULL;
2164 if (req_ref_put_and_test(req)) {
2165 nxt = io_req_find_next(req);
2171 static inline void io_put_req(struct io_kiocb *req)
2173 if (req_ref_put_and_test(req))
2177 static void io_put_req_deferred_cb(struct callback_head *cb)
2179 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2184 static void io_free_req_deferred(struct io_kiocb *req)
2186 req->task_work.func = io_put_req_deferred_cb;
2187 if (unlikely(io_req_task_work_add(req)))
2188 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2191 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2193 if (req_ref_sub_and_test(req, refs))
2194 io_free_req_deferred(req);
2197 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2199 /* See comment at the top of this file */
2201 return __io_cqring_events(ctx);
2204 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2206 struct io_rings *rings = ctx->rings;
2208 /* make sure SQ entry isn't read before tail */
2209 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2212 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2214 unsigned int cflags;
2216 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2217 cflags |= IORING_CQE_F_BUFFER;
2218 req->flags &= ~REQ_F_BUFFER_SELECTED;
2223 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2225 struct io_buffer *kbuf;
2227 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2228 return io_put_kbuf(req, kbuf);
2231 static inline bool io_run_task_work(void)
2234 * Not safe to run on exiting task, and the task_work handling will
2235 * not add work to such a task.
2237 if (unlikely(current->flags & PF_EXITING))
2239 if (current->task_works) {
2240 __set_current_state(TASK_RUNNING);
2249 * Find and free completed poll iocbs
2251 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2252 struct list_head *done)
2254 struct req_batch rb;
2255 struct io_kiocb *req;
2257 /* order with ->result store in io_complete_rw_iopoll() */
2260 io_init_req_batch(&rb);
2261 while (!list_empty(done)) {
2264 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2265 list_del(&req->inflight_entry);
2267 if (READ_ONCE(req->result) == -EAGAIN) {
2268 req->iopoll_completed = 0;
2269 if (io_rw_reissue(req))
2273 if (req->flags & REQ_F_BUFFER_SELECTED)
2274 cflags = io_put_rw_kbuf(req);
2276 __io_cqring_fill_event(req, req->result, cflags);
2279 if (req_ref_put_and_test(req))
2280 io_req_free_batch(&rb, req, &ctx->submit_state);
2283 io_commit_cqring(ctx);
2284 io_cqring_ev_posted_iopoll(ctx);
2285 io_req_free_batch_finish(ctx, &rb);
2288 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2291 struct io_kiocb *req, *tmp;
2297 * Only spin for completions if we don't have multiple devices hanging
2298 * off our complete list, and we're under the requested amount.
2300 spin = !ctx->poll_multi_file && *nr_events < min;
2303 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2304 struct kiocb *kiocb = &req->rw.kiocb;
2307 * Move completed and retryable entries to our local lists.
2308 * If we find a request that requires polling, break out
2309 * and complete those lists first, if we have entries there.
2311 if (READ_ONCE(req->iopoll_completed)) {
2312 list_move_tail(&req->inflight_entry, &done);
2315 if (!list_empty(&done))
2318 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2322 /* iopoll may have completed current req */
2323 if (READ_ONCE(req->iopoll_completed))
2324 list_move_tail(&req->inflight_entry, &done);
2331 if (!list_empty(&done))
2332 io_iopoll_complete(ctx, nr_events, &done);
2338 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2339 * non-spinning poll check - we'll still enter the driver poll loop, but only
2340 * as a non-spinning completion check.
2342 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2345 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2348 ret = io_do_iopoll(ctx, nr_events, min);
2351 if (*nr_events >= min)
2359 * We can't just wait for polled events to come to us, we have to actively
2360 * find and complete them.
2362 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2364 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2367 mutex_lock(&ctx->uring_lock);
2368 while (!list_empty(&ctx->iopoll_list)) {
2369 unsigned int nr_events = 0;
2371 io_do_iopoll(ctx, &nr_events, 0);
2373 /* let it sleep and repeat later if can't complete a request */
2377 * Ensure we allow local-to-the-cpu processing to take place,
2378 * in this case we need to ensure that we reap all events.
2379 * Also let task_work, etc. to progress by releasing the mutex
2381 if (need_resched()) {
2382 mutex_unlock(&ctx->uring_lock);
2384 mutex_lock(&ctx->uring_lock);
2387 mutex_unlock(&ctx->uring_lock);
2390 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2392 unsigned int nr_events = 0;
2393 int iters = 0, ret = 0;
2396 * We disallow the app entering submit/complete with polling, but we
2397 * still need to lock the ring to prevent racing with polled issue
2398 * that got punted to a workqueue.
2400 mutex_lock(&ctx->uring_lock);
2403 * Don't enter poll loop if we already have events pending.
2404 * If we do, we can potentially be spinning for commands that
2405 * already triggered a CQE (eg in error).
2407 if (test_bit(0, &ctx->cq_check_overflow))
2408 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2409 if (io_cqring_events(ctx))
2413 * If a submit got punted to a workqueue, we can have the
2414 * application entering polling for a command before it gets
2415 * issued. That app will hold the uring_lock for the duration
2416 * of the poll right here, so we need to take a breather every
2417 * now and then to ensure that the issue has a chance to add
2418 * the poll to the issued list. Otherwise we can spin here
2419 * forever, while the workqueue is stuck trying to acquire the
2422 if (!(++iters & 7)) {
2423 mutex_unlock(&ctx->uring_lock);
2425 mutex_lock(&ctx->uring_lock);
2428 ret = io_iopoll_getevents(ctx, &nr_events, min);
2432 } while (min && !nr_events && !need_resched());
2434 mutex_unlock(&ctx->uring_lock);
2438 static void kiocb_end_write(struct io_kiocb *req)
2441 * Tell lockdep we inherited freeze protection from submission
2444 if (req->flags & REQ_F_ISREG) {
2445 struct inode *inode = file_inode(req->file);
2447 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2449 file_end_write(req->file);
2453 static bool io_resubmit_prep(struct io_kiocb *req)
2455 /* either already prepared or successfully done */
2456 return req->async_data || !io_req_prep_async(req);
2459 static bool io_rw_should_reissue(struct io_kiocb *req)
2461 umode_t mode = file_inode(req->file)->i_mode;
2462 struct io_ring_ctx *ctx = req->ctx;
2464 if (!S_ISBLK(mode) && !S_ISREG(mode))
2466 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2467 !(ctx->flags & IORING_SETUP_IOPOLL)))
2470 * If ref is dying, we might be running poll reap from the exit work.
2471 * Don't attempt to reissue from that path, just let it fail with
2474 if (percpu_ref_is_dying(&ctx->refs))
2479 static bool io_rw_should_reissue(struct io_kiocb *req)
2485 static bool io_rw_reissue(struct io_kiocb *req)
2488 if (!io_rw_should_reissue(req))
2491 lockdep_assert_held(&req->ctx->uring_lock);
2493 if (io_resubmit_prep(req)) {
2495 io_queue_async_work(req);
2498 req_set_fail_links(req);
2503 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2504 unsigned int issue_flags)
2508 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2509 kiocb_end_write(req);
2510 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_should_reissue(req)) {
2511 req->flags |= REQ_F_REISSUE;
2514 if (res != req->result)
2515 req_set_fail_links(req);
2516 if (req->flags & REQ_F_BUFFER_SELECTED)
2517 cflags = io_put_rw_kbuf(req);
2518 __io_req_complete(req, issue_flags, res, cflags);
2521 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2523 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2525 __io_complete_rw(req, res, res2, 0);
2528 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2530 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2533 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2534 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2535 struct io_async_rw *rw = req->async_data;
2538 iov_iter_revert(&rw->iter,
2539 req->result - iov_iter_count(&rw->iter));
2540 else if (!io_resubmit_prep(req))
2545 if (kiocb->ki_flags & IOCB_WRITE)
2546 kiocb_end_write(req);
2548 if (res != -EAGAIN && res != req->result)
2549 req_set_fail_links(req);
2551 WRITE_ONCE(req->result, res);
2552 /* order with io_iopoll_complete() checking ->result */
2554 WRITE_ONCE(req->iopoll_completed, 1);
2558 * After the iocb has been issued, it's safe to be found on the poll list.
2559 * Adding the kiocb to the list AFTER submission ensures that we don't
2560 * find it from a io_iopoll_getevents() thread before the issuer is done
2561 * accessing the kiocb cookie.
2563 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2565 struct io_ring_ctx *ctx = req->ctx;
2568 * Track whether we have multiple files in our lists. This will impact
2569 * how we do polling eventually, not spinning if we're on potentially
2570 * different devices.
2572 if (list_empty(&ctx->iopoll_list)) {
2573 ctx->poll_multi_file = false;
2574 } else if (!ctx->poll_multi_file) {
2575 struct io_kiocb *list_req;
2577 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2579 if (list_req->file != req->file)
2580 ctx->poll_multi_file = true;
2584 * For fast devices, IO may have already completed. If it has, add
2585 * it to the front so we find it first.
2587 if (READ_ONCE(req->iopoll_completed))
2588 list_add(&req->inflight_entry, &ctx->iopoll_list);
2590 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2593 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2594 * task context or in io worker task context. If current task context is
2595 * sq thread, we don't need to check whether should wake up sq thread.
2597 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2598 wq_has_sleeper(&ctx->sq_data->wait))
2599 wake_up(&ctx->sq_data->wait);
2602 static inline void io_state_file_put(struct io_submit_state *state)
2604 if (state->file_refs) {
2605 fput_many(state->file, state->file_refs);
2606 state->file_refs = 0;
2611 * Get as many references to a file as we have IOs left in this submission,
2612 * assuming most submissions are for one file, or at least that each file
2613 * has more than one submission.
2615 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2620 if (state->file_refs) {
2621 if (state->fd == fd) {
2625 io_state_file_put(state);
2627 state->file = fget_many(fd, state->ios_left);
2628 if (unlikely(!state->file))
2632 state->file_refs = state->ios_left - 1;
2636 static bool io_bdev_nowait(struct block_device *bdev)
2638 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2642 * If we tracked the file through the SCM inflight mechanism, we could support
2643 * any file. For now, just ensure that anything potentially problematic is done
2646 static bool __io_file_supports_async(struct file *file, int rw)
2648 umode_t mode = file_inode(file)->i_mode;
2650 if (S_ISBLK(mode)) {
2651 if (IS_ENABLED(CONFIG_BLOCK) &&
2652 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2656 if (S_ISCHR(mode) || S_ISSOCK(mode))
2658 if (S_ISREG(mode)) {
2659 if (IS_ENABLED(CONFIG_BLOCK) &&
2660 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2661 file->f_op != &io_uring_fops)
2666 /* any ->read/write should understand O_NONBLOCK */
2667 if (file->f_flags & O_NONBLOCK)
2670 if (!(file->f_mode & FMODE_NOWAIT))
2674 return file->f_op->read_iter != NULL;
2676 return file->f_op->write_iter != NULL;
2679 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2681 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2683 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2686 return __io_file_supports_async(req->file, rw);
2689 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2691 struct io_ring_ctx *ctx = req->ctx;
2692 struct kiocb *kiocb = &req->rw.kiocb;
2693 struct file *file = req->file;
2697 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2698 req->flags |= REQ_F_ISREG;
2700 kiocb->ki_pos = READ_ONCE(sqe->off);
2701 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2702 req->flags |= REQ_F_CUR_POS;
2703 kiocb->ki_pos = file->f_pos;
2705 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2706 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2707 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2711 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2712 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2713 req->flags |= REQ_F_NOWAIT;
2715 ioprio = READ_ONCE(sqe->ioprio);
2717 ret = ioprio_check_cap(ioprio);
2721 kiocb->ki_ioprio = ioprio;
2723 kiocb->ki_ioprio = get_current_ioprio();
2725 if (ctx->flags & IORING_SETUP_IOPOLL) {
2726 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2727 !kiocb->ki_filp->f_op->iopoll)
2730 kiocb->ki_flags |= IOCB_HIPRI;
2731 kiocb->ki_complete = io_complete_rw_iopoll;
2732 req->iopoll_completed = 0;
2734 if (kiocb->ki_flags & IOCB_HIPRI)
2736 kiocb->ki_complete = io_complete_rw;
2739 req->rw.addr = READ_ONCE(sqe->addr);
2740 req->rw.len = READ_ONCE(sqe->len);
2741 req->buf_index = READ_ONCE(sqe->buf_index);
2745 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2751 case -ERESTARTNOINTR:
2752 case -ERESTARTNOHAND:
2753 case -ERESTART_RESTARTBLOCK:
2755 * We can't just restart the syscall, since previously
2756 * submitted sqes may already be in progress. Just fail this
2762 kiocb->ki_complete(kiocb, ret, 0);
2766 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2767 unsigned int issue_flags)
2769 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2770 struct io_async_rw *io = req->async_data;
2771 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2773 /* add previously done IO, if any */
2774 if (io && io->bytes_done > 0) {
2776 ret = io->bytes_done;
2778 ret += io->bytes_done;
2781 if (req->flags & REQ_F_CUR_POS)
2782 req->file->f_pos = kiocb->ki_pos;
2783 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2784 __io_complete_rw(req, ret, 0, issue_flags);
2786 io_rw_done(kiocb, ret);
2788 if (check_reissue && req->flags & REQ_F_REISSUE) {
2789 req->flags &= ~REQ_F_REISSUE;
2790 if (!io_rw_reissue(req)) {
2793 req_set_fail_links(req);
2794 if (req->flags & REQ_F_BUFFER_SELECTED)
2795 cflags = io_put_rw_kbuf(req);
2796 __io_req_complete(req, issue_flags, ret, cflags);
2801 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2803 struct io_ring_ctx *ctx = req->ctx;
2804 size_t len = req->rw.len;
2805 struct io_mapped_ubuf *imu;
2806 u16 index, buf_index = req->buf_index;
2810 if (unlikely(buf_index >= ctx->nr_user_bufs))
2812 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2813 imu = &ctx->user_bufs[index];
2814 buf_addr = req->rw.addr;
2817 if (buf_addr + len < buf_addr)
2819 /* not inside the mapped region */
2820 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2824 * May not be a start of buffer, set size appropriately
2825 * and advance us to the beginning.
2827 offset = buf_addr - imu->ubuf;
2828 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2832 * Don't use iov_iter_advance() here, as it's really slow for
2833 * using the latter parts of a big fixed buffer - it iterates
2834 * over each segment manually. We can cheat a bit here, because
2837 * 1) it's a BVEC iter, we set it up
2838 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2839 * first and last bvec
2841 * So just find our index, and adjust the iterator afterwards.
2842 * If the offset is within the first bvec (or the whole first
2843 * bvec, just use iov_iter_advance(). This makes it easier
2844 * since we can just skip the first segment, which may not
2845 * be PAGE_SIZE aligned.
2847 const struct bio_vec *bvec = imu->bvec;
2849 if (offset <= bvec->bv_len) {
2850 iov_iter_advance(iter, offset);
2852 unsigned long seg_skip;
2854 /* skip first vec */
2855 offset -= bvec->bv_len;
2856 seg_skip = 1 + (offset >> PAGE_SHIFT);
2858 iter->bvec = bvec + seg_skip;
2859 iter->nr_segs -= seg_skip;
2860 iter->count -= bvec->bv_len + offset;
2861 iter->iov_offset = offset & ~PAGE_MASK;
2868 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2871 mutex_unlock(&ctx->uring_lock);
2874 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2877 * "Normal" inline submissions always hold the uring_lock, since we
2878 * grab it from the system call. Same is true for the SQPOLL offload.
2879 * The only exception is when we've detached the request and issue it
2880 * from an async worker thread, grab the lock for that case.
2883 mutex_lock(&ctx->uring_lock);
2886 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2887 int bgid, struct io_buffer *kbuf,
2890 struct io_buffer *head;
2892 if (req->flags & REQ_F_BUFFER_SELECTED)
2895 io_ring_submit_lock(req->ctx, needs_lock);
2897 lockdep_assert_held(&req->ctx->uring_lock);
2899 head = xa_load(&req->ctx->io_buffers, bgid);
2901 if (!list_empty(&head->list)) {
2902 kbuf = list_last_entry(&head->list, struct io_buffer,
2904 list_del(&kbuf->list);
2907 xa_erase(&req->ctx->io_buffers, bgid);
2909 if (*len > kbuf->len)
2912 kbuf = ERR_PTR(-ENOBUFS);
2915 io_ring_submit_unlock(req->ctx, needs_lock);
2920 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2923 struct io_buffer *kbuf;
2926 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2927 bgid = req->buf_index;
2928 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2931 req->rw.addr = (u64) (unsigned long) kbuf;
2932 req->flags |= REQ_F_BUFFER_SELECTED;
2933 return u64_to_user_ptr(kbuf->addr);
2936 #ifdef CONFIG_COMPAT
2937 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2940 struct compat_iovec __user *uiov;
2941 compat_ssize_t clen;
2945 uiov = u64_to_user_ptr(req->rw.addr);
2946 if (!access_ok(uiov, sizeof(*uiov)))
2948 if (__get_user(clen, &uiov->iov_len))
2954 buf = io_rw_buffer_select(req, &len, needs_lock);
2956 return PTR_ERR(buf);
2957 iov[0].iov_base = buf;
2958 iov[0].iov_len = (compat_size_t) len;
2963 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2966 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2970 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2973 len = iov[0].iov_len;
2976 buf = io_rw_buffer_select(req, &len, needs_lock);
2978 return PTR_ERR(buf);
2979 iov[0].iov_base = buf;
2980 iov[0].iov_len = len;
2984 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2987 if (req->flags & REQ_F_BUFFER_SELECTED) {
2988 struct io_buffer *kbuf;
2990 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2991 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2992 iov[0].iov_len = kbuf->len;
2995 if (req->rw.len != 1)
2998 #ifdef CONFIG_COMPAT
2999 if (req->ctx->compat)
3000 return io_compat_import(req, iov, needs_lock);
3003 return __io_iov_buffer_select(req, iov, needs_lock);
3006 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3007 struct iov_iter *iter, bool needs_lock)
3009 void __user *buf = u64_to_user_ptr(req->rw.addr);
3010 size_t sqe_len = req->rw.len;
3011 u8 opcode = req->opcode;
3014 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3016 return io_import_fixed(req, rw, iter);
3019 /* buffer index only valid with fixed read/write, or buffer select */
3020 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3023 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3024 if (req->flags & REQ_F_BUFFER_SELECT) {
3025 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3027 return PTR_ERR(buf);
3028 req->rw.len = sqe_len;
3031 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3036 if (req->flags & REQ_F_BUFFER_SELECT) {
3037 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3039 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3044 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3048 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3050 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3054 * For files that don't have ->read_iter() and ->write_iter(), handle them
3055 * by looping over ->read() or ->write() manually.
3057 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3059 struct kiocb *kiocb = &req->rw.kiocb;
3060 struct file *file = req->file;
3064 * Don't support polled IO through this interface, and we can't
3065 * support non-blocking either. For the latter, this just causes
3066 * the kiocb to be handled from an async context.
3068 if (kiocb->ki_flags & IOCB_HIPRI)
3070 if (kiocb->ki_flags & IOCB_NOWAIT)
3073 while (iov_iter_count(iter)) {
3077 if (!iov_iter_is_bvec(iter)) {
3078 iovec = iov_iter_iovec(iter);
3080 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3081 iovec.iov_len = req->rw.len;
3085 nr = file->f_op->read(file, iovec.iov_base,
3086 iovec.iov_len, io_kiocb_ppos(kiocb));
3088 nr = file->f_op->write(file, iovec.iov_base,
3089 iovec.iov_len, io_kiocb_ppos(kiocb));
3098 if (nr != iovec.iov_len)
3102 iov_iter_advance(iter, nr);
3108 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3109 const struct iovec *fast_iov, struct iov_iter *iter)
3111 struct io_async_rw *rw = req->async_data;
3113 memcpy(&rw->iter, iter, sizeof(*iter));
3114 rw->free_iovec = iovec;
3116 /* can only be fixed buffers, no need to do anything */
3117 if (iov_iter_is_bvec(iter))
3120 unsigned iov_off = 0;
3122 rw->iter.iov = rw->fast_iov;
3123 if (iter->iov != fast_iov) {
3124 iov_off = iter->iov - fast_iov;
3125 rw->iter.iov += iov_off;
3127 if (rw->fast_iov != fast_iov)
3128 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3129 sizeof(struct iovec) * iter->nr_segs);
3131 req->flags |= REQ_F_NEED_CLEANUP;
3135 static inline int io_alloc_async_data(struct io_kiocb *req)
3137 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3138 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3139 return req->async_data == NULL;
3142 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3143 const struct iovec *fast_iov,
3144 struct iov_iter *iter, bool force)
3146 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3148 if (!req->async_data) {
3149 if (io_alloc_async_data(req)) {
3154 io_req_map_rw(req, iovec, fast_iov, iter);
3159 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3161 struct io_async_rw *iorw = req->async_data;
3162 struct iovec *iov = iorw->fast_iov;
3165 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3166 if (unlikely(ret < 0))
3169 iorw->bytes_done = 0;
3170 iorw->free_iovec = iov;
3172 req->flags |= REQ_F_NEED_CLEANUP;
3176 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3178 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3180 return io_prep_rw(req, sqe);
3184 * This is our waitqueue callback handler, registered through lock_page_async()
3185 * when we initially tried to do the IO with the iocb armed our waitqueue.
3186 * This gets called when the page is unlocked, and we generally expect that to
3187 * happen when the page IO is completed and the page is now uptodate. This will
3188 * queue a task_work based retry of the operation, attempting to copy the data
3189 * again. If the latter fails because the page was NOT uptodate, then we will
3190 * do a thread based blocking retry of the operation. That's the unexpected
3193 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3194 int sync, void *arg)
3196 struct wait_page_queue *wpq;
3197 struct io_kiocb *req = wait->private;
3198 struct wait_page_key *key = arg;
3200 wpq = container_of(wait, struct wait_page_queue, wait);
3202 if (!wake_page_match(wpq, key))
3205 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3206 list_del_init(&wait->entry);
3208 /* submit ref gets dropped, acquire a new one */
3210 io_req_task_queue(req);
3215 * This controls whether a given IO request should be armed for async page
3216 * based retry. If we return false here, the request is handed to the async
3217 * worker threads for retry. If we're doing buffered reads on a regular file,
3218 * we prepare a private wait_page_queue entry and retry the operation. This
3219 * will either succeed because the page is now uptodate and unlocked, or it
3220 * will register a callback when the page is unlocked at IO completion. Through
3221 * that callback, io_uring uses task_work to setup a retry of the operation.
3222 * That retry will attempt the buffered read again. The retry will generally
3223 * succeed, or in rare cases where it fails, we then fall back to using the
3224 * async worker threads for a blocking retry.
3226 static bool io_rw_should_retry(struct io_kiocb *req)
3228 struct io_async_rw *rw = req->async_data;
3229 struct wait_page_queue *wait = &rw->wpq;
3230 struct kiocb *kiocb = &req->rw.kiocb;
3232 /* never retry for NOWAIT, we just complete with -EAGAIN */
3233 if (req->flags & REQ_F_NOWAIT)
3236 /* Only for buffered IO */
3237 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3241 * just use poll if we can, and don't attempt if the fs doesn't
3242 * support callback based unlocks
3244 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3247 wait->wait.func = io_async_buf_func;
3248 wait->wait.private = req;
3249 wait->wait.flags = 0;
3250 INIT_LIST_HEAD(&wait->wait.entry);
3251 kiocb->ki_flags |= IOCB_WAITQ;
3252 kiocb->ki_flags &= ~IOCB_NOWAIT;
3253 kiocb->ki_waitq = wait;
3257 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3259 if (req->file->f_op->read_iter)
3260 return call_read_iter(req->file, &req->rw.kiocb, iter);
3261 else if (req->file->f_op->read)
3262 return loop_rw_iter(READ, req, iter);
3267 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3269 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3270 struct kiocb *kiocb = &req->rw.kiocb;
3271 struct iov_iter __iter, *iter = &__iter;
3272 struct io_async_rw *rw = req->async_data;
3273 ssize_t io_size, ret, ret2;
3274 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3280 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3284 io_size = iov_iter_count(iter);
3285 req->result = io_size;
3287 /* Ensure we clear previously set non-block flag */
3288 if (!force_nonblock)
3289 kiocb->ki_flags &= ~IOCB_NOWAIT;
3291 kiocb->ki_flags |= IOCB_NOWAIT;
3293 /* If the file doesn't support async, just async punt */
3294 if (force_nonblock && !io_file_supports_async(req, READ)) {
3295 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3296 return ret ?: -EAGAIN;
3299 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3300 if (unlikely(ret)) {
3305 ret = io_iter_do_read(req, iter);
3307 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3308 req->flags &= ~REQ_F_REISSUE;
3309 /* IOPOLL retry should happen for io-wq threads */
3310 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3312 /* no retry on NONBLOCK nor RWF_NOWAIT */
3313 if (req->flags & REQ_F_NOWAIT)
3315 /* some cases will consume bytes even on error returns */
3316 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3318 } else if (ret == -EIOCBQUEUED) {
3320 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3321 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3322 /* read all, failed, already did sync or don't want to retry */
3326 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3331 rw = req->async_data;
3332 /* now use our persistent iterator, if we aren't already */
3337 rw->bytes_done += ret;
3338 /* if we can retry, do so with the callbacks armed */
3339 if (!io_rw_should_retry(req)) {
3340 kiocb->ki_flags &= ~IOCB_WAITQ;
3345 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3346 * we get -EIOCBQUEUED, then we'll get a notification when the
3347 * desired page gets unlocked. We can also get a partial read
3348 * here, and if we do, then just retry at the new offset.
3350 ret = io_iter_do_read(req, iter);
3351 if (ret == -EIOCBQUEUED)
3353 /* we got some bytes, but not all. retry. */
3354 kiocb->ki_flags &= ~IOCB_WAITQ;
3355 } while (ret > 0 && ret < io_size);
3357 kiocb_done(kiocb, ret, issue_flags);
3359 /* it's faster to check here then delegate to kfree */
3365 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3367 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3369 return io_prep_rw(req, sqe);
3372 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3374 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3375 struct kiocb *kiocb = &req->rw.kiocb;
3376 struct iov_iter __iter, *iter = &__iter;
3377 struct io_async_rw *rw = req->async_data;
3378 ssize_t ret, ret2, io_size;
3379 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3385 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3389 io_size = iov_iter_count(iter);
3390 req->result = io_size;
3392 /* Ensure we clear previously set non-block flag */
3393 if (!force_nonblock)
3394 kiocb->ki_flags &= ~IOCB_NOWAIT;
3396 kiocb->ki_flags |= IOCB_NOWAIT;
3398 /* If the file doesn't support async, just async punt */
3399 if (force_nonblock && !io_file_supports_async(req, WRITE))
3402 /* file path doesn't support NOWAIT for non-direct_IO */
3403 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3404 (req->flags & REQ_F_ISREG))
3407 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3412 * Open-code file_start_write here to grab freeze protection,
3413 * which will be released by another thread in
3414 * io_complete_rw(). Fool lockdep by telling it the lock got
3415 * released so that it doesn't complain about the held lock when
3416 * we return to userspace.
3418 if (req->flags & REQ_F_ISREG) {
3419 sb_start_write(file_inode(req->file)->i_sb);
3420 __sb_writers_release(file_inode(req->file)->i_sb,
3423 kiocb->ki_flags |= IOCB_WRITE;
3425 if (req->file->f_op->write_iter)
3426 ret2 = call_write_iter(req->file, kiocb, iter);
3427 else if (req->file->f_op->write)
3428 ret2 = loop_rw_iter(WRITE, req, iter);
3432 if (req->flags & REQ_F_REISSUE) {
3433 req->flags &= ~REQ_F_REISSUE;
3438 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3439 * retry them without IOCB_NOWAIT.
3441 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3443 /* no retry on NONBLOCK nor RWF_NOWAIT */
3444 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3446 if (!force_nonblock || ret2 != -EAGAIN) {
3447 /* IOPOLL retry should happen for io-wq threads */
3448 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3451 kiocb_done(kiocb, ret2, issue_flags);
3454 /* some cases will consume bytes even on error returns */
3455 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3456 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3457 return ret ?: -EAGAIN;
3460 /* it's reportedly faster than delegating the null check to kfree() */
3466 static int io_renameat_prep(struct io_kiocb *req,
3467 const struct io_uring_sqe *sqe)
3469 struct io_rename *ren = &req->rename;
3470 const char __user *oldf, *newf;
3472 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3475 ren->old_dfd = READ_ONCE(sqe->fd);
3476 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3477 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3478 ren->new_dfd = READ_ONCE(sqe->len);
3479 ren->flags = READ_ONCE(sqe->rename_flags);
3481 ren->oldpath = getname(oldf);
3482 if (IS_ERR(ren->oldpath))
3483 return PTR_ERR(ren->oldpath);
3485 ren->newpath = getname(newf);
3486 if (IS_ERR(ren->newpath)) {
3487 putname(ren->oldpath);
3488 return PTR_ERR(ren->newpath);
3491 req->flags |= REQ_F_NEED_CLEANUP;
3495 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3497 struct io_rename *ren = &req->rename;
3500 if (issue_flags & IO_URING_F_NONBLOCK)
3503 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3504 ren->newpath, ren->flags);
3506 req->flags &= ~REQ_F_NEED_CLEANUP;
3508 req_set_fail_links(req);
3509 io_req_complete(req, ret);
3513 static int io_unlinkat_prep(struct io_kiocb *req,
3514 const struct io_uring_sqe *sqe)
3516 struct io_unlink *un = &req->unlink;
3517 const char __user *fname;
3519 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3522 un->dfd = READ_ONCE(sqe->fd);
3524 un->flags = READ_ONCE(sqe->unlink_flags);
3525 if (un->flags & ~AT_REMOVEDIR)
3528 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3529 un->filename = getname(fname);
3530 if (IS_ERR(un->filename))
3531 return PTR_ERR(un->filename);
3533 req->flags |= REQ_F_NEED_CLEANUP;
3537 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3539 struct io_unlink *un = &req->unlink;
3542 if (issue_flags & IO_URING_F_NONBLOCK)
3545 if (un->flags & AT_REMOVEDIR)
3546 ret = do_rmdir(un->dfd, un->filename);
3548 ret = do_unlinkat(un->dfd, un->filename);
3550 req->flags &= ~REQ_F_NEED_CLEANUP;
3552 req_set_fail_links(req);
3553 io_req_complete(req, ret);
3557 static int io_shutdown_prep(struct io_kiocb *req,
3558 const struct io_uring_sqe *sqe)
3560 #if defined(CONFIG_NET)
3561 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3563 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3567 req->shutdown.how = READ_ONCE(sqe->len);
3574 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3576 #if defined(CONFIG_NET)
3577 struct socket *sock;
3580 if (issue_flags & IO_URING_F_NONBLOCK)
3583 sock = sock_from_file(req->file);
3584 if (unlikely(!sock))
3587 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3589 req_set_fail_links(req);
3590 io_req_complete(req, ret);
3597 static int __io_splice_prep(struct io_kiocb *req,
3598 const struct io_uring_sqe *sqe)
3600 struct io_splice* sp = &req->splice;
3601 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3603 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3607 sp->len = READ_ONCE(sqe->len);
3608 sp->flags = READ_ONCE(sqe->splice_flags);
3610 if (unlikely(sp->flags & ~valid_flags))
3613 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3614 (sp->flags & SPLICE_F_FD_IN_FIXED));
3617 req->flags |= REQ_F_NEED_CLEANUP;
3619 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3621 * Splice operation will be punted aync, and here need to
3622 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3624 req->work.flags |= IO_WQ_WORK_UNBOUND;
3630 static int io_tee_prep(struct io_kiocb *req,
3631 const struct io_uring_sqe *sqe)
3633 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3635 return __io_splice_prep(req, sqe);
3638 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3640 struct io_splice *sp = &req->splice;
3641 struct file *in = sp->file_in;
3642 struct file *out = sp->file_out;
3643 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3646 if (issue_flags & IO_URING_F_NONBLOCK)
3649 ret = do_tee(in, out, sp->len, flags);
3651 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3652 req->flags &= ~REQ_F_NEED_CLEANUP;
3655 req_set_fail_links(req);
3656 io_req_complete(req, ret);
3660 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3662 struct io_splice* sp = &req->splice;
3664 sp->off_in = READ_ONCE(sqe->splice_off_in);
3665 sp->off_out = READ_ONCE(sqe->off);
3666 return __io_splice_prep(req, sqe);
3669 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3671 struct io_splice *sp = &req->splice;
3672 struct file *in = sp->file_in;
3673 struct file *out = sp->file_out;
3674 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3675 loff_t *poff_in, *poff_out;
3678 if (issue_flags & IO_URING_F_NONBLOCK)
3681 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3682 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3685 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3687 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3688 req->flags &= ~REQ_F_NEED_CLEANUP;
3691 req_set_fail_links(req);
3692 io_req_complete(req, ret);
3697 * IORING_OP_NOP just posts a completion event, nothing else.
3699 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3701 struct io_ring_ctx *ctx = req->ctx;
3703 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3706 __io_req_complete(req, issue_flags, 0, 0);
3710 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3712 struct io_ring_ctx *ctx = req->ctx;
3717 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3719 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3722 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3723 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3726 req->sync.off = READ_ONCE(sqe->off);
3727 req->sync.len = READ_ONCE(sqe->len);
3731 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3733 loff_t end = req->sync.off + req->sync.len;
3736 /* fsync always requires a blocking context */
3737 if (issue_flags & IO_URING_F_NONBLOCK)
3740 ret = vfs_fsync_range(req->file, req->sync.off,
3741 end > 0 ? end : LLONG_MAX,
3742 req->sync.flags & IORING_FSYNC_DATASYNC);
3744 req_set_fail_links(req);
3745 io_req_complete(req, ret);
3749 static int io_fallocate_prep(struct io_kiocb *req,
3750 const struct io_uring_sqe *sqe)
3752 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3754 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3757 req->sync.off = READ_ONCE(sqe->off);
3758 req->sync.len = READ_ONCE(sqe->addr);
3759 req->sync.mode = READ_ONCE(sqe->len);
3763 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3767 /* fallocate always requiring blocking context */
3768 if (issue_flags & IO_URING_F_NONBLOCK)
3770 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3773 req_set_fail_links(req);
3774 io_req_complete(req, ret);
3778 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3780 const char __user *fname;
3783 if (unlikely(sqe->ioprio || sqe->buf_index))
3785 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3788 /* open.how should be already initialised */
3789 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3790 req->open.how.flags |= O_LARGEFILE;
3792 req->open.dfd = READ_ONCE(sqe->fd);
3793 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3794 req->open.filename = getname(fname);
3795 if (IS_ERR(req->open.filename)) {
3796 ret = PTR_ERR(req->open.filename);
3797 req->open.filename = NULL;
3800 req->open.nofile = rlimit(RLIMIT_NOFILE);
3801 req->flags |= REQ_F_NEED_CLEANUP;
3805 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3809 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3811 mode = READ_ONCE(sqe->len);
3812 flags = READ_ONCE(sqe->open_flags);
3813 req->open.how = build_open_how(flags, mode);
3814 return __io_openat_prep(req, sqe);
3817 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3819 struct open_how __user *how;
3823 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3825 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3826 len = READ_ONCE(sqe->len);
3827 if (len < OPEN_HOW_SIZE_VER0)
3830 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3835 return __io_openat_prep(req, sqe);
3838 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3840 struct open_flags op;
3843 bool resolve_nonblock;
3846 ret = build_open_flags(&req->open.how, &op);
3849 nonblock_set = op.open_flag & O_NONBLOCK;
3850 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3851 if (issue_flags & IO_URING_F_NONBLOCK) {
3853 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3854 * it'll always -EAGAIN
3856 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3858 op.lookup_flags |= LOOKUP_CACHED;
3859 op.open_flag |= O_NONBLOCK;
3862 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3866 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3867 /* only retry if RESOLVE_CACHED wasn't already set by application */
3868 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3869 file == ERR_PTR(-EAGAIN)) {
3871 * We could hang on to this 'fd', but seems like marginal
3872 * gain for something that is now known to be a slower path.
3873 * So just put it, and we'll get a new one when we retry.
3881 ret = PTR_ERR(file);
3883 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3884 file->f_flags &= ~O_NONBLOCK;
3885 fsnotify_open(file);
3886 fd_install(ret, file);
3889 putname(req->open.filename);
3890 req->flags &= ~REQ_F_NEED_CLEANUP;
3892 req_set_fail_links(req);
3893 io_req_complete(req, ret);
3897 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3899 return io_openat2(req, issue_flags);
3902 static int io_remove_buffers_prep(struct io_kiocb *req,
3903 const struct io_uring_sqe *sqe)
3905 struct io_provide_buf *p = &req->pbuf;
3908 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3911 tmp = READ_ONCE(sqe->fd);
3912 if (!tmp || tmp > USHRT_MAX)
3915 memset(p, 0, sizeof(*p));
3917 p->bgid = READ_ONCE(sqe->buf_group);
3921 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3922 int bgid, unsigned nbufs)
3926 /* shouldn't happen */
3930 /* the head kbuf is the list itself */
3931 while (!list_empty(&buf->list)) {
3932 struct io_buffer *nxt;
3934 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3935 list_del(&nxt->list);
3942 xa_erase(&ctx->io_buffers, bgid);
3947 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3949 struct io_provide_buf *p = &req->pbuf;
3950 struct io_ring_ctx *ctx = req->ctx;
3951 struct io_buffer *head;
3953 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3955 io_ring_submit_lock(ctx, !force_nonblock);
3957 lockdep_assert_held(&ctx->uring_lock);
3960 head = xa_load(&ctx->io_buffers, p->bgid);
3962 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3964 req_set_fail_links(req);
3966 /* complete before unlock, IOPOLL may need the lock */
3967 __io_req_complete(req, issue_flags, ret, 0);
3968 io_ring_submit_unlock(ctx, !force_nonblock);
3972 static int io_provide_buffers_prep(struct io_kiocb *req,
3973 const struct io_uring_sqe *sqe)
3976 struct io_provide_buf *p = &req->pbuf;
3979 if (sqe->ioprio || sqe->rw_flags)
3982 tmp = READ_ONCE(sqe->fd);
3983 if (!tmp || tmp > USHRT_MAX)
3986 p->addr = READ_ONCE(sqe->addr);
3987 p->len = READ_ONCE(sqe->len);
3989 size = (unsigned long)p->len * p->nbufs;
3990 if (!access_ok(u64_to_user_ptr(p->addr), size))
3993 p->bgid = READ_ONCE(sqe->buf_group);
3994 tmp = READ_ONCE(sqe->off);
3995 if (tmp > USHRT_MAX)
4001 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4003 struct io_buffer *buf;
4004 u64 addr = pbuf->addr;
4005 int i, bid = pbuf->bid;
4007 for (i = 0; i < pbuf->nbufs; i++) {
4008 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4013 buf->len = pbuf->len;
4018 INIT_LIST_HEAD(&buf->list);
4021 list_add_tail(&buf->list, &(*head)->list);
4025 return i ? i : -ENOMEM;
4028 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4030 struct io_provide_buf *p = &req->pbuf;
4031 struct io_ring_ctx *ctx = req->ctx;
4032 struct io_buffer *head, *list;
4034 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4036 io_ring_submit_lock(ctx, !force_nonblock);
4038 lockdep_assert_held(&ctx->uring_lock);
4040 list = head = xa_load(&ctx->io_buffers, p->bgid);
4042 ret = io_add_buffers(p, &head);
4043 if (ret >= 0 && !list) {
4044 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4046 __io_remove_buffers(ctx, head, p->bgid, -1U);
4049 req_set_fail_links(req);
4050 /* complete before unlock, IOPOLL may need the lock */
4051 __io_req_complete(req, issue_flags, ret, 0);
4052 io_ring_submit_unlock(ctx, !force_nonblock);
4056 static int io_epoll_ctl_prep(struct io_kiocb *req,
4057 const struct io_uring_sqe *sqe)
4059 #if defined(CONFIG_EPOLL)
4060 if (sqe->ioprio || sqe->buf_index)
4062 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4065 req->epoll.epfd = READ_ONCE(sqe->fd);
4066 req->epoll.op = READ_ONCE(sqe->len);
4067 req->epoll.fd = READ_ONCE(sqe->off);
4069 if (ep_op_has_event(req->epoll.op)) {
4070 struct epoll_event __user *ev;
4072 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4073 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4083 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4085 #if defined(CONFIG_EPOLL)
4086 struct io_epoll *ie = &req->epoll;
4088 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4090 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4091 if (force_nonblock && ret == -EAGAIN)
4095 req_set_fail_links(req);
4096 __io_req_complete(req, issue_flags, ret, 0);
4103 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4105 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4106 if (sqe->ioprio || sqe->buf_index || sqe->off)
4108 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4111 req->madvise.addr = READ_ONCE(sqe->addr);
4112 req->madvise.len = READ_ONCE(sqe->len);
4113 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4120 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4122 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4123 struct io_madvise *ma = &req->madvise;
4126 if (issue_flags & IO_URING_F_NONBLOCK)
4129 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4131 req_set_fail_links(req);
4132 io_req_complete(req, ret);
4139 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4141 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4143 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4146 req->fadvise.offset = READ_ONCE(sqe->off);
4147 req->fadvise.len = READ_ONCE(sqe->len);
4148 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4152 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4154 struct io_fadvise *fa = &req->fadvise;
4157 if (issue_flags & IO_URING_F_NONBLOCK) {
4158 switch (fa->advice) {
4159 case POSIX_FADV_NORMAL:
4160 case POSIX_FADV_RANDOM:
4161 case POSIX_FADV_SEQUENTIAL:
4168 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4170 req_set_fail_links(req);
4171 io_req_complete(req, ret);
4175 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4177 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4179 if (sqe->ioprio || sqe->buf_index)
4181 if (req->flags & REQ_F_FIXED_FILE)
4184 req->statx.dfd = READ_ONCE(sqe->fd);
4185 req->statx.mask = READ_ONCE(sqe->len);
4186 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4187 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4188 req->statx.flags = READ_ONCE(sqe->statx_flags);
4193 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4195 struct io_statx *ctx = &req->statx;
4198 if (issue_flags & IO_URING_F_NONBLOCK) {
4199 /* only need file table for an actual valid fd */
4200 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4201 req->flags |= REQ_F_NO_FILE_TABLE;
4205 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4209 req_set_fail_links(req);
4210 io_req_complete(req, ret);
4214 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4216 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4218 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4219 sqe->rw_flags || sqe->buf_index)
4221 if (req->flags & REQ_F_FIXED_FILE)
4224 req->close.fd = READ_ONCE(sqe->fd);
4228 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4230 struct files_struct *files = current->files;
4231 struct io_close *close = &req->close;
4232 struct fdtable *fdt;
4238 spin_lock(&files->file_lock);
4239 fdt = files_fdtable(files);
4240 if (close->fd >= fdt->max_fds) {
4241 spin_unlock(&files->file_lock);
4244 file = fdt->fd[close->fd];
4246 spin_unlock(&files->file_lock);
4250 if (file->f_op == &io_uring_fops) {
4251 spin_unlock(&files->file_lock);
4256 /* if the file has a flush method, be safe and punt to async */
4257 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4258 spin_unlock(&files->file_lock);
4262 ret = __close_fd_get_file(close->fd, &file);
4263 spin_unlock(&files->file_lock);
4270 /* No ->flush() or already async, safely close from here */
4271 ret = filp_close(file, current->files);
4274 req_set_fail_links(req);
4277 __io_req_complete(req, issue_flags, ret, 0);
4281 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4283 struct io_ring_ctx *ctx = req->ctx;
4285 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4287 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4290 req->sync.off = READ_ONCE(sqe->off);
4291 req->sync.len = READ_ONCE(sqe->len);
4292 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4296 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4300 /* sync_file_range always requires a blocking context */
4301 if (issue_flags & IO_URING_F_NONBLOCK)
4304 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4307 req_set_fail_links(req);
4308 io_req_complete(req, ret);
4312 #if defined(CONFIG_NET)
4313 static int io_setup_async_msg(struct io_kiocb *req,
4314 struct io_async_msghdr *kmsg)
4316 struct io_async_msghdr *async_msg = req->async_data;
4320 if (io_alloc_async_data(req)) {
4321 kfree(kmsg->free_iov);
4324 async_msg = req->async_data;
4325 req->flags |= REQ_F_NEED_CLEANUP;
4326 memcpy(async_msg, kmsg, sizeof(*kmsg));
4327 async_msg->msg.msg_name = &async_msg->addr;
4328 /* if were using fast_iov, set it to the new one */
4329 if (!async_msg->free_iov)
4330 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4335 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4336 struct io_async_msghdr *iomsg)
4338 iomsg->msg.msg_name = &iomsg->addr;
4339 iomsg->free_iov = iomsg->fast_iov;
4340 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4341 req->sr_msg.msg_flags, &iomsg->free_iov);
4344 static int io_sendmsg_prep_async(struct io_kiocb *req)
4348 ret = io_sendmsg_copy_hdr(req, req->async_data);
4350 req->flags |= REQ_F_NEED_CLEANUP;
4354 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4356 struct io_sr_msg *sr = &req->sr_msg;
4358 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4361 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4362 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4363 sr->len = READ_ONCE(sqe->len);
4365 #ifdef CONFIG_COMPAT
4366 if (req->ctx->compat)
4367 sr->msg_flags |= MSG_CMSG_COMPAT;
4372 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4374 struct io_async_msghdr iomsg, *kmsg;
4375 struct socket *sock;
4380 sock = sock_from_file(req->file);
4381 if (unlikely(!sock))
4384 kmsg = req->async_data;
4386 ret = io_sendmsg_copy_hdr(req, &iomsg);
4392 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4393 if (flags & MSG_DONTWAIT)
4394 req->flags |= REQ_F_NOWAIT;
4395 else if (issue_flags & IO_URING_F_NONBLOCK)
4396 flags |= MSG_DONTWAIT;
4398 if (flags & MSG_WAITALL)
4399 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4401 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4402 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4403 return io_setup_async_msg(req, kmsg);
4404 if (ret == -ERESTARTSYS)
4407 /* fast path, check for non-NULL to avoid function call */
4409 kfree(kmsg->free_iov);
4410 req->flags &= ~REQ_F_NEED_CLEANUP;
4412 req_set_fail_links(req);
4413 __io_req_complete(req, issue_flags, ret, 0);
4417 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4419 struct io_sr_msg *sr = &req->sr_msg;
4422 struct socket *sock;
4427 sock = sock_from_file(req->file);
4428 if (unlikely(!sock))
4431 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4435 msg.msg_name = NULL;
4436 msg.msg_control = NULL;
4437 msg.msg_controllen = 0;
4438 msg.msg_namelen = 0;
4440 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4441 if (flags & MSG_DONTWAIT)
4442 req->flags |= REQ_F_NOWAIT;
4443 else if (issue_flags & IO_URING_F_NONBLOCK)
4444 flags |= MSG_DONTWAIT;
4446 if (flags & MSG_WAITALL)
4447 min_ret = iov_iter_count(&msg.msg_iter);
4449 msg.msg_flags = flags;
4450 ret = sock_sendmsg(sock, &msg);
4451 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4453 if (ret == -ERESTARTSYS)
4457 req_set_fail_links(req);
4458 __io_req_complete(req, issue_flags, ret, 0);
4462 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4463 struct io_async_msghdr *iomsg)
4465 struct io_sr_msg *sr = &req->sr_msg;
4466 struct iovec __user *uiov;
4470 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4471 &iomsg->uaddr, &uiov, &iov_len);
4475 if (req->flags & REQ_F_BUFFER_SELECT) {
4478 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4480 sr->len = iomsg->fast_iov[0].iov_len;
4481 iomsg->free_iov = NULL;
4483 iomsg->free_iov = iomsg->fast_iov;
4484 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4485 &iomsg->free_iov, &iomsg->msg.msg_iter,
4494 #ifdef CONFIG_COMPAT
4495 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4496 struct io_async_msghdr *iomsg)
4498 struct compat_msghdr __user *msg_compat;
4499 struct io_sr_msg *sr = &req->sr_msg;
4500 struct compat_iovec __user *uiov;
4505 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4506 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4511 uiov = compat_ptr(ptr);
4512 if (req->flags & REQ_F_BUFFER_SELECT) {
4513 compat_ssize_t clen;
4517 if (!access_ok(uiov, sizeof(*uiov)))
4519 if (__get_user(clen, &uiov->iov_len))
4524 iomsg->free_iov = NULL;
4526 iomsg->free_iov = iomsg->fast_iov;
4527 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4528 UIO_FASTIOV, &iomsg->free_iov,
4529 &iomsg->msg.msg_iter, true);
4538 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4539 struct io_async_msghdr *iomsg)
4541 iomsg->msg.msg_name = &iomsg->addr;
4543 #ifdef CONFIG_COMPAT
4544 if (req->ctx->compat)
4545 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4548 return __io_recvmsg_copy_hdr(req, iomsg);
4551 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4554 struct io_sr_msg *sr = &req->sr_msg;
4555 struct io_buffer *kbuf;
4557 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4562 req->flags |= REQ_F_BUFFER_SELECTED;
4566 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4568 return io_put_kbuf(req, req->sr_msg.kbuf);
4571 static int io_recvmsg_prep_async(struct io_kiocb *req)
4575 ret = io_recvmsg_copy_hdr(req, req->async_data);
4577 req->flags |= REQ_F_NEED_CLEANUP;
4581 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4583 struct io_sr_msg *sr = &req->sr_msg;
4585 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4588 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4589 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4590 sr->len = READ_ONCE(sqe->len);
4591 sr->bgid = READ_ONCE(sqe->buf_group);
4593 #ifdef CONFIG_COMPAT
4594 if (req->ctx->compat)
4595 sr->msg_flags |= MSG_CMSG_COMPAT;
4600 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4602 struct io_async_msghdr iomsg, *kmsg;
4603 struct socket *sock;
4604 struct io_buffer *kbuf;
4607 int ret, cflags = 0;
4608 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4610 sock = sock_from_file(req->file);
4611 if (unlikely(!sock))
4614 kmsg = req->async_data;
4616 ret = io_recvmsg_copy_hdr(req, &iomsg);
4622 if (req->flags & REQ_F_BUFFER_SELECT) {
4623 kbuf = io_recv_buffer_select(req, !force_nonblock);
4625 return PTR_ERR(kbuf);
4626 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4627 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4628 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4629 1, req->sr_msg.len);
4632 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4633 if (flags & MSG_DONTWAIT)
4634 req->flags |= REQ_F_NOWAIT;
4635 else if (force_nonblock)
4636 flags |= MSG_DONTWAIT;
4638 if (flags & MSG_WAITALL)
4639 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4641 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4642 kmsg->uaddr, flags);
4643 if (force_nonblock && ret == -EAGAIN)
4644 return io_setup_async_msg(req, kmsg);
4645 if (ret == -ERESTARTSYS)
4648 if (req->flags & REQ_F_BUFFER_SELECTED)
4649 cflags = io_put_recv_kbuf(req);
4650 /* fast path, check for non-NULL to avoid function call */
4652 kfree(kmsg->free_iov);
4653 req->flags &= ~REQ_F_NEED_CLEANUP;
4654 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4655 req_set_fail_links(req);
4656 __io_req_complete(req, issue_flags, ret, cflags);
4660 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4662 struct io_buffer *kbuf;
4663 struct io_sr_msg *sr = &req->sr_msg;
4665 void __user *buf = sr->buf;
4666 struct socket *sock;
4670 int ret, cflags = 0;
4671 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4673 sock = sock_from_file(req->file);
4674 if (unlikely(!sock))
4677 if (req->flags & REQ_F_BUFFER_SELECT) {
4678 kbuf = io_recv_buffer_select(req, !force_nonblock);
4680 return PTR_ERR(kbuf);
4681 buf = u64_to_user_ptr(kbuf->addr);
4684 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4688 msg.msg_name = NULL;
4689 msg.msg_control = NULL;
4690 msg.msg_controllen = 0;
4691 msg.msg_namelen = 0;
4692 msg.msg_iocb = NULL;
4695 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4696 if (flags & MSG_DONTWAIT)
4697 req->flags |= REQ_F_NOWAIT;
4698 else if (force_nonblock)
4699 flags |= MSG_DONTWAIT;
4701 if (flags & MSG_WAITALL)
4702 min_ret = iov_iter_count(&msg.msg_iter);
4704 ret = sock_recvmsg(sock, &msg, flags);
4705 if (force_nonblock && ret == -EAGAIN)
4707 if (ret == -ERESTARTSYS)
4710 if (req->flags & REQ_F_BUFFER_SELECTED)
4711 cflags = io_put_recv_kbuf(req);
4712 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4713 req_set_fail_links(req);
4714 __io_req_complete(req, issue_flags, ret, cflags);
4718 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4720 struct io_accept *accept = &req->accept;
4722 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4724 if (sqe->ioprio || sqe->len || sqe->buf_index)
4727 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4728 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4729 accept->flags = READ_ONCE(sqe->accept_flags);
4730 accept->nofile = rlimit(RLIMIT_NOFILE);
4734 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4736 struct io_accept *accept = &req->accept;
4737 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4738 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4741 if (req->file->f_flags & O_NONBLOCK)
4742 req->flags |= REQ_F_NOWAIT;
4744 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4745 accept->addr_len, accept->flags,
4747 if (ret == -EAGAIN && force_nonblock)
4750 if (ret == -ERESTARTSYS)
4752 req_set_fail_links(req);
4754 __io_req_complete(req, issue_flags, ret, 0);
4758 static int io_connect_prep_async(struct io_kiocb *req)
4760 struct io_async_connect *io = req->async_data;
4761 struct io_connect *conn = &req->connect;
4763 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4766 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4768 struct io_connect *conn = &req->connect;
4770 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4772 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4775 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4776 conn->addr_len = READ_ONCE(sqe->addr2);
4780 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4782 struct io_async_connect __io, *io;
4783 unsigned file_flags;
4785 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4787 if (req->async_data) {
4788 io = req->async_data;
4790 ret = move_addr_to_kernel(req->connect.addr,
4791 req->connect.addr_len,
4798 file_flags = force_nonblock ? O_NONBLOCK : 0;
4800 ret = __sys_connect_file(req->file, &io->address,
4801 req->connect.addr_len, file_flags);
4802 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4803 if (req->async_data)
4805 if (io_alloc_async_data(req)) {
4809 memcpy(req->async_data, &__io, sizeof(__io));
4812 if (ret == -ERESTARTSYS)
4816 req_set_fail_links(req);
4817 __io_req_complete(req, issue_flags, ret, 0);
4820 #else /* !CONFIG_NET */
4821 #define IO_NETOP_FN(op) \
4822 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4824 return -EOPNOTSUPP; \
4827 #define IO_NETOP_PREP(op) \
4829 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4831 return -EOPNOTSUPP; \
4834 #define IO_NETOP_PREP_ASYNC(op) \
4836 static int io_##op##_prep_async(struct io_kiocb *req) \
4838 return -EOPNOTSUPP; \
4841 IO_NETOP_PREP_ASYNC(sendmsg);
4842 IO_NETOP_PREP_ASYNC(recvmsg);
4843 IO_NETOP_PREP_ASYNC(connect);
4844 IO_NETOP_PREP(accept);
4847 #endif /* CONFIG_NET */
4849 struct io_poll_table {
4850 struct poll_table_struct pt;
4851 struct io_kiocb *req;
4855 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4856 __poll_t mask, task_work_func_t func)
4860 /* for instances that support it check for an event match first: */
4861 if (mask && !(mask & poll->events))
4864 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4866 list_del_init(&poll->wait.entry);
4869 req->task_work.func = func;
4872 * If this fails, then the task is exiting. When a task exits, the
4873 * work gets canceled, so just cancel this request as well instead
4874 * of executing it. We can't safely execute it anyway, as we may not
4875 * have the needed state needed for it anyway.
4877 ret = io_req_task_work_add(req);
4878 if (unlikely(ret)) {
4879 WRITE_ONCE(poll->canceled, true);
4880 io_req_task_work_add_fallback(req, func);
4885 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4886 __acquires(&req->ctx->completion_lock)
4888 struct io_ring_ctx *ctx = req->ctx;
4890 if (!req->result && !READ_ONCE(poll->canceled)) {
4891 struct poll_table_struct pt = { ._key = poll->events };
4893 req->result = vfs_poll(req->file, &pt) & poll->events;
4896 spin_lock_irq(&ctx->completion_lock);
4897 if (!req->result && !READ_ONCE(poll->canceled)) {
4898 add_wait_queue(poll->head, &poll->wait);
4905 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4907 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4908 if (req->opcode == IORING_OP_POLL_ADD)
4909 return req->async_data;
4910 return req->apoll->double_poll;
4913 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4915 if (req->opcode == IORING_OP_POLL_ADD)
4917 return &req->apoll->poll;
4920 static void io_poll_remove_double(struct io_kiocb *req)
4922 struct io_poll_iocb *poll = io_poll_get_double(req);
4924 lockdep_assert_held(&req->ctx->completion_lock);
4926 if (poll && poll->head) {
4927 struct wait_queue_head *head = poll->head;
4929 spin_lock(&head->lock);
4930 list_del_init(&poll->wait.entry);
4931 if (poll->wait.private)
4934 spin_unlock(&head->lock);
4938 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4940 struct io_ring_ctx *ctx = req->ctx;
4942 if (!error && req->poll.canceled)
4945 io_poll_remove_double(req);
4946 req->poll.done = true;
4947 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4948 io_commit_cqring(ctx);
4951 static void io_poll_task_func(struct callback_head *cb)
4953 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4954 struct io_ring_ctx *ctx = req->ctx;
4955 struct io_kiocb *nxt;
4957 if (io_poll_rewait(req, &req->poll)) {
4958 spin_unlock_irq(&ctx->completion_lock);
4960 hash_del(&req->hash_node);
4961 io_poll_complete(req, req->result, 0);
4962 spin_unlock_irq(&ctx->completion_lock);
4964 nxt = io_put_req_find_next(req);
4965 io_cqring_ev_posted(ctx);
4967 __io_req_task_submit(nxt);
4971 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4972 int sync, void *key)
4974 struct io_kiocb *req = wait->private;
4975 struct io_poll_iocb *poll = io_poll_get_single(req);
4976 __poll_t mask = key_to_poll(key);
4978 /* for instances that support it check for an event match first: */
4979 if (mask && !(mask & poll->events))
4982 list_del_init(&wait->entry);
4984 if (poll && poll->head) {
4987 spin_lock(&poll->head->lock);
4988 done = list_empty(&poll->wait.entry);
4990 list_del_init(&poll->wait.entry);
4991 /* make sure double remove sees this as being gone */
4992 wait->private = NULL;
4993 spin_unlock(&poll->head->lock);
4995 /* use wait func handler, so it matches the rq type */
4996 poll->wait.func(&poll->wait, mode, sync, key);
5003 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5004 wait_queue_func_t wake_func)
5008 poll->canceled = false;
5009 poll->events = events;
5010 INIT_LIST_HEAD(&poll->wait.entry);
5011 init_waitqueue_func_entry(&poll->wait, wake_func);
5014 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5015 struct wait_queue_head *head,
5016 struct io_poll_iocb **poll_ptr)
5018 struct io_kiocb *req = pt->req;
5021 * If poll->head is already set, it's because the file being polled
5022 * uses multiple waitqueues for poll handling (eg one for read, one
5023 * for write). Setup a separate io_poll_iocb if this happens.
5025 if (unlikely(poll->head)) {
5026 struct io_poll_iocb *poll_one = poll;
5028 /* already have a 2nd entry, fail a third attempt */
5030 pt->error = -EINVAL;
5033 /* double add on the same waitqueue head, ignore */
5034 if (poll->head == head)
5036 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5038 pt->error = -ENOMEM;
5041 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5043 poll->wait.private = req;
5050 if (poll->events & EPOLLEXCLUSIVE)
5051 add_wait_queue_exclusive(head, &poll->wait);
5053 add_wait_queue(head, &poll->wait);
5056 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5057 struct poll_table_struct *p)
5059 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5060 struct async_poll *apoll = pt->req->apoll;
5062 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5065 static void io_async_task_func(struct callback_head *cb)
5067 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5068 struct async_poll *apoll = req->apoll;
5069 struct io_ring_ctx *ctx = req->ctx;
5071 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5073 if (io_poll_rewait(req, &apoll->poll)) {
5074 spin_unlock_irq(&ctx->completion_lock);
5078 /* If req is still hashed, it cannot have been canceled. Don't check. */
5079 if (hash_hashed(&req->hash_node))
5080 hash_del(&req->hash_node);
5082 io_poll_remove_double(req);
5083 spin_unlock_irq(&ctx->completion_lock);
5085 if (!READ_ONCE(apoll->poll.canceled))
5086 __io_req_task_submit(req);
5088 io_req_complete_failed(req, -ECANCELED);
5090 kfree(apoll->double_poll);
5094 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5097 struct io_kiocb *req = wait->private;
5098 struct io_poll_iocb *poll = &req->apoll->poll;
5100 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5103 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5106 static void io_poll_req_insert(struct io_kiocb *req)
5108 struct io_ring_ctx *ctx = req->ctx;
5109 struct hlist_head *list;
5111 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5112 hlist_add_head(&req->hash_node, list);
5115 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5116 struct io_poll_iocb *poll,
5117 struct io_poll_table *ipt, __poll_t mask,
5118 wait_queue_func_t wake_func)
5119 __acquires(&ctx->completion_lock)
5121 struct io_ring_ctx *ctx = req->ctx;
5122 bool cancel = false;
5124 INIT_HLIST_NODE(&req->hash_node);
5125 io_init_poll_iocb(poll, mask, wake_func);
5126 poll->file = req->file;
5127 poll->wait.private = req;
5129 ipt->pt._key = mask;
5131 ipt->error = -EINVAL;
5133 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5135 spin_lock_irq(&ctx->completion_lock);
5136 if (likely(poll->head)) {
5137 spin_lock(&poll->head->lock);
5138 if (unlikely(list_empty(&poll->wait.entry))) {
5144 if (mask || ipt->error)
5145 list_del_init(&poll->wait.entry);
5147 WRITE_ONCE(poll->canceled, true);
5148 else if (!poll->done) /* actually waiting for an event */
5149 io_poll_req_insert(req);
5150 spin_unlock(&poll->head->lock);
5156 static bool io_arm_poll_handler(struct io_kiocb *req)
5158 const struct io_op_def *def = &io_op_defs[req->opcode];
5159 struct io_ring_ctx *ctx = req->ctx;
5160 struct async_poll *apoll;
5161 struct io_poll_table ipt;
5165 if (!req->file || !file_can_poll(req->file))
5167 if (req->flags & REQ_F_POLLED)
5171 else if (def->pollout)
5175 /* if we can't nonblock try, then no point in arming a poll handler */
5176 if (!io_file_supports_async(req, rw))
5179 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5180 if (unlikely(!apoll))
5182 apoll->double_poll = NULL;
5184 req->flags |= REQ_F_POLLED;
5189 mask |= POLLIN | POLLRDNORM;
5191 mask |= POLLOUT | POLLWRNORM;
5193 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5194 if ((req->opcode == IORING_OP_RECVMSG) &&
5195 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5198 mask |= POLLERR | POLLPRI;
5200 ipt.pt._qproc = io_async_queue_proc;
5202 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5204 if (ret || ipt.error) {
5205 io_poll_remove_double(req);
5206 spin_unlock_irq(&ctx->completion_lock);
5207 kfree(apoll->double_poll);
5211 spin_unlock_irq(&ctx->completion_lock);
5212 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5213 apoll->poll.events);
5217 static bool __io_poll_remove_one(struct io_kiocb *req,
5218 struct io_poll_iocb *poll)
5220 bool do_complete = false;
5222 spin_lock(&poll->head->lock);
5223 WRITE_ONCE(poll->canceled, true);
5224 if (!list_empty(&poll->wait.entry)) {
5225 list_del_init(&poll->wait.entry);
5228 spin_unlock(&poll->head->lock);
5229 hash_del(&req->hash_node);
5233 static bool io_poll_remove_one(struct io_kiocb *req)
5237 io_poll_remove_double(req);
5239 if (req->opcode == IORING_OP_POLL_ADD) {
5240 do_complete = __io_poll_remove_one(req, &req->poll);
5242 struct async_poll *apoll = req->apoll;
5244 /* non-poll requests have submit ref still */
5245 do_complete = __io_poll_remove_one(req, &apoll->poll);
5248 kfree(apoll->double_poll);
5254 io_cqring_fill_event(req, -ECANCELED);
5255 io_commit_cqring(req->ctx);
5256 req_set_fail_links(req);
5257 io_put_req_deferred(req, 1);
5264 * Returns true if we found and killed one or more poll requests
5266 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5267 struct files_struct *files)
5269 struct hlist_node *tmp;
5270 struct io_kiocb *req;
5273 spin_lock_irq(&ctx->completion_lock);
5274 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5275 struct hlist_head *list;
5277 list = &ctx->cancel_hash[i];
5278 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5279 if (io_match_task(req, tsk, files))
5280 posted += io_poll_remove_one(req);
5283 spin_unlock_irq(&ctx->completion_lock);
5286 io_cqring_ev_posted(ctx);
5291 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5293 struct hlist_head *list;
5294 struct io_kiocb *req;
5296 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5297 hlist_for_each_entry(req, list, hash_node) {
5298 if (sqe_addr != req->user_data)
5300 if (io_poll_remove_one(req))
5308 static int io_poll_remove_prep(struct io_kiocb *req,
5309 const struct io_uring_sqe *sqe)
5311 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5313 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5317 req->poll_remove.addr = READ_ONCE(sqe->addr);
5322 * Find a running poll command that matches one specified in sqe->addr,
5323 * and remove it if found.
5325 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5327 struct io_ring_ctx *ctx = req->ctx;
5330 spin_lock_irq(&ctx->completion_lock);
5331 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5332 spin_unlock_irq(&ctx->completion_lock);
5335 req_set_fail_links(req);
5336 io_req_complete(req, ret);
5340 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5343 struct io_kiocb *req = wait->private;
5344 struct io_poll_iocb *poll = &req->poll;
5346 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5349 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5350 struct poll_table_struct *p)
5352 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5354 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5357 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5359 struct io_poll_iocb *poll = &req->poll;
5362 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5364 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5367 events = READ_ONCE(sqe->poll32_events);
5369 events = swahw32(events);
5371 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5372 (events & EPOLLEXCLUSIVE);
5376 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5378 struct io_poll_iocb *poll = &req->poll;
5379 struct io_ring_ctx *ctx = req->ctx;
5380 struct io_poll_table ipt;
5383 ipt.pt._qproc = io_poll_queue_proc;
5385 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5388 if (mask) { /* no async, we'd stolen it */
5390 io_poll_complete(req, mask, 0);
5392 spin_unlock_irq(&ctx->completion_lock);
5395 io_cqring_ev_posted(ctx);
5401 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5403 struct io_timeout_data *data = container_of(timer,
5404 struct io_timeout_data, timer);
5405 struct io_kiocb *req = data->req;
5406 struct io_ring_ctx *ctx = req->ctx;
5407 unsigned long flags;
5409 spin_lock_irqsave(&ctx->completion_lock, flags);
5410 list_del_init(&req->timeout.list);
5411 atomic_set(&req->ctx->cq_timeouts,
5412 atomic_read(&req->ctx->cq_timeouts) + 1);
5414 io_cqring_fill_event(req, -ETIME);
5415 io_commit_cqring(ctx);
5416 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5418 io_cqring_ev_posted(ctx);
5419 req_set_fail_links(req);
5421 return HRTIMER_NORESTART;
5424 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5427 struct io_timeout_data *io;
5428 struct io_kiocb *req;
5431 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5432 if (user_data == req->user_data) {
5439 return ERR_PTR(ret);
5441 io = req->async_data;
5442 ret = hrtimer_try_to_cancel(&io->timer);
5444 return ERR_PTR(-EALREADY);
5445 list_del_init(&req->timeout.list);
5449 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5451 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5454 return PTR_ERR(req);
5456 req_set_fail_links(req);
5457 io_cqring_fill_event(req, -ECANCELED);
5458 io_put_req_deferred(req, 1);
5462 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5463 struct timespec64 *ts, enum hrtimer_mode mode)
5465 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5466 struct io_timeout_data *data;
5469 return PTR_ERR(req);
5471 req->timeout.off = 0; /* noseq */
5472 data = req->async_data;
5473 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5474 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5475 data->timer.function = io_timeout_fn;
5476 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5480 static int io_timeout_remove_prep(struct io_kiocb *req,
5481 const struct io_uring_sqe *sqe)
5483 struct io_timeout_rem *tr = &req->timeout_rem;
5485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5487 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5489 if (sqe->ioprio || sqe->buf_index || sqe->len)
5492 tr->addr = READ_ONCE(sqe->addr);
5493 tr->flags = READ_ONCE(sqe->timeout_flags);
5494 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5495 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5497 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5499 } else if (tr->flags) {
5500 /* timeout removal doesn't support flags */
5507 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5509 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5514 * Remove or update an existing timeout command
5516 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5518 struct io_timeout_rem *tr = &req->timeout_rem;
5519 struct io_ring_ctx *ctx = req->ctx;
5522 spin_lock_irq(&ctx->completion_lock);
5523 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5524 ret = io_timeout_cancel(ctx, tr->addr);
5526 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5527 io_translate_timeout_mode(tr->flags));
5529 io_cqring_fill_event(req, ret);
5530 io_commit_cqring(ctx);
5531 spin_unlock_irq(&ctx->completion_lock);
5532 io_cqring_ev_posted(ctx);
5534 req_set_fail_links(req);
5539 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5540 bool is_timeout_link)
5542 struct io_timeout_data *data;
5544 u32 off = READ_ONCE(sqe->off);
5546 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5548 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5550 if (off && is_timeout_link)
5552 flags = READ_ONCE(sqe->timeout_flags);
5553 if (flags & ~IORING_TIMEOUT_ABS)
5556 req->timeout.off = off;
5558 if (!req->async_data && io_alloc_async_data(req))
5561 data = req->async_data;
5564 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5567 data->mode = io_translate_timeout_mode(flags);
5568 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5569 if (is_timeout_link)
5570 io_req_track_inflight(req);
5574 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5576 struct io_ring_ctx *ctx = req->ctx;
5577 struct io_timeout_data *data = req->async_data;
5578 struct list_head *entry;
5579 u32 tail, off = req->timeout.off;
5581 spin_lock_irq(&ctx->completion_lock);
5584 * sqe->off holds how many events that need to occur for this
5585 * timeout event to be satisfied. If it isn't set, then this is
5586 * a pure timeout request, sequence isn't used.
5588 if (io_is_timeout_noseq(req)) {
5589 entry = ctx->timeout_list.prev;
5593 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5594 req->timeout.target_seq = tail + off;
5596 /* Update the last seq here in case io_flush_timeouts() hasn't.
5597 * This is safe because ->completion_lock is held, and submissions
5598 * and completions are never mixed in the same ->completion_lock section.
5600 ctx->cq_last_tm_flush = tail;
5603 * Insertion sort, ensuring the first entry in the list is always
5604 * the one we need first.
5606 list_for_each_prev(entry, &ctx->timeout_list) {
5607 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5610 if (io_is_timeout_noseq(nxt))
5612 /* nxt.seq is behind @tail, otherwise would've been completed */
5613 if (off >= nxt->timeout.target_seq - tail)
5617 list_add(&req->timeout.list, entry);
5618 data->timer.function = io_timeout_fn;
5619 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5620 spin_unlock_irq(&ctx->completion_lock);
5624 struct io_cancel_data {
5625 struct io_ring_ctx *ctx;
5629 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5631 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5632 struct io_cancel_data *cd = data;
5634 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5637 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5638 struct io_ring_ctx *ctx)
5640 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5641 enum io_wq_cancel cancel_ret;
5644 if (!tctx || !tctx->io_wq)
5647 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5648 switch (cancel_ret) {
5649 case IO_WQ_CANCEL_OK:
5652 case IO_WQ_CANCEL_RUNNING:
5655 case IO_WQ_CANCEL_NOTFOUND:
5663 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5664 struct io_kiocb *req, __u64 sqe_addr,
5667 unsigned long flags;
5670 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5671 if (ret != -ENOENT) {
5672 spin_lock_irqsave(&ctx->completion_lock, flags);
5676 spin_lock_irqsave(&ctx->completion_lock, flags);
5677 ret = io_timeout_cancel(ctx, sqe_addr);
5680 ret = io_poll_cancel(ctx, sqe_addr);
5684 io_cqring_fill_event(req, ret);
5685 io_commit_cqring(ctx);
5686 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5687 io_cqring_ev_posted(ctx);
5690 req_set_fail_links(req);
5694 static int io_async_cancel_prep(struct io_kiocb *req,
5695 const struct io_uring_sqe *sqe)
5697 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5699 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5701 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5704 req->cancel.addr = READ_ONCE(sqe->addr);
5708 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5710 struct io_ring_ctx *ctx = req->ctx;
5711 u64 sqe_addr = req->cancel.addr;
5712 struct io_tctx_node *node;
5715 /* tasks should wait for their io-wq threads, so safe w/o sync */
5716 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5717 spin_lock_irq(&ctx->completion_lock);
5720 ret = io_timeout_cancel(ctx, sqe_addr);
5723 ret = io_poll_cancel(ctx, sqe_addr);
5726 spin_unlock_irq(&ctx->completion_lock);
5728 /* slow path, try all io-wq's */
5729 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5731 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5732 struct io_uring_task *tctx = node->task->io_uring;
5734 if (!tctx || !tctx->io_wq)
5736 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5740 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5742 spin_lock_irq(&ctx->completion_lock);
5744 io_cqring_fill_event(req, ret);
5745 io_commit_cqring(ctx);
5746 spin_unlock_irq(&ctx->completion_lock);
5747 io_cqring_ev_posted(ctx);
5750 req_set_fail_links(req);
5755 static int io_rsrc_update_prep(struct io_kiocb *req,
5756 const struct io_uring_sqe *sqe)
5758 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5760 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5762 if (sqe->ioprio || sqe->rw_flags)
5765 req->rsrc_update.offset = READ_ONCE(sqe->off);
5766 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5767 if (!req->rsrc_update.nr_args)
5769 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5773 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5775 struct io_ring_ctx *ctx = req->ctx;
5776 struct io_uring_rsrc_update up;
5779 if (issue_flags & IO_URING_F_NONBLOCK)
5782 up.offset = req->rsrc_update.offset;
5783 up.data = req->rsrc_update.arg;
5785 mutex_lock(&ctx->uring_lock);
5786 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5787 mutex_unlock(&ctx->uring_lock);
5790 req_set_fail_links(req);
5791 __io_req_complete(req, issue_flags, ret, 0);
5795 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5797 switch (req->opcode) {
5800 case IORING_OP_READV:
5801 case IORING_OP_READ_FIXED:
5802 case IORING_OP_READ:
5803 return io_read_prep(req, sqe);
5804 case IORING_OP_WRITEV:
5805 case IORING_OP_WRITE_FIXED:
5806 case IORING_OP_WRITE:
5807 return io_write_prep(req, sqe);
5808 case IORING_OP_POLL_ADD:
5809 return io_poll_add_prep(req, sqe);
5810 case IORING_OP_POLL_REMOVE:
5811 return io_poll_remove_prep(req, sqe);
5812 case IORING_OP_FSYNC:
5813 return io_fsync_prep(req, sqe);
5814 case IORING_OP_SYNC_FILE_RANGE:
5815 return io_sfr_prep(req, sqe);
5816 case IORING_OP_SENDMSG:
5817 case IORING_OP_SEND:
5818 return io_sendmsg_prep(req, sqe);
5819 case IORING_OP_RECVMSG:
5820 case IORING_OP_RECV:
5821 return io_recvmsg_prep(req, sqe);
5822 case IORING_OP_CONNECT:
5823 return io_connect_prep(req, sqe);
5824 case IORING_OP_TIMEOUT:
5825 return io_timeout_prep(req, sqe, false);
5826 case IORING_OP_TIMEOUT_REMOVE:
5827 return io_timeout_remove_prep(req, sqe);
5828 case IORING_OP_ASYNC_CANCEL:
5829 return io_async_cancel_prep(req, sqe);
5830 case IORING_OP_LINK_TIMEOUT:
5831 return io_timeout_prep(req, sqe, true);
5832 case IORING_OP_ACCEPT:
5833 return io_accept_prep(req, sqe);
5834 case IORING_OP_FALLOCATE:
5835 return io_fallocate_prep(req, sqe);
5836 case IORING_OP_OPENAT:
5837 return io_openat_prep(req, sqe);
5838 case IORING_OP_CLOSE:
5839 return io_close_prep(req, sqe);
5840 case IORING_OP_FILES_UPDATE:
5841 return io_rsrc_update_prep(req, sqe);
5842 case IORING_OP_STATX:
5843 return io_statx_prep(req, sqe);
5844 case IORING_OP_FADVISE:
5845 return io_fadvise_prep(req, sqe);
5846 case IORING_OP_MADVISE:
5847 return io_madvise_prep(req, sqe);
5848 case IORING_OP_OPENAT2:
5849 return io_openat2_prep(req, sqe);
5850 case IORING_OP_EPOLL_CTL:
5851 return io_epoll_ctl_prep(req, sqe);
5852 case IORING_OP_SPLICE:
5853 return io_splice_prep(req, sqe);
5854 case IORING_OP_PROVIDE_BUFFERS:
5855 return io_provide_buffers_prep(req, sqe);
5856 case IORING_OP_REMOVE_BUFFERS:
5857 return io_remove_buffers_prep(req, sqe);
5859 return io_tee_prep(req, sqe);
5860 case IORING_OP_SHUTDOWN:
5861 return io_shutdown_prep(req, sqe);
5862 case IORING_OP_RENAMEAT:
5863 return io_renameat_prep(req, sqe);
5864 case IORING_OP_UNLINKAT:
5865 return io_unlinkat_prep(req, sqe);
5868 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5873 static int io_req_prep_async(struct io_kiocb *req)
5875 if (!io_op_defs[req->opcode].needs_async_setup)
5877 if (WARN_ON_ONCE(req->async_data))
5879 if (io_alloc_async_data(req))
5882 switch (req->opcode) {
5883 case IORING_OP_READV:
5884 return io_rw_prep_async(req, READ);
5885 case IORING_OP_WRITEV:
5886 return io_rw_prep_async(req, WRITE);
5887 case IORING_OP_SENDMSG:
5888 return io_sendmsg_prep_async(req);
5889 case IORING_OP_RECVMSG:
5890 return io_recvmsg_prep_async(req);
5891 case IORING_OP_CONNECT:
5892 return io_connect_prep_async(req);
5894 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5899 static u32 io_get_sequence(struct io_kiocb *req)
5901 struct io_kiocb *pos;
5902 struct io_ring_ctx *ctx = req->ctx;
5903 u32 total_submitted, nr_reqs = 0;
5905 io_for_each_link(pos, req)
5908 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5909 return total_submitted - nr_reqs;
5912 static int io_req_defer(struct io_kiocb *req)
5914 struct io_ring_ctx *ctx = req->ctx;
5915 struct io_defer_entry *de;
5919 /* Still need defer if there is pending req in defer list. */
5920 if (likely(list_empty_careful(&ctx->defer_list) &&
5921 !(req->flags & REQ_F_IO_DRAIN)))
5924 seq = io_get_sequence(req);
5925 /* Still a chance to pass the sequence check */
5926 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5929 ret = io_req_prep_async(req);
5932 io_prep_async_link(req);
5933 de = kmalloc(sizeof(*de), GFP_KERNEL);
5937 spin_lock_irq(&ctx->completion_lock);
5938 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5939 spin_unlock_irq(&ctx->completion_lock);
5941 io_queue_async_work(req);
5942 return -EIOCBQUEUED;
5945 trace_io_uring_defer(ctx, req, req->user_data);
5948 list_add_tail(&de->list, &ctx->defer_list);
5949 spin_unlock_irq(&ctx->completion_lock);
5950 return -EIOCBQUEUED;
5953 static void io_clean_op(struct io_kiocb *req)
5955 if (req->flags & REQ_F_BUFFER_SELECTED) {
5956 switch (req->opcode) {
5957 case IORING_OP_READV:
5958 case IORING_OP_READ_FIXED:
5959 case IORING_OP_READ:
5960 kfree((void *)(unsigned long)req->rw.addr);
5962 case IORING_OP_RECVMSG:
5963 case IORING_OP_RECV:
5964 kfree(req->sr_msg.kbuf);
5967 req->flags &= ~REQ_F_BUFFER_SELECTED;
5970 if (req->flags & REQ_F_NEED_CLEANUP) {
5971 switch (req->opcode) {
5972 case IORING_OP_READV:
5973 case IORING_OP_READ_FIXED:
5974 case IORING_OP_READ:
5975 case IORING_OP_WRITEV:
5976 case IORING_OP_WRITE_FIXED:
5977 case IORING_OP_WRITE: {
5978 struct io_async_rw *io = req->async_data;
5980 kfree(io->free_iovec);
5983 case IORING_OP_RECVMSG:
5984 case IORING_OP_SENDMSG: {
5985 struct io_async_msghdr *io = req->async_data;
5987 kfree(io->free_iov);
5990 case IORING_OP_SPLICE:
5992 io_put_file(req, req->splice.file_in,
5993 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5995 case IORING_OP_OPENAT:
5996 case IORING_OP_OPENAT2:
5997 if (req->open.filename)
5998 putname(req->open.filename);
6000 case IORING_OP_RENAMEAT:
6001 putname(req->rename.oldpath);
6002 putname(req->rename.newpath);
6004 case IORING_OP_UNLINKAT:
6005 putname(req->unlink.filename);
6008 req->flags &= ~REQ_F_NEED_CLEANUP;
6012 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6014 struct io_ring_ctx *ctx = req->ctx;
6015 const struct cred *creds = NULL;
6018 if (req->work.creds && req->work.creds != current_cred())
6019 creds = override_creds(req->work.creds);
6021 switch (req->opcode) {
6023 ret = io_nop(req, issue_flags);
6025 case IORING_OP_READV:
6026 case IORING_OP_READ_FIXED:
6027 case IORING_OP_READ:
6028 ret = io_read(req, issue_flags);
6030 case IORING_OP_WRITEV:
6031 case IORING_OP_WRITE_FIXED:
6032 case IORING_OP_WRITE:
6033 ret = io_write(req, issue_flags);
6035 case IORING_OP_FSYNC:
6036 ret = io_fsync(req, issue_flags);
6038 case IORING_OP_POLL_ADD:
6039 ret = io_poll_add(req, issue_flags);
6041 case IORING_OP_POLL_REMOVE:
6042 ret = io_poll_remove(req, issue_flags);
6044 case IORING_OP_SYNC_FILE_RANGE:
6045 ret = io_sync_file_range(req, issue_flags);
6047 case IORING_OP_SENDMSG:
6048 ret = io_sendmsg(req, issue_flags);
6050 case IORING_OP_SEND:
6051 ret = io_send(req, issue_flags);
6053 case IORING_OP_RECVMSG:
6054 ret = io_recvmsg(req, issue_flags);
6056 case IORING_OP_RECV:
6057 ret = io_recv(req, issue_flags);
6059 case IORING_OP_TIMEOUT:
6060 ret = io_timeout(req, issue_flags);
6062 case IORING_OP_TIMEOUT_REMOVE:
6063 ret = io_timeout_remove(req, issue_flags);
6065 case IORING_OP_ACCEPT:
6066 ret = io_accept(req, issue_flags);
6068 case IORING_OP_CONNECT:
6069 ret = io_connect(req, issue_flags);
6071 case IORING_OP_ASYNC_CANCEL:
6072 ret = io_async_cancel(req, issue_flags);
6074 case IORING_OP_FALLOCATE:
6075 ret = io_fallocate(req, issue_flags);
6077 case IORING_OP_OPENAT:
6078 ret = io_openat(req, issue_flags);
6080 case IORING_OP_CLOSE:
6081 ret = io_close(req, issue_flags);
6083 case IORING_OP_FILES_UPDATE:
6084 ret = io_files_update(req, issue_flags);
6086 case IORING_OP_STATX:
6087 ret = io_statx(req, issue_flags);
6089 case IORING_OP_FADVISE:
6090 ret = io_fadvise(req, issue_flags);
6092 case IORING_OP_MADVISE:
6093 ret = io_madvise(req, issue_flags);
6095 case IORING_OP_OPENAT2:
6096 ret = io_openat2(req, issue_flags);
6098 case IORING_OP_EPOLL_CTL:
6099 ret = io_epoll_ctl(req, issue_flags);
6101 case IORING_OP_SPLICE:
6102 ret = io_splice(req, issue_flags);
6104 case IORING_OP_PROVIDE_BUFFERS:
6105 ret = io_provide_buffers(req, issue_flags);
6107 case IORING_OP_REMOVE_BUFFERS:
6108 ret = io_remove_buffers(req, issue_flags);
6111 ret = io_tee(req, issue_flags);
6113 case IORING_OP_SHUTDOWN:
6114 ret = io_shutdown(req, issue_flags);
6116 case IORING_OP_RENAMEAT:
6117 ret = io_renameat(req, issue_flags);
6119 case IORING_OP_UNLINKAT:
6120 ret = io_unlinkat(req, issue_flags);
6128 revert_creds(creds);
6133 /* If the op doesn't have a file, we're not polling for it */
6134 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6135 const bool in_async = io_wq_current_is_worker();
6137 /* workqueue context doesn't hold uring_lock, grab it now */
6139 mutex_lock(&ctx->uring_lock);
6141 io_iopoll_req_issued(req, in_async);
6144 mutex_unlock(&ctx->uring_lock);
6150 static void io_wq_submit_work(struct io_wq_work *work)
6152 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6153 struct io_kiocb *timeout;
6156 timeout = io_prep_linked_timeout(req);
6158 io_queue_linked_timeout(timeout);
6160 if (work->flags & IO_WQ_WORK_CANCEL)
6165 ret = io_issue_sqe(req, 0);
6167 * We can get EAGAIN for polled IO even though we're
6168 * forcing a sync submission from here, since we can't
6169 * wait for request slots on the block side.
6177 /* avoid locking problems by failing it from a clean context */
6179 /* io-wq is going to take one down */
6181 io_req_task_queue_fail(req, ret);
6185 #define FFS_ASYNC_READ 0x1UL
6186 #define FFS_ASYNC_WRITE 0x2UL
6188 #define FFS_ISREG 0x4UL
6190 #define FFS_ISREG 0x0UL
6192 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6194 static inline struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
6197 struct fixed_rsrc_table *table;
6199 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6200 return &table->files[i & IORING_FILE_TABLE_MASK];
6203 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6206 struct file **file_slot = io_fixed_file_slot(ctx->file_data, index);
6208 return (struct file *) ((unsigned long) *file_slot & FFS_MASK);
6211 static struct file *io_file_get(struct io_submit_state *state,
6212 struct io_kiocb *req, int fd, bool fixed)
6214 struct io_ring_ctx *ctx = req->ctx;
6218 unsigned long file_ptr;
6220 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6222 fd = array_index_nospec(fd, ctx->nr_user_files);
6223 file_ptr = (unsigned long) *io_fixed_file_slot(ctx->file_data, fd);
6224 file = (struct file *) (file_ptr & FFS_MASK);
6225 file_ptr &= ~FFS_MASK;
6226 /* mask in overlapping REQ_F and FFS bits */
6227 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6228 io_set_resource_node(req);
6230 trace_io_uring_file_get(ctx, fd);
6231 file = __io_file_get(state, fd);
6233 /* we don't allow fixed io_uring files */
6234 if (file && unlikely(file->f_op == &io_uring_fops))
6235 io_req_track_inflight(req);
6241 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6243 struct io_timeout_data *data = container_of(timer,
6244 struct io_timeout_data, timer);
6245 struct io_kiocb *prev, *req = data->req;
6246 struct io_ring_ctx *ctx = req->ctx;
6247 unsigned long flags;
6249 spin_lock_irqsave(&ctx->completion_lock, flags);
6250 prev = req->timeout.head;
6251 req->timeout.head = NULL;
6254 * We don't expect the list to be empty, that will only happen if we
6255 * race with the completion of the linked work.
6257 if (prev && req_ref_inc_not_zero(prev))
6258 io_remove_next_linked(prev);
6261 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6264 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6265 io_put_req_deferred(prev, 1);
6267 io_req_complete_post(req, -ETIME, 0);
6268 io_put_req_deferred(req, 1);
6270 return HRTIMER_NORESTART;
6273 static void io_queue_linked_timeout(struct io_kiocb *req)
6275 struct io_ring_ctx *ctx = req->ctx;
6277 spin_lock_irq(&ctx->completion_lock);
6279 * If the back reference is NULL, then our linked request finished
6280 * before we got a chance to setup the timer
6282 if (req->timeout.head) {
6283 struct io_timeout_data *data = req->async_data;
6285 data->timer.function = io_link_timeout_fn;
6286 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6289 spin_unlock_irq(&ctx->completion_lock);
6290 /* drop submission reference */
6294 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6296 struct io_kiocb *nxt = req->link;
6298 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6299 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6302 nxt->timeout.head = req;
6303 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6304 req->flags |= REQ_F_LINK_TIMEOUT;
6308 static void __io_queue_sqe(struct io_kiocb *req)
6310 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6313 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6316 * We async punt it if the file wasn't marked NOWAIT, or if the file
6317 * doesn't support non-blocking read/write attempts
6320 /* drop submission reference */
6321 if (req->flags & REQ_F_COMPLETE_INLINE) {
6322 struct io_ring_ctx *ctx = req->ctx;
6323 struct io_comp_state *cs = &ctx->submit_state.comp;
6325 cs->reqs[cs->nr++] = req;
6326 if (cs->nr == ARRAY_SIZE(cs->reqs))
6327 io_submit_flush_completions(cs, ctx);
6331 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6332 if (!io_arm_poll_handler(req)) {
6334 * Queued up for async execution, worker will release
6335 * submit reference when the iocb is actually submitted.
6337 io_queue_async_work(req);
6340 io_req_complete_failed(req, ret);
6343 io_queue_linked_timeout(linked_timeout);
6346 static void io_queue_sqe(struct io_kiocb *req)
6350 ret = io_req_defer(req);
6352 if (ret != -EIOCBQUEUED) {
6354 io_req_complete_failed(req, ret);
6356 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6357 ret = io_req_prep_async(req);
6360 io_queue_async_work(req);
6362 __io_queue_sqe(req);
6367 * Check SQE restrictions (opcode and flags).
6369 * Returns 'true' if SQE is allowed, 'false' otherwise.
6371 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6372 struct io_kiocb *req,
6373 unsigned int sqe_flags)
6375 if (!ctx->restricted)
6378 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6381 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6382 ctx->restrictions.sqe_flags_required)
6385 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6386 ctx->restrictions.sqe_flags_required))
6392 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6393 const struct io_uring_sqe *sqe)
6395 struct io_submit_state *state;
6396 unsigned int sqe_flags;
6397 int personality, ret = 0;
6399 req->opcode = READ_ONCE(sqe->opcode);
6400 /* same numerical values with corresponding REQ_F_*, safe to copy */
6401 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6402 req->user_data = READ_ONCE(sqe->user_data);
6403 req->async_data = NULL;
6407 req->fixed_rsrc_refs = NULL;
6408 /* one is dropped after submission, the other at completion */
6409 atomic_set(&req->refs, 2);
6410 req->task = current;
6412 req->work.list.next = NULL;
6413 req->work.creds = NULL;
6414 req->work.flags = 0;
6416 /* enforce forwards compatibility on users */
6417 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6422 if (unlikely(req->opcode >= IORING_OP_LAST))
6425 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6428 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6429 !io_op_defs[req->opcode].buffer_select)
6432 personality = READ_ONCE(sqe->personality);
6434 req->work.creds = xa_load(&ctx->personalities, personality);
6435 if (!req->work.creds)
6437 get_cred(req->work.creds);
6439 state = &ctx->submit_state;
6442 * Plug now if we have more than 1 IO left after this, and the target
6443 * is potentially a read/write to block based storage.
6445 if (!state->plug_started && state->ios_left > 1 &&
6446 io_op_defs[req->opcode].plug) {
6447 blk_start_plug(&state->plug);
6448 state->plug_started = true;
6451 if (io_op_defs[req->opcode].needs_file) {
6452 bool fixed = req->flags & REQ_F_FIXED_FILE;
6454 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6455 if (unlikely(!req->file))
6463 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6464 const struct io_uring_sqe *sqe)
6466 struct io_submit_link *link = &ctx->submit_state.link;
6469 ret = io_init_req(ctx, req, sqe);
6470 if (unlikely(ret)) {
6473 /* fail even hard links since we don't submit */
6474 link->head->flags |= REQ_F_FAIL_LINK;
6475 io_req_complete_failed(link->head, -ECANCELED);
6478 io_req_complete_failed(req, ret);
6481 ret = io_req_prep(req, sqe);
6485 /* don't need @sqe from now on */
6486 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6487 true, ctx->flags & IORING_SETUP_SQPOLL);
6490 * If we already have a head request, queue this one for async
6491 * submittal once the head completes. If we don't have a head but
6492 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6493 * submitted sync once the chain is complete. If none of those
6494 * conditions are true (normal request), then just queue it.
6497 struct io_kiocb *head = link->head;
6500 * Taking sequential execution of a link, draining both sides
6501 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6502 * requests in the link. So, it drains the head and the
6503 * next after the link request. The last one is done via
6504 * drain_next flag to persist the effect across calls.
6506 if (req->flags & REQ_F_IO_DRAIN) {
6507 head->flags |= REQ_F_IO_DRAIN;
6508 ctx->drain_next = 1;
6510 ret = io_req_prep_async(req);
6513 trace_io_uring_link(ctx, req, head);
6514 link->last->link = req;
6517 /* last request of a link, enqueue the link */
6518 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6523 if (unlikely(ctx->drain_next)) {
6524 req->flags |= REQ_F_IO_DRAIN;
6525 ctx->drain_next = 0;
6527 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6539 * Batched submission is done, ensure local IO is flushed out.
6541 static void io_submit_state_end(struct io_submit_state *state,
6542 struct io_ring_ctx *ctx)
6544 if (state->link.head)
6545 io_queue_sqe(state->link.head);
6547 io_submit_flush_completions(&state->comp, ctx);
6548 if (state->plug_started)
6549 blk_finish_plug(&state->plug);
6550 io_state_file_put(state);
6554 * Start submission side cache.
6556 static void io_submit_state_start(struct io_submit_state *state,
6557 unsigned int max_ios)
6559 state->plug_started = false;
6560 state->ios_left = max_ios;
6561 /* set only head, no need to init link_last in advance */
6562 state->link.head = NULL;
6565 static void io_commit_sqring(struct io_ring_ctx *ctx)
6567 struct io_rings *rings = ctx->rings;
6570 * Ensure any loads from the SQEs are done at this point,
6571 * since once we write the new head, the application could
6572 * write new data to them.
6574 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6578 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6579 * that is mapped by userspace. This means that care needs to be taken to
6580 * ensure that reads are stable, as we cannot rely on userspace always
6581 * being a good citizen. If members of the sqe are validated and then later
6582 * used, it's important that those reads are done through READ_ONCE() to
6583 * prevent a re-load down the line.
6585 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6587 u32 *sq_array = ctx->sq_array;
6591 * The cached sq head (or cq tail) serves two purposes:
6593 * 1) allows us to batch the cost of updating the user visible
6595 * 2) allows the kernel side to track the head on its own, even
6596 * though the application is the one updating it.
6598 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6599 if (likely(head < ctx->sq_entries))
6600 return &ctx->sq_sqes[head];
6602 /* drop invalid entries */
6603 ctx->cached_sq_dropped++;
6604 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6608 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6612 /* if we have a backlog and couldn't flush it all, return BUSY */
6613 if (test_bit(0, &ctx->sq_check_overflow)) {
6614 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6618 /* make sure SQ entry isn't read before tail */
6619 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6621 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6624 percpu_counter_add(¤t->io_uring->inflight, nr);
6625 refcount_add(nr, ¤t->usage);
6626 io_submit_state_start(&ctx->submit_state, nr);
6628 while (submitted < nr) {
6629 const struct io_uring_sqe *sqe;
6630 struct io_kiocb *req;
6632 req = io_alloc_req(ctx);
6633 if (unlikely(!req)) {
6635 submitted = -EAGAIN;
6638 sqe = io_get_sqe(ctx);
6639 if (unlikely(!sqe)) {
6640 kmem_cache_free(req_cachep, req);
6643 /* will complete beyond this point, count as submitted */
6645 if (io_submit_sqe(ctx, req, sqe))
6649 if (unlikely(submitted != nr)) {
6650 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6651 struct io_uring_task *tctx = current->io_uring;
6652 int unused = nr - ref_used;
6654 percpu_ref_put_many(&ctx->refs, unused);
6655 percpu_counter_sub(&tctx->inflight, unused);
6656 put_task_struct_many(current, unused);
6659 io_submit_state_end(&ctx->submit_state, ctx);
6660 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6661 io_commit_sqring(ctx);
6666 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6668 /* Tell userspace we may need a wakeup call */
6669 spin_lock_irq(&ctx->completion_lock);
6670 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6671 spin_unlock_irq(&ctx->completion_lock);
6674 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6676 spin_lock_irq(&ctx->completion_lock);
6677 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6678 spin_unlock_irq(&ctx->completion_lock);
6681 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6683 unsigned int to_submit;
6686 to_submit = io_sqring_entries(ctx);
6687 /* if we're handling multiple rings, cap submit size for fairness */
6688 if (cap_entries && to_submit > 8)
6691 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6692 unsigned nr_events = 0;
6694 mutex_lock(&ctx->uring_lock);
6695 if (!list_empty(&ctx->iopoll_list))
6696 io_do_iopoll(ctx, &nr_events, 0);
6698 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6699 !(ctx->flags & IORING_SETUP_R_DISABLED))
6700 ret = io_submit_sqes(ctx, to_submit);
6701 mutex_unlock(&ctx->uring_lock);
6704 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6705 wake_up(&ctx->sqo_sq_wait);
6710 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6712 struct io_ring_ctx *ctx;
6713 unsigned sq_thread_idle = 0;
6715 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6716 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6717 sqd->sq_thread_idle = sq_thread_idle;
6720 static int io_sq_thread(void *data)
6722 struct io_sq_data *sqd = data;
6723 struct io_ring_ctx *ctx;
6724 unsigned long timeout = 0;
6725 char buf[TASK_COMM_LEN];
6728 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6729 set_task_comm(current, buf);
6730 current->pf_io_worker = NULL;
6732 if (sqd->sq_cpu != -1)
6733 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6735 set_cpus_allowed_ptr(current, cpu_online_mask);
6736 current->flags |= PF_NO_SETAFFINITY;
6738 mutex_lock(&sqd->lock);
6739 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6741 bool cap_entries, sqt_spin, needs_sched;
6743 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6744 signal_pending(current)) {
6745 bool did_sig = false;
6747 mutex_unlock(&sqd->lock);
6748 if (signal_pending(current)) {
6749 struct ksignal ksig;
6751 did_sig = get_signal(&ksig);
6754 mutex_lock(&sqd->lock);
6758 io_run_task_work_head(&sqd->park_task_work);
6759 timeout = jiffies + sqd->sq_thread_idle;
6763 cap_entries = !list_is_singular(&sqd->ctx_list);
6764 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6765 const struct cred *creds = NULL;
6767 if (ctx->sq_creds != current_cred())
6768 creds = override_creds(ctx->sq_creds);
6769 ret = __io_sq_thread(ctx, cap_entries);
6771 revert_creds(creds);
6772 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6776 if (sqt_spin || !time_after(jiffies, timeout)) {
6780 timeout = jiffies + sqd->sq_thread_idle;
6785 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6786 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6787 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6788 !list_empty_careful(&ctx->iopoll_list)) {
6789 needs_sched = false;
6792 if (io_sqring_entries(ctx)) {
6793 needs_sched = false;
6798 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6799 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6800 io_ring_set_wakeup_flag(ctx);
6802 mutex_unlock(&sqd->lock);
6804 mutex_lock(&sqd->lock);
6805 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6806 io_ring_clear_wakeup_flag(ctx);
6809 finish_wait(&sqd->wait, &wait);
6810 io_run_task_work_head(&sqd->park_task_work);
6811 timeout = jiffies + sqd->sq_thread_idle;
6814 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6815 io_uring_cancel_sqpoll(ctx);
6817 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6818 io_ring_set_wakeup_flag(ctx);
6819 mutex_unlock(&sqd->lock);
6822 io_run_task_work_head(&sqd->park_task_work);
6823 complete(&sqd->exited);
6827 struct io_wait_queue {
6828 struct wait_queue_entry wq;
6829 struct io_ring_ctx *ctx;
6831 unsigned nr_timeouts;
6834 static inline bool io_should_wake(struct io_wait_queue *iowq)
6836 struct io_ring_ctx *ctx = iowq->ctx;
6839 * Wake up if we have enough events, or if a timeout occurred since we
6840 * started waiting. For timeouts, we always want to return to userspace,
6841 * regardless of event count.
6843 return io_cqring_events(ctx) >= iowq->to_wait ||
6844 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6847 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6848 int wake_flags, void *key)
6850 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6854 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6855 * the task, and the next invocation will do it.
6857 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6858 return autoremove_wake_function(curr, mode, wake_flags, key);
6862 static int io_run_task_work_sig(void)
6864 if (io_run_task_work())
6866 if (!signal_pending(current))
6868 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6869 return -ERESTARTSYS;
6873 /* when returns >0, the caller should retry */
6874 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6875 struct io_wait_queue *iowq,
6876 signed long *timeout)
6880 /* make sure we run task_work before checking for signals */
6881 ret = io_run_task_work_sig();
6882 if (ret || io_should_wake(iowq))
6884 /* let the caller flush overflows, retry */
6885 if (test_bit(0, &ctx->cq_check_overflow))
6888 *timeout = schedule_timeout(*timeout);
6889 return !*timeout ? -ETIME : 1;
6893 * Wait until events become available, if we don't already have some. The
6894 * application must reap them itself, as they reside on the shared cq ring.
6896 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6897 const sigset_t __user *sig, size_t sigsz,
6898 struct __kernel_timespec __user *uts)
6900 struct io_wait_queue iowq = {
6903 .func = io_wake_function,
6904 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6907 .to_wait = min_events,
6909 struct io_rings *rings = ctx->rings;
6910 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6914 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6915 if (io_cqring_events(ctx) >= min_events)
6917 if (!io_run_task_work())
6922 #ifdef CONFIG_COMPAT
6923 if (in_compat_syscall())
6924 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6928 ret = set_user_sigmask(sig, sigsz);
6935 struct timespec64 ts;
6937 if (get_timespec64(&ts, uts))
6939 timeout = timespec64_to_jiffies(&ts);
6942 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6943 trace_io_uring_cqring_wait(ctx, min_events);
6945 /* if we can't even flush overflow, don't wait for more */
6946 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6950 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6951 TASK_INTERRUPTIBLE);
6952 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6953 finish_wait(&ctx->wait, &iowq.wq);
6957 restore_saved_sigmask_unless(ret == -EINTR);
6959 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6962 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6964 #if defined(CONFIG_UNIX)
6965 if (ctx->ring_sock) {
6966 struct sock *sock = ctx->ring_sock->sk;
6967 struct sk_buff *skb;
6969 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6975 for (i = 0; i < ctx->nr_user_files; i++) {
6978 file = io_file_from_index(ctx, i);
6985 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6987 struct fixed_rsrc_data *data;
6989 data = container_of(ref, struct fixed_rsrc_data, refs);
6990 complete(&data->done);
6993 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6995 spin_lock_bh(&ctx->rsrc_ref_lock);
6998 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7000 spin_unlock_bh(&ctx->rsrc_ref_lock);
7003 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7004 struct fixed_rsrc_data *rsrc_data,
7005 struct fixed_rsrc_ref_node *ref_node)
7007 io_rsrc_ref_lock(ctx);
7008 rsrc_data->node = ref_node;
7009 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7010 io_rsrc_ref_unlock(ctx);
7011 percpu_ref_get(&rsrc_data->refs);
7014 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7016 struct fixed_rsrc_ref_node *ref_node = NULL;
7018 io_rsrc_ref_lock(ctx);
7019 ref_node = data->node;
7021 io_rsrc_ref_unlock(ctx);
7023 percpu_ref_kill(&ref_node->refs);
7026 static int io_rsrc_refnode_prealloc(struct io_ring_ctx *ctx)
7028 if (ctx->rsrc_backup_node)
7030 ctx->rsrc_backup_node = alloc_fixed_rsrc_ref_node(ctx);
7031 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7034 static struct fixed_rsrc_ref_node *
7035 io_rsrc_refnode_get(struct io_ring_ctx *ctx,
7036 struct fixed_rsrc_data *rsrc_data,
7037 void (*rsrc_put)(struct io_ring_ctx *ctx,
7038 struct io_rsrc_put *prsrc))
7040 struct fixed_rsrc_ref_node *node = ctx->rsrc_backup_node;
7042 WARN_ON_ONCE(!node);
7044 ctx->rsrc_backup_node = NULL;
7045 node->rsrc_data = rsrc_data;
7046 node->rsrc_put = rsrc_put;
7050 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7051 struct io_ring_ctx *ctx,
7052 void (*rsrc_put)(struct io_ring_ctx *ctx,
7053 struct io_rsrc_put *prsrc))
7055 struct fixed_rsrc_ref_node *node;
7061 data->quiesce = true;
7063 ret = io_rsrc_refnode_prealloc(ctx);
7066 io_sqe_rsrc_kill_node(ctx, data);
7067 percpu_ref_kill(&data->refs);
7068 flush_delayed_work(&ctx->rsrc_put_work);
7070 ret = wait_for_completion_interruptible(&data->done);
7074 percpu_ref_resurrect(&data->refs);
7075 node = io_rsrc_refnode_get(ctx, data, rsrc_put);
7076 io_sqe_rsrc_set_node(ctx, data, node);
7077 reinit_completion(&data->done);
7079 mutex_unlock(&ctx->uring_lock);
7080 ret = io_run_task_work_sig();
7081 mutex_lock(&ctx->uring_lock);
7083 data->quiesce = false;
7088 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7090 struct fixed_rsrc_data *data;
7092 data = kzalloc(sizeof(*data), GFP_KERNEL);
7096 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7097 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7102 init_completion(&data->done);
7106 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7108 percpu_ref_exit(&data->refs);
7113 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7115 struct fixed_rsrc_data *data = ctx->file_data;
7116 unsigned nr_tables, i;
7120 * percpu_ref_is_dying() is to stop parallel files unregister
7121 * Since we possibly drop uring lock later in this function to
7124 if (!data || percpu_ref_is_dying(&data->refs))
7126 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7130 __io_sqe_files_unregister(ctx);
7131 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7132 for (i = 0; i < nr_tables; i++)
7133 kfree(data->table[i].files);
7134 free_fixed_rsrc_data(data);
7135 ctx->file_data = NULL;
7136 ctx->nr_user_files = 0;
7140 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7141 __releases(&sqd->lock)
7143 WARN_ON_ONCE(sqd->thread == current);
7146 * Do the dance but not conditional clear_bit() because it'd race with
7147 * other threads incrementing park_pending and setting the bit.
7149 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7150 if (atomic_dec_return(&sqd->park_pending))
7151 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7152 mutex_unlock(&sqd->lock);
7155 static void io_sq_thread_park(struct io_sq_data *sqd)
7156 __acquires(&sqd->lock)
7158 WARN_ON_ONCE(sqd->thread == current);
7160 atomic_inc(&sqd->park_pending);
7161 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7162 mutex_lock(&sqd->lock);
7164 wake_up_process(sqd->thread);
7167 static void io_sq_thread_stop(struct io_sq_data *sqd)
7169 WARN_ON_ONCE(sqd->thread == current);
7171 mutex_lock(&sqd->lock);
7172 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7174 wake_up_process(sqd->thread);
7175 mutex_unlock(&sqd->lock);
7176 wait_for_completion(&sqd->exited);
7179 static void io_put_sq_data(struct io_sq_data *sqd)
7181 if (refcount_dec_and_test(&sqd->refs)) {
7182 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7184 io_sq_thread_stop(sqd);
7189 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7191 struct io_sq_data *sqd = ctx->sq_data;
7194 io_sq_thread_park(sqd);
7195 list_del_init(&ctx->sqd_list);
7196 io_sqd_update_thread_idle(sqd);
7197 io_sq_thread_unpark(sqd);
7199 io_put_sq_data(sqd);
7200 ctx->sq_data = NULL;
7202 put_cred(ctx->sq_creds);
7206 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7208 struct io_ring_ctx *ctx_attach;
7209 struct io_sq_data *sqd;
7212 f = fdget(p->wq_fd);
7214 return ERR_PTR(-ENXIO);
7215 if (f.file->f_op != &io_uring_fops) {
7217 return ERR_PTR(-EINVAL);
7220 ctx_attach = f.file->private_data;
7221 sqd = ctx_attach->sq_data;
7224 return ERR_PTR(-EINVAL);
7226 if (sqd->task_tgid != current->tgid) {
7228 return ERR_PTR(-EPERM);
7231 refcount_inc(&sqd->refs);
7236 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7239 struct io_sq_data *sqd;
7242 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7243 sqd = io_attach_sq_data(p);
7248 /* fall through for EPERM case, setup new sqd/task */
7249 if (PTR_ERR(sqd) != -EPERM)
7253 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7255 return ERR_PTR(-ENOMEM);
7257 atomic_set(&sqd->park_pending, 0);
7258 refcount_set(&sqd->refs, 1);
7259 INIT_LIST_HEAD(&sqd->ctx_list);
7260 mutex_init(&sqd->lock);
7261 init_waitqueue_head(&sqd->wait);
7262 init_completion(&sqd->exited);
7266 #if defined(CONFIG_UNIX)
7268 * Ensure the UNIX gc is aware of our file set, so we are certain that
7269 * the io_uring can be safely unregistered on process exit, even if we have
7270 * loops in the file referencing.
7272 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7274 struct sock *sk = ctx->ring_sock->sk;
7275 struct scm_fp_list *fpl;
7276 struct sk_buff *skb;
7279 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7283 skb = alloc_skb(0, GFP_KERNEL);
7292 fpl->user = get_uid(current_user());
7293 for (i = 0; i < nr; i++) {
7294 struct file *file = io_file_from_index(ctx, i + offset);
7298 fpl->fp[nr_files] = get_file(file);
7299 unix_inflight(fpl->user, fpl->fp[nr_files]);
7304 fpl->max = SCM_MAX_FD;
7305 fpl->count = nr_files;
7306 UNIXCB(skb).fp = fpl;
7307 skb->destructor = unix_destruct_scm;
7308 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7309 skb_queue_head(&sk->sk_receive_queue, skb);
7311 for (i = 0; i < nr_files; i++)
7322 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7323 * causes regular reference counting to break down. We rely on the UNIX
7324 * garbage collection to take care of this problem for us.
7326 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7328 unsigned left, total;
7332 left = ctx->nr_user_files;
7334 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7336 ret = __io_sqe_files_scm(ctx, this_files, total);
7340 total += this_files;
7346 while (total < ctx->nr_user_files) {
7347 struct file *file = io_file_from_index(ctx, total);
7357 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7363 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7364 unsigned nr_tables, unsigned nr_files)
7368 for (i = 0; i < nr_tables; i++) {
7369 struct fixed_rsrc_table *table = &file_data->table[i];
7370 unsigned this_files;
7372 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7373 table->files = kcalloc(this_files, sizeof(struct file *),
7377 nr_files -= this_files;
7383 for (i = 0; i < nr_tables; i++) {
7384 struct fixed_rsrc_table *table = &file_data->table[i];
7385 kfree(table->files);
7390 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7392 struct file *file = prsrc->file;
7393 #if defined(CONFIG_UNIX)
7394 struct sock *sock = ctx->ring_sock->sk;
7395 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7396 struct sk_buff *skb;
7399 __skb_queue_head_init(&list);
7402 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7403 * remove this entry and rearrange the file array.
7405 skb = skb_dequeue(head);
7407 struct scm_fp_list *fp;
7409 fp = UNIXCB(skb).fp;
7410 for (i = 0; i < fp->count; i++) {
7413 if (fp->fp[i] != file)
7416 unix_notinflight(fp->user, fp->fp[i]);
7417 left = fp->count - 1 - i;
7419 memmove(&fp->fp[i], &fp->fp[i + 1],
7420 left * sizeof(struct file *));
7427 __skb_queue_tail(&list, skb);
7437 __skb_queue_tail(&list, skb);
7439 skb = skb_dequeue(head);
7442 if (skb_peek(&list)) {
7443 spin_lock_irq(&head->lock);
7444 while ((skb = __skb_dequeue(&list)) != NULL)
7445 __skb_queue_tail(head, skb);
7446 spin_unlock_irq(&head->lock);
7453 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7455 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7456 struct io_ring_ctx *ctx = rsrc_data->ctx;
7457 struct io_rsrc_put *prsrc, *tmp;
7459 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7460 list_del(&prsrc->list);
7461 ref_node->rsrc_put(ctx, prsrc);
7465 percpu_ref_exit(&ref_node->refs);
7467 percpu_ref_put(&rsrc_data->refs);
7470 static void io_rsrc_put_work(struct work_struct *work)
7472 struct io_ring_ctx *ctx;
7473 struct llist_node *node;
7475 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7476 node = llist_del_all(&ctx->rsrc_put_llist);
7479 struct fixed_rsrc_ref_node *ref_node;
7480 struct llist_node *next = node->next;
7482 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7483 __io_rsrc_put_work(ref_node);
7488 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7490 struct fixed_rsrc_ref_node *ref_node;
7491 struct fixed_rsrc_data *data;
7492 struct io_ring_ctx *ctx;
7493 bool first_add = false;
7496 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7497 data = ref_node->rsrc_data;
7500 io_rsrc_ref_lock(ctx);
7501 ref_node->done = true;
7503 while (!list_empty(&ctx->rsrc_ref_list)) {
7504 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7505 struct fixed_rsrc_ref_node, node);
7506 /* recycle ref nodes in order */
7507 if (!ref_node->done)
7509 list_del(&ref_node->node);
7510 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7512 io_rsrc_ref_unlock(ctx);
7514 if (percpu_ref_is_dying(&data->refs))
7518 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7520 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7523 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7524 struct io_ring_ctx *ctx)
7526 struct fixed_rsrc_ref_node *ref_node;
7528 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7532 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7537 INIT_LIST_HEAD(&ref_node->node);
7538 INIT_LIST_HEAD(&ref_node->rsrc_list);
7539 ref_node->done = false;
7543 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7544 struct fixed_rsrc_ref_node *ref_node)
7546 ref_node->rsrc_data = ctx->file_data;
7547 ref_node->rsrc_put = io_ring_file_put;
7550 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7552 percpu_ref_exit(&ref_node->refs);
7557 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7560 __s32 __user *fds = (__s32 __user *) arg;
7561 unsigned nr_tables, i;
7563 int fd, ret = -ENOMEM;
7564 struct fixed_rsrc_ref_node *ref_node;
7565 struct fixed_rsrc_data *file_data;
7571 if (nr_args > IORING_MAX_FIXED_FILES)
7574 file_data = alloc_fixed_rsrc_data(ctx);
7577 ctx->file_data = file_data;
7579 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7580 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7582 if (!file_data->table)
7585 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7588 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7589 unsigned long file_ptr;
7591 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7595 /* allow sparse sets */
7605 * Don't allow io_uring instances to be registered. If UNIX
7606 * isn't enabled, then this causes a reference cycle and this
7607 * instance can never get freed. If UNIX is enabled we'll
7608 * handle it just fine, but there's still no point in allowing
7609 * a ring fd as it doesn't support regular read/write anyway.
7611 if (file->f_op == &io_uring_fops) {
7615 file_ptr = (unsigned long) file;
7616 if (__io_file_supports_async(file, READ))
7617 file_ptr |= FFS_ASYNC_READ;
7618 if (__io_file_supports_async(file, WRITE))
7619 file_ptr |= FFS_ASYNC_WRITE;
7620 if (S_ISREG(file_inode(file)->i_mode))
7621 file_ptr |= FFS_ISREG;
7622 *io_fixed_file_slot(file_data, i) = (struct file *) file_ptr;
7625 ret = io_sqe_files_scm(ctx);
7627 io_sqe_files_unregister(ctx);
7631 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7633 io_sqe_files_unregister(ctx);
7636 init_fixed_file_ref_node(ctx, ref_node);
7638 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7641 for (i = 0; i < ctx->nr_user_files; i++) {
7642 file = io_file_from_index(ctx, i);
7646 for (i = 0; i < nr_tables; i++)
7647 kfree(file_data->table[i].files);
7648 ctx->nr_user_files = 0;
7650 free_fixed_rsrc_data(ctx->file_data);
7651 ctx->file_data = NULL;
7655 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7658 #if defined(CONFIG_UNIX)
7659 struct sock *sock = ctx->ring_sock->sk;
7660 struct sk_buff_head *head = &sock->sk_receive_queue;
7661 struct sk_buff *skb;
7664 * See if we can merge this file into an existing skb SCM_RIGHTS
7665 * file set. If there's no room, fall back to allocating a new skb
7666 * and filling it in.
7668 spin_lock_irq(&head->lock);
7669 skb = skb_peek(head);
7671 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7673 if (fpl->count < SCM_MAX_FD) {
7674 __skb_unlink(skb, head);
7675 spin_unlock_irq(&head->lock);
7676 fpl->fp[fpl->count] = get_file(file);
7677 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7679 spin_lock_irq(&head->lock);
7680 __skb_queue_head(head, skb);
7685 spin_unlock_irq(&head->lock);
7692 return __io_sqe_files_scm(ctx, 1, index);
7698 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7700 struct io_rsrc_put *prsrc;
7701 struct fixed_rsrc_ref_node *ref_node = data->node;
7703 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7708 list_add(&prsrc->list, &ref_node->rsrc_list);
7713 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7716 return io_queue_rsrc_removal(data, (void *)file);
7719 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7720 struct io_uring_rsrc_update *up,
7723 struct fixed_rsrc_data *data = ctx->file_data;
7724 struct fixed_rsrc_ref_node *ref_node;
7725 struct file *file, **file_slot;
7729 bool needs_switch = false;
7731 if (check_add_overflow(up->offset, nr_args, &done))
7733 if (done > ctx->nr_user_files)
7735 err = io_rsrc_refnode_prealloc(ctx);
7739 fds = u64_to_user_ptr(up->data);
7740 for (done = 0; done < nr_args; done++) {
7742 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7746 if (fd == IORING_REGISTER_FILES_SKIP)
7749 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7750 file_slot = io_fixed_file_slot(ctx->file_data, i);
7753 file = (struct file *) ((unsigned long) *file_slot & FFS_MASK);
7754 err = io_queue_file_removal(data, file);
7758 needs_switch = true;
7767 * Don't allow io_uring instances to be registered. If
7768 * UNIX isn't enabled, then this causes a reference
7769 * cycle and this instance can never get freed. If UNIX
7770 * is enabled we'll handle it just fine, but there's
7771 * still no point in allowing a ring fd as it doesn't
7772 * support regular read/write anyway.
7774 if (file->f_op == &io_uring_fops) {
7780 err = io_sqe_file_register(ctx, file, i);
7790 percpu_ref_kill(&data->node->refs);
7791 ref_node = io_rsrc_refnode_get(ctx, data, io_ring_file_put);
7792 io_sqe_rsrc_set_node(ctx, data, ref_node);
7794 return done ? done : err;
7797 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7800 struct io_uring_rsrc_update up;
7802 if (!ctx->file_data)
7806 if (copy_from_user(&up, arg, sizeof(up)))
7811 return __io_sqe_files_update(ctx, &up, nr_args);
7814 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7816 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7818 req = io_put_req_find_next(req);
7819 return req ? &req->work : NULL;
7822 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7824 struct io_wq_hash *hash;
7825 struct io_wq_data data;
7826 unsigned int concurrency;
7828 hash = ctx->hash_map;
7830 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7832 return ERR_PTR(-ENOMEM);
7833 refcount_set(&hash->refs, 1);
7834 init_waitqueue_head(&hash->wait);
7835 ctx->hash_map = hash;
7839 data.free_work = io_free_work;
7840 data.do_work = io_wq_submit_work;
7842 /* Do QD, or 4 * CPUS, whatever is smallest */
7843 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7845 return io_wq_create(concurrency, &data);
7848 static int io_uring_alloc_task_context(struct task_struct *task,
7849 struct io_ring_ctx *ctx)
7851 struct io_uring_task *tctx;
7854 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7855 if (unlikely(!tctx))
7858 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7859 if (unlikely(ret)) {
7864 tctx->io_wq = io_init_wq_offload(ctx);
7865 if (IS_ERR(tctx->io_wq)) {
7866 ret = PTR_ERR(tctx->io_wq);
7867 percpu_counter_destroy(&tctx->inflight);
7873 init_waitqueue_head(&tctx->wait);
7875 atomic_set(&tctx->in_idle, 0);
7876 task->io_uring = tctx;
7877 spin_lock_init(&tctx->task_lock);
7878 INIT_WQ_LIST(&tctx->task_list);
7879 tctx->task_state = 0;
7880 init_task_work(&tctx->task_work, tctx_task_work);
7884 void __io_uring_free(struct task_struct *tsk)
7886 struct io_uring_task *tctx = tsk->io_uring;
7888 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7889 WARN_ON_ONCE(tctx->io_wq);
7891 percpu_counter_destroy(&tctx->inflight);
7893 tsk->io_uring = NULL;
7896 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7897 struct io_uring_params *p)
7901 /* Retain compatibility with failing for an invalid attach attempt */
7902 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7903 IORING_SETUP_ATTACH_WQ) {
7906 f = fdget(p->wq_fd);
7909 if (f.file->f_op != &io_uring_fops) {
7915 if (ctx->flags & IORING_SETUP_SQPOLL) {
7916 struct task_struct *tsk;
7917 struct io_sq_data *sqd;
7921 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7924 sqd = io_get_sq_data(p, &attached);
7930 ctx->sq_creds = get_current_cred();
7932 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7933 if (!ctx->sq_thread_idle)
7934 ctx->sq_thread_idle = HZ;
7937 io_sq_thread_park(sqd);
7938 list_add(&ctx->sqd_list, &sqd->ctx_list);
7939 io_sqd_update_thread_idle(sqd);
7940 /* don't attach to a dying SQPOLL thread, would be racy */
7941 if (attached && !sqd->thread)
7943 io_sq_thread_unpark(sqd);
7950 if (p->flags & IORING_SETUP_SQ_AFF) {
7951 int cpu = p->sq_thread_cpu;
7954 if (cpu >= nr_cpu_ids)
7956 if (!cpu_online(cpu))
7964 sqd->task_pid = current->pid;
7965 sqd->task_tgid = current->tgid;
7966 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7973 ret = io_uring_alloc_task_context(tsk, ctx);
7974 wake_up_new_task(tsk);
7977 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7978 /* Can't have SQ_AFF without SQPOLL */
7985 io_sq_thread_finish(ctx);
7988 complete(&ctx->sq_data->exited);
7992 static inline void __io_unaccount_mem(struct user_struct *user,
7993 unsigned long nr_pages)
7995 atomic_long_sub(nr_pages, &user->locked_vm);
7998 static inline int __io_account_mem(struct user_struct *user,
7999 unsigned long nr_pages)
8001 unsigned long page_limit, cur_pages, new_pages;
8003 /* Don't allow more pages than we can safely lock */
8004 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8007 cur_pages = atomic_long_read(&user->locked_vm);
8008 new_pages = cur_pages + nr_pages;
8009 if (new_pages > page_limit)
8011 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8012 new_pages) != cur_pages);
8017 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8020 __io_unaccount_mem(ctx->user, nr_pages);
8022 if (ctx->mm_account)
8023 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8026 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8031 ret = __io_account_mem(ctx->user, nr_pages);
8036 if (ctx->mm_account)
8037 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8042 static void io_mem_free(void *ptr)
8049 page = virt_to_head_page(ptr);
8050 if (put_page_testzero(page))
8051 free_compound_page(page);
8054 static void *io_mem_alloc(size_t size)
8056 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8057 __GFP_NORETRY | __GFP_ACCOUNT;
8059 return (void *) __get_free_pages(gfp_flags, get_order(size));
8062 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8065 struct io_rings *rings;
8066 size_t off, sq_array_size;
8068 off = struct_size(rings, cqes, cq_entries);
8069 if (off == SIZE_MAX)
8073 off = ALIGN(off, SMP_CACHE_BYTES);
8081 sq_array_size = array_size(sizeof(u32), sq_entries);
8082 if (sq_array_size == SIZE_MAX)
8085 if (check_add_overflow(off, sq_array_size, &off))
8091 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8095 if (!ctx->user_bufs)
8098 for (i = 0; i < ctx->nr_user_bufs; i++) {
8099 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8101 for (j = 0; j < imu->nr_bvecs; j++)
8102 unpin_user_page(imu->bvec[j].bv_page);
8104 if (imu->acct_pages)
8105 io_unaccount_mem(ctx, imu->acct_pages);
8110 kfree(ctx->user_bufs);
8111 ctx->user_bufs = NULL;
8112 ctx->nr_user_bufs = 0;
8116 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8117 void __user *arg, unsigned index)
8119 struct iovec __user *src;
8121 #ifdef CONFIG_COMPAT
8123 struct compat_iovec __user *ciovs;
8124 struct compat_iovec ciov;
8126 ciovs = (struct compat_iovec __user *) arg;
8127 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8130 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8131 dst->iov_len = ciov.iov_len;
8135 src = (struct iovec __user *) arg;
8136 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8142 * Not super efficient, but this is just a registration time. And we do cache
8143 * the last compound head, so generally we'll only do a full search if we don't
8146 * We check if the given compound head page has already been accounted, to
8147 * avoid double accounting it. This allows us to account the full size of the
8148 * page, not just the constituent pages of a huge page.
8150 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8151 int nr_pages, struct page *hpage)
8155 /* check current page array */
8156 for (i = 0; i < nr_pages; i++) {
8157 if (!PageCompound(pages[i]))
8159 if (compound_head(pages[i]) == hpage)
8163 /* check previously registered pages */
8164 for (i = 0; i < ctx->nr_user_bufs; i++) {
8165 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8167 for (j = 0; j < imu->nr_bvecs; j++) {
8168 if (!PageCompound(imu->bvec[j].bv_page))
8170 if (compound_head(imu->bvec[j].bv_page) == hpage)
8178 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8179 int nr_pages, struct io_mapped_ubuf *imu,
8180 struct page **last_hpage)
8184 for (i = 0; i < nr_pages; i++) {
8185 if (!PageCompound(pages[i])) {
8190 hpage = compound_head(pages[i]);
8191 if (hpage == *last_hpage)
8193 *last_hpage = hpage;
8194 if (headpage_already_acct(ctx, pages, i, hpage))
8196 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8200 if (!imu->acct_pages)
8203 ret = io_account_mem(ctx, imu->acct_pages);
8205 imu->acct_pages = 0;
8209 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8210 struct io_mapped_ubuf *imu,
8211 struct page **last_hpage)
8213 struct vm_area_struct **vmas = NULL;
8214 struct page **pages = NULL;
8215 unsigned long off, start, end, ubuf;
8217 int ret, pret, nr_pages, i;
8219 ubuf = (unsigned long) iov->iov_base;
8220 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8221 start = ubuf >> PAGE_SHIFT;
8222 nr_pages = end - start;
8226 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8230 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8235 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8241 mmap_read_lock(current->mm);
8242 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8244 if (pret == nr_pages) {
8245 /* don't support file backed memory */
8246 for (i = 0; i < nr_pages; i++) {
8247 struct vm_area_struct *vma = vmas[i];
8250 !is_file_hugepages(vma->vm_file)) {
8256 ret = pret < 0 ? pret : -EFAULT;
8258 mmap_read_unlock(current->mm);
8261 * if we did partial map, or found file backed vmas,
8262 * release any pages we did get
8265 unpin_user_pages(pages, pret);
8270 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8272 unpin_user_pages(pages, pret);
8277 off = ubuf & ~PAGE_MASK;
8278 size = iov->iov_len;
8279 for (i = 0; i < nr_pages; i++) {
8282 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8283 imu->bvec[i].bv_page = pages[i];
8284 imu->bvec[i].bv_len = vec_len;
8285 imu->bvec[i].bv_offset = off;
8289 /* store original address for later verification */
8291 imu->len = iov->iov_len;
8292 imu->nr_bvecs = nr_pages;
8300 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8304 if (!nr_args || nr_args > UIO_MAXIOV)
8307 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8309 if (!ctx->user_bufs)
8315 static int io_buffer_validate(struct iovec *iov)
8318 * Don't impose further limits on the size and buffer
8319 * constraints here, we'll -EINVAL later when IO is
8320 * submitted if they are wrong.
8322 if (!iov->iov_base || !iov->iov_len)
8325 /* arbitrary limit, but we need something */
8326 if (iov->iov_len > SZ_1G)
8332 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8333 unsigned int nr_args)
8337 struct page *last_hpage = NULL;
8339 ret = io_buffers_map_alloc(ctx, nr_args);
8343 for (i = 0; i < nr_args; i++) {
8344 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8346 ret = io_copy_iov(ctx, &iov, arg, i);
8350 ret = io_buffer_validate(&iov);
8354 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8358 ctx->nr_user_bufs++;
8362 io_sqe_buffers_unregister(ctx);
8367 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8369 __s32 __user *fds = arg;
8375 if (copy_from_user(&fd, fds, sizeof(*fds)))
8378 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8379 if (IS_ERR(ctx->cq_ev_fd)) {
8380 int ret = PTR_ERR(ctx->cq_ev_fd);
8381 ctx->cq_ev_fd = NULL;
8388 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8390 if (ctx->cq_ev_fd) {
8391 eventfd_ctx_put(ctx->cq_ev_fd);
8392 ctx->cq_ev_fd = NULL;
8399 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8401 struct io_buffer *buf;
8402 unsigned long index;
8404 xa_for_each(&ctx->io_buffers, index, buf)
8405 __io_remove_buffers(ctx, buf, index, -1U);
8408 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8410 struct io_kiocb *req, *nxt;
8412 list_for_each_entry_safe(req, nxt, list, compl.list) {
8413 if (tsk && req->task != tsk)
8415 list_del(&req->compl.list);
8416 kmem_cache_free(req_cachep, req);
8420 static void io_req_caches_free(struct io_ring_ctx *ctx)
8422 struct io_submit_state *submit_state = &ctx->submit_state;
8423 struct io_comp_state *cs = &ctx->submit_state.comp;
8425 mutex_lock(&ctx->uring_lock);
8427 if (submit_state->free_reqs) {
8428 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8429 submit_state->reqs);
8430 submit_state->free_reqs = 0;
8433 io_flush_cached_locked_reqs(ctx, cs);
8434 io_req_cache_free(&cs->free_list, NULL);
8435 mutex_unlock(&ctx->uring_lock);
8438 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8441 * Some may use context even when all refs and requests have been put,
8442 * and they are free to do so while still holding uring_lock or
8443 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8445 mutex_lock(&ctx->uring_lock);
8446 mutex_unlock(&ctx->uring_lock);
8447 spin_lock_irq(&ctx->completion_lock);
8448 spin_unlock_irq(&ctx->completion_lock);
8450 io_sq_thread_finish(ctx);
8451 io_sqe_buffers_unregister(ctx);
8453 if (ctx->mm_account) {
8454 mmdrop(ctx->mm_account);
8455 ctx->mm_account = NULL;
8458 mutex_lock(&ctx->uring_lock);
8459 io_sqe_files_unregister(ctx);
8460 mutex_unlock(&ctx->uring_lock);
8461 io_eventfd_unregister(ctx);
8462 io_destroy_buffers(ctx);
8464 if (ctx->rsrc_backup_node)
8465 destroy_fixed_rsrc_ref_node(ctx->rsrc_backup_node);
8467 #if defined(CONFIG_UNIX)
8468 if (ctx->ring_sock) {
8469 ctx->ring_sock->file = NULL; /* so that iput() is called */
8470 sock_release(ctx->ring_sock);
8474 io_mem_free(ctx->rings);
8475 io_mem_free(ctx->sq_sqes);
8477 percpu_ref_exit(&ctx->refs);
8478 free_uid(ctx->user);
8479 io_req_caches_free(ctx);
8481 io_wq_put_hash(ctx->hash_map);
8482 kfree(ctx->cancel_hash);
8486 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8488 struct io_ring_ctx *ctx = file->private_data;
8491 poll_wait(file, &ctx->cq_wait, wait);
8493 * synchronizes with barrier from wq_has_sleeper call in
8497 if (!io_sqring_full(ctx))
8498 mask |= EPOLLOUT | EPOLLWRNORM;
8501 * Don't flush cqring overflow list here, just do a simple check.
8502 * Otherwise there could possible be ABBA deadlock:
8505 * lock(&ctx->uring_lock);
8507 * lock(&ctx->uring_lock);
8510 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8511 * pushs them to do the flush.
8513 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8514 mask |= EPOLLIN | EPOLLRDNORM;
8519 static int io_uring_fasync(int fd, struct file *file, int on)
8521 struct io_ring_ctx *ctx = file->private_data;
8523 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8526 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8528 const struct cred *creds;
8530 creds = xa_erase(&ctx->personalities, id);
8539 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8541 return io_run_task_work_head(&ctx->exit_task_work);
8544 struct io_tctx_exit {
8545 struct callback_head task_work;
8546 struct completion completion;
8547 struct io_ring_ctx *ctx;
8550 static void io_tctx_exit_cb(struct callback_head *cb)
8552 struct io_uring_task *tctx = current->io_uring;
8553 struct io_tctx_exit *work;
8555 work = container_of(cb, struct io_tctx_exit, task_work);
8557 * When @in_idle, we're in cancellation and it's racy to remove the
8558 * node. It'll be removed by the end of cancellation, just ignore it.
8560 if (!atomic_read(&tctx->in_idle))
8561 io_uring_del_task_file((unsigned long)work->ctx);
8562 complete(&work->completion);
8565 static void io_ring_exit_work(struct work_struct *work)
8567 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8568 unsigned long timeout = jiffies + HZ * 60 * 5;
8569 struct io_tctx_exit exit;
8570 struct io_tctx_node *node;
8573 /* prevent SQPOLL from submitting new requests */
8575 io_sq_thread_park(ctx->sq_data);
8576 list_del_init(&ctx->sqd_list);
8577 io_sqd_update_thread_idle(ctx->sq_data);
8578 io_sq_thread_unpark(ctx->sq_data);
8582 * If we're doing polled IO and end up having requests being
8583 * submitted async (out-of-line), then completions can come in while
8584 * we're waiting for refs to drop. We need to reap these manually,
8585 * as nobody else will be looking for them.
8588 io_uring_try_cancel_requests(ctx, NULL, NULL);
8590 WARN_ON_ONCE(time_after(jiffies, timeout));
8591 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8593 mutex_lock(&ctx->uring_lock);
8594 while (!list_empty(&ctx->tctx_list)) {
8595 WARN_ON_ONCE(time_after(jiffies, timeout));
8597 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8600 init_completion(&exit.completion);
8601 init_task_work(&exit.task_work, io_tctx_exit_cb);
8602 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8603 if (WARN_ON_ONCE(ret))
8605 wake_up_process(node->task);
8607 mutex_unlock(&ctx->uring_lock);
8608 wait_for_completion(&exit.completion);
8610 mutex_lock(&ctx->uring_lock);
8612 mutex_unlock(&ctx->uring_lock);
8614 io_ring_ctx_free(ctx);
8617 /* Returns true if we found and killed one or more timeouts */
8618 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8619 struct files_struct *files)
8621 struct io_kiocb *req, *tmp;
8624 spin_lock_irq(&ctx->completion_lock);
8625 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8626 if (io_match_task(req, tsk, files)) {
8627 io_kill_timeout(req, -ECANCELED);
8632 io_commit_cqring(ctx);
8633 spin_unlock_irq(&ctx->completion_lock);
8635 io_cqring_ev_posted(ctx);
8636 return canceled != 0;
8639 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8641 unsigned long index;
8642 struct creds *creds;
8644 mutex_lock(&ctx->uring_lock);
8645 percpu_ref_kill(&ctx->refs);
8646 /* if force is set, the ring is going away. always drop after that */
8647 ctx->cq_overflow_flushed = 1;
8649 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8650 xa_for_each(&ctx->personalities, index, creds)
8651 io_unregister_personality(ctx, index);
8652 mutex_unlock(&ctx->uring_lock);
8654 io_kill_timeouts(ctx, NULL, NULL);
8655 io_poll_remove_all(ctx, NULL, NULL);
8657 /* if we failed setting up the ctx, we might not have any rings */
8658 io_iopoll_try_reap_events(ctx);
8660 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8662 * Use system_unbound_wq to avoid spawning tons of event kworkers
8663 * if we're exiting a ton of rings at the same time. It just adds
8664 * noise and overhead, there's no discernable change in runtime
8665 * over using system_wq.
8667 queue_work(system_unbound_wq, &ctx->exit_work);
8670 static int io_uring_release(struct inode *inode, struct file *file)
8672 struct io_ring_ctx *ctx = file->private_data;
8674 file->private_data = NULL;
8675 io_ring_ctx_wait_and_kill(ctx);
8679 struct io_task_cancel {
8680 struct task_struct *task;
8681 struct files_struct *files;
8684 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8686 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8687 struct io_task_cancel *cancel = data;
8690 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8691 unsigned long flags;
8692 struct io_ring_ctx *ctx = req->ctx;
8694 /* protect against races with linked timeouts */
8695 spin_lock_irqsave(&ctx->completion_lock, flags);
8696 ret = io_match_task(req, cancel->task, cancel->files);
8697 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8699 ret = io_match_task(req, cancel->task, cancel->files);
8704 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8705 struct task_struct *task,
8706 struct files_struct *files)
8708 struct io_defer_entry *de;
8711 spin_lock_irq(&ctx->completion_lock);
8712 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8713 if (io_match_task(de->req, task, files)) {
8714 list_cut_position(&list, &ctx->defer_list, &de->list);
8718 spin_unlock_irq(&ctx->completion_lock);
8719 if (list_empty(&list))
8722 while (!list_empty(&list)) {
8723 de = list_first_entry(&list, struct io_defer_entry, list);
8724 list_del_init(&de->list);
8725 io_req_complete_failed(de->req, -ECANCELED);
8731 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8733 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8735 return req->ctx == data;
8738 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8740 struct io_tctx_node *node;
8741 enum io_wq_cancel cret;
8744 mutex_lock(&ctx->uring_lock);
8745 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8746 struct io_uring_task *tctx = node->task->io_uring;
8749 * io_wq will stay alive while we hold uring_lock, because it's
8750 * killed after ctx nodes, which requires to take the lock.
8752 if (!tctx || !tctx->io_wq)
8754 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8755 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8757 mutex_unlock(&ctx->uring_lock);
8762 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8763 struct task_struct *task,
8764 struct files_struct *files)
8766 struct io_task_cancel cancel = { .task = task, .files = files, };
8767 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8770 enum io_wq_cancel cret;
8774 ret |= io_uring_try_cancel_iowq(ctx);
8775 } else if (tctx && tctx->io_wq) {
8777 * Cancels requests of all rings, not only @ctx, but
8778 * it's fine as the task is in exit/exec.
8780 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8782 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8785 /* SQPOLL thread does its own polling */
8786 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8787 (ctx->sq_data && ctx->sq_data->thread == current)) {
8788 while (!list_empty_careful(&ctx->iopoll_list)) {
8789 io_iopoll_try_reap_events(ctx);
8794 ret |= io_cancel_defer_files(ctx, task, files);
8795 ret |= io_poll_remove_all(ctx, task, files);
8796 ret |= io_kill_timeouts(ctx, task, files);
8797 ret |= io_run_task_work();
8798 ret |= io_run_ctx_fallback(ctx);
8799 io_cqring_overflow_flush(ctx, true, task, files);
8806 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8807 struct task_struct *task,
8808 struct files_struct *files)
8810 struct io_kiocb *req;
8813 spin_lock_irq(&ctx->inflight_lock);
8814 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8815 cnt += io_match_task(req, task, files);
8816 spin_unlock_irq(&ctx->inflight_lock);
8820 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8821 struct task_struct *task,
8822 struct files_struct *files)
8824 while (!list_empty_careful(&ctx->inflight_list)) {
8828 inflight = io_uring_count_inflight(ctx, task, files);
8832 io_uring_try_cancel_requests(ctx, task, files);
8834 prepare_to_wait(&task->io_uring->wait, &wait,
8835 TASK_UNINTERRUPTIBLE);
8836 if (inflight == io_uring_count_inflight(ctx, task, files))
8838 finish_wait(&task->io_uring->wait, &wait);
8842 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8844 struct io_uring_task *tctx = current->io_uring;
8845 struct io_tctx_node *node;
8848 if (unlikely(!tctx)) {
8849 ret = io_uring_alloc_task_context(current, ctx);
8852 tctx = current->io_uring;
8854 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8855 node = kmalloc(sizeof(*node), GFP_KERNEL);
8859 node->task = current;
8861 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8868 mutex_lock(&ctx->uring_lock);
8869 list_add(&node->ctx_node, &ctx->tctx_list);
8870 mutex_unlock(&ctx->uring_lock);
8877 * Note that this task has used io_uring. We use it for cancelation purposes.
8879 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
8881 struct io_uring_task *tctx = current->io_uring;
8883 if (likely(tctx && tctx->last == ctx))
8885 return __io_uring_add_task_file(ctx);
8889 * Remove this io_uring_file -> task mapping.
8891 static void io_uring_del_task_file(unsigned long index)
8893 struct io_uring_task *tctx = current->io_uring;
8894 struct io_tctx_node *node;
8898 node = xa_erase(&tctx->xa, index);
8902 WARN_ON_ONCE(current != node->task);
8903 WARN_ON_ONCE(list_empty(&node->ctx_node));
8905 mutex_lock(&node->ctx->uring_lock);
8906 list_del(&node->ctx_node);
8907 mutex_unlock(&node->ctx->uring_lock);
8909 if (tctx->last == node->ctx)
8914 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8916 struct io_tctx_node *node;
8917 unsigned long index;
8919 xa_for_each(&tctx->xa, index, node)
8920 io_uring_del_task_file(index);
8922 io_wq_put_and_exit(tctx->io_wq);
8927 static s64 tctx_inflight(struct io_uring_task *tctx)
8929 return percpu_counter_sum(&tctx->inflight);
8932 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8934 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8935 struct io_ring_ctx *ctx = work->ctx;
8936 struct io_sq_data *sqd = ctx->sq_data;
8939 io_uring_cancel_sqpoll(ctx);
8940 complete(&work->completion);
8943 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8945 struct io_sq_data *sqd = ctx->sq_data;
8946 struct io_tctx_exit work = { .ctx = ctx, };
8947 struct task_struct *task;
8949 io_sq_thread_park(sqd);
8950 list_del_init(&ctx->sqd_list);
8951 io_sqd_update_thread_idle(sqd);
8954 init_completion(&work.completion);
8955 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8956 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8957 wake_up_process(task);
8959 io_sq_thread_unpark(sqd);
8962 wait_for_completion(&work.completion);
8965 void __io_uring_files_cancel(struct files_struct *files)
8967 struct io_uring_task *tctx = current->io_uring;
8968 struct io_tctx_node *node;
8969 unsigned long index;
8971 /* make sure overflow events are dropped */
8972 atomic_inc(&tctx->in_idle);
8973 xa_for_each(&tctx->xa, index, node) {
8974 struct io_ring_ctx *ctx = node->ctx;
8977 io_sqpoll_cancel_sync(ctx);
8980 io_uring_cancel_files(ctx, current, files);
8982 io_uring_try_cancel_requests(ctx, current, NULL);
8984 atomic_dec(&tctx->in_idle);
8987 io_uring_clean_tctx(tctx);
8990 /* should only be called by SQPOLL task */
8991 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8993 struct io_sq_data *sqd = ctx->sq_data;
8994 struct io_uring_task *tctx = current->io_uring;
8998 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
9000 atomic_inc(&tctx->in_idle);
9002 /* read completions before cancelations */
9003 inflight = tctx_inflight(tctx);
9006 io_uring_try_cancel_requests(ctx, current, NULL);
9008 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9010 * If we've seen completions, retry without waiting. This
9011 * avoids a race where a completion comes in before we did
9012 * prepare_to_wait().
9014 if (inflight == tctx_inflight(tctx))
9016 finish_wait(&tctx->wait, &wait);
9018 atomic_dec(&tctx->in_idle);
9022 * Find any io_uring fd that this task has registered or done IO on, and cancel
9025 void __io_uring_task_cancel(void)
9027 struct io_uring_task *tctx = current->io_uring;
9031 /* make sure overflow events are dropped */
9032 atomic_inc(&tctx->in_idle);
9033 __io_uring_files_cancel(NULL);
9036 /* read completions before cancelations */
9037 inflight = tctx_inflight(tctx);
9040 __io_uring_files_cancel(NULL);
9042 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9045 * If we've seen completions, retry without waiting. This
9046 * avoids a race where a completion comes in before we did
9047 * prepare_to_wait().
9049 if (inflight == tctx_inflight(tctx))
9051 finish_wait(&tctx->wait, &wait);
9054 atomic_dec(&tctx->in_idle);
9056 io_uring_clean_tctx(tctx);
9057 /* all current's requests should be gone, we can kill tctx */
9058 __io_uring_free(current);
9061 static void *io_uring_validate_mmap_request(struct file *file,
9062 loff_t pgoff, size_t sz)
9064 struct io_ring_ctx *ctx = file->private_data;
9065 loff_t offset = pgoff << PAGE_SHIFT;
9070 case IORING_OFF_SQ_RING:
9071 case IORING_OFF_CQ_RING:
9074 case IORING_OFF_SQES:
9078 return ERR_PTR(-EINVAL);
9081 page = virt_to_head_page(ptr);
9082 if (sz > page_size(page))
9083 return ERR_PTR(-EINVAL);
9090 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9092 size_t sz = vma->vm_end - vma->vm_start;
9096 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9098 return PTR_ERR(ptr);
9100 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9101 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9104 #else /* !CONFIG_MMU */
9106 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9108 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9111 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9113 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9116 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9117 unsigned long addr, unsigned long len,
9118 unsigned long pgoff, unsigned long flags)
9122 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9124 return PTR_ERR(ptr);
9126 return (unsigned long) ptr;
9129 #endif /* !CONFIG_MMU */
9131 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9136 if (!io_sqring_full(ctx))
9138 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9140 if (!io_sqring_full(ctx))
9143 } while (!signal_pending(current));
9145 finish_wait(&ctx->sqo_sq_wait, &wait);
9149 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9150 struct __kernel_timespec __user **ts,
9151 const sigset_t __user **sig)
9153 struct io_uring_getevents_arg arg;
9156 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9157 * is just a pointer to the sigset_t.
9159 if (!(flags & IORING_ENTER_EXT_ARG)) {
9160 *sig = (const sigset_t __user *) argp;
9166 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9167 * timespec and sigset_t pointers if good.
9169 if (*argsz != sizeof(arg))
9171 if (copy_from_user(&arg, argp, sizeof(arg)))
9173 *sig = u64_to_user_ptr(arg.sigmask);
9174 *argsz = arg.sigmask_sz;
9175 *ts = u64_to_user_ptr(arg.ts);
9179 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9180 u32, min_complete, u32, flags, const void __user *, argp,
9183 struct io_ring_ctx *ctx;
9190 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9191 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9195 if (unlikely(!f.file))
9199 if (unlikely(f.file->f_op != &io_uring_fops))
9203 ctx = f.file->private_data;
9204 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9208 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9212 * For SQ polling, the thread will do all submissions and completions.
9213 * Just return the requested submit count, and wake the thread if
9217 if (ctx->flags & IORING_SETUP_SQPOLL) {
9218 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9221 if (unlikely(ctx->sq_data->thread == NULL)) {
9224 if (flags & IORING_ENTER_SQ_WAKEUP)
9225 wake_up(&ctx->sq_data->wait);
9226 if (flags & IORING_ENTER_SQ_WAIT) {
9227 ret = io_sqpoll_wait_sq(ctx);
9231 submitted = to_submit;
9232 } else if (to_submit) {
9233 ret = io_uring_add_task_file(ctx);
9236 mutex_lock(&ctx->uring_lock);
9237 submitted = io_submit_sqes(ctx, to_submit);
9238 mutex_unlock(&ctx->uring_lock);
9240 if (submitted != to_submit)
9243 if (flags & IORING_ENTER_GETEVENTS) {
9244 const sigset_t __user *sig;
9245 struct __kernel_timespec __user *ts;
9247 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9251 min_complete = min(min_complete, ctx->cq_entries);
9254 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9255 * space applications don't need to do io completion events
9256 * polling again, they can rely on io_sq_thread to do polling
9257 * work, which can reduce cpu usage and uring_lock contention.
9259 if (ctx->flags & IORING_SETUP_IOPOLL &&
9260 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9261 ret = io_iopoll_check(ctx, min_complete);
9263 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9268 percpu_ref_put(&ctx->refs);
9271 return submitted ? submitted : ret;
9274 #ifdef CONFIG_PROC_FS
9275 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9276 const struct cred *cred)
9278 struct user_namespace *uns = seq_user_ns(m);
9279 struct group_info *gi;
9284 seq_printf(m, "%5d\n", id);
9285 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9286 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9287 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9288 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9289 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9290 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9291 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9292 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9293 seq_puts(m, "\n\tGroups:\t");
9294 gi = cred->group_info;
9295 for (g = 0; g < gi->ngroups; g++) {
9296 seq_put_decimal_ull(m, g ? " " : "",
9297 from_kgid_munged(uns, gi->gid[g]));
9299 seq_puts(m, "\n\tCapEff:\t");
9300 cap = cred->cap_effective;
9301 CAP_FOR_EACH_U32(__capi)
9302 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9307 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9309 struct io_sq_data *sq = NULL;
9314 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9315 * since fdinfo case grabs it in the opposite direction of normal use
9316 * cases. If we fail to get the lock, we just don't iterate any
9317 * structures that could be going away outside the io_uring mutex.
9319 has_lock = mutex_trylock(&ctx->uring_lock);
9321 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9327 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9328 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9329 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9330 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9331 struct file *f = io_file_from_index(ctx, i);
9334 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9336 seq_printf(m, "%5u: <none>\n", i);
9338 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9339 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9340 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9342 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9343 (unsigned int) buf->len);
9345 if (has_lock && !xa_empty(&ctx->personalities)) {
9346 unsigned long index;
9347 const struct cred *cred;
9349 seq_printf(m, "Personalities:\n");
9350 xa_for_each(&ctx->personalities, index, cred)
9351 io_uring_show_cred(m, index, cred);
9353 seq_printf(m, "PollList:\n");
9354 spin_lock_irq(&ctx->completion_lock);
9355 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9356 struct hlist_head *list = &ctx->cancel_hash[i];
9357 struct io_kiocb *req;
9359 hlist_for_each_entry(req, list, hash_node)
9360 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9361 req->task->task_works != NULL);
9363 spin_unlock_irq(&ctx->completion_lock);
9365 mutex_unlock(&ctx->uring_lock);
9368 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9370 struct io_ring_ctx *ctx = f->private_data;
9372 if (percpu_ref_tryget(&ctx->refs)) {
9373 __io_uring_show_fdinfo(ctx, m);
9374 percpu_ref_put(&ctx->refs);
9379 static const struct file_operations io_uring_fops = {
9380 .release = io_uring_release,
9381 .mmap = io_uring_mmap,
9383 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9384 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9386 .poll = io_uring_poll,
9387 .fasync = io_uring_fasync,
9388 #ifdef CONFIG_PROC_FS
9389 .show_fdinfo = io_uring_show_fdinfo,
9393 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9394 struct io_uring_params *p)
9396 struct io_rings *rings;
9397 size_t size, sq_array_offset;
9399 /* make sure these are sane, as we already accounted them */
9400 ctx->sq_entries = p->sq_entries;
9401 ctx->cq_entries = p->cq_entries;
9403 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9404 if (size == SIZE_MAX)
9407 rings = io_mem_alloc(size);
9412 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9413 rings->sq_ring_mask = p->sq_entries - 1;
9414 rings->cq_ring_mask = p->cq_entries - 1;
9415 rings->sq_ring_entries = p->sq_entries;
9416 rings->cq_ring_entries = p->cq_entries;
9417 ctx->sq_mask = rings->sq_ring_mask;
9418 ctx->cq_mask = rings->cq_ring_mask;
9420 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9421 if (size == SIZE_MAX) {
9422 io_mem_free(ctx->rings);
9427 ctx->sq_sqes = io_mem_alloc(size);
9428 if (!ctx->sq_sqes) {
9429 io_mem_free(ctx->rings);
9437 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9441 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9445 ret = io_uring_add_task_file(ctx);
9450 fd_install(fd, file);
9455 * Allocate an anonymous fd, this is what constitutes the application
9456 * visible backing of an io_uring instance. The application mmaps this
9457 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9458 * we have to tie this fd to a socket for file garbage collection purposes.
9460 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9463 #if defined(CONFIG_UNIX)
9466 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9469 return ERR_PTR(ret);
9472 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9473 O_RDWR | O_CLOEXEC);
9474 #if defined(CONFIG_UNIX)
9476 sock_release(ctx->ring_sock);
9477 ctx->ring_sock = NULL;
9479 ctx->ring_sock->file = file;
9485 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9486 struct io_uring_params __user *params)
9488 struct io_ring_ctx *ctx;
9494 if (entries > IORING_MAX_ENTRIES) {
9495 if (!(p->flags & IORING_SETUP_CLAMP))
9497 entries = IORING_MAX_ENTRIES;
9501 * Use twice as many entries for the CQ ring. It's possible for the
9502 * application to drive a higher depth than the size of the SQ ring,
9503 * since the sqes are only used at submission time. This allows for
9504 * some flexibility in overcommitting a bit. If the application has
9505 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9506 * of CQ ring entries manually.
9508 p->sq_entries = roundup_pow_of_two(entries);
9509 if (p->flags & IORING_SETUP_CQSIZE) {
9511 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9512 * to a power-of-two, if it isn't already. We do NOT impose
9513 * any cq vs sq ring sizing.
9517 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9518 if (!(p->flags & IORING_SETUP_CLAMP))
9520 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9522 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9523 if (p->cq_entries < p->sq_entries)
9526 p->cq_entries = 2 * p->sq_entries;
9529 ctx = io_ring_ctx_alloc(p);
9532 ctx->compat = in_compat_syscall();
9533 if (!capable(CAP_IPC_LOCK))
9534 ctx->user = get_uid(current_user());
9537 * This is just grabbed for accounting purposes. When a process exits,
9538 * the mm is exited and dropped before the files, hence we need to hang
9539 * on to this mm purely for the purposes of being able to unaccount
9540 * memory (locked/pinned vm). It's not used for anything else.
9542 mmgrab(current->mm);
9543 ctx->mm_account = current->mm;
9545 ret = io_allocate_scq_urings(ctx, p);
9549 ret = io_sq_offload_create(ctx, p);
9553 memset(&p->sq_off, 0, sizeof(p->sq_off));
9554 p->sq_off.head = offsetof(struct io_rings, sq.head);
9555 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9556 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9557 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9558 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9559 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9560 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9562 memset(&p->cq_off, 0, sizeof(p->cq_off));
9563 p->cq_off.head = offsetof(struct io_rings, cq.head);
9564 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9565 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9566 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9567 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9568 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9569 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9571 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9572 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9573 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9574 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9575 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9577 if (copy_to_user(params, p, sizeof(*p))) {
9582 file = io_uring_get_file(ctx);
9584 ret = PTR_ERR(file);
9589 * Install ring fd as the very last thing, so we don't risk someone
9590 * having closed it before we finish setup
9592 ret = io_uring_install_fd(ctx, file);
9594 /* fput will clean it up */
9599 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9602 io_ring_ctx_wait_and_kill(ctx);
9607 * Sets up an aio uring context, and returns the fd. Applications asks for a
9608 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9609 * params structure passed in.
9611 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9613 struct io_uring_params p;
9616 if (copy_from_user(&p, params, sizeof(p)))
9618 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9623 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9624 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9625 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9626 IORING_SETUP_R_DISABLED))
9629 return io_uring_create(entries, &p, params);
9632 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9633 struct io_uring_params __user *, params)
9635 return io_uring_setup(entries, params);
9638 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9640 struct io_uring_probe *p;
9644 size = struct_size(p, ops, nr_args);
9645 if (size == SIZE_MAX)
9647 p = kzalloc(size, GFP_KERNEL);
9652 if (copy_from_user(p, arg, size))
9655 if (memchr_inv(p, 0, size))
9658 p->last_op = IORING_OP_LAST - 1;
9659 if (nr_args > IORING_OP_LAST)
9660 nr_args = IORING_OP_LAST;
9662 for (i = 0; i < nr_args; i++) {
9664 if (!io_op_defs[i].not_supported)
9665 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9670 if (copy_to_user(arg, p, size))
9677 static int io_register_personality(struct io_ring_ctx *ctx)
9679 const struct cred *creds;
9683 creds = get_current_cred();
9685 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9686 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9693 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9694 unsigned int nr_args)
9696 struct io_uring_restriction *res;
9700 /* Restrictions allowed only if rings started disabled */
9701 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9704 /* We allow only a single restrictions registration */
9705 if (ctx->restrictions.registered)
9708 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9711 size = array_size(nr_args, sizeof(*res));
9712 if (size == SIZE_MAX)
9715 res = memdup_user(arg, size);
9717 return PTR_ERR(res);
9721 for (i = 0; i < nr_args; i++) {
9722 switch (res[i].opcode) {
9723 case IORING_RESTRICTION_REGISTER_OP:
9724 if (res[i].register_op >= IORING_REGISTER_LAST) {
9729 __set_bit(res[i].register_op,
9730 ctx->restrictions.register_op);
9732 case IORING_RESTRICTION_SQE_OP:
9733 if (res[i].sqe_op >= IORING_OP_LAST) {
9738 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9740 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9741 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9743 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9744 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9753 /* Reset all restrictions if an error happened */
9755 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9757 ctx->restrictions.registered = true;
9763 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9765 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9768 if (ctx->restrictions.registered)
9769 ctx->restricted = 1;
9771 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9772 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9773 wake_up(&ctx->sq_data->wait);
9777 static bool io_register_op_must_quiesce(int op)
9780 case IORING_UNREGISTER_FILES:
9781 case IORING_REGISTER_FILES_UPDATE:
9782 case IORING_REGISTER_PROBE:
9783 case IORING_REGISTER_PERSONALITY:
9784 case IORING_UNREGISTER_PERSONALITY:
9791 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9792 void __user *arg, unsigned nr_args)
9793 __releases(ctx->uring_lock)
9794 __acquires(ctx->uring_lock)
9799 * We're inside the ring mutex, if the ref is already dying, then
9800 * someone else killed the ctx or is already going through
9801 * io_uring_register().
9803 if (percpu_ref_is_dying(&ctx->refs))
9806 if (io_register_op_must_quiesce(opcode)) {
9807 percpu_ref_kill(&ctx->refs);
9810 * Drop uring mutex before waiting for references to exit. If
9811 * another thread is currently inside io_uring_enter() it might
9812 * need to grab the uring_lock to make progress. If we hold it
9813 * here across the drain wait, then we can deadlock. It's safe
9814 * to drop the mutex here, since no new references will come in
9815 * after we've killed the percpu ref.
9817 mutex_unlock(&ctx->uring_lock);
9819 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9822 ret = io_run_task_work_sig();
9827 mutex_lock(&ctx->uring_lock);
9830 percpu_ref_resurrect(&ctx->refs);
9835 if (ctx->restricted) {
9836 if (opcode >= IORING_REGISTER_LAST) {
9841 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9848 case IORING_REGISTER_BUFFERS:
9849 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9851 case IORING_UNREGISTER_BUFFERS:
9855 ret = io_sqe_buffers_unregister(ctx);
9857 case IORING_REGISTER_FILES:
9858 ret = io_sqe_files_register(ctx, arg, nr_args);
9860 case IORING_UNREGISTER_FILES:
9864 ret = io_sqe_files_unregister(ctx);
9866 case IORING_REGISTER_FILES_UPDATE:
9867 ret = io_sqe_files_update(ctx, arg, nr_args);
9869 case IORING_REGISTER_EVENTFD:
9870 case IORING_REGISTER_EVENTFD_ASYNC:
9874 ret = io_eventfd_register(ctx, arg);
9877 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9878 ctx->eventfd_async = 1;
9880 ctx->eventfd_async = 0;
9882 case IORING_UNREGISTER_EVENTFD:
9886 ret = io_eventfd_unregister(ctx);
9888 case IORING_REGISTER_PROBE:
9890 if (!arg || nr_args > 256)
9892 ret = io_probe(ctx, arg, nr_args);
9894 case IORING_REGISTER_PERSONALITY:
9898 ret = io_register_personality(ctx);
9900 case IORING_UNREGISTER_PERSONALITY:
9904 ret = io_unregister_personality(ctx, nr_args);
9906 case IORING_REGISTER_ENABLE_RINGS:
9910 ret = io_register_enable_rings(ctx);
9912 case IORING_REGISTER_RESTRICTIONS:
9913 ret = io_register_restrictions(ctx, arg, nr_args);
9921 if (io_register_op_must_quiesce(opcode)) {
9922 /* bring the ctx back to life */
9923 percpu_ref_reinit(&ctx->refs);
9925 reinit_completion(&ctx->ref_comp);
9930 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9931 void __user *, arg, unsigned int, nr_args)
9933 struct io_ring_ctx *ctx;
9942 if (f.file->f_op != &io_uring_fops)
9945 ctx = f.file->private_data;
9949 mutex_lock(&ctx->uring_lock);
9950 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9951 mutex_unlock(&ctx->uring_lock);
9952 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9953 ctx->cq_ev_fd != NULL, ret);
9959 static int __init io_uring_init(void)
9961 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9962 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9963 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9966 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9967 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9968 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9969 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9970 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9971 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9972 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9973 BUILD_BUG_SQE_ELEM(8, __u64, off);
9974 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9975 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9976 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9977 BUILD_BUG_SQE_ELEM(24, __u32, len);
9978 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9979 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9980 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9981 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9982 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9983 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9984 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9985 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9986 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9987 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9988 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9989 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9990 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9991 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9992 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9993 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9994 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9995 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9996 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9998 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9999 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10000 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10004 __initcall(io_uring_init);