1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
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 IORING_MAX_REG_BUFFERS (1U << 14)
105 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
110 u32 head ____cacheline_aligned_in_smp;
111 u32 tail ____cacheline_aligned_in_smp;
115 * This data is shared with the application through the mmap at offsets
116 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
118 * The offsets to the member fields are published through struct
119 * io_sqring_offsets when calling io_uring_setup.
123 * Head and tail offsets into the ring; the offsets need to be
124 * masked to get valid indices.
126 * The kernel controls head of the sq ring and the tail of the cq ring,
127 * and the application controls tail of the sq ring and the head of the
130 struct io_uring sq, cq;
132 * Bitmasks to apply to head and tail offsets (constant, equals
135 u32 sq_ring_mask, cq_ring_mask;
136 /* Ring sizes (constant, power of 2) */
137 u32 sq_ring_entries, cq_ring_entries;
139 * Number of invalid entries dropped by the kernel due to
140 * invalid index stored in array
142 * Written by the kernel, shouldn't be modified by the
143 * application (i.e. get number of "new events" by comparing to
146 * After a new SQ head value was read by the application this
147 * counter includes all submissions that were dropped reaching
148 * the new SQ head (and possibly more).
154 * Written by the kernel, shouldn't be modified by the
157 * The application needs a full memory barrier before checking
158 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
164 * Written by the application, shouldn't be modified by the
169 * Number of completion events lost because the queue was full;
170 * this should be avoided by the application by making sure
171 * there are not more requests pending than there is space in
172 * the completion queue.
174 * Written by the kernel, shouldn't be modified by the
175 * application (i.e. get number of "new events" by comparing to
178 * As completion events come in out of order this counter is not
179 * ordered with any other data.
183 * Ring buffer of completion events.
185 * The kernel writes completion events fresh every time they are
186 * produced, so the application is allowed to modify pending
189 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
192 enum io_uring_cmd_flags {
193 IO_URING_F_NONBLOCK = 1,
194 IO_URING_F_COMPLETE_DEFER = 2,
197 struct io_mapped_ubuf {
200 unsigned int nr_bvecs;
201 unsigned long acct_pages;
202 struct bio_vec bvec[];
207 struct io_overflow_cqe {
208 struct io_uring_cqe cqe;
209 struct list_head list;
212 struct io_fixed_file {
213 /* file * with additional FFS_* flags */
214 unsigned long file_ptr;
218 struct list_head list;
223 struct io_mapped_ubuf *buf;
227 struct io_file_table {
228 /* two level table */
229 struct io_fixed_file **files;
232 struct io_rsrc_node {
233 struct percpu_ref refs;
234 struct list_head node;
235 struct list_head rsrc_list;
236 struct io_rsrc_data *rsrc_data;
237 struct llist_node llist;
241 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
243 struct io_rsrc_data {
244 struct io_ring_ctx *ctx;
249 struct completion done;
254 struct list_head list;
260 struct io_restriction {
261 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
262 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
263 u8 sqe_flags_allowed;
264 u8 sqe_flags_required;
269 IO_SQ_THREAD_SHOULD_STOP = 0,
270 IO_SQ_THREAD_SHOULD_PARK,
275 atomic_t park_pending;
278 /* ctx's that are using this sqd */
279 struct list_head ctx_list;
281 struct task_struct *thread;
282 struct wait_queue_head wait;
284 unsigned sq_thread_idle;
290 struct completion exited;
293 #define IO_IOPOLL_BATCH 8
294 #define IO_COMPL_BATCH 32
295 #define IO_REQ_CACHE_SIZE 32
296 #define IO_REQ_ALLOC_BATCH 8
298 struct io_comp_state {
299 struct io_kiocb *reqs[IO_COMPL_BATCH];
301 /* inline/task_work completion list, under ->uring_lock */
302 struct list_head free_list;
305 struct io_submit_link {
306 struct io_kiocb *head;
307 struct io_kiocb *last;
310 struct io_submit_state {
311 struct blk_plug plug;
312 struct io_submit_link link;
315 * io_kiocb alloc cache
317 void *reqs[IO_REQ_CACHE_SIZE];
318 unsigned int free_reqs;
323 * Batch completion logic
325 struct io_comp_state comp;
328 * File reference cache
332 unsigned int file_refs;
333 unsigned int ios_left;
338 struct percpu_ref refs;
339 } ____cacheline_aligned_in_smp;
343 unsigned int compat: 1;
344 unsigned int drain_next: 1;
345 unsigned int eventfd_async: 1;
346 unsigned int restricted: 1;
349 * Ring buffer of indices into array of io_uring_sqe, which is
350 * mmapped by the application using the IORING_OFF_SQES offset.
352 * This indirection could e.g. be used to assign fixed
353 * io_uring_sqe entries to operations and only submit them to
354 * the queue when needed.
356 * The kernel modifies neither the indices array nor the entries
360 unsigned cached_sq_head;
362 unsigned sq_thread_idle;
363 unsigned cached_sq_dropped;
364 unsigned long sq_check_overflow;
366 struct list_head defer_list;
367 struct list_head timeout_list;
368 struct list_head cq_overflow_list;
370 struct io_uring_sqe *sq_sqes;
371 } ____cacheline_aligned_in_smp;
374 struct mutex uring_lock;
375 wait_queue_head_t wait;
376 } ____cacheline_aligned_in_smp;
378 struct io_submit_state submit_state;
379 /* IRQ completion list, under ->completion_lock */
380 struct list_head locked_free_list;
381 unsigned int locked_free_nr;
383 struct io_rings *rings;
385 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
386 struct io_sq_data *sq_data; /* if using sq thread polling */
388 struct wait_queue_head sqo_sq_wait;
389 struct list_head sqd_list;
392 * Fixed resources fast path, should be accessed only under uring_lock,
393 * and updated through io_uring_register(2)
395 struct io_rsrc_node *rsrc_node;
397 struct io_file_table file_table;
398 unsigned nr_user_files;
399 unsigned nr_user_bufs;
400 struct io_mapped_ubuf **user_bufs;
402 struct xarray io_buffers;
403 struct xarray personalities;
407 unsigned cached_cq_tail;
409 atomic_t cq_timeouts;
410 unsigned cq_last_tm_flush;
412 unsigned long cq_check_overflow;
413 struct wait_queue_head cq_wait;
414 struct fasync_struct *cq_fasync;
415 struct eventfd_ctx *cq_ev_fd;
416 } ____cacheline_aligned_in_smp;
419 spinlock_t completion_lock;
422 * ->iopoll_list is protected by the ctx->uring_lock for
423 * io_uring instances that don't use IORING_SETUP_SQPOLL.
424 * For SQPOLL, only the single threaded io_sq_thread() will
425 * manipulate the list, hence no extra locking is needed there.
427 struct list_head iopoll_list;
428 struct hlist_head *cancel_hash;
429 unsigned cancel_hash_bits;
430 bool poll_multi_file;
431 } ____cacheline_aligned_in_smp;
433 struct io_restriction restrictions;
435 /* slow path rsrc auxilary data, used by update/register */
437 struct io_rsrc_node *rsrc_backup_node;
438 struct io_mapped_ubuf *dummy_ubuf;
439 struct io_rsrc_data *file_data;
440 struct io_rsrc_data *buf_data;
442 struct delayed_work rsrc_put_work;
443 struct llist_head rsrc_put_llist;
444 struct list_head rsrc_ref_list;
445 spinlock_t rsrc_ref_lock;
448 /* Keep this last, we don't need it for the fast path */
450 #if defined(CONFIG_UNIX)
451 struct socket *ring_sock;
453 /* hashed buffered write serialization */
454 struct io_wq_hash *hash_map;
456 /* Only used for accounting purposes */
457 struct user_struct *user;
458 struct mm_struct *mm_account;
460 /* ctx exit and cancelation */
461 struct callback_head *exit_task_work;
462 struct work_struct exit_work;
463 struct list_head tctx_list;
464 struct completion ref_comp;
468 struct io_uring_task {
469 /* submission side */
471 struct wait_queue_head wait;
472 const struct io_ring_ctx *last;
474 struct percpu_counter inflight;
475 atomic_t inflight_tracked;
478 spinlock_t task_lock;
479 struct io_wq_work_list task_list;
480 unsigned long task_state;
481 struct callback_head task_work;
485 * First field must be the file pointer in all the
486 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
488 struct io_poll_iocb {
490 struct wait_queue_head *head;
494 struct wait_queue_entry wait;
497 struct io_poll_update {
503 bool update_user_data;
511 struct io_timeout_data {
512 struct io_kiocb *req;
513 struct hrtimer timer;
514 struct timespec64 ts;
515 enum hrtimer_mode mode;
520 struct sockaddr __user *addr;
521 int __user *addr_len;
523 unsigned long nofile;
543 struct list_head list;
544 /* head of the link, used by linked timeouts only */
545 struct io_kiocb *head;
548 struct io_timeout_rem {
553 struct timespec64 ts;
558 /* NOTE: kiocb has the file as the first member, so don't do it here */
566 struct sockaddr __user *addr;
573 struct compat_msghdr __user *umsg_compat;
574 struct user_msghdr __user *umsg;
580 struct io_buffer *kbuf;
586 struct filename *filename;
588 unsigned long nofile;
591 struct io_rsrc_update {
617 struct epoll_event event;
621 struct file *file_out;
622 struct file *file_in;
629 struct io_provide_buf {
643 const char __user *filename;
644 struct statx __user *buffer;
656 struct filename *oldpath;
657 struct filename *newpath;
665 struct filename *filename;
668 struct io_completion {
670 struct list_head list;
674 struct io_async_connect {
675 struct sockaddr_storage address;
678 struct io_async_msghdr {
679 struct iovec fast_iov[UIO_FASTIOV];
680 /* points to an allocated iov, if NULL we use fast_iov instead */
681 struct iovec *free_iov;
682 struct sockaddr __user *uaddr;
684 struct sockaddr_storage addr;
688 struct iovec fast_iov[UIO_FASTIOV];
689 const struct iovec *free_iovec;
690 struct iov_iter iter;
692 struct wait_page_queue wpq;
696 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
697 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
698 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
699 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
700 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
701 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
703 /* first byte is taken by user flags, shift it to not overlap */
708 REQ_F_LINK_TIMEOUT_BIT,
709 REQ_F_NEED_CLEANUP_BIT,
711 REQ_F_BUFFER_SELECTED_BIT,
712 REQ_F_LTIMEOUT_ACTIVE_BIT,
713 REQ_F_COMPLETE_INLINE_BIT,
715 REQ_F_DONT_REISSUE_BIT,
716 /* keep async read/write and isreg together and in order */
717 REQ_F_ASYNC_READ_BIT,
718 REQ_F_ASYNC_WRITE_BIT,
721 /* not a real bit, just to check we're not overflowing the space */
727 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
728 /* drain existing IO first */
729 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
731 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
732 /* doesn't sever on completion < 0 */
733 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
735 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
736 /* IOSQE_BUFFER_SELECT */
737 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
739 /* fail rest of links */
740 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
741 /* on inflight list, should be cancelled and waited on exit reliably */
742 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
743 /* read/write uses file position */
744 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
745 /* must not punt to workers */
746 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
747 /* has or had linked timeout */
748 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
750 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
751 /* already went through poll handler */
752 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
753 /* buffer already selected */
754 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
755 /* linked timeout is active, i.e. prepared by link's head */
756 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
757 /* completion is deferred through io_comp_state */
758 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
759 /* caller should reissue async */
760 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
761 /* don't attempt request reissue, see io_rw_reissue() */
762 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
763 /* supports async reads */
764 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
765 /* supports async writes */
766 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
768 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
772 struct io_poll_iocb poll;
773 struct io_poll_iocb *double_poll;
776 struct io_task_work {
777 struct io_wq_work_node node;
778 task_work_func_t func;
782 IORING_RSRC_FILE = 0,
783 IORING_RSRC_BUFFER = 1,
787 * NOTE! Each of the iocb union members has the file pointer
788 * as the first entry in their struct definition. So you can
789 * access the file pointer through any of the sub-structs,
790 * or directly as just 'ki_filp' in this struct.
796 struct io_poll_iocb poll;
797 struct io_poll_update poll_update;
798 struct io_accept accept;
800 struct io_cancel cancel;
801 struct io_timeout timeout;
802 struct io_timeout_rem timeout_rem;
803 struct io_connect connect;
804 struct io_sr_msg sr_msg;
806 struct io_close close;
807 struct io_rsrc_update rsrc_update;
808 struct io_fadvise fadvise;
809 struct io_madvise madvise;
810 struct io_epoll epoll;
811 struct io_splice splice;
812 struct io_provide_buf pbuf;
813 struct io_statx statx;
814 struct io_shutdown shutdown;
815 struct io_rename rename;
816 struct io_unlink unlink;
817 /* use only after cleaning per-op data, see io_clean_op() */
818 struct io_completion compl;
821 /* opcode allocated if it needs to store data for async defer */
824 /* polled IO has completed */
830 struct io_ring_ctx *ctx;
833 struct task_struct *task;
836 struct io_kiocb *link;
837 struct percpu_ref *fixed_rsrc_refs;
839 /* used with ctx->iopoll_list with reads/writes */
840 struct list_head inflight_entry;
842 struct io_task_work io_task_work;
843 struct callback_head task_work;
845 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
846 struct hlist_node hash_node;
847 struct async_poll *apoll;
848 struct io_wq_work work;
849 /* store used ubuf, so we can prevent reloading */
850 struct io_mapped_ubuf *imu;
853 struct io_tctx_node {
854 struct list_head ctx_node;
855 struct task_struct *task;
856 struct io_ring_ctx *ctx;
859 struct io_defer_entry {
860 struct list_head list;
861 struct io_kiocb *req;
866 /* needs req->file assigned */
867 unsigned needs_file : 1;
868 /* hash wq insertion if file is a regular file */
869 unsigned hash_reg_file : 1;
870 /* unbound wq insertion if file is a non-regular file */
871 unsigned unbound_nonreg_file : 1;
872 /* opcode is not supported by this kernel */
873 unsigned not_supported : 1;
874 /* set if opcode supports polled "wait" */
876 unsigned pollout : 1;
877 /* op supports buffer selection */
878 unsigned buffer_select : 1;
879 /* do prep async if is going to be punted */
880 unsigned needs_async_setup : 1;
881 /* should block plug */
883 /* size of async data needed, if any */
884 unsigned short async_size;
887 static const struct io_op_def io_op_defs[] = {
888 [IORING_OP_NOP] = {},
889 [IORING_OP_READV] = {
891 .unbound_nonreg_file = 1,
894 .needs_async_setup = 1,
896 .async_size = sizeof(struct io_async_rw),
898 [IORING_OP_WRITEV] = {
901 .unbound_nonreg_file = 1,
903 .needs_async_setup = 1,
905 .async_size = sizeof(struct io_async_rw),
907 [IORING_OP_FSYNC] = {
910 [IORING_OP_READ_FIXED] = {
912 .unbound_nonreg_file = 1,
915 .async_size = sizeof(struct io_async_rw),
917 [IORING_OP_WRITE_FIXED] = {
920 .unbound_nonreg_file = 1,
923 .async_size = sizeof(struct io_async_rw),
925 [IORING_OP_POLL_ADD] = {
927 .unbound_nonreg_file = 1,
929 [IORING_OP_POLL_REMOVE] = {},
930 [IORING_OP_SYNC_FILE_RANGE] = {
933 [IORING_OP_SENDMSG] = {
935 .unbound_nonreg_file = 1,
937 .needs_async_setup = 1,
938 .async_size = sizeof(struct io_async_msghdr),
940 [IORING_OP_RECVMSG] = {
942 .unbound_nonreg_file = 1,
945 .needs_async_setup = 1,
946 .async_size = sizeof(struct io_async_msghdr),
948 [IORING_OP_TIMEOUT] = {
949 .async_size = sizeof(struct io_timeout_data),
951 [IORING_OP_TIMEOUT_REMOVE] = {
952 /* used by timeout updates' prep() */
954 [IORING_OP_ACCEPT] = {
956 .unbound_nonreg_file = 1,
959 [IORING_OP_ASYNC_CANCEL] = {},
960 [IORING_OP_LINK_TIMEOUT] = {
961 .async_size = sizeof(struct io_timeout_data),
963 [IORING_OP_CONNECT] = {
965 .unbound_nonreg_file = 1,
967 .needs_async_setup = 1,
968 .async_size = sizeof(struct io_async_connect),
970 [IORING_OP_FALLOCATE] = {
973 [IORING_OP_OPENAT] = {},
974 [IORING_OP_CLOSE] = {},
975 [IORING_OP_FILES_UPDATE] = {},
976 [IORING_OP_STATX] = {},
979 .unbound_nonreg_file = 1,
983 .async_size = sizeof(struct io_async_rw),
985 [IORING_OP_WRITE] = {
987 .unbound_nonreg_file = 1,
990 .async_size = sizeof(struct io_async_rw),
992 [IORING_OP_FADVISE] = {
995 [IORING_OP_MADVISE] = {},
998 .unbound_nonreg_file = 1,
1001 [IORING_OP_RECV] = {
1003 .unbound_nonreg_file = 1,
1007 [IORING_OP_OPENAT2] = {
1009 [IORING_OP_EPOLL_CTL] = {
1010 .unbound_nonreg_file = 1,
1012 [IORING_OP_SPLICE] = {
1015 .unbound_nonreg_file = 1,
1017 [IORING_OP_PROVIDE_BUFFERS] = {},
1018 [IORING_OP_REMOVE_BUFFERS] = {},
1022 .unbound_nonreg_file = 1,
1024 [IORING_OP_SHUTDOWN] = {
1027 [IORING_OP_RENAMEAT] = {},
1028 [IORING_OP_UNLINKAT] = {},
1031 static bool io_disarm_next(struct io_kiocb *req);
1032 static void io_uring_del_tctx_node(unsigned long index);
1033 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1034 struct task_struct *task,
1036 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd);
1037 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1039 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1040 long res, unsigned int cflags);
1041 static void io_put_req(struct io_kiocb *req);
1042 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1043 static void io_dismantle_req(struct io_kiocb *req);
1044 static void io_put_task(struct task_struct *task, int nr);
1045 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1046 static void io_queue_linked_timeout(struct io_kiocb *req);
1047 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1048 struct io_uring_rsrc_update2 *up,
1050 static void io_clean_op(struct io_kiocb *req);
1051 static struct file *io_file_get(struct io_submit_state *state,
1052 struct io_kiocb *req, int fd, bool fixed);
1053 static void __io_queue_sqe(struct io_kiocb *req);
1054 static void io_rsrc_put_work(struct work_struct *work);
1056 static void io_req_task_queue(struct io_kiocb *req);
1057 static void io_submit_flush_completions(struct io_comp_state *cs,
1058 struct io_ring_ctx *ctx);
1059 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1060 static int io_req_prep_async(struct io_kiocb *req);
1062 static struct kmem_cache *req_cachep;
1064 static const struct file_operations io_uring_fops;
1066 struct sock *io_uring_get_socket(struct file *file)
1068 #if defined(CONFIG_UNIX)
1069 if (file->f_op == &io_uring_fops) {
1070 struct io_ring_ctx *ctx = file->private_data;
1072 return ctx->ring_sock->sk;
1077 EXPORT_SYMBOL(io_uring_get_socket);
1079 #define io_for_each_link(pos, head) \
1080 for (pos = (head); pos; pos = pos->link)
1082 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1084 struct io_ring_ctx *ctx = req->ctx;
1086 if (!req->fixed_rsrc_refs) {
1087 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1088 percpu_ref_get(req->fixed_rsrc_refs);
1092 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1094 bool got = percpu_ref_tryget(ref);
1096 /* already at zero, wait for ->release() */
1098 wait_for_completion(compl);
1099 percpu_ref_resurrect(ref);
1101 percpu_ref_put(ref);
1104 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1107 struct io_kiocb *req;
1109 if (task && head->task != task)
1114 io_for_each_link(req, head) {
1115 if (req->flags & REQ_F_INFLIGHT)
1121 static inline void req_set_fail(struct io_kiocb *req)
1123 req->flags |= REQ_F_FAIL;
1126 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1128 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1130 complete(&ctx->ref_comp);
1133 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1135 return !req->timeout.off;
1138 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1140 struct io_ring_ctx *ctx;
1143 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1148 * Use 5 bits less than the max cq entries, that should give us around
1149 * 32 entries per hash list if totally full and uniformly spread.
1151 hash_bits = ilog2(p->cq_entries);
1155 ctx->cancel_hash_bits = hash_bits;
1156 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1158 if (!ctx->cancel_hash)
1160 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1162 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1163 if (!ctx->dummy_ubuf)
1165 /* set invalid range, so io_import_fixed() fails meeting it */
1166 ctx->dummy_ubuf->ubuf = -1UL;
1168 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1169 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1172 ctx->flags = p->flags;
1173 init_waitqueue_head(&ctx->sqo_sq_wait);
1174 INIT_LIST_HEAD(&ctx->sqd_list);
1175 init_waitqueue_head(&ctx->cq_wait);
1176 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1177 init_completion(&ctx->ref_comp);
1178 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1179 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1180 mutex_init(&ctx->uring_lock);
1181 init_waitqueue_head(&ctx->wait);
1182 spin_lock_init(&ctx->completion_lock);
1183 INIT_LIST_HEAD(&ctx->iopoll_list);
1184 INIT_LIST_HEAD(&ctx->defer_list);
1185 INIT_LIST_HEAD(&ctx->timeout_list);
1186 spin_lock_init(&ctx->rsrc_ref_lock);
1187 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1188 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1189 init_llist_head(&ctx->rsrc_put_llist);
1190 INIT_LIST_HEAD(&ctx->tctx_list);
1191 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1192 INIT_LIST_HEAD(&ctx->locked_free_list);
1195 kfree(ctx->dummy_ubuf);
1196 kfree(ctx->cancel_hash);
1201 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1203 struct io_rings *r = ctx->rings;
1205 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1209 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1211 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1212 struct io_ring_ctx *ctx = req->ctx;
1214 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1220 static void io_req_track_inflight(struct io_kiocb *req)
1222 if (!(req->flags & REQ_F_INFLIGHT)) {
1223 req->flags |= REQ_F_INFLIGHT;
1224 atomic_inc(¤t->io_uring->inflight_tracked);
1228 static void io_prep_async_work(struct io_kiocb *req)
1230 const struct io_op_def *def = &io_op_defs[req->opcode];
1231 struct io_ring_ctx *ctx = req->ctx;
1233 if (!req->work.creds)
1234 req->work.creds = get_current_cred();
1236 req->work.list.next = NULL;
1237 req->work.flags = 0;
1238 if (req->flags & REQ_F_FORCE_ASYNC)
1239 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1241 if (req->flags & REQ_F_ISREG) {
1242 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1243 io_wq_hash_work(&req->work, file_inode(req->file));
1244 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1245 if (def->unbound_nonreg_file)
1246 req->work.flags |= IO_WQ_WORK_UNBOUND;
1249 switch (req->opcode) {
1250 case IORING_OP_SPLICE:
1252 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1253 req->work.flags |= IO_WQ_WORK_UNBOUND;
1258 static void io_prep_async_link(struct io_kiocb *req)
1260 struct io_kiocb *cur;
1262 io_for_each_link(cur, req)
1263 io_prep_async_work(cur);
1266 static void io_queue_async_work(struct io_kiocb *req)
1268 struct io_ring_ctx *ctx = req->ctx;
1269 struct io_kiocb *link = io_prep_linked_timeout(req);
1270 struct io_uring_task *tctx = req->task->io_uring;
1273 BUG_ON(!tctx->io_wq);
1275 /* init ->work of the whole link before punting */
1276 io_prep_async_link(req);
1277 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1278 &req->work, req->flags);
1279 io_wq_enqueue(tctx->io_wq, &req->work);
1281 io_queue_linked_timeout(link);
1284 static void io_kill_timeout(struct io_kiocb *req, int status)
1285 __must_hold(&req->ctx->completion_lock)
1287 struct io_timeout_data *io = req->async_data;
1289 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1290 atomic_set(&req->ctx->cq_timeouts,
1291 atomic_read(&req->ctx->cq_timeouts) + 1);
1292 list_del_init(&req->timeout.list);
1293 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1294 io_put_req_deferred(req, 1);
1298 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1301 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1302 struct io_defer_entry, list);
1304 if (req_need_defer(de->req, de->seq))
1306 list_del_init(&de->list);
1307 io_req_task_queue(de->req);
1309 } while (!list_empty(&ctx->defer_list));
1312 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1316 if (list_empty(&ctx->timeout_list))
1319 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1322 u32 events_needed, events_got;
1323 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1324 struct io_kiocb, timeout.list);
1326 if (io_is_timeout_noseq(req))
1330 * Since seq can easily wrap around over time, subtract
1331 * the last seq at which timeouts were flushed before comparing.
1332 * Assuming not more than 2^31-1 events have happened since,
1333 * these subtractions won't have wrapped, so we can check if
1334 * target is in [last_seq, current_seq] by comparing the two.
1336 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1337 events_got = seq - ctx->cq_last_tm_flush;
1338 if (events_got < events_needed)
1341 list_del_init(&req->timeout.list);
1342 io_kill_timeout(req, 0);
1343 } while (!list_empty(&ctx->timeout_list));
1345 ctx->cq_last_tm_flush = seq;
1348 static void io_commit_cqring(struct io_ring_ctx *ctx)
1350 io_flush_timeouts(ctx);
1352 /* order cqe stores with ring update */
1353 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1355 if (unlikely(!list_empty(&ctx->defer_list)))
1356 __io_queue_deferred(ctx);
1359 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1361 struct io_rings *r = ctx->rings;
1363 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1366 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1368 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1371 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1373 struct io_rings *rings = ctx->rings;
1374 unsigned tail, mask = ctx->cq_entries - 1;
1377 * writes to the cq entry need to come after reading head; the
1378 * control dependency is enough as we're using WRITE_ONCE to
1381 if (__io_cqring_events(ctx) == ctx->cq_entries)
1384 tail = ctx->cached_cq_tail++;
1385 return &rings->cqes[tail & mask];
1388 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1390 if (likely(!ctx->cq_ev_fd))
1392 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1394 return !ctx->eventfd_async || io_wq_current_is_worker();
1397 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1399 /* see waitqueue_active() comment */
1402 if (waitqueue_active(&ctx->wait))
1403 wake_up(&ctx->wait);
1404 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1405 wake_up(&ctx->sq_data->wait);
1406 if (io_should_trigger_evfd(ctx))
1407 eventfd_signal(ctx->cq_ev_fd, 1);
1408 if (waitqueue_active(&ctx->cq_wait)) {
1409 wake_up_interruptible(&ctx->cq_wait);
1410 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1414 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1416 /* see waitqueue_active() comment */
1419 if (ctx->flags & IORING_SETUP_SQPOLL) {
1420 if (waitqueue_active(&ctx->wait))
1421 wake_up(&ctx->wait);
1423 if (io_should_trigger_evfd(ctx))
1424 eventfd_signal(ctx->cq_ev_fd, 1);
1425 if (waitqueue_active(&ctx->cq_wait)) {
1426 wake_up_interruptible(&ctx->cq_wait);
1427 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1431 /* Returns true if there are no backlogged entries after the flush */
1432 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1434 unsigned long flags;
1435 bool all_flushed, posted;
1437 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1441 spin_lock_irqsave(&ctx->completion_lock, flags);
1442 while (!list_empty(&ctx->cq_overflow_list)) {
1443 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1444 struct io_overflow_cqe *ocqe;
1448 ocqe = list_first_entry(&ctx->cq_overflow_list,
1449 struct io_overflow_cqe, list);
1451 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1453 io_account_cq_overflow(ctx);
1456 list_del(&ocqe->list);
1460 all_flushed = list_empty(&ctx->cq_overflow_list);
1462 clear_bit(0, &ctx->sq_check_overflow);
1463 clear_bit(0, &ctx->cq_check_overflow);
1464 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1468 io_commit_cqring(ctx);
1469 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1471 io_cqring_ev_posted(ctx);
1475 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1479 if (test_bit(0, &ctx->cq_check_overflow)) {
1480 /* iopoll syncs against uring_lock, not completion_lock */
1481 if (ctx->flags & IORING_SETUP_IOPOLL)
1482 mutex_lock(&ctx->uring_lock);
1483 ret = __io_cqring_overflow_flush(ctx, force);
1484 if (ctx->flags & IORING_SETUP_IOPOLL)
1485 mutex_unlock(&ctx->uring_lock);
1492 * Shamelessly stolen from the mm implementation of page reference checking,
1493 * see commit f958d7b528b1 for details.
1495 #define req_ref_zero_or_close_to_overflow(req) \
1496 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1498 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1500 return atomic_inc_not_zero(&req->refs);
1503 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1505 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1506 return atomic_sub_and_test(refs, &req->refs);
1509 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1511 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1512 return atomic_dec_and_test(&req->refs);
1515 static inline void req_ref_put(struct io_kiocb *req)
1517 WARN_ON_ONCE(req_ref_put_and_test(req));
1520 static inline void req_ref_get(struct io_kiocb *req)
1522 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1523 atomic_inc(&req->refs);
1526 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1527 long res, unsigned int cflags)
1529 struct io_overflow_cqe *ocqe;
1531 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1534 * If we're in ring overflow flush mode, or in task cancel mode,
1535 * or cannot allocate an overflow entry, then we need to drop it
1538 io_account_cq_overflow(ctx);
1541 if (list_empty(&ctx->cq_overflow_list)) {
1542 set_bit(0, &ctx->sq_check_overflow);
1543 set_bit(0, &ctx->cq_check_overflow);
1544 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1546 ocqe->cqe.user_data = user_data;
1547 ocqe->cqe.res = res;
1548 ocqe->cqe.flags = cflags;
1549 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1553 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1554 long res, unsigned int cflags)
1556 struct io_uring_cqe *cqe;
1558 trace_io_uring_complete(ctx, user_data, res, cflags);
1561 * If we can't get a cq entry, userspace overflowed the
1562 * submission (by quite a lot). Increment the overflow count in
1565 cqe = io_get_cqe(ctx);
1567 WRITE_ONCE(cqe->user_data, user_data);
1568 WRITE_ONCE(cqe->res, res);
1569 WRITE_ONCE(cqe->flags, cflags);
1572 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1575 /* not as hot to bloat with inlining */
1576 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1577 long res, unsigned int cflags)
1579 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1582 static void io_req_complete_post(struct io_kiocb *req, long res,
1583 unsigned int cflags)
1585 struct io_ring_ctx *ctx = req->ctx;
1586 unsigned long flags;
1588 spin_lock_irqsave(&ctx->completion_lock, flags);
1589 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1591 * If we're the last reference to this request, add to our locked
1594 if (req_ref_put_and_test(req)) {
1595 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1596 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1597 io_disarm_next(req);
1599 io_req_task_queue(req->link);
1603 io_dismantle_req(req);
1604 io_put_task(req->task, 1);
1605 list_add(&req->compl.list, &ctx->locked_free_list);
1606 ctx->locked_free_nr++;
1608 if (!percpu_ref_tryget(&ctx->refs))
1611 io_commit_cqring(ctx);
1612 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1615 io_cqring_ev_posted(ctx);
1616 percpu_ref_put(&ctx->refs);
1620 static inline bool io_req_needs_clean(struct io_kiocb *req)
1622 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1623 REQ_F_POLLED | REQ_F_INFLIGHT);
1626 static void io_req_complete_state(struct io_kiocb *req, long res,
1627 unsigned int cflags)
1629 if (io_req_needs_clean(req))
1632 req->compl.cflags = cflags;
1633 req->flags |= REQ_F_COMPLETE_INLINE;
1636 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1637 long res, unsigned cflags)
1639 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1640 io_req_complete_state(req, res, cflags);
1642 io_req_complete_post(req, res, cflags);
1645 static inline void io_req_complete(struct io_kiocb *req, long res)
1647 __io_req_complete(req, 0, res, 0);
1650 static void io_req_complete_failed(struct io_kiocb *req, long res)
1654 io_req_complete_post(req, res, 0);
1657 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1658 struct io_comp_state *cs)
1660 spin_lock_irq(&ctx->completion_lock);
1661 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1662 ctx->locked_free_nr = 0;
1663 spin_unlock_irq(&ctx->completion_lock);
1666 /* Returns true IFF there are requests in the cache */
1667 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1669 struct io_submit_state *state = &ctx->submit_state;
1670 struct io_comp_state *cs = &state->comp;
1674 * If we have more than a batch's worth of requests in our IRQ side
1675 * locked cache, grab the lock and move them over to our submission
1678 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1679 io_flush_cached_locked_reqs(ctx, cs);
1681 nr = state->free_reqs;
1682 while (!list_empty(&cs->free_list)) {
1683 struct io_kiocb *req = list_first_entry(&cs->free_list,
1684 struct io_kiocb, compl.list);
1686 list_del(&req->compl.list);
1687 state->reqs[nr++] = req;
1688 if (nr == ARRAY_SIZE(state->reqs))
1692 state->free_reqs = nr;
1696 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1698 struct io_submit_state *state = &ctx->submit_state;
1700 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1702 if (!state->free_reqs) {
1703 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1706 if (io_flush_cached_reqs(ctx))
1709 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1713 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1714 * retry single alloc to be on the safe side.
1716 if (unlikely(ret <= 0)) {
1717 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1718 if (!state->reqs[0])
1722 state->free_reqs = ret;
1726 return state->reqs[state->free_reqs];
1729 static inline void io_put_file(struct file *file)
1735 static void io_dismantle_req(struct io_kiocb *req)
1737 unsigned int flags = req->flags;
1739 if (io_req_needs_clean(req))
1741 if (!(flags & REQ_F_FIXED_FILE))
1742 io_put_file(req->file);
1743 if (req->fixed_rsrc_refs)
1744 percpu_ref_put(req->fixed_rsrc_refs);
1745 if (req->async_data)
1746 kfree(req->async_data);
1747 if (req->work.creds) {
1748 put_cred(req->work.creds);
1749 req->work.creds = NULL;
1753 /* must to be called somewhat shortly after putting a request */
1754 static inline void io_put_task(struct task_struct *task, int nr)
1756 struct io_uring_task *tctx = task->io_uring;
1758 percpu_counter_sub(&tctx->inflight, nr);
1759 if (unlikely(atomic_read(&tctx->in_idle)))
1760 wake_up(&tctx->wait);
1761 put_task_struct_many(task, nr);
1764 static void __io_free_req(struct io_kiocb *req)
1766 struct io_ring_ctx *ctx = req->ctx;
1768 io_dismantle_req(req);
1769 io_put_task(req->task, 1);
1771 kmem_cache_free(req_cachep, req);
1772 percpu_ref_put(&ctx->refs);
1775 static inline void io_remove_next_linked(struct io_kiocb *req)
1777 struct io_kiocb *nxt = req->link;
1779 req->link = nxt->link;
1783 static bool io_kill_linked_timeout(struct io_kiocb *req)
1784 __must_hold(&req->ctx->completion_lock)
1786 struct io_kiocb *link = req->link;
1789 * Can happen if a linked timeout fired and link had been like
1790 * req -> link t-out -> link t-out [-> ...]
1792 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1793 struct io_timeout_data *io = link->async_data;
1795 io_remove_next_linked(req);
1796 link->timeout.head = NULL;
1797 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1798 io_cqring_fill_event(link->ctx, link->user_data,
1800 io_put_req_deferred(link, 1);
1807 static void io_fail_links(struct io_kiocb *req)
1808 __must_hold(&req->ctx->completion_lock)
1810 struct io_kiocb *nxt, *link = req->link;
1817 trace_io_uring_fail_link(req, link);
1818 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1819 io_put_req_deferred(link, 2);
1824 static bool io_disarm_next(struct io_kiocb *req)
1825 __must_hold(&req->ctx->completion_lock)
1827 bool posted = false;
1829 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1830 posted = io_kill_linked_timeout(req);
1831 if (unlikely((req->flags & REQ_F_FAIL) &&
1832 !(req->flags & REQ_F_HARDLINK))) {
1833 posted |= (req->link != NULL);
1839 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1841 struct io_kiocb *nxt;
1844 * If LINK is set, we have dependent requests in this chain. If we
1845 * didn't fail this request, queue the first one up, moving any other
1846 * dependencies to the next request. In case of failure, fail the rest
1849 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1850 struct io_ring_ctx *ctx = req->ctx;
1851 unsigned long flags;
1854 spin_lock_irqsave(&ctx->completion_lock, flags);
1855 posted = io_disarm_next(req);
1857 io_commit_cqring(req->ctx);
1858 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1860 io_cqring_ev_posted(ctx);
1867 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1869 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1871 return __io_req_find_next(req);
1874 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1878 if (ctx->submit_state.comp.nr) {
1879 mutex_lock(&ctx->uring_lock);
1880 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1881 mutex_unlock(&ctx->uring_lock);
1883 percpu_ref_put(&ctx->refs);
1886 static bool __tctx_task_work(struct io_uring_task *tctx)
1888 struct io_ring_ctx *ctx = NULL;
1889 struct io_wq_work_list list;
1890 struct io_wq_work_node *node;
1892 if (wq_list_empty(&tctx->task_list))
1895 spin_lock_irq(&tctx->task_lock);
1896 list = tctx->task_list;
1897 INIT_WQ_LIST(&tctx->task_list);
1898 spin_unlock_irq(&tctx->task_lock);
1902 struct io_wq_work_node *next = node->next;
1903 struct io_kiocb *req;
1905 req = container_of(node, struct io_kiocb, io_task_work.node);
1906 if (req->ctx != ctx) {
1907 ctx_flush_and_put(ctx);
1909 percpu_ref_get(&ctx->refs);
1912 req->task_work.func(&req->task_work);
1916 ctx_flush_and_put(ctx);
1917 return list.first != NULL;
1920 static void tctx_task_work(struct callback_head *cb)
1922 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1924 clear_bit(0, &tctx->task_state);
1926 while (__tctx_task_work(tctx))
1930 static int io_req_task_work_add(struct io_kiocb *req)
1932 struct task_struct *tsk = req->task;
1933 struct io_uring_task *tctx = tsk->io_uring;
1934 enum task_work_notify_mode notify;
1935 struct io_wq_work_node *node, *prev;
1936 unsigned long flags;
1939 if (unlikely(tsk->flags & PF_EXITING))
1942 WARN_ON_ONCE(!tctx);
1944 spin_lock_irqsave(&tctx->task_lock, flags);
1945 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1946 spin_unlock_irqrestore(&tctx->task_lock, flags);
1948 /* task_work already pending, we're done */
1949 if (test_bit(0, &tctx->task_state) ||
1950 test_and_set_bit(0, &tctx->task_state))
1954 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1955 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1956 * processing task_work. There's no reliable way to tell if TWA_RESUME
1959 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1961 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1962 wake_up_process(tsk);
1967 * Slow path - we failed, find and delete work. if the work is not
1968 * in the list, it got run and we're fine.
1970 spin_lock_irqsave(&tctx->task_lock, flags);
1971 wq_list_for_each(node, prev, &tctx->task_list) {
1972 if (&req->io_task_work.node == node) {
1973 wq_list_del(&tctx->task_list, node, prev);
1978 spin_unlock_irqrestore(&tctx->task_lock, flags);
1979 clear_bit(0, &tctx->task_state);
1983 static bool io_run_task_work_head(struct callback_head **work_head)
1985 struct callback_head *work, *next;
1986 bool executed = false;
1989 work = xchg(work_head, NULL);
2005 static void io_task_work_add_head(struct callback_head **work_head,
2006 struct callback_head *task_work)
2008 struct callback_head *head;
2011 head = READ_ONCE(*work_head);
2012 task_work->next = head;
2013 } while (cmpxchg(work_head, head, task_work) != head);
2016 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2017 task_work_func_t cb)
2019 init_task_work(&req->task_work, cb);
2020 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2023 static void io_req_task_cancel(struct callback_head *cb)
2025 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2026 struct io_ring_ctx *ctx = req->ctx;
2028 /* ctx is guaranteed to stay alive while we hold uring_lock */
2029 mutex_lock(&ctx->uring_lock);
2030 io_req_complete_failed(req, req->result);
2031 mutex_unlock(&ctx->uring_lock);
2034 static void __io_req_task_submit(struct io_kiocb *req)
2036 struct io_ring_ctx *ctx = req->ctx;
2038 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2039 mutex_lock(&ctx->uring_lock);
2040 if (!(current->flags & PF_EXITING) && !current->in_execve)
2041 __io_queue_sqe(req);
2043 io_req_complete_failed(req, -EFAULT);
2044 mutex_unlock(&ctx->uring_lock);
2047 static void io_req_task_submit(struct callback_head *cb)
2049 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2051 __io_req_task_submit(req);
2054 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2057 req->task_work.func = io_req_task_cancel;
2059 if (unlikely(io_req_task_work_add(req)))
2060 io_req_task_work_add_fallback(req, io_req_task_cancel);
2063 static void io_req_task_queue(struct io_kiocb *req)
2065 req->task_work.func = io_req_task_submit;
2067 if (unlikely(io_req_task_work_add(req)))
2068 io_req_task_queue_fail(req, -ECANCELED);
2071 static inline void io_queue_next(struct io_kiocb *req)
2073 struct io_kiocb *nxt = io_req_find_next(req);
2076 io_req_task_queue(nxt);
2079 static void io_free_req(struct io_kiocb *req)
2086 struct task_struct *task;
2091 static inline void io_init_req_batch(struct req_batch *rb)
2098 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2099 struct req_batch *rb)
2102 io_put_task(rb->task, rb->task_refs);
2104 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2107 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2108 struct io_submit_state *state)
2111 io_dismantle_req(req);
2113 if (req->task != rb->task) {
2115 io_put_task(rb->task, rb->task_refs);
2116 rb->task = req->task;
2122 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2123 state->reqs[state->free_reqs++] = req;
2125 list_add(&req->compl.list, &state->comp.free_list);
2128 static void io_submit_flush_completions(struct io_comp_state *cs,
2129 struct io_ring_ctx *ctx)
2132 struct io_kiocb *req;
2133 struct req_batch rb;
2135 io_init_req_batch(&rb);
2136 spin_lock_irq(&ctx->completion_lock);
2137 for (i = 0; i < nr; i++) {
2139 __io_cqring_fill_event(ctx, req->user_data, req->result,
2142 io_commit_cqring(ctx);
2143 spin_unlock_irq(&ctx->completion_lock);
2145 io_cqring_ev_posted(ctx);
2146 for (i = 0; i < nr; i++) {
2149 /* submission and completion refs */
2150 if (req_ref_sub_and_test(req, 2))
2151 io_req_free_batch(&rb, req, &ctx->submit_state);
2154 io_req_free_batch_finish(ctx, &rb);
2159 * Drop reference to request, return next in chain (if there is one) if this
2160 * was the last reference to this request.
2162 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2164 struct io_kiocb *nxt = NULL;
2166 if (req_ref_put_and_test(req)) {
2167 nxt = io_req_find_next(req);
2173 static inline void io_put_req(struct io_kiocb *req)
2175 if (req_ref_put_and_test(req))
2179 static void io_put_req_deferred_cb(struct callback_head *cb)
2181 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2186 static void io_free_req_deferred(struct io_kiocb *req)
2188 req->task_work.func = io_put_req_deferred_cb;
2189 if (unlikely(io_req_task_work_add(req)))
2190 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2193 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2195 if (req_ref_sub_and_test(req, refs))
2196 io_free_req_deferred(req);
2199 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2201 /* See comment at the top of this file */
2203 return __io_cqring_events(ctx);
2206 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2208 struct io_rings *rings = ctx->rings;
2210 /* make sure SQ entry isn't read before tail */
2211 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2214 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2216 unsigned int cflags;
2218 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2219 cflags |= IORING_CQE_F_BUFFER;
2220 req->flags &= ~REQ_F_BUFFER_SELECTED;
2225 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2227 struct io_buffer *kbuf;
2229 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2230 return io_put_kbuf(req, kbuf);
2233 static inline bool io_run_task_work(void)
2236 * Not safe to run on exiting task, and the task_work handling will
2237 * not add work to such a task.
2239 if (unlikely(current->flags & PF_EXITING))
2241 if (current->task_works) {
2242 __set_current_state(TASK_RUNNING);
2251 * Find and free completed poll iocbs
2253 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2254 struct list_head *done)
2256 struct req_batch rb;
2257 struct io_kiocb *req;
2259 /* order with ->result store in io_complete_rw_iopoll() */
2262 io_init_req_batch(&rb);
2263 while (!list_empty(done)) {
2266 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2267 list_del(&req->inflight_entry);
2269 if (READ_ONCE(req->result) == -EAGAIN &&
2270 !(req->flags & REQ_F_DONT_REISSUE)) {
2271 req->iopoll_completed = 0;
2273 io_queue_async_work(req);
2277 if (req->flags & REQ_F_BUFFER_SELECTED)
2278 cflags = io_put_rw_kbuf(req);
2280 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2283 if (req_ref_put_and_test(req))
2284 io_req_free_batch(&rb, req, &ctx->submit_state);
2287 io_commit_cqring(ctx);
2288 io_cqring_ev_posted_iopoll(ctx);
2289 io_req_free_batch_finish(ctx, &rb);
2292 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2295 struct io_kiocb *req, *tmp;
2301 * Only spin for completions if we don't have multiple devices hanging
2302 * off our complete list, and we're under the requested amount.
2304 spin = !ctx->poll_multi_file && *nr_events < min;
2307 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2308 struct kiocb *kiocb = &req->rw.kiocb;
2311 * Move completed and retryable entries to our local lists.
2312 * If we find a request that requires polling, break out
2313 * and complete those lists first, if we have entries there.
2315 if (READ_ONCE(req->iopoll_completed)) {
2316 list_move_tail(&req->inflight_entry, &done);
2319 if (!list_empty(&done))
2322 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2326 /* iopoll may have completed current req */
2327 if (READ_ONCE(req->iopoll_completed))
2328 list_move_tail(&req->inflight_entry, &done);
2335 if (!list_empty(&done))
2336 io_iopoll_complete(ctx, nr_events, &done);
2342 * We can't just wait for polled events to come to us, we have to actively
2343 * find and complete them.
2345 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2347 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2350 mutex_lock(&ctx->uring_lock);
2351 while (!list_empty(&ctx->iopoll_list)) {
2352 unsigned int nr_events = 0;
2354 io_do_iopoll(ctx, &nr_events, 0);
2356 /* let it sleep and repeat later if can't complete a request */
2360 * Ensure we allow local-to-the-cpu processing to take place,
2361 * in this case we need to ensure that we reap all events.
2362 * Also let task_work, etc. to progress by releasing the mutex
2364 if (need_resched()) {
2365 mutex_unlock(&ctx->uring_lock);
2367 mutex_lock(&ctx->uring_lock);
2370 mutex_unlock(&ctx->uring_lock);
2373 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2375 unsigned int nr_events = 0;
2379 * We disallow the app entering submit/complete with polling, but we
2380 * still need to lock the ring to prevent racing with polled issue
2381 * that got punted to a workqueue.
2383 mutex_lock(&ctx->uring_lock);
2385 * Don't enter poll loop if we already have events pending.
2386 * If we do, we can potentially be spinning for commands that
2387 * already triggered a CQE (eg in error).
2389 if (test_bit(0, &ctx->cq_check_overflow))
2390 __io_cqring_overflow_flush(ctx, false);
2391 if (io_cqring_events(ctx))
2395 * If a submit got punted to a workqueue, we can have the
2396 * application entering polling for a command before it gets
2397 * issued. That app will hold the uring_lock for the duration
2398 * of the poll right here, so we need to take a breather every
2399 * now and then to ensure that the issue has a chance to add
2400 * the poll to the issued list. Otherwise we can spin here
2401 * forever, while the workqueue is stuck trying to acquire the
2404 if (list_empty(&ctx->iopoll_list)) {
2405 mutex_unlock(&ctx->uring_lock);
2407 mutex_lock(&ctx->uring_lock);
2409 if (list_empty(&ctx->iopoll_list))
2412 ret = io_do_iopoll(ctx, &nr_events, min);
2413 } while (!ret && nr_events < min && !need_resched());
2415 mutex_unlock(&ctx->uring_lock);
2419 static void kiocb_end_write(struct io_kiocb *req)
2422 * Tell lockdep we inherited freeze protection from submission
2425 if (req->flags & REQ_F_ISREG) {
2426 struct super_block *sb = file_inode(req->file)->i_sb;
2428 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2434 static bool io_resubmit_prep(struct io_kiocb *req)
2436 struct io_async_rw *rw = req->async_data;
2439 return !io_req_prep_async(req);
2440 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2441 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2445 static bool io_rw_should_reissue(struct io_kiocb *req)
2447 umode_t mode = file_inode(req->file)->i_mode;
2448 struct io_ring_ctx *ctx = req->ctx;
2450 if (!S_ISBLK(mode) && !S_ISREG(mode))
2452 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2453 !(ctx->flags & IORING_SETUP_IOPOLL)))
2456 * If ref is dying, we might be running poll reap from the exit work.
2457 * Don't attempt to reissue from that path, just let it fail with
2460 if (percpu_ref_is_dying(&ctx->refs))
2465 static bool io_resubmit_prep(struct io_kiocb *req)
2469 static bool io_rw_should_reissue(struct io_kiocb *req)
2475 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2476 unsigned int issue_flags)
2480 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2481 kiocb_end_write(req);
2482 if (res != req->result) {
2483 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2484 io_rw_should_reissue(req)) {
2485 req->flags |= REQ_F_REISSUE;
2490 if (req->flags & REQ_F_BUFFER_SELECTED)
2491 cflags = io_put_rw_kbuf(req);
2492 __io_req_complete(req, issue_flags, res, cflags);
2495 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2497 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2499 __io_complete_rw(req, res, res2, 0);
2502 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2504 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2506 if (kiocb->ki_flags & IOCB_WRITE)
2507 kiocb_end_write(req);
2508 if (unlikely(res != req->result)) {
2509 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2510 io_resubmit_prep(req))) {
2512 req->flags |= REQ_F_DONT_REISSUE;
2516 WRITE_ONCE(req->result, res);
2517 /* order with io_iopoll_complete() checking ->result */
2519 WRITE_ONCE(req->iopoll_completed, 1);
2523 * After the iocb has been issued, it's safe to be found on the poll list.
2524 * Adding the kiocb to the list AFTER submission ensures that we don't
2525 * find it from a io_do_iopoll() thread before the issuer is done
2526 * accessing the kiocb cookie.
2528 static void io_iopoll_req_issued(struct io_kiocb *req)
2530 struct io_ring_ctx *ctx = req->ctx;
2531 const bool in_async = io_wq_current_is_worker();
2533 /* workqueue context doesn't hold uring_lock, grab it now */
2534 if (unlikely(in_async))
2535 mutex_lock(&ctx->uring_lock);
2538 * Track whether we have multiple files in our lists. This will impact
2539 * how we do polling eventually, not spinning if we're on potentially
2540 * different devices.
2542 if (list_empty(&ctx->iopoll_list)) {
2543 ctx->poll_multi_file = false;
2544 } else if (!ctx->poll_multi_file) {
2545 struct io_kiocb *list_req;
2547 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2549 if (list_req->file != req->file)
2550 ctx->poll_multi_file = true;
2554 * For fast devices, IO may have already completed. If it has, add
2555 * it to the front so we find it first.
2557 if (READ_ONCE(req->iopoll_completed))
2558 list_add(&req->inflight_entry, &ctx->iopoll_list);
2560 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2562 if (unlikely(in_async)) {
2564 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2565 * in sq thread task context or in io worker task context. If
2566 * current task context is sq thread, we don't need to check
2567 * whether should wake up sq thread.
2569 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2570 wq_has_sleeper(&ctx->sq_data->wait))
2571 wake_up(&ctx->sq_data->wait);
2573 mutex_unlock(&ctx->uring_lock);
2577 static inline void io_state_file_put(struct io_submit_state *state)
2579 if (state->file_refs) {
2580 fput_many(state->file, state->file_refs);
2581 state->file_refs = 0;
2586 * Get as many references to a file as we have IOs left in this submission,
2587 * assuming most submissions are for one file, or at least that each file
2588 * has more than one submission.
2590 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2595 if (state->file_refs) {
2596 if (state->fd == fd) {
2600 io_state_file_put(state);
2602 state->file = fget_many(fd, state->ios_left);
2603 if (unlikely(!state->file))
2607 state->file_refs = state->ios_left - 1;
2611 static bool io_bdev_nowait(struct block_device *bdev)
2613 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2617 * If we tracked the file through the SCM inflight mechanism, we could support
2618 * any file. For now, just ensure that anything potentially problematic is done
2621 static bool __io_file_supports_async(struct file *file, int rw)
2623 umode_t mode = file_inode(file)->i_mode;
2625 if (S_ISBLK(mode)) {
2626 if (IS_ENABLED(CONFIG_BLOCK) &&
2627 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2633 if (S_ISREG(mode)) {
2634 if (IS_ENABLED(CONFIG_BLOCK) &&
2635 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2636 file->f_op != &io_uring_fops)
2641 /* any ->read/write should understand O_NONBLOCK */
2642 if (file->f_flags & O_NONBLOCK)
2645 if (!(file->f_mode & FMODE_NOWAIT))
2649 return file->f_op->read_iter != NULL;
2651 return file->f_op->write_iter != NULL;
2654 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2656 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2658 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2661 return __io_file_supports_async(req->file, rw);
2664 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2666 struct io_ring_ctx *ctx = req->ctx;
2667 struct kiocb *kiocb = &req->rw.kiocb;
2668 struct file *file = req->file;
2672 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2673 req->flags |= REQ_F_ISREG;
2675 kiocb->ki_pos = READ_ONCE(sqe->off);
2676 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2677 req->flags |= REQ_F_CUR_POS;
2678 kiocb->ki_pos = file->f_pos;
2680 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2681 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2682 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2686 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2687 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2688 req->flags |= REQ_F_NOWAIT;
2690 ioprio = READ_ONCE(sqe->ioprio);
2692 ret = ioprio_check_cap(ioprio);
2696 kiocb->ki_ioprio = ioprio;
2698 kiocb->ki_ioprio = get_current_ioprio();
2700 if (ctx->flags & IORING_SETUP_IOPOLL) {
2701 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2702 !kiocb->ki_filp->f_op->iopoll)
2705 kiocb->ki_flags |= IOCB_HIPRI;
2706 kiocb->ki_complete = io_complete_rw_iopoll;
2707 req->iopoll_completed = 0;
2709 if (kiocb->ki_flags & IOCB_HIPRI)
2711 kiocb->ki_complete = io_complete_rw;
2714 if (req->opcode == IORING_OP_READ_FIXED ||
2715 req->opcode == IORING_OP_WRITE_FIXED) {
2717 io_req_set_rsrc_node(req);
2720 req->rw.addr = READ_ONCE(sqe->addr);
2721 req->rw.len = READ_ONCE(sqe->len);
2722 req->buf_index = READ_ONCE(sqe->buf_index);
2726 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2732 case -ERESTARTNOINTR:
2733 case -ERESTARTNOHAND:
2734 case -ERESTART_RESTARTBLOCK:
2736 * We can't just restart the syscall, since previously
2737 * submitted sqes may already be in progress. Just fail this
2743 kiocb->ki_complete(kiocb, ret, 0);
2747 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2748 unsigned int issue_flags)
2750 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2751 struct io_async_rw *io = req->async_data;
2752 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2754 /* add previously done IO, if any */
2755 if (io && io->bytes_done > 0) {
2757 ret = io->bytes_done;
2759 ret += io->bytes_done;
2762 if (req->flags & REQ_F_CUR_POS)
2763 req->file->f_pos = kiocb->ki_pos;
2764 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2765 __io_complete_rw(req, ret, 0, issue_flags);
2767 io_rw_done(kiocb, ret);
2769 if (check_reissue && req->flags & REQ_F_REISSUE) {
2770 req->flags &= ~REQ_F_REISSUE;
2771 if (io_resubmit_prep(req)) {
2773 io_queue_async_work(req);
2778 if (req->flags & REQ_F_BUFFER_SELECTED)
2779 cflags = io_put_rw_kbuf(req);
2780 __io_req_complete(req, issue_flags, ret, cflags);
2785 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2786 struct io_mapped_ubuf *imu)
2788 size_t len = req->rw.len;
2789 u64 buf_end, buf_addr = req->rw.addr;
2792 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2794 /* not inside the mapped region */
2795 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2799 * May not be a start of buffer, set size appropriately
2800 * and advance us to the beginning.
2802 offset = buf_addr - imu->ubuf;
2803 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2807 * Don't use iov_iter_advance() here, as it's really slow for
2808 * using the latter parts of a big fixed buffer - it iterates
2809 * over each segment manually. We can cheat a bit here, because
2812 * 1) it's a BVEC iter, we set it up
2813 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2814 * first and last bvec
2816 * So just find our index, and adjust the iterator afterwards.
2817 * If the offset is within the first bvec (or the whole first
2818 * bvec, just use iov_iter_advance(). This makes it easier
2819 * since we can just skip the first segment, which may not
2820 * be PAGE_SIZE aligned.
2822 const struct bio_vec *bvec = imu->bvec;
2824 if (offset <= bvec->bv_len) {
2825 iov_iter_advance(iter, offset);
2827 unsigned long seg_skip;
2829 /* skip first vec */
2830 offset -= bvec->bv_len;
2831 seg_skip = 1 + (offset >> PAGE_SHIFT);
2833 iter->bvec = bvec + seg_skip;
2834 iter->nr_segs -= seg_skip;
2835 iter->count -= bvec->bv_len + offset;
2836 iter->iov_offset = offset & ~PAGE_MASK;
2843 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2845 struct io_ring_ctx *ctx = req->ctx;
2846 struct io_mapped_ubuf *imu = req->imu;
2847 u16 index, buf_index = req->buf_index;
2850 if (unlikely(buf_index >= ctx->nr_user_bufs))
2852 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2853 imu = READ_ONCE(ctx->user_bufs[index]);
2856 return __io_import_fixed(req, rw, iter, imu);
2859 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2862 mutex_unlock(&ctx->uring_lock);
2865 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2868 * "Normal" inline submissions always hold the uring_lock, since we
2869 * grab it from the system call. Same is true for the SQPOLL offload.
2870 * The only exception is when we've detached the request and issue it
2871 * from an async worker thread, grab the lock for that case.
2874 mutex_lock(&ctx->uring_lock);
2877 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2878 int bgid, struct io_buffer *kbuf,
2881 struct io_buffer *head;
2883 if (req->flags & REQ_F_BUFFER_SELECTED)
2886 io_ring_submit_lock(req->ctx, needs_lock);
2888 lockdep_assert_held(&req->ctx->uring_lock);
2890 head = xa_load(&req->ctx->io_buffers, bgid);
2892 if (!list_empty(&head->list)) {
2893 kbuf = list_last_entry(&head->list, struct io_buffer,
2895 list_del(&kbuf->list);
2898 xa_erase(&req->ctx->io_buffers, bgid);
2900 if (*len > kbuf->len)
2903 kbuf = ERR_PTR(-ENOBUFS);
2906 io_ring_submit_unlock(req->ctx, needs_lock);
2911 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2914 struct io_buffer *kbuf;
2917 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2918 bgid = req->buf_index;
2919 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2922 req->rw.addr = (u64) (unsigned long) kbuf;
2923 req->flags |= REQ_F_BUFFER_SELECTED;
2924 return u64_to_user_ptr(kbuf->addr);
2927 #ifdef CONFIG_COMPAT
2928 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2931 struct compat_iovec __user *uiov;
2932 compat_ssize_t clen;
2936 uiov = u64_to_user_ptr(req->rw.addr);
2937 if (!access_ok(uiov, sizeof(*uiov)))
2939 if (__get_user(clen, &uiov->iov_len))
2945 buf = io_rw_buffer_select(req, &len, needs_lock);
2947 return PTR_ERR(buf);
2948 iov[0].iov_base = buf;
2949 iov[0].iov_len = (compat_size_t) len;
2954 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2957 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2961 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2964 len = iov[0].iov_len;
2967 buf = io_rw_buffer_select(req, &len, needs_lock);
2969 return PTR_ERR(buf);
2970 iov[0].iov_base = buf;
2971 iov[0].iov_len = len;
2975 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2978 if (req->flags & REQ_F_BUFFER_SELECTED) {
2979 struct io_buffer *kbuf;
2981 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2982 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2983 iov[0].iov_len = kbuf->len;
2986 if (req->rw.len != 1)
2989 #ifdef CONFIG_COMPAT
2990 if (req->ctx->compat)
2991 return io_compat_import(req, iov, needs_lock);
2994 return __io_iov_buffer_select(req, iov, needs_lock);
2997 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2998 struct iov_iter *iter, bool needs_lock)
3000 void __user *buf = u64_to_user_ptr(req->rw.addr);
3001 size_t sqe_len = req->rw.len;
3002 u8 opcode = req->opcode;
3005 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3007 return io_import_fixed(req, rw, iter);
3010 /* buffer index only valid with fixed read/write, or buffer select */
3011 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3014 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3015 if (req->flags & REQ_F_BUFFER_SELECT) {
3016 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3018 return PTR_ERR(buf);
3019 req->rw.len = sqe_len;
3022 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3027 if (req->flags & REQ_F_BUFFER_SELECT) {
3028 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3030 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3035 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3039 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3041 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3045 * For files that don't have ->read_iter() and ->write_iter(), handle them
3046 * by looping over ->read() or ->write() manually.
3048 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3050 struct kiocb *kiocb = &req->rw.kiocb;
3051 struct file *file = req->file;
3055 * Don't support polled IO through this interface, and we can't
3056 * support non-blocking either. For the latter, this just causes
3057 * the kiocb to be handled from an async context.
3059 if (kiocb->ki_flags & IOCB_HIPRI)
3061 if (kiocb->ki_flags & IOCB_NOWAIT)
3064 while (iov_iter_count(iter)) {
3068 if (!iov_iter_is_bvec(iter)) {
3069 iovec = iov_iter_iovec(iter);
3071 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3072 iovec.iov_len = req->rw.len;
3076 nr = file->f_op->read(file, iovec.iov_base,
3077 iovec.iov_len, io_kiocb_ppos(kiocb));
3079 nr = file->f_op->write(file, iovec.iov_base,
3080 iovec.iov_len, io_kiocb_ppos(kiocb));
3089 if (nr != iovec.iov_len)
3093 iov_iter_advance(iter, nr);
3099 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3100 const struct iovec *fast_iov, struct iov_iter *iter)
3102 struct io_async_rw *rw = req->async_data;
3104 memcpy(&rw->iter, iter, sizeof(*iter));
3105 rw->free_iovec = iovec;
3107 /* can only be fixed buffers, no need to do anything */
3108 if (iov_iter_is_bvec(iter))
3111 unsigned iov_off = 0;
3113 rw->iter.iov = rw->fast_iov;
3114 if (iter->iov != fast_iov) {
3115 iov_off = iter->iov - fast_iov;
3116 rw->iter.iov += iov_off;
3118 if (rw->fast_iov != fast_iov)
3119 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3120 sizeof(struct iovec) * iter->nr_segs);
3122 req->flags |= REQ_F_NEED_CLEANUP;
3126 static inline int io_alloc_async_data(struct io_kiocb *req)
3128 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3129 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3130 return req->async_data == NULL;
3133 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3134 const struct iovec *fast_iov,
3135 struct iov_iter *iter, bool force)
3137 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3139 if (!req->async_data) {
3140 if (io_alloc_async_data(req)) {
3145 io_req_map_rw(req, iovec, fast_iov, iter);
3150 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3152 struct io_async_rw *iorw = req->async_data;
3153 struct iovec *iov = iorw->fast_iov;
3156 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3157 if (unlikely(ret < 0))
3160 iorw->bytes_done = 0;
3161 iorw->free_iovec = iov;
3163 req->flags |= REQ_F_NEED_CLEANUP;
3167 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3169 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3171 return io_prep_rw(req, sqe);
3175 * This is our waitqueue callback handler, registered through lock_page_async()
3176 * when we initially tried to do the IO with the iocb armed our waitqueue.
3177 * This gets called when the page is unlocked, and we generally expect that to
3178 * happen when the page IO is completed and the page is now uptodate. This will
3179 * queue a task_work based retry of the operation, attempting to copy the data
3180 * again. If the latter fails because the page was NOT uptodate, then we will
3181 * do a thread based blocking retry of the operation. That's the unexpected
3184 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3185 int sync, void *arg)
3187 struct wait_page_queue *wpq;
3188 struct io_kiocb *req = wait->private;
3189 struct wait_page_key *key = arg;
3191 wpq = container_of(wait, struct wait_page_queue, wait);
3193 if (!wake_page_match(wpq, key))
3196 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3197 list_del_init(&wait->entry);
3199 /* submit ref gets dropped, acquire a new one */
3201 io_req_task_queue(req);
3206 * This controls whether a given IO request should be armed for async page
3207 * based retry. If we return false here, the request is handed to the async
3208 * worker threads for retry. If we're doing buffered reads on a regular file,
3209 * we prepare a private wait_page_queue entry and retry the operation. This
3210 * will either succeed because the page is now uptodate and unlocked, or it
3211 * will register a callback when the page is unlocked at IO completion. Through
3212 * that callback, io_uring uses task_work to setup a retry of the operation.
3213 * That retry will attempt the buffered read again. The retry will generally
3214 * succeed, or in rare cases where it fails, we then fall back to using the
3215 * async worker threads for a blocking retry.
3217 static bool io_rw_should_retry(struct io_kiocb *req)
3219 struct io_async_rw *rw = req->async_data;
3220 struct wait_page_queue *wait = &rw->wpq;
3221 struct kiocb *kiocb = &req->rw.kiocb;
3223 /* never retry for NOWAIT, we just complete with -EAGAIN */
3224 if (req->flags & REQ_F_NOWAIT)
3227 /* Only for buffered IO */
3228 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3232 * just use poll if we can, and don't attempt if the fs doesn't
3233 * support callback based unlocks
3235 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3238 wait->wait.func = io_async_buf_func;
3239 wait->wait.private = req;
3240 wait->wait.flags = 0;
3241 INIT_LIST_HEAD(&wait->wait.entry);
3242 kiocb->ki_flags |= IOCB_WAITQ;
3243 kiocb->ki_flags &= ~IOCB_NOWAIT;
3244 kiocb->ki_waitq = wait;
3248 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3250 if (req->file->f_op->read_iter)
3251 return call_read_iter(req->file, &req->rw.kiocb, iter);
3252 else if (req->file->f_op->read)
3253 return loop_rw_iter(READ, req, iter);
3258 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3260 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3261 struct kiocb *kiocb = &req->rw.kiocb;
3262 struct iov_iter __iter, *iter = &__iter;
3263 struct io_async_rw *rw = req->async_data;
3264 ssize_t io_size, ret, ret2;
3265 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3271 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3275 io_size = iov_iter_count(iter);
3276 req->result = io_size;
3278 /* Ensure we clear previously set non-block flag */
3279 if (!force_nonblock)
3280 kiocb->ki_flags &= ~IOCB_NOWAIT;
3282 kiocb->ki_flags |= IOCB_NOWAIT;
3284 /* If the file doesn't support async, just async punt */
3285 if (force_nonblock && !io_file_supports_async(req, READ)) {
3286 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3287 return ret ?: -EAGAIN;
3290 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3291 if (unlikely(ret)) {
3296 ret = io_iter_do_read(req, iter);
3298 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3299 req->flags &= ~REQ_F_REISSUE;
3300 /* IOPOLL retry should happen for io-wq threads */
3301 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3303 /* no retry on NONBLOCK nor RWF_NOWAIT */
3304 if (req->flags & REQ_F_NOWAIT)
3306 /* some cases will consume bytes even on error returns */
3307 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3309 } else if (ret == -EIOCBQUEUED) {
3311 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3312 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3313 /* read all, failed, already did sync or don't want to retry */
3317 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3322 rw = req->async_data;
3323 /* now use our persistent iterator, if we aren't already */
3328 rw->bytes_done += ret;
3329 /* if we can retry, do so with the callbacks armed */
3330 if (!io_rw_should_retry(req)) {
3331 kiocb->ki_flags &= ~IOCB_WAITQ;
3336 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3337 * we get -EIOCBQUEUED, then we'll get a notification when the
3338 * desired page gets unlocked. We can also get a partial read
3339 * here, and if we do, then just retry at the new offset.
3341 ret = io_iter_do_read(req, iter);
3342 if (ret == -EIOCBQUEUED)
3344 /* we got some bytes, but not all. retry. */
3345 kiocb->ki_flags &= ~IOCB_WAITQ;
3346 } while (ret > 0 && ret < io_size);
3348 kiocb_done(kiocb, ret, issue_flags);
3350 /* it's faster to check here then delegate to kfree */
3356 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3358 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3360 return io_prep_rw(req, sqe);
3363 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3365 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3366 struct kiocb *kiocb = &req->rw.kiocb;
3367 struct iov_iter __iter, *iter = &__iter;
3368 struct io_async_rw *rw = req->async_data;
3369 ssize_t ret, ret2, io_size;
3370 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3376 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3380 io_size = iov_iter_count(iter);
3381 req->result = io_size;
3383 /* Ensure we clear previously set non-block flag */
3384 if (!force_nonblock)
3385 kiocb->ki_flags &= ~IOCB_NOWAIT;
3387 kiocb->ki_flags |= IOCB_NOWAIT;
3389 /* If the file doesn't support async, just async punt */
3390 if (force_nonblock && !io_file_supports_async(req, WRITE))
3393 /* file path doesn't support NOWAIT for non-direct_IO */
3394 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3395 (req->flags & REQ_F_ISREG))
3398 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3403 * Open-code file_start_write here to grab freeze protection,
3404 * which will be released by another thread in
3405 * io_complete_rw(). Fool lockdep by telling it the lock got
3406 * released so that it doesn't complain about the held lock when
3407 * we return to userspace.
3409 if (req->flags & REQ_F_ISREG) {
3410 sb_start_write(file_inode(req->file)->i_sb);
3411 __sb_writers_release(file_inode(req->file)->i_sb,
3414 kiocb->ki_flags |= IOCB_WRITE;
3416 if (req->file->f_op->write_iter)
3417 ret2 = call_write_iter(req->file, kiocb, iter);
3418 else if (req->file->f_op->write)
3419 ret2 = loop_rw_iter(WRITE, req, iter);
3423 if (req->flags & REQ_F_REISSUE) {
3424 req->flags &= ~REQ_F_REISSUE;
3429 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3430 * retry them without IOCB_NOWAIT.
3432 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3434 /* no retry on NONBLOCK nor RWF_NOWAIT */
3435 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3437 if (!force_nonblock || ret2 != -EAGAIN) {
3438 /* IOPOLL retry should happen for io-wq threads */
3439 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3442 kiocb_done(kiocb, ret2, issue_flags);
3445 /* some cases will consume bytes even on error returns */
3446 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3447 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3448 return ret ?: -EAGAIN;
3451 /* it's reportedly faster than delegating the null check to kfree() */
3457 static int io_renameat_prep(struct io_kiocb *req,
3458 const struct io_uring_sqe *sqe)
3460 struct io_rename *ren = &req->rename;
3461 const char __user *oldf, *newf;
3463 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3466 ren->old_dfd = READ_ONCE(sqe->fd);
3467 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3468 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3469 ren->new_dfd = READ_ONCE(sqe->len);
3470 ren->flags = READ_ONCE(sqe->rename_flags);
3472 ren->oldpath = getname(oldf);
3473 if (IS_ERR(ren->oldpath))
3474 return PTR_ERR(ren->oldpath);
3476 ren->newpath = getname(newf);
3477 if (IS_ERR(ren->newpath)) {
3478 putname(ren->oldpath);
3479 return PTR_ERR(ren->newpath);
3482 req->flags |= REQ_F_NEED_CLEANUP;
3486 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3488 struct io_rename *ren = &req->rename;
3491 if (issue_flags & IO_URING_F_NONBLOCK)
3494 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3495 ren->newpath, ren->flags);
3497 req->flags &= ~REQ_F_NEED_CLEANUP;
3500 io_req_complete(req, ret);
3504 static int io_unlinkat_prep(struct io_kiocb *req,
3505 const struct io_uring_sqe *sqe)
3507 struct io_unlink *un = &req->unlink;
3508 const char __user *fname;
3510 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3513 un->dfd = READ_ONCE(sqe->fd);
3515 un->flags = READ_ONCE(sqe->unlink_flags);
3516 if (un->flags & ~AT_REMOVEDIR)
3519 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3520 un->filename = getname(fname);
3521 if (IS_ERR(un->filename))
3522 return PTR_ERR(un->filename);
3524 req->flags |= REQ_F_NEED_CLEANUP;
3528 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3530 struct io_unlink *un = &req->unlink;
3533 if (issue_flags & IO_URING_F_NONBLOCK)
3536 if (un->flags & AT_REMOVEDIR)
3537 ret = do_rmdir(un->dfd, un->filename);
3539 ret = do_unlinkat(un->dfd, un->filename);
3541 req->flags &= ~REQ_F_NEED_CLEANUP;
3544 io_req_complete(req, ret);
3548 static int io_shutdown_prep(struct io_kiocb *req,
3549 const struct io_uring_sqe *sqe)
3551 #if defined(CONFIG_NET)
3552 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3554 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3558 req->shutdown.how = READ_ONCE(sqe->len);
3565 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3567 #if defined(CONFIG_NET)
3568 struct socket *sock;
3571 if (issue_flags & IO_URING_F_NONBLOCK)
3574 sock = sock_from_file(req->file);
3575 if (unlikely(!sock))
3578 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3581 io_req_complete(req, ret);
3588 static int __io_splice_prep(struct io_kiocb *req,
3589 const struct io_uring_sqe *sqe)
3591 struct io_splice* sp = &req->splice;
3592 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3594 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3598 sp->len = READ_ONCE(sqe->len);
3599 sp->flags = READ_ONCE(sqe->splice_flags);
3601 if (unlikely(sp->flags & ~valid_flags))
3604 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3605 (sp->flags & SPLICE_F_FD_IN_FIXED));
3608 req->flags |= REQ_F_NEED_CLEANUP;
3612 static int io_tee_prep(struct io_kiocb *req,
3613 const struct io_uring_sqe *sqe)
3615 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3617 return __io_splice_prep(req, sqe);
3620 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3622 struct io_splice *sp = &req->splice;
3623 struct file *in = sp->file_in;
3624 struct file *out = sp->file_out;
3625 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3628 if (issue_flags & IO_URING_F_NONBLOCK)
3631 ret = do_tee(in, out, sp->len, flags);
3633 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3635 req->flags &= ~REQ_F_NEED_CLEANUP;
3639 io_req_complete(req, ret);
3643 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3645 struct io_splice* sp = &req->splice;
3647 sp->off_in = READ_ONCE(sqe->splice_off_in);
3648 sp->off_out = READ_ONCE(sqe->off);
3649 return __io_splice_prep(req, sqe);
3652 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3654 struct io_splice *sp = &req->splice;
3655 struct file *in = sp->file_in;
3656 struct file *out = sp->file_out;
3657 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3658 loff_t *poff_in, *poff_out;
3661 if (issue_flags & IO_URING_F_NONBLOCK)
3664 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3665 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3668 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3670 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3672 req->flags &= ~REQ_F_NEED_CLEANUP;
3676 io_req_complete(req, ret);
3681 * IORING_OP_NOP just posts a completion event, nothing else.
3683 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3685 struct io_ring_ctx *ctx = req->ctx;
3687 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3690 __io_req_complete(req, issue_flags, 0, 0);
3694 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3696 struct io_ring_ctx *ctx = req->ctx;
3701 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3703 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3706 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3707 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3710 req->sync.off = READ_ONCE(sqe->off);
3711 req->sync.len = READ_ONCE(sqe->len);
3715 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3717 loff_t end = req->sync.off + req->sync.len;
3720 /* fsync always requires a blocking context */
3721 if (issue_flags & IO_URING_F_NONBLOCK)
3724 ret = vfs_fsync_range(req->file, req->sync.off,
3725 end > 0 ? end : LLONG_MAX,
3726 req->sync.flags & IORING_FSYNC_DATASYNC);
3729 io_req_complete(req, ret);
3733 static int io_fallocate_prep(struct io_kiocb *req,
3734 const struct io_uring_sqe *sqe)
3736 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3738 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3741 req->sync.off = READ_ONCE(sqe->off);
3742 req->sync.len = READ_ONCE(sqe->addr);
3743 req->sync.mode = READ_ONCE(sqe->len);
3747 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3751 /* fallocate always requiring blocking context */
3752 if (issue_flags & IO_URING_F_NONBLOCK)
3754 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3758 io_req_complete(req, ret);
3762 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3764 const char __user *fname;
3767 if (unlikely(sqe->ioprio || sqe->buf_index))
3769 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3772 /* open.how should be already initialised */
3773 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3774 req->open.how.flags |= O_LARGEFILE;
3776 req->open.dfd = READ_ONCE(sqe->fd);
3777 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3778 req->open.filename = getname(fname);
3779 if (IS_ERR(req->open.filename)) {
3780 ret = PTR_ERR(req->open.filename);
3781 req->open.filename = NULL;
3784 req->open.nofile = rlimit(RLIMIT_NOFILE);
3785 req->flags |= REQ_F_NEED_CLEANUP;
3789 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3793 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3795 mode = READ_ONCE(sqe->len);
3796 flags = READ_ONCE(sqe->open_flags);
3797 req->open.how = build_open_how(flags, mode);
3798 return __io_openat_prep(req, sqe);
3801 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3803 struct open_how __user *how;
3807 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3809 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3810 len = READ_ONCE(sqe->len);
3811 if (len < OPEN_HOW_SIZE_VER0)
3814 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3819 return __io_openat_prep(req, sqe);
3822 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3824 struct open_flags op;
3827 bool resolve_nonblock;
3830 ret = build_open_flags(&req->open.how, &op);
3833 nonblock_set = op.open_flag & O_NONBLOCK;
3834 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3835 if (issue_flags & IO_URING_F_NONBLOCK) {
3837 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3838 * it'll always -EAGAIN
3840 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3842 op.lookup_flags |= LOOKUP_CACHED;
3843 op.open_flag |= O_NONBLOCK;
3846 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3850 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3851 /* only retry if RESOLVE_CACHED wasn't already set by application */
3852 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3853 file == ERR_PTR(-EAGAIN)) {
3855 * We could hang on to this 'fd', but seems like marginal
3856 * gain for something that is now known to be a slower path.
3857 * So just put it, and we'll get a new one when we retry.
3865 ret = PTR_ERR(file);
3867 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3868 file->f_flags &= ~O_NONBLOCK;
3869 fsnotify_open(file);
3870 fd_install(ret, file);
3873 putname(req->open.filename);
3874 req->flags &= ~REQ_F_NEED_CLEANUP;
3877 __io_req_complete(req, issue_flags, ret, 0);
3881 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3883 return io_openat2(req, issue_flags);
3886 static int io_remove_buffers_prep(struct io_kiocb *req,
3887 const struct io_uring_sqe *sqe)
3889 struct io_provide_buf *p = &req->pbuf;
3892 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3895 tmp = READ_ONCE(sqe->fd);
3896 if (!tmp || tmp > USHRT_MAX)
3899 memset(p, 0, sizeof(*p));
3901 p->bgid = READ_ONCE(sqe->buf_group);
3905 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3906 int bgid, unsigned nbufs)
3910 /* shouldn't happen */
3914 /* the head kbuf is the list itself */
3915 while (!list_empty(&buf->list)) {
3916 struct io_buffer *nxt;
3918 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3919 list_del(&nxt->list);
3926 xa_erase(&ctx->io_buffers, bgid);
3931 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3933 struct io_provide_buf *p = &req->pbuf;
3934 struct io_ring_ctx *ctx = req->ctx;
3935 struct io_buffer *head;
3937 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3939 io_ring_submit_lock(ctx, !force_nonblock);
3941 lockdep_assert_held(&ctx->uring_lock);
3944 head = xa_load(&ctx->io_buffers, p->bgid);
3946 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3950 /* complete before unlock, IOPOLL may need the lock */
3951 __io_req_complete(req, issue_flags, ret, 0);
3952 io_ring_submit_unlock(ctx, !force_nonblock);
3956 static int io_provide_buffers_prep(struct io_kiocb *req,
3957 const struct io_uring_sqe *sqe)
3959 unsigned long size, tmp_check;
3960 struct io_provide_buf *p = &req->pbuf;
3963 if (sqe->ioprio || sqe->rw_flags)
3966 tmp = READ_ONCE(sqe->fd);
3967 if (!tmp || tmp > USHRT_MAX)
3970 p->addr = READ_ONCE(sqe->addr);
3971 p->len = READ_ONCE(sqe->len);
3973 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3976 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3979 size = (unsigned long)p->len * p->nbufs;
3980 if (!access_ok(u64_to_user_ptr(p->addr), size))
3983 p->bgid = READ_ONCE(sqe->buf_group);
3984 tmp = READ_ONCE(sqe->off);
3985 if (tmp > USHRT_MAX)
3991 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3993 struct io_buffer *buf;
3994 u64 addr = pbuf->addr;
3995 int i, bid = pbuf->bid;
3997 for (i = 0; i < pbuf->nbufs; i++) {
3998 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4003 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4008 INIT_LIST_HEAD(&buf->list);
4011 list_add_tail(&buf->list, &(*head)->list);
4015 return i ? i : -ENOMEM;
4018 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4020 struct io_provide_buf *p = &req->pbuf;
4021 struct io_ring_ctx *ctx = req->ctx;
4022 struct io_buffer *head, *list;
4024 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4026 io_ring_submit_lock(ctx, !force_nonblock);
4028 lockdep_assert_held(&ctx->uring_lock);
4030 list = head = xa_load(&ctx->io_buffers, p->bgid);
4032 ret = io_add_buffers(p, &head);
4033 if (ret >= 0 && !list) {
4034 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4036 __io_remove_buffers(ctx, head, p->bgid, -1U);
4040 /* complete before unlock, IOPOLL may need the lock */
4041 __io_req_complete(req, issue_flags, ret, 0);
4042 io_ring_submit_unlock(ctx, !force_nonblock);
4046 static int io_epoll_ctl_prep(struct io_kiocb *req,
4047 const struct io_uring_sqe *sqe)
4049 #if defined(CONFIG_EPOLL)
4050 if (sqe->ioprio || sqe->buf_index)
4052 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4055 req->epoll.epfd = READ_ONCE(sqe->fd);
4056 req->epoll.op = READ_ONCE(sqe->len);
4057 req->epoll.fd = READ_ONCE(sqe->off);
4059 if (ep_op_has_event(req->epoll.op)) {
4060 struct epoll_event __user *ev;
4062 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4063 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4073 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4075 #if defined(CONFIG_EPOLL)
4076 struct io_epoll *ie = &req->epoll;
4078 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4080 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4081 if (force_nonblock && ret == -EAGAIN)
4086 __io_req_complete(req, issue_flags, ret, 0);
4093 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4095 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4096 if (sqe->ioprio || sqe->buf_index || sqe->off)
4098 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4101 req->madvise.addr = READ_ONCE(sqe->addr);
4102 req->madvise.len = READ_ONCE(sqe->len);
4103 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4110 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4112 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4113 struct io_madvise *ma = &req->madvise;
4116 if (issue_flags & IO_URING_F_NONBLOCK)
4119 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4122 io_req_complete(req, ret);
4129 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4131 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4133 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4136 req->fadvise.offset = READ_ONCE(sqe->off);
4137 req->fadvise.len = READ_ONCE(sqe->len);
4138 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4142 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4144 struct io_fadvise *fa = &req->fadvise;
4147 if (issue_flags & IO_URING_F_NONBLOCK) {
4148 switch (fa->advice) {
4149 case POSIX_FADV_NORMAL:
4150 case POSIX_FADV_RANDOM:
4151 case POSIX_FADV_SEQUENTIAL:
4158 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4161 __io_req_complete(req, issue_flags, ret, 0);
4165 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4167 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4169 if (sqe->ioprio || sqe->buf_index)
4171 if (req->flags & REQ_F_FIXED_FILE)
4174 req->statx.dfd = READ_ONCE(sqe->fd);
4175 req->statx.mask = READ_ONCE(sqe->len);
4176 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4177 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4178 req->statx.flags = READ_ONCE(sqe->statx_flags);
4183 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4185 struct io_statx *ctx = &req->statx;
4188 if (issue_flags & IO_URING_F_NONBLOCK)
4191 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4196 io_req_complete(req, ret);
4200 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4202 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4204 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4205 sqe->rw_flags || sqe->buf_index)
4207 if (req->flags & REQ_F_FIXED_FILE)
4210 req->close.fd = READ_ONCE(sqe->fd);
4214 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4216 struct files_struct *files = current->files;
4217 struct io_close *close = &req->close;
4218 struct fdtable *fdt;
4219 struct file *file = NULL;
4222 spin_lock(&files->file_lock);
4223 fdt = files_fdtable(files);
4224 if (close->fd >= fdt->max_fds) {
4225 spin_unlock(&files->file_lock);
4228 file = fdt->fd[close->fd];
4229 if (!file || file->f_op == &io_uring_fops) {
4230 spin_unlock(&files->file_lock);
4235 /* if the file has a flush method, be safe and punt to async */
4236 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4237 spin_unlock(&files->file_lock);
4241 ret = __close_fd_get_file(close->fd, &file);
4242 spin_unlock(&files->file_lock);
4249 /* No ->flush() or already async, safely close from here */
4250 ret = filp_close(file, current->files);
4256 __io_req_complete(req, issue_flags, ret, 0);
4260 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4262 struct io_ring_ctx *ctx = req->ctx;
4264 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4266 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4269 req->sync.off = READ_ONCE(sqe->off);
4270 req->sync.len = READ_ONCE(sqe->len);
4271 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4275 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4279 /* sync_file_range always requires a blocking context */
4280 if (issue_flags & IO_URING_F_NONBLOCK)
4283 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4287 io_req_complete(req, ret);
4291 #if defined(CONFIG_NET)
4292 static int io_setup_async_msg(struct io_kiocb *req,
4293 struct io_async_msghdr *kmsg)
4295 struct io_async_msghdr *async_msg = req->async_data;
4299 if (io_alloc_async_data(req)) {
4300 kfree(kmsg->free_iov);
4303 async_msg = req->async_data;
4304 req->flags |= REQ_F_NEED_CLEANUP;
4305 memcpy(async_msg, kmsg, sizeof(*kmsg));
4306 async_msg->msg.msg_name = &async_msg->addr;
4307 /* if were using fast_iov, set it to the new one */
4308 if (!async_msg->free_iov)
4309 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4314 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4315 struct io_async_msghdr *iomsg)
4317 iomsg->msg.msg_name = &iomsg->addr;
4318 iomsg->free_iov = iomsg->fast_iov;
4319 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4320 req->sr_msg.msg_flags, &iomsg->free_iov);
4323 static int io_sendmsg_prep_async(struct io_kiocb *req)
4327 ret = io_sendmsg_copy_hdr(req, req->async_data);
4329 req->flags |= REQ_F_NEED_CLEANUP;
4333 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4335 struct io_sr_msg *sr = &req->sr_msg;
4337 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4340 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4341 sr->len = READ_ONCE(sqe->len);
4342 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4343 if (sr->msg_flags & MSG_DONTWAIT)
4344 req->flags |= REQ_F_NOWAIT;
4346 #ifdef CONFIG_COMPAT
4347 if (req->ctx->compat)
4348 sr->msg_flags |= MSG_CMSG_COMPAT;
4353 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4355 struct io_async_msghdr iomsg, *kmsg;
4356 struct socket *sock;
4361 sock = sock_from_file(req->file);
4362 if (unlikely(!sock))
4365 kmsg = req->async_data;
4367 ret = io_sendmsg_copy_hdr(req, &iomsg);
4373 flags = req->sr_msg.msg_flags;
4374 if (issue_flags & IO_URING_F_NONBLOCK)
4375 flags |= MSG_DONTWAIT;
4376 if (flags & MSG_WAITALL)
4377 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4379 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4380 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4381 return io_setup_async_msg(req, kmsg);
4382 if (ret == -ERESTARTSYS)
4385 /* fast path, check for non-NULL to avoid function call */
4387 kfree(kmsg->free_iov);
4388 req->flags &= ~REQ_F_NEED_CLEANUP;
4391 __io_req_complete(req, issue_flags, ret, 0);
4395 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4397 struct io_sr_msg *sr = &req->sr_msg;
4400 struct socket *sock;
4405 sock = sock_from_file(req->file);
4406 if (unlikely(!sock))
4409 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4413 msg.msg_name = NULL;
4414 msg.msg_control = NULL;
4415 msg.msg_controllen = 0;
4416 msg.msg_namelen = 0;
4418 flags = req->sr_msg.msg_flags;
4419 if (issue_flags & IO_URING_F_NONBLOCK)
4420 flags |= MSG_DONTWAIT;
4421 if (flags & MSG_WAITALL)
4422 min_ret = iov_iter_count(&msg.msg_iter);
4424 msg.msg_flags = flags;
4425 ret = sock_sendmsg(sock, &msg);
4426 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4428 if (ret == -ERESTARTSYS)
4433 __io_req_complete(req, issue_flags, ret, 0);
4437 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4438 struct io_async_msghdr *iomsg)
4440 struct io_sr_msg *sr = &req->sr_msg;
4441 struct iovec __user *uiov;
4445 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4446 &iomsg->uaddr, &uiov, &iov_len);
4450 if (req->flags & REQ_F_BUFFER_SELECT) {
4453 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4455 sr->len = iomsg->fast_iov[0].iov_len;
4456 iomsg->free_iov = NULL;
4458 iomsg->free_iov = iomsg->fast_iov;
4459 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4460 &iomsg->free_iov, &iomsg->msg.msg_iter,
4469 #ifdef CONFIG_COMPAT
4470 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4471 struct io_async_msghdr *iomsg)
4473 struct io_sr_msg *sr = &req->sr_msg;
4474 struct compat_iovec __user *uiov;
4479 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4484 uiov = compat_ptr(ptr);
4485 if (req->flags & REQ_F_BUFFER_SELECT) {
4486 compat_ssize_t clen;
4490 if (!access_ok(uiov, sizeof(*uiov)))
4492 if (__get_user(clen, &uiov->iov_len))
4497 iomsg->free_iov = NULL;
4499 iomsg->free_iov = iomsg->fast_iov;
4500 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4501 UIO_FASTIOV, &iomsg->free_iov,
4502 &iomsg->msg.msg_iter, true);
4511 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4512 struct io_async_msghdr *iomsg)
4514 iomsg->msg.msg_name = &iomsg->addr;
4516 #ifdef CONFIG_COMPAT
4517 if (req->ctx->compat)
4518 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4521 return __io_recvmsg_copy_hdr(req, iomsg);
4524 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4527 struct io_sr_msg *sr = &req->sr_msg;
4528 struct io_buffer *kbuf;
4530 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4535 req->flags |= REQ_F_BUFFER_SELECTED;
4539 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4541 return io_put_kbuf(req, req->sr_msg.kbuf);
4544 static int io_recvmsg_prep_async(struct io_kiocb *req)
4548 ret = io_recvmsg_copy_hdr(req, req->async_data);
4550 req->flags |= REQ_F_NEED_CLEANUP;
4554 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4556 struct io_sr_msg *sr = &req->sr_msg;
4558 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4561 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4562 sr->len = READ_ONCE(sqe->len);
4563 sr->bgid = READ_ONCE(sqe->buf_group);
4564 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4565 if (sr->msg_flags & MSG_DONTWAIT)
4566 req->flags |= REQ_F_NOWAIT;
4568 #ifdef CONFIG_COMPAT
4569 if (req->ctx->compat)
4570 sr->msg_flags |= MSG_CMSG_COMPAT;
4575 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4577 struct io_async_msghdr iomsg, *kmsg;
4578 struct socket *sock;
4579 struct io_buffer *kbuf;
4582 int ret, cflags = 0;
4583 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4585 sock = sock_from_file(req->file);
4586 if (unlikely(!sock))
4589 kmsg = req->async_data;
4591 ret = io_recvmsg_copy_hdr(req, &iomsg);
4597 if (req->flags & REQ_F_BUFFER_SELECT) {
4598 kbuf = io_recv_buffer_select(req, !force_nonblock);
4600 return PTR_ERR(kbuf);
4601 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4602 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4603 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4604 1, req->sr_msg.len);
4607 flags = req->sr_msg.msg_flags;
4609 flags |= MSG_DONTWAIT;
4610 if (flags & MSG_WAITALL)
4611 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4613 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4614 kmsg->uaddr, flags);
4615 if (force_nonblock && ret == -EAGAIN)
4616 return io_setup_async_msg(req, kmsg);
4617 if (ret == -ERESTARTSYS)
4620 if (req->flags & REQ_F_BUFFER_SELECTED)
4621 cflags = io_put_recv_kbuf(req);
4622 /* fast path, check for non-NULL to avoid function call */
4624 kfree(kmsg->free_iov);
4625 req->flags &= ~REQ_F_NEED_CLEANUP;
4626 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4628 __io_req_complete(req, issue_flags, ret, cflags);
4632 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4634 struct io_buffer *kbuf;
4635 struct io_sr_msg *sr = &req->sr_msg;
4637 void __user *buf = sr->buf;
4638 struct socket *sock;
4642 int ret, cflags = 0;
4643 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4645 sock = sock_from_file(req->file);
4646 if (unlikely(!sock))
4649 if (req->flags & REQ_F_BUFFER_SELECT) {
4650 kbuf = io_recv_buffer_select(req, !force_nonblock);
4652 return PTR_ERR(kbuf);
4653 buf = u64_to_user_ptr(kbuf->addr);
4656 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4660 msg.msg_name = NULL;
4661 msg.msg_control = NULL;
4662 msg.msg_controllen = 0;
4663 msg.msg_namelen = 0;
4664 msg.msg_iocb = NULL;
4667 flags = req->sr_msg.msg_flags;
4669 flags |= MSG_DONTWAIT;
4670 if (flags & MSG_WAITALL)
4671 min_ret = iov_iter_count(&msg.msg_iter);
4673 ret = sock_recvmsg(sock, &msg, flags);
4674 if (force_nonblock && ret == -EAGAIN)
4676 if (ret == -ERESTARTSYS)
4679 if (req->flags & REQ_F_BUFFER_SELECTED)
4680 cflags = io_put_recv_kbuf(req);
4681 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4683 __io_req_complete(req, issue_flags, ret, cflags);
4687 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4689 struct io_accept *accept = &req->accept;
4691 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4693 if (sqe->ioprio || sqe->len || sqe->buf_index)
4696 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4697 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4698 accept->flags = READ_ONCE(sqe->accept_flags);
4699 accept->nofile = rlimit(RLIMIT_NOFILE);
4703 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4705 struct io_accept *accept = &req->accept;
4706 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4707 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4710 if (req->file->f_flags & O_NONBLOCK)
4711 req->flags |= REQ_F_NOWAIT;
4713 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4714 accept->addr_len, accept->flags,
4716 if (ret == -EAGAIN && force_nonblock)
4719 if (ret == -ERESTARTSYS)
4723 __io_req_complete(req, issue_flags, ret, 0);
4727 static int io_connect_prep_async(struct io_kiocb *req)
4729 struct io_async_connect *io = req->async_data;
4730 struct io_connect *conn = &req->connect;
4732 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4735 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4737 struct io_connect *conn = &req->connect;
4739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4741 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4744 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4745 conn->addr_len = READ_ONCE(sqe->addr2);
4749 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4751 struct io_async_connect __io, *io;
4752 unsigned file_flags;
4754 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4756 if (req->async_data) {
4757 io = req->async_data;
4759 ret = move_addr_to_kernel(req->connect.addr,
4760 req->connect.addr_len,
4767 file_flags = force_nonblock ? O_NONBLOCK : 0;
4769 ret = __sys_connect_file(req->file, &io->address,
4770 req->connect.addr_len, file_flags);
4771 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4772 if (req->async_data)
4774 if (io_alloc_async_data(req)) {
4778 memcpy(req->async_data, &__io, sizeof(__io));
4781 if (ret == -ERESTARTSYS)
4786 __io_req_complete(req, issue_flags, ret, 0);
4789 #else /* !CONFIG_NET */
4790 #define IO_NETOP_FN(op) \
4791 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4793 return -EOPNOTSUPP; \
4796 #define IO_NETOP_PREP(op) \
4798 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4800 return -EOPNOTSUPP; \
4803 #define IO_NETOP_PREP_ASYNC(op) \
4805 static int io_##op##_prep_async(struct io_kiocb *req) \
4807 return -EOPNOTSUPP; \
4810 IO_NETOP_PREP_ASYNC(sendmsg);
4811 IO_NETOP_PREP_ASYNC(recvmsg);
4812 IO_NETOP_PREP_ASYNC(connect);
4813 IO_NETOP_PREP(accept);
4816 #endif /* CONFIG_NET */
4818 struct io_poll_table {
4819 struct poll_table_struct pt;
4820 struct io_kiocb *req;
4824 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4825 __poll_t mask, task_work_func_t func)
4829 /* for instances that support it check for an event match first: */
4830 if (mask && !(mask & poll->events))
4833 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4835 list_del_init(&poll->wait.entry);
4838 req->task_work.func = func;
4841 * If this fails, then the task is exiting. When a task exits, the
4842 * work gets canceled, so just cancel this request as well instead
4843 * of executing it. We can't safely execute it anyway, as we may not
4844 * have the needed state needed for it anyway.
4846 ret = io_req_task_work_add(req);
4847 if (unlikely(ret)) {
4848 WRITE_ONCE(poll->canceled, true);
4849 io_req_task_work_add_fallback(req, func);
4854 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4855 __acquires(&req->ctx->completion_lock)
4857 struct io_ring_ctx *ctx = req->ctx;
4859 if (!req->result && !READ_ONCE(poll->canceled)) {
4860 struct poll_table_struct pt = { ._key = poll->events };
4862 req->result = vfs_poll(req->file, &pt) & poll->events;
4865 spin_lock_irq(&ctx->completion_lock);
4866 if (!req->result && !READ_ONCE(poll->canceled)) {
4867 add_wait_queue(poll->head, &poll->wait);
4874 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4876 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4877 if (req->opcode == IORING_OP_POLL_ADD)
4878 return req->async_data;
4879 return req->apoll->double_poll;
4882 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4884 if (req->opcode == IORING_OP_POLL_ADD)
4886 return &req->apoll->poll;
4889 static void io_poll_remove_double(struct io_kiocb *req)
4890 __must_hold(&req->ctx->completion_lock)
4892 struct io_poll_iocb *poll = io_poll_get_double(req);
4894 lockdep_assert_held(&req->ctx->completion_lock);
4896 if (poll && poll->head) {
4897 struct wait_queue_head *head = poll->head;
4899 spin_lock(&head->lock);
4900 list_del_init(&poll->wait.entry);
4901 if (poll->wait.private)
4904 spin_unlock(&head->lock);
4908 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4909 __must_hold(&req->ctx->completion_lock)
4911 struct io_ring_ctx *ctx = req->ctx;
4912 unsigned flags = IORING_CQE_F_MORE;
4915 if (READ_ONCE(req->poll.canceled)) {
4917 req->poll.events |= EPOLLONESHOT;
4919 error = mangle_poll(mask);
4921 if (req->poll.events & EPOLLONESHOT)
4923 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4924 io_poll_remove_waitqs(req);
4925 req->poll.done = true;
4928 if (flags & IORING_CQE_F_MORE)
4931 io_commit_cqring(ctx);
4932 return !(flags & IORING_CQE_F_MORE);
4935 static void io_poll_task_func(struct callback_head *cb)
4937 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4938 struct io_ring_ctx *ctx = req->ctx;
4939 struct io_kiocb *nxt;
4941 if (io_poll_rewait(req, &req->poll)) {
4942 spin_unlock_irq(&ctx->completion_lock);
4946 done = io_poll_complete(req, req->result);
4948 hash_del(&req->hash_node);
4951 add_wait_queue(req->poll.head, &req->poll.wait);
4953 spin_unlock_irq(&ctx->completion_lock);
4954 io_cqring_ev_posted(ctx);
4957 nxt = io_put_req_find_next(req);
4959 __io_req_task_submit(nxt);
4964 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4965 int sync, void *key)
4967 struct io_kiocb *req = wait->private;
4968 struct io_poll_iocb *poll = io_poll_get_single(req);
4969 __poll_t mask = key_to_poll(key);
4971 /* for instances that support it check for an event match first: */
4972 if (mask && !(mask & poll->events))
4974 if (!(poll->events & EPOLLONESHOT))
4975 return poll->wait.func(&poll->wait, mode, sync, key);
4977 list_del_init(&wait->entry);
4979 if (poll && poll->head) {
4982 spin_lock(&poll->head->lock);
4983 done = list_empty(&poll->wait.entry);
4985 list_del_init(&poll->wait.entry);
4986 /* make sure double remove sees this as being gone */
4987 wait->private = NULL;
4988 spin_unlock(&poll->head->lock);
4990 /* use wait func handler, so it matches the rq type */
4991 poll->wait.func(&poll->wait, mode, sync, key);
4998 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4999 wait_queue_func_t wake_func)
5003 poll->canceled = false;
5004 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5005 /* mask in events that we always want/need */
5006 poll->events = events | IO_POLL_UNMASK;
5007 INIT_LIST_HEAD(&poll->wait.entry);
5008 init_waitqueue_func_entry(&poll->wait, wake_func);
5011 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5012 struct wait_queue_head *head,
5013 struct io_poll_iocb **poll_ptr)
5015 struct io_kiocb *req = pt->req;
5018 * If poll->head is already set, it's because the file being polled
5019 * uses multiple waitqueues for poll handling (eg one for read, one
5020 * for write). Setup a separate io_poll_iocb if this happens.
5022 if (unlikely(poll->head)) {
5023 struct io_poll_iocb *poll_one = poll;
5025 /* already have a 2nd entry, fail a third attempt */
5027 pt->error = -EINVAL;
5031 * Can't handle multishot for double wait for now, turn it
5032 * into one-shot mode.
5034 if (!(poll_one->events & EPOLLONESHOT))
5035 poll_one->events |= EPOLLONESHOT;
5036 /* double add on the same waitqueue head, ignore */
5037 if (poll_one->head == head)
5039 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5041 pt->error = -ENOMEM;
5044 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5046 poll->wait.private = req;
5053 if (poll->events & EPOLLEXCLUSIVE)
5054 add_wait_queue_exclusive(head, &poll->wait);
5056 add_wait_queue(head, &poll->wait);
5059 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5060 struct poll_table_struct *p)
5062 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5063 struct async_poll *apoll = pt->req->apoll;
5065 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5068 static void io_async_task_func(struct callback_head *cb)
5070 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5071 struct async_poll *apoll = req->apoll;
5072 struct io_ring_ctx *ctx = req->ctx;
5074 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5076 if (io_poll_rewait(req, &apoll->poll)) {
5077 spin_unlock_irq(&ctx->completion_lock);
5081 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);
5091 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5094 struct io_kiocb *req = wait->private;
5095 struct io_poll_iocb *poll = &req->apoll->poll;
5097 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5100 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5103 static void io_poll_req_insert(struct io_kiocb *req)
5105 struct io_ring_ctx *ctx = req->ctx;
5106 struct hlist_head *list;
5108 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5109 hlist_add_head(&req->hash_node, list);
5112 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5113 struct io_poll_iocb *poll,
5114 struct io_poll_table *ipt, __poll_t mask,
5115 wait_queue_func_t wake_func)
5116 __acquires(&ctx->completion_lock)
5118 struct io_ring_ctx *ctx = req->ctx;
5119 bool cancel = false;
5121 INIT_HLIST_NODE(&req->hash_node);
5122 io_init_poll_iocb(poll, mask, wake_func);
5123 poll->file = req->file;
5124 poll->wait.private = req;
5126 ipt->pt._key = mask;
5128 ipt->error = -EINVAL;
5130 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5132 spin_lock_irq(&ctx->completion_lock);
5133 if (likely(poll->head)) {
5134 spin_lock(&poll->head->lock);
5135 if (unlikely(list_empty(&poll->wait.entry))) {
5141 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5142 list_del_init(&poll->wait.entry);
5144 WRITE_ONCE(poll->canceled, true);
5145 else if (!poll->done) /* actually waiting for an event */
5146 io_poll_req_insert(req);
5147 spin_unlock(&poll->head->lock);
5153 static bool io_arm_poll_handler(struct io_kiocb *req)
5155 const struct io_op_def *def = &io_op_defs[req->opcode];
5156 struct io_ring_ctx *ctx = req->ctx;
5157 struct async_poll *apoll;
5158 struct io_poll_table ipt;
5162 if (!req->file || !file_can_poll(req->file))
5164 if (req->flags & REQ_F_POLLED)
5168 else if (def->pollout)
5172 /* if we can't nonblock try, then no point in arming a poll handler */
5173 if (!io_file_supports_async(req, rw))
5176 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5177 if (unlikely(!apoll))
5179 apoll->double_poll = NULL;
5181 req->flags |= REQ_F_POLLED;
5184 mask = EPOLLONESHOT;
5186 mask |= POLLIN | POLLRDNORM;
5188 mask |= POLLOUT | POLLWRNORM;
5190 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5191 if ((req->opcode == IORING_OP_RECVMSG) &&
5192 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5195 mask |= POLLERR | POLLPRI;
5197 ipt.pt._qproc = io_async_queue_proc;
5199 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5201 if (ret || ipt.error) {
5202 io_poll_remove_double(req);
5203 spin_unlock_irq(&ctx->completion_lock);
5206 spin_unlock_irq(&ctx->completion_lock);
5207 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5208 apoll->poll.events);
5212 static bool __io_poll_remove_one(struct io_kiocb *req,
5213 struct io_poll_iocb *poll, bool do_cancel)
5214 __must_hold(&req->ctx->completion_lock)
5216 bool do_complete = false;
5220 spin_lock(&poll->head->lock);
5222 WRITE_ONCE(poll->canceled, true);
5223 if (!list_empty(&poll->wait.entry)) {
5224 list_del_init(&poll->wait.entry);
5227 spin_unlock(&poll->head->lock);
5228 hash_del(&req->hash_node);
5232 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5233 __must_hold(&req->ctx->completion_lock)
5237 io_poll_remove_double(req);
5238 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5240 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5241 /* non-poll requests have submit ref still */
5247 static bool io_poll_remove_one(struct io_kiocb *req)
5248 __must_hold(&req->ctx->completion_lock)
5252 do_complete = io_poll_remove_waitqs(req);
5254 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5255 io_commit_cqring(req->ctx);
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,
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, cancel_all))
5280 posted += io_poll_remove_one(req);
5283 spin_unlock_irq(&ctx->completion_lock);
5286 io_cqring_ev_posted(ctx);
5291 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5293 __must_hold(&ctx->completion_lock)
5295 struct hlist_head *list;
5296 struct io_kiocb *req;
5298 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5299 hlist_for_each_entry(req, list, hash_node) {
5300 if (sqe_addr != req->user_data)
5302 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5309 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5311 __must_hold(&ctx->completion_lock)
5313 struct io_kiocb *req;
5315 req = io_poll_find(ctx, sqe_addr, poll_only);
5318 if (io_poll_remove_one(req))
5324 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5329 events = READ_ONCE(sqe->poll32_events);
5331 events = swahw32(events);
5333 if (!(flags & IORING_POLL_ADD_MULTI))
5334 events |= EPOLLONESHOT;
5335 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5338 static int io_poll_update_prep(struct io_kiocb *req,
5339 const struct io_uring_sqe *sqe)
5341 struct io_poll_update *upd = &req->poll_update;
5344 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5346 if (sqe->ioprio || sqe->buf_index)
5348 flags = READ_ONCE(sqe->len);
5349 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5350 IORING_POLL_ADD_MULTI))
5352 /* meaningless without update */
5353 if (flags == IORING_POLL_ADD_MULTI)
5356 upd->old_user_data = READ_ONCE(sqe->addr);
5357 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5358 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5360 upd->new_user_data = READ_ONCE(sqe->off);
5361 if (!upd->update_user_data && upd->new_user_data)
5363 if (upd->update_events)
5364 upd->events = io_poll_parse_events(sqe, flags);
5365 else if (sqe->poll32_events)
5371 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5374 struct io_kiocb *req = wait->private;
5375 struct io_poll_iocb *poll = &req->poll;
5377 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5380 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5381 struct poll_table_struct *p)
5383 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5385 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5388 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5390 struct io_poll_iocb *poll = &req->poll;
5393 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5395 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5397 flags = READ_ONCE(sqe->len);
5398 if (flags & ~IORING_POLL_ADD_MULTI)
5401 poll->events = io_poll_parse_events(sqe, flags);
5405 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5407 struct io_poll_iocb *poll = &req->poll;
5408 struct io_ring_ctx *ctx = req->ctx;
5409 struct io_poll_table ipt;
5412 ipt.pt._qproc = io_poll_queue_proc;
5414 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5417 if (mask) { /* no async, we'd stolen it */
5419 io_poll_complete(req, mask);
5421 spin_unlock_irq(&ctx->completion_lock);
5424 io_cqring_ev_posted(ctx);
5425 if (poll->events & EPOLLONESHOT)
5431 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5433 struct io_ring_ctx *ctx = req->ctx;
5434 struct io_kiocb *preq;
5438 spin_lock_irq(&ctx->completion_lock);
5439 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5445 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5447 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5452 * Don't allow racy completion with singleshot, as we cannot safely
5453 * update those. For multishot, if we're racing with completion, just
5454 * let completion re-add it.
5456 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5457 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5461 /* we now have a detached poll request. reissue. */
5465 spin_unlock_irq(&ctx->completion_lock);
5467 io_req_complete(req, ret);
5470 /* only mask one event flags, keep behavior flags */
5471 if (req->poll_update.update_events) {
5472 preq->poll.events &= ~0xffff;
5473 preq->poll.events |= req->poll_update.events & 0xffff;
5474 preq->poll.events |= IO_POLL_UNMASK;
5476 if (req->poll_update.update_user_data)
5477 preq->user_data = req->poll_update.new_user_data;
5478 spin_unlock_irq(&ctx->completion_lock);
5480 /* complete update request, we're done with it */
5481 io_req_complete(req, ret);
5484 ret = io_poll_add(preq, issue_flags);
5487 io_req_complete(preq, ret);
5493 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5495 struct io_timeout_data *data = container_of(timer,
5496 struct io_timeout_data, timer);
5497 struct io_kiocb *req = data->req;
5498 struct io_ring_ctx *ctx = req->ctx;
5499 unsigned long flags;
5501 spin_lock_irqsave(&ctx->completion_lock, flags);
5502 list_del_init(&req->timeout.list);
5503 atomic_set(&req->ctx->cq_timeouts,
5504 atomic_read(&req->ctx->cq_timeouts) + 1);
5506 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5507 io_commit_cqring(ctx);
5508 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5510 io_cqring_ev_posted(ctx);
5513 return HRTIMER_NORESTART;
5516 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5518 __must_hold(&ctx->completion_lock)
5520 struct io_timeout_data *io;
5521 struct io_kiocb *req;
5524 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5525 found = user_data == req->user_data;
5530 return ERR_PTR(-ENOENT);
5532 io = req->async_data;
5533 if (hrtimer_try_to_cancel(&io->timer) == -1)
5534 return ERR_PTR(-EALREADY);
5535 list_del_init(&req->timeout.list);
5539 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5540 __must_hold(&ctx->completion_lock)
5542 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5545 return PTR_ERR(req);
5548 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5549 io_put_req_deferred(req, 1);
5553 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5554 struct timespec64 *ts, enum hrtimer_mode mode)
5555 __must_hold(&ctx->completion_lock)
5557 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5558 struct io_timeout_data *data;
5561 return PTR_ERR(req);
5563 req->timeout.off = 0; /* noseq */
5564 data = req->async_data;
5565 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5566 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5567 data->timer.function = io_timeout_fn;
5568 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5572 static int io_timeout_remove_prep(struct io_kiocb *req,
5573 const struct io_uring_sqe *sqe)
5575 struct io_timeout_rem *tr = &req->timeout_rem;
5577 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5579 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5581 if (sqe->ioprio || sqe->buf_index || sqe->len)
5584 tr->addr = READ_ONCE(sqe->addr);
5585 tr->flags = READ_ONCE(sqe->timeout_flags);
5586 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5587 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5589 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5591 } else if (tr->flags) {
5592 /* timeout removal doesn't support flags */
5599 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5601 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5606 * Remove or update an existing timeout command
5608 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5610 struct io_timeout_rem *tr = &req->timeout_rem;
5611 struct io_ring_ctx *ctx = req->ctx;
5614 spin_lock_irq(&ctx->completion_lock);
5615 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5616 ret = io_timeout_cancel(ctx, tr->addr);
5618 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5619 io_translate_timeout_mode(tr->flags));
5621 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5622 io_commit_cqring(ctx);
5623 spin_unlock_irq(&ctx->completion_lock);
5624 io_cqring_ev_posted(ctx);
5631 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5632 bool is_timeout_link)
5634 struct io_timeout_data *data;
5636 u32 off = READ_ONCE(sqe->off);
5638 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5640 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5642 if (off && is_timeout_link)
5644 flags = READ_ONCE(sqe->timeout_flags);
5645 if (flags & ~IORING_TIMEOUT_ABS)
5648 req->timeout.off = off;
5650 if (!req->async_data && io_alloc_async_data(req))
5653 data = req->async_data;
5656 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5659 data->mode = io_translate_timeout_mode(flags);
5660 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5661 if (is_timeout_link)
5662 io_req_track_inflight(req);
5666 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5668 struct io_ring_ctx *ctx = req->ctx;
5669 struct io_timeout_data *data = req->async_data;
5670 struct list_head *entry;
5671 u32 tail, off = req->timeout.off;
5673 spin_lock_irq(&ctx->completion_lock);
5676 * sqe->off holds how many events that need to occur for this
5677 * timeout event to be satisfied. If it isn't set, then this is
5678 * a pure timeout request, sequence isn't used.
5680 if (io_is_timeout_noseq(req)) {
5681 entry = ctx->timeout_list.prev;
5685 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5686 req->timeout.target_seq = tail + off;
5688 /* Update the last seq here in case io_flush_timeouts() hasn't.
5689 * This is safe because ->completion_lock is held, and submissions
5690 * and completions are never mixed in the same ->completion_lock section.
5692 ctx->cq_last_tm_flush = tail;
5695 * Insertion sort, ensuring the first entry in the list is always
5696 * the one we need first.
5698 list_for_each_prev(entry, &ctx->timeout_list) {
5699 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5702 if (io_is_timeout_noseq(nxt))
5704 /* nxt.seq is behind @tail, otherwise would've been completed */
5705 if (off >= nxt->timeout.target_seq - tail)
5709 list_add(&req->timeout.list, entry);
5710 data->timer.function = io_timeout_fn;
5711 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5712 spin_unlock_irq(&ctx->completion_lock);
5716 struct io_cancel_data {
5717 struct io_ring_ctx *ctx;
5721 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5723 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5724 struct io_cancel_data *cd = data;
5726 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5729 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5730 struct io_ring_ctx *ctx)
5732 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5733 enum io_wq_cancel cancel_ret;
5736 if (!tctx || !tctx->io_wq)
5739 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5740 switch (cancel_ret) {
5741 case IO_WQ_CANCEL_OK:
5744 case IO_WQ_CANCEL_RUNNING:
5747 case IO_WQ_CANCEL_NOTFOUND:
5755 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5756 struct io_kiocb *req, __u64 sqe_addr,
5759 unsigned long flags;
5762 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5763 spin_lock_irqsave(&ctx->completion_lock, flags);
5766 ret = io_timeout_cancel(ctx, sqe_addr);
5769 ret = io_poll_cancel(ctx, sqe_addr, false);
5773 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5774 io_commit_cqring(ctx);
5775 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5776 io_cqring_ev_posted(ctx);
5782 static int io_async_cancel_prep(struct io_kiocb *req,
5783 const struct io_uring_sqe *sqe)
5785 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5787 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5789 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5792 req->cancel.addr = READ_ONCE(sqe->addr);
5796 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5798 struct io_ring_ctx *ctx = req->ctx;
5799 u64 sqe_addr = req->cancel.addr;
5800 struct io_tctx_node *node;
5803 /* tasks should wait for their io-wq threads, so safe w/o sync */
5804 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5805 spin_lock_irq(&ctx->completion_lock);
5808 ret = io_timeout_cancel(ctx, sqe_addr);
5811 ret = io_poll_cancel(ctx, sqe_addr, false);
5814 spin_unlock_irq(&ctx->completion_lock);
5816 /* slow path, try all io-wq's */
5817 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5819 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5820 struct io_uring_task *tctx = node->task->io_uring;
5822 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5826 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5828 spin_lock_irq(&ctx->completion_lock);
5830 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5831 io_commit_cqring(ctx);
5832 spin_unlock_irq(&ctx->completion_lock);
5833 io_cqring_ev_posted(ctx);
5841 static int io_rsrc_update_prep(struct io_kiocb *req,
5842 const struct io_uring_sqe *sqe)
5844 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5846 if (sqe->ioprio || sqe->rw_flags)
5849 req->rsrc_update.offset = READ_ONCE(sqe->off);
5850 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5851 if (!req->rsrc_update.nr_args)
5853 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5857 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5859 struct io_ring_ctx *ctx = req->ctx;
5860 struct io_uring_rsrc_update2 up;
5863 if (issue_flags & IO_URING_F_NONBLOCK)
5866 up.offset = req->rsrc_update.offset;
5867 up.data = req->rsrc_update.arg;
5872 mutex_lock(&ctx->uring_lock);
5873 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5874 &up, req->rsrc_update.nr_args);
5875 mutex_unlock(&ctx->uring_lock);
5879 __io_req_complete(req, issue_flags, ret, 0);
5883 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5885 switch (req->opcode) {
5888 case IORING_OP_READV:
5889 case IORING_OP_READ_FIXED:
5890 case IORING_OP_READ:
5891 return io_read_prep(req, sqe);
5892 case IORING_OP_WRITEV:
5893 case IORING_OP_WRITE_FIXED:
5894 case IORING_OP_WRITE:
5895 return io_write_prep(req, sqe);
5896 case IORING_OP_POLL_ADD:
5897 return io_poll_add_prep(req, sqe);
5898 case IORING_OP_POLL_REMOVE:
5899 return io_poll_update_prep(req, sqe);
5900 case IORING_OP_FSYNC:
5901 return io_fsync_prep(req, sqe);
5902 case IORING_OP_SYNC_FILE_RANGE:
5903 return io_sfr_prep(req, sqe);
5904 case IORING_OP_SENDMSG:
5905 case IORING_OP_SEND:
5906 return io_sendmsg_prep(req, sqe);
5907 case IORING_OP_RECVMSG:
5908 case IORING_OP_RECV:
5909 return io_recvmsg_prep(req, sqe);
5910 case IORING_OP_CONNECT:
5911 return io_connect_prep(req, sqe);
5912 case IORING_OP_TIMEOUT:
5913 return io_timeout_prep(req, sqe, false);
5914 case IORING_OP_TIMEOUT_REMOVE:
5915 return io_timeout_remove_prep(req, sqe);
5916 case IORING_OP_ASYNC_CANCEL:
5917 return io_async_cancel_prep(req, sqe);
5918 case IORING_OP_LINK_TIMEOUT:
5919 return io_timeout_prep(req, sqe, true);
5920 case IORING_OP_ACCEPT:
5921 return io_accept_prep(req, sqe);
5922 case IORING_OP_FALLOCATE:
5923 return io_fallocate_prep(req, sqe);
5924 case IORING_OP_OPENAT:
5925 return io_openat_prep(req, sqe);
5926 case IORING_OP_CLOSE:
5927 return io_close_prep(req, sqe);
5928 case IORING_OP_FILES_UPDATE:
5929 return io_rsrc_update_prep(req, sqe);
5930 case IORING_OP_STATX:
5931 return io_statx_prep(req, sqe);
5932 case IORING_OP_FADVISE:
5933 return io_fadvise_prep(req, sqe);
5934 case IORING_OP_MADVISE:
5935 return io_madvise_prep(req, sqe);
5936 case IORING_OP_OPENAT2:
5937 return io_openat2_prep(req, sqe);
5938 case IORING_OP_EPOLL_CTL:
5939 return io_epoll_ctl_prep(req, sqe);
5940 case IORING_OP_SPLICE:
5941 return io_splice_prep(req, sqe);
5942 case IORING_OP_PROVIDE_BUFFERS:
5943 return io_provide_buffers_prep(req, sqe);
5944 case IORING_OP_REMOVE_BUFFERS:
5945 return io_remove_buffers_prep(req, sqe);
5947 return io_tee_prep(req, sqe);
5948 case IORING_OP_SHUTDOWN:
5949 return io_shutdown_prep(req, sqe);
5950 case IORING_OP_RENAMEAT:
5951 return io_renameat_prep(req, sqe);
5952 case IORING_OP_UNLINKAT:
5953 return io_unlinkat_prep(req, sqe);
5956 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5961 static int io_req_prep_async(struct io_kiocb *req)
5963 if (!io_op_defs[req->opcode].needs_async_setup)
5965 if (WARN_ON_ONCE(req->async_data))
5967 if (io_alloc_async_data(req))
5970 switch (req->opcode) {
5971 case IORING_OP_READV:
5972 return io_rw_prep_async(req, READ);
5973 case IORING_OP_WRITEV:
5974 return io_rw_prep_async(req, WRITE);
5975 case IORING_OP_SENDMSG:
5976 return io_sendmsg_prep_async(req);
5977 case IORING_OP_RECVMSG:
5978 return io_recvmsg_prep_async(req);
5979 case IORING_OP_CONNECT:
5980 return io_connect_prep_async(req);
5982 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5987 static u32 io_get_sequence(struct io_kiocb *req)
5989 struct io_kiocb *pos;
5990 struct io_ring_ctx *ctx = req->ctx;
5991 u32 total_submitted, nr_reqs = 0;
5993 io_for_each_link(pos, req)
5996 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5997 return total_submitted - nr_reqs;
6000 static int io_req_defer(struct io_kiocb *req)
6002 struct io_ring_ctx *ctx = req->ctx;
6003 struct io_defer_entry *de;
6007 /* Still need defer if there is pending req in defer list. */
6008 if (likely(list_empty_careful(&ctx->defer_list) &&
6009 !(req->flags & REQ_F_IO_DRAIN)))
6012 seq = io_get_sequence(req);
6013 /* Still a chance to pass the sequence check */
6014 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6017 ret = io_req_prep_async(req);
6020 io_prep_async_link(req);
6021 de = kmalloc(sizeof(*de), GFP_KERNEL);
6025 spin_lock_irq(&ctx->completion_lock);
6026 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6027 spin_unlock_irq(&ctx->completion_lock);
6029 io_queue_async_work(req);
6030 return -EIOCBQUEUED;
6033 trace_io_uring_defer(ctx, req, req->user_data);
6036 list_add_tail(&de->list, &ctx->defer_list);
6037 spin_unlock_irq(&ctx->completion_lock);
6038 return -EIOCBQUEUED;
6041 static void io_clean_op(struct io_kiocb *req)
6043 if (req->flags & REQ_F_BUFFER_SELECTED) {
6044 switch (req->opcode) {
6045 case IORING_OP_READV:
6046 case IORING_OP_READ_FIXED:
6047 case IORING_OP_READ:
6048 kfree((void *)(unsigned long)req->rw.addr);
6050 case IORING_OP_RECVMSG:
6051 case IORING_OP_RECV:
6052 kfree(req->sr_msg.kbuf);
6055 req->flags &= ~REQ_F_BUFFER_SELECTED;
6058 if (req->flags & REQ_F_NEED_CLEANUP) {
6059 switch (req->opcode) {
6060 case IORING_OP_READV:
6061 case IORING_OP_READ_FIXED:
6062 case IORING_OP_READ:
6063 case IORING_OP_WRITEV:
6064 case IORING_OP_WRITE_FIXED:
6065 case IORING_OP_WRITE: {
6066 struct io_async_rw *io = req->async_data;
6068 kfree(io->free_iovec);
6071 case IORING_OP_RECVMSG:
6072 case IORING_OP_SENDMSG: {
6073 struct io_async_msghdr *io = req->async_data;
6075 kfree(io->free_iov);
6078 case IORING_OP_SPLICE:
6080 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6081 io_put_file(req->splice.file_in);
6083 case IORING_OP_OPENAT:
6084 case IORING_OP_OPENAT2:
6085 if (req->open.filename)
6086 putname(req->open.filename);
6088 case IORING_OP_RENAMEAT:
6089 putname(req->rename.oldpath);
6090 putname(req->rename.newpath);
6092 case IORING_OP_UNLINKAT:
6093 putname(req->unlink.filename);
6096 req->flags &= ~REQ_F_NEED_CLEANUP;
6098 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6099 kfree(req->apoll->double_poll);
6103 if (req->flags & REQ_F_INFLIGHT) {
6104 struct io_uring_task *tctx = req->task->io_uring;
6106 atomic_dec(&tctx->inflight_tracked);
6107 req->flags &= ~REQ_F_INFLIGHT;
6111 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6113 struct io_ring_ctx *ctx = req->ctx;
6114 const struct cred *creds = NULL;
6117 if (req->work.creds && req->work.creds != current_cred())
6118 creds = override_creds(req->work.creds);
6120 switch (req->opcode) {
6122 ret = io_nop(req, issue_flags);
6124 case IORING_OP_READV:
6125 case IORING_OP_READ_FIXED:
6126 case IORING_OP_READ:
6127 ret = io_read(req, issue_flags);
6129 case IORING_OP_WRITEV:
6130 case IORING_OP_WRITE_FIXED:
6131 case IORING_OP_WRITE:
6132 ret = io_write(req, issue_flags);
6134 case IORING_OP_FSYNC:
6135 ret = io_fsync(req, issue_flags);
6137 case IORING_OP_POLL_ADD:
6138 ret = io_poll_add(req, issue_flags);
6140 case IORING_OP_POLL_REMOVE:
6141 ret = io_poll_update(req, issue_flags);
6143 case IORING_OP_SYNC_FILE_RANGE:
6144 ret = io_sync_file_range(req, issue_flags);
6146 case IORING_OP_SENDMSG:
6147 ret = io_sendmsg(req, issue_flags);
6149 case IORING_OP_SEND:
6150 ret = io_send(req, issue_flags);
6152 case IORING_OP_RECVMSG:
6153 ret = io_recvmsg(req, issue_flags);
6155 case IORING_OP_RECV:
6156 ret = io_recv(req, issue_flags);
6158 case IORING_OP_TIMEOUT:
6159 ret = io_timeout(req, issue_flags);
6161 case IORING_OP_TIMEOUT_REMOVE:
6162 ret = io_timeout_remove(req, issue_flags);
6164 case IORING_OP_ACCEPT:
6165 ret = io_accept(req, issue_flags);
6167 case IORING_OP_CONNECT:
6168 ret = io_connect(req, issue_flags);
6170 case IORING_OP_ASYNC_CANCEL:
6171 ret = io_async_cancel(req, issue_flags);
6173 case IORING_OP_FALLOCATE:
6174 ret = io_fallocate(req, issue_flags);
6176 case IORING_OP_OPENAT:
6177 ret = io_openat(req, issue_flags);
6179 case IORING_OP_CLOSE:
6180 ret = io_close(req, issue_flags);
6182 case IORING_OP_FILES_UPDATE:
6183 ret = io_files_update(req, issue_flags);
6185 case IORING_OP_STATX:
6186 ret = io_statx(req, issue_flags);
6188 case IORING_OP_FADVISE:
6189 ret = io_fadvise(req, issue_flags);
6191 case IORING_OP_MADVISE:
6192 ret = io_madvise(req, issue_flags);
6194 case IORING_OP_OPENAT2:
6195 ret = io_openat2(req, issue_flags);
6197 case IORING_OP_EPOLL_CTL:
6198 ret = io_epoll_ctl(req, issue_flags);
6200 case IORING_OP_SPLICE:
6201 ret = io_splice(req, issue_flags);
6203 case IORING_OP_PROVIDE_BUFFERS:
6204 ret = io_provide_buffers(req, issue_flags);
6206 case IORING_OP_REMOVE_BUFFERS:
6207 ret = io_remove_buffers(req, issue_flags);
6210 ret = io_tee(req, issue_flags);
6212 case IORING_OP_SHUTDOWN:
6213 ret = io_shutdown(req, issue_flags);
6215 case IORING_OP_RENAMEAT:
6216 ret = io_renameat(req, issue_flags);
6218 case IORING_OP_UNLINKAT:
6219 ret = io_unlinkat(req, issue_flags);
6227 revert_creds(creds);
6230 /* If the op doesn't have a file, we're not polling for it */
6231 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6232 io_iopoll_req_issued(req);
6237 static void io_wq_submit_work(struct io_wq_work *work)
6239 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6240 struct io_kiocb *timeout;
6243 timeout = io_prep_linked_timeout(req);
6245 io_queue_linked_timeout(timeout);
6247 if (work->flags & IO_WQ_WORK_CANCEL)
6252 ret = io_issue_sqe(req, 0);
6254 * We can get EAGAIN for polled IO even though we're
6255 * forcing a sync submission from here, since we can't
6256 * wait for request slots on the block side.
6264 /* avoid locking problems by failing it from a clean context */
6266 /* io-wq is going to take one down */
6268 io_req_task_queue_fail(req, ret);
6272 #define FFS_ASYNC_READ 0x1UL
6273 #define FFS_ASYNC_WRITE 0x2UL
6275 #define FFS_ISREG 0x4UL
6277 #define FFS_ISREG 0x0UL
6279 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6281 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6284 struct io_fixed_file *table_l2;
6286 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6287 return &table_l2[i & IORING_FILE_TABLE_MASK];
6290 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6293 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6295 return (struct file *) (slot->file_ptr & FFS_MASK);
6298 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6300 unsigned long file_ptr = (unsigned long) file;
6302 if (__io_file_supports_async(file, READ))
6303 file_ptr |= FFS_ASYNC_READ;
6304 if (__io_file_supports_async(file, WRITE))
6305 file_ptr |= FFS_ASYNC_WRITE;
6306 if (S_ISREG(file_inode(file)->i_mode))
6307 file_ptr |= FFS_ISREG;
6308 file_slot->file_ptr = file_ptr;
6311 static struct file *io_file_get(struct io_submit_state *state,
6312 struct io_kiocb *req, int fd, bool fixed)
6314 struct io_ring_ctx *ctx = req->ctx;
6318 unsigned long file_ptr;
6320 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6322 fd = array_index_nospec(fd, ctx->nr_user_files);
6323 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6324 file = (struct file *) (file_ptr & FFS_MASK);
6325 file_ptr &= ~FFS_MASK;
6326 /* mask in overlapping REQ_F and FFS bits */
6327 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6328 io_req_set_rsrc_node(req);
6330 trace_io_uring_file_get(ctx, fd);
6331 file = __io_file_get(state, fd);
6333 /* we don't allow fixed io_uring files */
6334 if (file && unlikely(file->f_op == &io_uring_fops))
6335 io_req_track_inflight(req);
6341 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6343 struct io_timeout_data *data = container_of(timer,
6344 struct io_timeout_data, timer);
6345 struct io_kiocb *prev, *req = data->req;
6346 struct io_ring_ctx *ctx = req->ctx;
6347 unsigned long flags;
6349 spin_lock_irqsave(&ctx->completion_lock, flags);
6350 prev = req->timeout.head;
6351 req->timeout.head = NULL;
6354 * We don't expect the list to be empty, that will only happen if we
6355 * race with the completion of the linked work.
6358 io_remove_next_linked(prev);
6359 if (!req_ref_inc_not_zero(prev))
6362 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6365 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6366 io_put_req_deferred(prev, 1);
6367 io_put_req_deferred(req, 1);
6369 io_req_complete_post(req, -ETIME, 0);
6371 return HRTIMER_NORESTART;
6374 static void io_queue_linked_timeout(struct io_kiocb *req)
6376 struct io_ring_ctx *ctx = req->ctx;
6378 spin_lock_irq(&ctx->completion_lock);
6380 * If the back reference is NULL, then our linked request finished
6381 * before we got a chance to setup the timer
6383 if (req->timeout.head) {
6384 struct io_timeout_data *data = req->async_data;
6386 data->timer.function = io_link_timeout_fn;
6387 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6390 spin_unlock_irq(&ctx->completion_lock);
6391 /* drop submission reference */
6395 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6397 struct io_kiocb *nxt = req->link;
6399 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6400 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6403 nxt->timeout.head = req;
6404 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6405 req->flags |= REQ_F_LINK_TIMEOUT;
6409 static void __io_queue_sqe(struct io_kiocb *req)
6411 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6414 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6417 * We async punt it if the file wasn't marked NOWAIT, or if the file
6418 * doesn't support non-blocking read/write attempts
6421 /* drop submission reference */
6422 if (req->flags & REQ_F_COMPLETE_INLINE) {
6423 struct io_ring_ctx *ctx = req->ctx;
6424 struct io_comp_state *cs = &ctx->submit_state.comp;
6426 cs->reqs[cs->nr++] = req;
6427 if (cs->nr == ARRAY_SIZE(cs->reqs))
6428 io_submit_flush_completions(cs, ctx);
6432 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6433 if (!io_arm_poll_handler(req)) {
6435 * Queued up for async execution, worker will release
6436 * submit reference when the iocb is actually submitted.
6438 io_queue_async_work(req);
6441 io_req_complete_failed(req, ret);
6444 io_queue_linked_timeout(linked_timeout);
6447 static void io_queue_sqe(struct io_kiocb *req)
6451 ret = io_req_defer(req);
6453 if (ret != -EIOCBQUEUED) {
6455 io_req_complete_failed(req, ret);
6457 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6458 ret = io_req_prep_async(req);
6461 io_queue_async_work(req);
6463 __io_queue_sqe(req);
6468 * Check SQE restrictions (opcode and flags).
6470 * Returns 'true' if SQE is allowed, 'false' otherwise.
6472 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6473 struct io_kiocb *req,
6474 unsigned int sqe_flags)
6476 if (!ctx->restricted)
6479 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6482 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6483 ctx->restrictions.sqe_flags_required)
6486 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6487 ctx->restrictions.sqe_flags_required))
6493 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6494 const struct io_uring_sqe *sqe)
6496 struct io_submit_state *state;
6497 unsigned int sqe_flags;
6498 int personality, ret = 0;
6500 req->opcode = READ_ONCE(sqe->opcode);
6501 /* same numerical values with corresponding REQ_F_*, safe to copy */
6502 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6503 req->user_data = READ_ONCE(sqe->user_data);
6504 req->async_data = NULL;
6508 req->fixed_rsrc_refs = NULL;
6509 /* one is dropped after submission, the other at completion */
6510 atomic_set(&req->refs, 2);
6511 req->task = current;
6513 req->work.creds = NULL;
6515 /* enforce forwards compatibility on users */
6516 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6518 if (unlikely(req->opcode >= IORING_OP_LAST))
6520 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6523 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6524 !io_op_defs[req->opcode].buffer_select)
6527 personality = READ_ONCE(sqe->personality);
6529 req->work.creds = xa_load(&ctx->personalities, personality);
6530 if (!req->work.creds)
6532 get_cred(req->work.creds);
6534 state = &ctx->submit_state;
6537 * Plug now if we have more than 1 IO left after this, and the target
6538 * is potentially a read/write to block based storage.
6540 if (!state->plug_started && state->ios_left > 1 &&
6541 io_op_defs[req->opcode].plug) {
6542 blk_start_plug(&state->plug);
6543 state->plug_started = true;
6546 if (io_op_defs[req->opcode].needs_file) {
6547 bool fixed = req->flags & REQ_F_FIXED_FILE;
6549 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6550 if (unlikely(!req->file))
6558 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6559 const struct io_uring_sqe *sqe)
6561 struct io_submit_link *link = &ctx->submit_state.link;
6564 ret = io_init_req(ctx, req, sqe);
6565 if (unlikely(ret)) {
6568 /* fail even hard links since we don't submit */
6569 req_set_fail(link->head);
6570 io_req_complete_failed(link->head, -ECANCELED);
6573 io_req_complete_failed(req, ret);
6576 ret = io_req_prep(req, sqe);
6580 /* don't need @sqe from now on */
6581 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6582 true, ctx->flags & IORING_SETUP_SQPOLL);
6585 * If we already have a head request, queue this one for async
6586 * submittal once the head completes. If we don't have a head but
6587 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6588 * submitted sync once the chain is complete. If none of those
6589 * conditions are true (normal request), then just queue it.
6592 struct io_kiocb *head = link->head;
6595 * Taking sequential execution of a link, draining both sides
6596 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6597 * requests in the link. So, it drains the head and the
6598 * next after the link request. The last one is done via
6599 * drain_next flag to persist the effect across calls.
6601 if (req->flags & REQ_F_IO_DRAIN) {
6602 head->flags |= REQ_F_IO_DRAIN;
6603 ctx->drain_next = 1;
6605 ret = io_req_prep_async(req);
6608 trace_io_uring_link(ctx, req, head);
6609 link->last->link = req;
6612 /* last request of a link, enqueue the link */
6613 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6618 if (unlikely(ctx->drain_next)) {
6619 req->flags |= REQ_F_IO_DRAIN;
6620 ctx->drain_next = 0;
6622 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6634 * Batched submission is done, ensure local IO is flushed out.
6636 static void io_submit_state_end(struct io_submit_state *state,
6637 struct io_ring_ctx *ctx)
6639 if (state->link.head)
6640 io_queue_sqe(state->link.head);
6642 io_submit_flush_completions(&state->comp, ctx);
6643 if (state->plug_started)
6644 blk_finish_plug(&state->plug);
6645 io_state_file_put(state);
6649 * Start submission side cache.
6651 static void io_submit_state_start(struct io_submit_state *state,
6652 unsigned int max_ios)
6654 state->plug_started = false;
6655 state->ios_left = max_ios;
6656 /* set only head, no need to init link_last in advance */
6657 state->link.head = NULL;
6660 static void io_commit_sqring(struct io_ring_ctx *ctx)
6662 struct io_rings *rings = ctx->rings;
6665 * Ensure any loads from the SQEs are done at this point,
6666 * since once we write the new head, the application could
6667 * write new data to them.
6669 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6673 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6674 * that is mapped by userspace. This means that care needs to be taken to
6675 * ensure that reads are stable, as we cannot rely on userspace always
6676 * being a good citizen. If members of the sqe are validated and then later
6677 * used, it's important that those reads are done through READ_ONCE() to
6678 * prevent a re-load down the line.
6680 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6682 u32 *sq_array = ctx->sq_array;
6683 unsigned head, mask = ctx->sq_entries - 1;
6686 * The cached sq head (or cq tail) serves two purposes:
6688 * 1) allows us to batch the cost of updating the user visible
6690 * 2) allows the kernel side to track the head on its own, even
6691 * though the application is the one updating it.
6693 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & mask]);
6694 if (likely(head < ctx->sq_entries))
6695 return &ctx->sq_sqes[head];
6697 /* drop invalid entries */
6698 ctx->cached_sq_dropped++;
6699 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6703 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6707 /* make sure SQ entry isn't read before tail */
6708 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6710 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6713 percpu_counter_add(¤t->io_uring->inflight, nr);
6714 refcount_add(nr, ¤t->usage);
6715 io_submit_state_start(&ctx->submit_state, nr);
6717 while (submitted < nr) {
6718 const struct io_uring_sqe *sqe;
6719 struct io_kiocb *req;
6721 req = io_alloc_req(ctx);
6722 if (unlikely(!req)) {
6724 submitted = -EAGAIN;
6727 sqe = io_get_sqe(ctx);
6728 if (unlikely(!sqe)) {
6729 kmem_cache_free(req_cachep, req);
6732 /* will complete beyond this point, count as submitted */
6734 if (io_submit_sqe(ctx, req, sqe))
6738 if (unlikely(submitted != nr)) {
6739 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6740 struct io_uring_task *tctx = current->io_uring;
6741 int unused = nr - ref_used;
6743 percpu_ref_put_many(&ctx->refs, unused);
6744 percpu_counter_sub(&tctx->inflight, unused);
6745 put_task_struct_many(current, unused);
6748 io_submit_state_end(&ctx->submit_state, ctx);
6749 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6750 io_commit_sqring(ctx);
6755 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6757 return READ_ONCE(sqd->state);
6760 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6762 /* Tell userspace we may need a wakeup call */
6763 spin_lock_irq(&ctx->completion_lock);
6764 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6765 spin_unlock_irq(&ctx->completion_lock);
6768 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6770 spin_lock_irq(&ctx->completion_lock);
6771 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6772 spin_unlock_irq(&ctx->completion_lock);
6775 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6777 unsigned int to_submit;
6780 to_submit = io_sqring_entries(ctx);
6781 /* if we're handling multiple rings, cap submit size for fairness */
6782 if (cap_entries && to_submit > 8)
6785 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6786 unsigned nr_events = 0;
6788 mutex_lock(&ctx->uring_lock);
6789 if (!list_empty(&ctx->iopoll_list))
6790 io_do_iopoll(ctx, &nr_events, 0);
6793 * Don't submit if refs are dying, good for io_uring_register(),
6794 * but also it is relied upon by io_ring_exit_work()
6796 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6797 !(ctx->flags & IORING_SETUP_R_DISABLED))
6798 ret = io_submit_sqes(ctx, to_submit);
6799 mutex_unlock(&ctx->uring_lock);
6801 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6802 wake_up(&ctx->sqo_sq_wait);
6808 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6810 struct io_ring_ctx *ctx;
6811 unsigned sq_thread_idle = 0;
6813 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6814 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6815 sqd->sq_thread_idle = sq_thread_idle;
6818 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6820 bool did_sig = false;
6821 struct ksignal ksig;
6823 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6824 signal_pending(current)) {
6825 mutex_unlock(&sqd->lock);
6826 if (signal_pending(current))
6827 did_sig = get_signal(&ksig);
6829 mutex_lock(&sqd->lock);
6832 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6835 static int io_sq_thread(void *data)
6837 struct io_sq_data *sqd = data;
6838 struct io_ring_ctx *ctx;
6839 unsigned long timeout = 0;
6840 char buf[TASK_COMM_LEN];
6843 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6844 set_task_comm(current, buf);
6846 if (sqd->sq_cpu != -1)
6847 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6849 set_cpus_allowed_ptr(current, cpu_online_mask);
6850 current->flags |= PF_NO_SETAFFINITY;
6852 mutex_lock(&sqd->lock);
6855 bool cap_entries, sqt_spin, needs_sched;
6857 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6858 if (io_sqd_handle_event(sqd))
6860 timeout = jiffies + sqd->sq_thread_idle;
6865 cap_entries = !list_is_singular(&sqd->ctx_list);
6866 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6867 const struct cred *creds = NULL;
6869 if (ctx->sq_creds != current_cred())
6870 creds = override_creds(ctx->sq_creds);
6871 ret = __io_sq_thread(ctx, cap_entries);
6873 revert_creds(creds);
6874 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6878 if (sqt_spin || !time_after(jiffies, timeout)) {
6882 timeout = jiffies + sqd->sq_thread_idle;
6886 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6887 if (!io_sqd_events_pending(sqd)) {
6889 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6890 io_ring_set_wakeup_flag(ctx);
6892 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6893 !list_empty_careful(&ctx->iopoll_list)) {
6894 needs_sched = false;
6897 if (io_sqring_entries(ctx)) {
6898 needs_sched = false;
6904 mutex_unlock(&sqd->lock);
6906 mutex_lock(&sqd->lock);
6908 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6909 io_ring_clear_wakeup_flag(ctx);
6912 finish_wait(&sqd->wait, &wait);
6913 timeout = jiffies + sqd->sq_thread_idle;
6916 io_uring_cancel_sqpoll(sqd);
6918 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6919 io_ring_set_wakeup_flag(ctx);
6921 mutex_unlock(&sqd->lock);
6923 complete(&sqd->exited);
6927 struct io_wait_queue {
6928 struct wait_queue_entry wq;
6929 struct io_ring_ctx *ctx;
6931 unsigned nr_timeouts;
6934 static inline bool io_should_wake(struct io_wait_queue *iowq)
6936 struct io_ring_ctx *ctx = iowq->ctx;
6939 * Wake up if we have enough events, or if a timeout occurred since we
6940 * started waiting. For timeouts, we always want to return to userspace,
6941 * regardless of event count.
6943 return io_cqring_events(ctx) >= iowq->to_wait ||
6944 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6947 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6948 int wake_flags, void *key)
6950 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6954 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6955 * the task, and the next invocation will do it.
6957 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6958 return autoremove_wake_function(curr, mode, wake_flags, key);
6962 static int io_run_task_work_sig(void)
6964 if (io_run_task_work())
6966 if (!signal_pending(current))
6968 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6969 return -ERESTARTSYS;
6973 /* when returns >0, the caller should retry */
6974 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6975 struct io_wait_queue *iowq,
6976 signed long *timeout)
6980 /* make sure we run task_work before checking for signals */
6981 ret = io_run_task_work_sig();
6982 if (ret || io_should_wake(iowq))
6984 /* let the caller flush overflows, retry */
6985 if (test_bit(0, &ctx->cq_check_overflow))
6988 *timeout = schedule_timeout(*timeout);
6989 return !*timeout ? -ETIME : 1;
6993 * Wait until events become available, if we don't already have some. The
6994 * application must reap them itself, as they reside on the shared cq ring.
6996 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6997 const sigset_t __user *sig, size_t sigsz,
6998 struct __kernel_timespec __user *uts)
7000 struct io_wait_queue iowq = {
7003 .func = io_wake_function,
7004 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7007 .to_wait = min_events,
7009 struct io_rings *rings = ctx->rings;
7010 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7014 io_cqring_overflow_flush(ctx, false);
7015 if (io_cqring_events(ctx) >= min_events)
7017 if (!io_run_task_work())
7022 #ifdef CONFIG_COMPAT
7023 if (in_compat_syscall())
7024 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7028 ret = set_user_sigmask(sig, sigsz);
7035 struct timespec64 ts;
7037 if (get_timespec64(&ts, uts))
7039 timeout = timespec64_to_jiffies(&ts);
7042 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7043 trace_io_uring_cqring_wait(ctx, min_events);
7045 /* if we can't even flush overflow, don't wait for more */
7046 if (!io_cqring_overflow_flush(ctx, false)) {
7050 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7051 TASK_INTERRUPTIBLE);
7052 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7053 finish_wait(&ctx->wait, &iowq.wq);
7057 restore_saved_sigmask_unless(ret == -EINTR);
7059 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7062 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7064 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7066 for (i = 0; i < nr_tables; i++)
7067 kfree(table->files[i]);
7068 kfree(table->files);
7069 table->files = NULL;
7072 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7074 spin_lock_bh(&ctx->rsrc_ref_lock);
7077 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7079 spin_unlock_bh(&ctx->rsrc_ref_lock);
7082 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7084 percpu_ref_exit(&ref_node->refs);
7088 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7089 struct io_rsrc_data *data_to_kill)
7091 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7092 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7095 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7097 rsrc_node->rsrc_data = data_to_kill;
7098 io_rsrc_ref_lock(ctx);
7099 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7100 io_rsrc_ref_unlock(ctx);
7102 atomic_inc(&data_to_kill->refs);
7103 percpu_ref_kill(&rsrc_node->refs);
7104 ctx->rsrc_node = NULL;
7107 if (!ctx->rsrc_node) {
7108 ctx->rsrc_node = ctx->rsrc_backup_node;
7109 ctx->rsrc_backup_node = NULL;
7113 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7115 if (ctx->rsrc_backup_node)
7117 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7118 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7121 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7125 /* As we may drop ->uring_lock, other task may have started quiesce */
7129 data->quiesce = true;
7131 ret = io_rsrc_node_switch_start(ctx);
7134 io_rsrc_node_switch(ctx, data);
7136 /* kill initial ref, already quiesced if zero */
7137 if (atomic_dec_and_test(&data->refs))
7139 flush_delayed_work(&ctx->rsrc_put_work);
7140 ret = wait_for_completion_interruptible(&data->done);
7144 atomic_inc(&data->refs);
7145 /* wait for all works potentially completing data->done */
7146 flush_delayed_work(&ctx->rsrc_put_work);
7147 reinit_completion(&data->done);
7149 mutex_unlock(&ctx->uring_lock);
7150 ret = io_run_task_work_sig();
7151 mutex_lock(&ctx->uring_lock);
7153 data->quiesce = false;
7158 static void io_rsrc_data_free(struct io_rsrc_data *data)
7164 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7165 rsrc_put_fn *do_put,
7168 struct io_rsrc_data *data;
7170 data = kzalloc(sizeof(*data), GFP_KERNEL);
7174 data->tags = kvcalloc(nr, sizeof(*data->tags), GFP_KERNEL);
7180 atomic_set(&data->refs, 1);
7182 data->do_put = do_put;
7183 init_completion(&data->done);
7187 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7189 #if defined(CONFIG_UNIX)
7190 if (ctx->ring_sock) {
7191 struct sock *sock = ctx->ring_sock->sk;
7192 struct sk_buff *skb;
7194 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7200 for (i = 0; i < ctx->nr_user_files; i++) {
7203 file = io_file_from_index(ctx, i);
7208 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7209 io_rsrc_data_free(ctx->file_data);
7210 ctx->file_data = NULL;
7211 ctx->nr_user_files = 0;
7214 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7218 if (!ctx->file_data)
7220 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7222 __io_sqe_files_unregister(ctx);
7226 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7227 __releases(&sqd->lock)
7229 WARN_ON_ONCE(sqd->thread == current);
7232 * Do the dance but not conditional clear_bit() because it'd race with
7233 * other threads incrementing park_pending and setting the bit.
7235 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7236 if (atomic_dec_return(&sqd->park_pending))
7237 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7238 mutex_unlock(&sqd->lock);
7241 static void io_sq_thread_park(struct io_sq_data *sqd)
7242 __acquires(&sqd->lock)
7244 WARN_ON_ONCE(sqd->thread == current);
7246 atomic_inc(&sqd->park_pending);
7247 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7248 mutex_lock(&sqd->lock);
7250 wake_up_process(sqd->thread);
7253 static void io_sq_thread_stop(struct io_sq_data *sqd)
7255 WARN_ON_ONCE(sqd->thread == current);
7256 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7258 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7259 mutex_lock(&sqd->lock);
7261 wake_up_process(sqd->thread);
7262 mutex_unlock(&sqd->lock);
7263 wait_for_completion(&sqd->exited);
7266 static void io_put_sq_data(struct io_sq_data *sqd)
7268 if (refcount_dec_and_test(&sqd->refs)) {
7269 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7271 io_sq_thread_stop(sqd);
7276 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7278 struct io_sq_data *sqd = ctx->sq_data;
7281 io_sq_thread_park(sqd);
7282 list_del_init(&ctx->sqd_list);
7283 io_sqd_update_thread_idle(sqd);
7284 io_sq_thread_unpark(sqd);
7286 io_put_sq_data(sqd);
7287 ctx->sq_data = NULL;
7291 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7293 struct io_ring_ctx *ctx_attach;
7294 struct io_sq_data *sqd;
7297 f = fdget(p->wq_fd);
7299 return ERR_PTR(-ENXIO);
7300 if (f.file->f_op != &io_uring_fops) {
7302 return ERR_PTR(-EINVAL);
7305 ctx_attach = f.file->private_data;
7306 sqd = ctx_attach->sq_data;
7309 return ERR_PTR(-EINVAL);
7311 if (sqd->task_tgid != current->tgid) {
7313 return ERR_PTR(-EPERM);
7316 refcount_inc(&sqd->refs);
7321 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7324 struct io_sq_data *sqd;
7327 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7328 sqd = io_attach_sq_data(p);
7333 /* fall through for EPERM case, setup new sqd/task */
7334 if (PTR_ERR(sqd) != -EPERM)
7338 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7340 return ERR_PTR(-ENOMEM);
7342 atomic_set(&sqd->park_pending, 0);
7343 refcount_set(&sqd->refs, 1);
7344 INIT_LIST_HEAD(&sqd->ctx_list);
7345 mutex_init(&sqd->lock);
7346 init_waitqueue_head(&sqd->wait);
7347 init_completion(&sqd->exited);
7351 #if defined(CONFIG_UNIX)
7353 * Ensure the UNIX gc is aware of our file set, so we are certain that
7354 * the io_uring can be safely unregistered on process exit, even if we have
7355 * loops in the file referencing.
7357 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7359 struct sock *sk = ctx->ring_sock->sk;
7360 struct scm_fp_list *fpl;
7361 struct sk_buff *skb;
7364 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7368 skb = alloc_skb(0, GFP_KERNEL);
7377 fpl->user = get_uid(current_user());
7378 for (i = 0; i < nr; i++) {
7379 struct file *file = io_file_from_index(ctx, i + offset);
7383 fpl->fp[nr_files] = get_file(file);
7384 unix_inflight(fpl->user, fpl->fp[nr_files]);
7389 fpl->max = SCM_MAX_FD;
7390 fpl->count = nr_files;
7391 UNIXCB(skb).fp = fpl;
7392 skb->destructor = unix_destruct_scm;
7393 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7394 skb_queue_head(&sk->sk_receive_queue, skb);
7396 for (i = 0; i < nr_files; i++)
7407 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7408 * causes regular reference counting to break down. We rely on the UNIX
7409 * garbage collection to take care of this problem for us.
7411 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7413 unsigned left, total;
7417 left = ctx->nr_user_files;
7419 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7421 ret = __io_sqe_files_scm(ctx, this_files, total);
7425 total += this_files;
7431 while (total < ctx->nr_user_files) {
7432 struct file *file = io_file_from_index(ctx, total);
7442 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7448 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7450 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7452 table->files = kcalloc(nr_tables, sizeof(*table->files), GFP_KERNEL);
7456 for (i = 0; i < nr_tables; i++) {
7457 unsigned int this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7459 table->files[i] = kcalloc(this_files, sizeof(*table->files[i]),
7461 if (!table->files[i])
7463 nr_files -= this_files;
7469 io_free_file_tables(table, nr_tables * IORING_MAX_FILES_TABLE);
7473 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7475 struct file *file = prsrc->file;
7476 #if defined(CONFIG_UNIX)
7477 struct sock *sock = ctx->ring_sock->sk;
7478 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7479 struct sk_buff *skb;
7482 __skb_queue_head_init(&list);
7485 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7486 * remove this entry and rearrange the file array.
7488 skb = skb_dequeue(head);
7490 struct scm_fp_list *fp;
7492 fp = UNIXCB(skb).fp;
7493 for (i = 0; i < fp->count; i++) {
7496 if (fp->fp[i] != file)
7499 unix_notinflight(fp->user, fp->fp[i]);
7500 left = fp->count - 1 - i;
7502 memmove(&fp->fp[i], &fp->fp[i + 1],
7503 left * sizeof(struct file *));
7510 __skb_queue_tail(&list, skb);
7520 __skb_queue_tail(&list, skb);
7522 skb = skb_dequeue(head);
7525 if (skb_peek(&list)) {
7526 spin_lock_irq(&head->lock);
7527 while ((skb = __skb_dequeue(&list)) != NULL)
7528 __skb_queue_tail(head, skb);
7529 spin_unlock_irq(&head->lock);
7536 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7538 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7539 struct io_ring_ctx *ctx = rsrc_data->ctx;
7540 struct io_rsrc_put *prsrc, *tmp;
7542 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7543 list_del(&prsrc->list);
7546 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7547 unsigned long flags;
7549 io_ring_submit_lock(ctx, lock_ring);
7550 spin_lock_irqsave(&ctx->completion_lock, flags);
7551 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7553 io_commit_cqring(ctx);
7554 spin_unlock_irqrestore(&ctx->completion_lock, flags);
7555 io_cqring_ev_posted(ctx);
7556 io_ring_submit_unlock(ctx, lock_ring);
7559 rsrc_data->do_put(ctx, prsrc);
7563 io_rsrc_node_destroy(ref_node);
7564 if (atomic_dec_and_test(&rsrc_data->refs))
7565 complete(&rsrc_data->done);
7568 static void io_rsrc_put_work(struct work_struct *work)
7570 struct io_ring_ctx *ctx;
7571 struct llist_node *node;
7573 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7574 node = llist_del_all(&ctx->rsrc_put_llist);
7577 struct io_rsrc_node *ref_node;
7578 struct llist_node *next = node->next;
7580 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7581 __io_rsrc_put_work(ref_node);
7586 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7588 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7589 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7590 bool first_add = false;
7592 io_rsrc_ref_lock(ctx);
7595 while (!list_empty(&ctx->rsrc_ref_list)) {
7596 node = list_first_entry(&ctx->rsrc_ref_list,
7597 struct io_rsrc_node, node);
7598 /* recycle ref nodes in order */
7601 list_del(&node->node);
7602 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7604 io_rsrc_ref_unlock(ctx);
7607 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7610 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7612 struct io_rsrc_node *ref_node;
7614 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7618 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7623 INIT_LIST_HEAD(&ref_node->node);
7624 INIT_LIST_HEAD(&ref_node->rsrc_list);
7625 ref_node->done = false;
7629 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7630 unsigned nr_args, u64 __user *tags)
7632 __s32 __user *fds = (__s32 __user *) arg;
7636 struct io_rsrc_data *file_data;
7642 if (nr_args > IORING_MAX_FIXED_FILES)
7644 ret = io_rsrc_node_switch_start(ctx);
7648 file_data = io_rsrc_data_alloc(ctx, io_rsrc_file_put, nr_args);
7651 ctx->file_data = file_data;
7653 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7656 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7659 if ((tags && copy_from_user(&tag, &tags[i], sizeof(tag))) ||
7660 copy_from_user(&fd, &fds[i], sizeof(fd))) {
7664 /* allow sparse sets */
7674 if (unlikely(!file))
7678 * Don't allow io_uring instances to be registered. If UNIX
7679 * isn't enabled, then this causes a reference cycle and this
7680 * instance can never get freed. If UNIX is enabled we'll
7681 * handle it just fine, but there's still no point in allowing
7682 * a ring fd as it doesn't support regular read/write anyway.
7684 if (file->f_op == &io_uring_fops) {
7688 ctx->file_data->tags[i] = tag;
7689 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7692 ret = io_sqe_files_scm(ctx);
7694 __io_sqe_files_unregister(ctx);
7698 io_rsrc_node_switch(ctx, NULL);
7701 for (i = 0; i < ctx->nr_user_files; i++) {
7702 file = io_file_from_index(ctx, i);
7706 io_free_file_tables(&ctx->file_table, nr_args);
7707 ctx->nr_user_files = 0;
7709 io_rsrc_data_free(ctx->file_data);
7710 ctx->file_data = NULL;
7714 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7717 #if defined(CONFIG_UNIX)
7718 struct sock *sock = ctx->ring_sock->sk;
7719 struct sk_buff_head *head = &sock->sk_receive_queue;
7720 struct sk_buff *skb;
7723 * See if we can merge this file into an existing skb SCM_RIGHTS
7724 * file set. If there's no room, fall back to allocating a new skb
7725 * and filling it in.
7727 spin_lock_irq(&head->lock);
7728 skb = skb_peek(head);
7730 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7732 if (fpl->count < SCM_MAX_FD) {
7733 __skb_unlink(skb, head);
7734 spin_unlock_irq(&head->lock);
7735 fpl->fp[fpl->count] = get_file(file);
7736 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7738 spin_lock_irq(&head->lock);
7739 __skb_queue_head(head, skb);
7744 spin_unlock_irq(&head->lock);
7751 return __io_sqe_files_scm(ctx, 1, index);
7757 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7758 struct io_rsrc_node *node, void *rsrc)
7760 struct io_rsrc_put *prsrc;
7762 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7766 prsrc->tag = data->tags[idx];
7768 list_add(&prsrc->list, &node->rsrc_list);
7772 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7773 struct io_uring_rsrc_update2 *up,
7776 u64 __user *tags = u64_to_user_ptr(up->tags);
7777 __s32 __user *fds = u64_to_user_ptr(up->data);
7778 struct io_rsrc_data *data = ctx->file_data;
7779 struct io_fixed_file *file_slot;
7783 bool needs_switch = false;
7785 if (!ctx->file_data)
7787 if (up->offset + nr_args > ctx->nr_user_files)
7790 for (done = 0; done < nr_args; done++) {
7793 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7794 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7798 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7802 if (fd == IORING_REGISTER_FILES_SKIP)
7805 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7806 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7808 if (file_slot->file_ptr) {
7809 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7810 err = io_queue_rsrc_removal(data, up->offset + done,
7811 ctx->rsrc_node, file);
7814 file_slot->file_ptr = 0;
7815 needs_switch = true;
7824 * Don't allow io_uring instances to be registered. If
7825 * UNIX isn't enabled, then this causes a reference
7826 * cycle and this instance can never get freed. If UNIX
7827 * is enabled we'll handle it just fine, but there's
7828 * still no point in allowing a ring fd as it doesn't
7829 * support regular read/write anyway.
7831 if (file->f_op == &io_uring_fops) {
7836 data->tags[up->offset + done] = tag;
7837 io_fixed_file_set(file_slot, file);
7838 err = io_sqe_file_register(ctx, file, i);
7840 file_slot->file_ptr = 0;
7848 io_rsrc_node_switch(ctx, data);
7849 return done ? done : err;
7852 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7854 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7856 req = io_put_req_find_next(req);
7857 return req ? &req->work : NULL;
7860 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7861 struct task_struct *task)
7863 struct io_wq_hash *hash;
7864 struct io_wq_data data;
7865 unsigned int concurrency;
7867 hash = ctx->hash_map;
7869 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7871 return ERR_PTR(-ENOMEM);
7872 refcount_set(&hash->refs, 1);
7873 init_waitqueue_head(&hash->wait);
7874 ctx->hash_map = hash;
7879 data.free_work = io_free_work;
7880 data.do_work = io_wq_submit_work;
7882 /* Do QD, or 4 * CPUS, whatever is smallest */
7883 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7885 return io_wq_create(concurrency, &data);
7888 static int io_uring_alloc_task_context(struct task_struct *task,
7889 struct io_ring_ctx *ctx)
7891 struct io_uring_task *tctx;
7894 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7895 if (unlikely(!tctx))
7898 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7899 if (unlikely(ret)) {
7904 tctx->io_wq = io_init_wq_offload(ctx, task);
7905 if (IS_ERR(tctx->io_wq)) {
7906 ret = PTR_ERR(tctx->io_wq);
7907 percpu_counter_destroy(&tctx->inflight);
7913 init_waitqueue_head(&tctx->wait);
7915 atomic_set(&tctx->in_idle, 0);
7916 atomic_set(&tctx->inflight_tracked, 0);
7917 task->io_uring = tctx;
7918 spin_lock_init(&tctx->task_lock);
7919 INIT_WQ_LIST(&tctx->task_list);
7920 tctx->task_state = 0;
7921 init_task_work(&tctx->task_work, tctx_task_work);
7925 void __io_uring_free(struct task_struct *tsk)
7927 struct io_uring_task *tctx = tsk->io_uring;
7929 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7930 WARN_ON_ONCE(tctx->io_wq);
7932 percpu_counter_destroy(&tctx->inflight);
7934 tsk->io_uring = NULL;
7937 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7938 struct io_uring_params *p)
7942 /* Retain compatibility with failing for an invalid attach attempt */
7943 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7944 IORING_SETUP_ATTACH_WQ) {
7947 f = fdget(p->wq_fd);
7951 if (f.file->f_op != &io_uring_fops)
7954 if (ctx->flags & IORING_SETUP_SQPOLL) {
7955 struct task_struct *tsk;
7956 struct io_sq_data *sqd;
7959 sqd = io_get_sq_data(p, &attached);
7965 ctx->sq_creds = get_current_cred();
7967 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7968 if (!ctx->sq_thread_idle)
7969 ctx->sq_thread_idle = HZ;
7971 io_sq_thread_park(sqd);
7972 list_add(&ctx->sqd_list, &sqd->ctx_list);
7973 io_sqd_update_thread_idle(sqd);
7974 /* don't attach to a dying SQPOLL thread, would be racy */
7975 ret = (attached && !sqd->thread) ? -ENXIO : 0;
7976 io_sq_thread_unpark(sqd);
7983 if (p->flags & IORING_SETUP_SQ_AFF) {
7984 int cpu = p->sq_thread_cpu;
7987 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
7994 sqd->task_pid = current->pid;
7995 sqd->task_tgid = current->tgid;
7996 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8003 ret = io_uring_alloc_task_context(tsk, ctx);
8004 wake_up_new_task(tsk);
8007 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8008 /* Can't have SQ_AFF without SQPOLL */
8015 complete(&ctx->sq_data->exited);
8017 io_sq_thread_finish(ctx);
8021 static inline void __io_unaccount_mem(struct user_struct *user,
8022 unsigned long nr_pages)
8024 atomic_long_sub(nr_pages, &user->locked_vm);
8027 static inline int __io_account_mem(struct user_struct *user,
8028 unsigned long nr_pages)
8030 unsigned long page_limit, cur_pages, new_pages;
8032 /* Don't allow more pages than we can safely lock */
8033 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8036 cur_pages = atomic_long_read(&user->locked_vm);
8037 new_pages = cur_pages + nr_pages;
8038 if (new_pages > page_limit)
8040 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8041 new_pages) != cur_pages);
8046 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8049 __io_unaccount_mem(ctx->user, nr_pages);
8051 if (ctx->mm_account)
8052 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8055 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8060 ret = __io_account_mem(ctx->user, nr_pages);
8065 if (ctx->mm_account)
8066 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8071 static void io_mem_free(void *ptr)
8078 page = virt_to_head_page(ptr);
8079 if (put_page_testzero(page))
8080 free_compound_page(page);
8083 static void *io_mem_alloc(size_t size)
8085 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8086 __GFP_NORETRY | __GFP_ACCOUNT;
8088 return (void *) __get_free_pages(gfp_flags, get_order(size));
8091 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8094 struct io_rings *rings;
8095 size_t off, sq_array_size;
8097 off = struct_size(rings, cqes, cq_entries);
8098 if (off == SIZE_MAX)
8102 off = ALIGN(off, SMP_CACHE_BYTES);
8110 sq_array_size = array_size(sizeof(u32), sq_entries);
8111 if (sq_array_size == SIZE_MAX)
8114 if (check_add_overflow(off, sq_array_size, &off))
8120 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8122 struct io_mapped_ubuf *imu = *slot;
8125 if (imu != ctx->dummy_ubuf) {
8126 for (i = 0; i < imu->nr_bvecs; i++)
8127 unpin_user_page(imu->bvec[i].bv_page);
8128 if (imu->acct_pages)
8129 io_unaccount_mem(ctx, imu->acct_pages);
8135 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8137 io_buffer_unmap(ctx, &prsrc->buf);
8141 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8145 for (i = 0; i < ctx->nr_user_bufs; i++)
8146 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8147 kfree(ctx->user_bufs);
8148 io_rsrc_data_free(ctx->buf_data);
8149 ctx->user_bufs = NULL;
8150 ctx->buf_data = NULL;
8151 ctx->nr_user_bufs = 0;
8154 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8161 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8163 __io_sqe_buffers_unregister(ctx);
8167 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8168 void __user *arg, unsigned index)
8170 struct iovec __user *src;
8172 #ifdef CONFIG_COMPAT
8174 struct compat_iovec __user *ciovs;
8175 struct compat_iovec ciov;
8177 ciovs = (struct compat_iovec __user *) arg;
8178 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8181 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8182 dst->iov_len = ciov.iov_len;
8186 src = (struct iovec __user *) arg;
8187 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8193 * Not super efficient, but this is just a registration time. And we do cache
8194 * the last compound head, so generally we'll only do a full search if we don't
8197 * We check if the given compound head page has already been accounted, to
8198 * avoid double accounting it. This allows us to account the full size of the
8199 * page, not just the constituent pages of a huge page.
8201 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8202 int nr_pages, struct page *hpage)
8206 /* check current page array */
8207 for (i = 0; i < nr_pages; i++) {
8208 if (!PageCompound(pages[i]))
8210 if (compound_head(pages[i]) == hpage)
8214 /* check previously registered pages */
8215 for (i = 0; i < ctx->nr_user_bufs; i++) {
8216 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8218 for (j = 0; j < imu->nr_bvecs; j++) {
8219 if (!PageCompound(imu->bvec[j].bv_page))
8221 if (compound_head(imu->bvec[j].bv_page) == hpage)
8229 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8230 int nr_pages, struct io_mapped_ubuf *imu,
8231 struct page **last_hpage)
8235 imu->acct_pages = 0;
8236 for (i = 0; i < nr_pages; i++) {
8237 if (!PageCompound(pages[i])) {
8242 hpage = compound_head(pages[i]);
8243 if (hpage == *last_hpage)
8245 *last_hpage = hpage;
8246 if (headpage_already_acct(ctx, pages, i, hpage))
8248 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8252 if (!imu->acct_pages)
8255 ret = io_account_mem(ctx, imu->acct_pages);
8257 imu->acct_pages = 0;
8261 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8262 struct io_mapped_ubuf **pimu,
8263 struct page **last_hpage)
8265 struct io_mapped_ubuf *imu = NULL;
8266 struct vm_area_struct **vmas = NULL;
8267 struct page **pages = NULL;
8268 unsigned long off, start, end, ubuf;
8270 int ret, pret, nr_pages, i;
8272 if (!iov->iov_base) {
8273 *pimu = ctx->dummy_ubuf;
8277 ubuf = (unsigned long) iov->iov_base;
8278 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8279 start = ubuf >> PAGE_SHIFT;
8280 nr_pages = end - start;
8285 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8289 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8294 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8299 mmap_read_lock(current->mm);
8300 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8302 if (pret == nr_pages) {
8303 /* don't support file backed memory */
8304 for (i = 0; i < nr_pages; i++) {
8305 struct vm_area_struct *vma = vmas[i];
8307 if (vma_is_shmem(vma))
8310 !is_file_hugepages(vma->vm_file)) {
8316 ret = pret < 0 ? pret : -EFAULT;
8318 mmap_read_unlock(current->mm);
8321 * if we did partial map, or found file backed vmas,
8322 * release any pages we did get
8325 unpin_user_pages(pages, pret);
8329 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8331 unpin_user_pages(pages, pret);
8335 off = ubuf & ~PAGE_MASK;
8336 size = iov->iov_len;
8337 for (i = 0; i < nr_pages; i++) {
8340 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8341 imu->bvec[i].bv_page = pages[i];
8342 imu->bvec[i].bv_len = vec_len;
8343 imu->bvec[i].bv_offset = off;
8347 /* store original address for later verification */
8349 imu->ubuf_end = ubuf + iov->iov_len;
8350 imu->nr_bvecs = nr_pages;
8361 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8363 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8364 return ctx->user_bufs ? 0 : -ENOMEM;
8367 static int io_buffer_validate(struct iovec *iov)
8369 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8372 * Don't impose further limits on the size and buffer
8373 * constraints here, we'll -EINVAL later when IO is
8374 * submitted if they are wrong.
8377 return iov->iov_len ? -EFAULT : 0;
8381 /* arbitrary limit, but we need something */
8382 if (iov->iov_len > SZ_1G)
8385 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8391 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8392 unsigned int nr_args, u64 __user *tags)
8394 struct page *last_hpage = NULL;
8395 struct io_rsrc_data *data;
8401 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8403 ret = io_rsrc_node_switch_start(ctx);
8406 data = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, nr_args);
8409 ret = io_buffers_map_alloc(ctx, nr_args);
8411 io_rsrc_data_free(data);
8415 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8418 if (tags && copy_from_user(&tag, &tags[i], sizeof(tag))) {
8422 ret = io_copy_iov(ctx, &iov, arg, i);
8425 ret = io_buffer_validate(&iov);
8428 if (!iov.iov_base && tag) {
8433 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8437 data->tags[i] = tag;
8440 WARN_ON_ONCE(ctx->buf_data);
8442 ctx->buf_data = data;
8444 __io_sqe_buffers_unregister(ctx);
8446 io_rsrc_node_switch(ctx, NULL);
8450 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8451 struct io_uring_rsrc_update2 *up,
8452 unsigned int nr_args)
8454 u64 __user *tags = u64_to_user_ptr(up->tags);
8455 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8456 struct page *last_hpage = NULL;
8457 bool needs_switch = false;
8463 if (up->offset + nr_args > ctx->nr_user_bufs)
8466 for (done = 0; done < nr_args; done++) {
8467 struct io_mapped_ubuf *imu;
8468 int offset = up->offset + done;
8471 err = io_copy_iov(ctx, &iov, iovs, done);
8474 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8478 err = io_buffer_validate(&iov);
8481 if (!iov.iov_base && tag) {
8485 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8489 i = array_index_nospec(offset, ctx->nr_user_bufs);
8490 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8491 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8492 ctx->rsrc_node, ctx->user_bufs[i]);
8493 if (unlikely(err)) {
8494 io_buffer_unmap(ctx, &imu);
8497 ctx->user_bufs[i] = NULL;
8498 needs_switch = true;
8501 ctx->user_bufs[i] = imu;
8502 ctx->buf_data->tags[offset] = tag;
8506 io_rsrc_node_switch(ctx, ctx->buf_data);
8507 return done ? done : err;
8510 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8512 __s32 __user *fds = arg;
8518 if (copy_from_user(&fd, fds, sizeof(*fds)))
8521 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8522 if (IS_ERR(ctx->cq_ev_fd)) {
8523 int ret = PTR_ERR(ctx->cq_ev_fd);
8524 ctx->cq_ev_fd = NULL;
8531 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8533 if (ctx->cq_ev_fd) {
8534 eventfd_ctx_put(ctx->cq_ev_fd);
8535 ctx->cq_ev_fd = NULL;
8542 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8544 struct io_buffer *buf;
8545 unsigned long index;
8547 xa_for_each(&ctx->io_buffers, index, buf)
8548 __io_remove_buffers(ctx, buf, index, -1U);
8551 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8553 struct io_kiocb *req, *nxt;
8555 list_for_each_entry_safe(req, nxt, list, compl.list) {
8556 if (tsk && req->task != tsk)
8558 list_del(&req->compl.list);
8559 kmem_cache_free(req_cachep, req);
8563 static void io_req_caches_free(struct io_ring_ctx *ctx)
8565 struct io_submit_state *submit_state = &ctx->submit_state;
8566 struct io_comp_state *cs = &ctx->submit_state.comp;
8568 mutex_lock(&ctx->uring_lock);
8570 if (submit_state->free_reqs) {
8571 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8572 submit_state->reqs);
8573 submit_state->free_reqs = 0;
8576 io_flush_cached_locked_reqs(ctx, cs);
8577 io_req_cache_free(&cs->free_list, NULL);
8578 mutex_unlock(&ctx->uring_lock);
8581 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8585 if (!atomic_dec_and_test(&data->refs))
8586 wait_for_completion(&data->done);
8590 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8592 io_sq_thread_finish(ctx);
8594 if (ctx->mm_account) {
8595 mmdrop(ctx->mm_account);
8596 ctx->mm_account = NULL;
8599 mutex_lock(&ctx->uring_lock);
8600 if (io_wait_rsrc_data(ctx->buf_data))
8601 __io_sqe_buffers_unregister(ctx);
8602 if (io_wait_rsrc_data(ctx->file_data))
8603 __io_sqe_files_unregister(ctx);
8605 __io_cqring_overflow_flush(ctx, true);
8606 mutex_unlock(&ctx->uring_lock);
8607 io_eventfd_unregister(ctx);
8608 io_destroy_buffers(ctx);
8610 put_cred(ctx->sq_creds);
8612 /* there are no registered resources left, nobody uses it */
8614 io_rsrc_node_destroy(ctx->rsrc_node);
8615 if (ctx->rsrc_backup_node)
8616 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8617 flush_delayed_work(&ctx->rsrc_put_work);
8619 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8620 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8622 #if defined(CONFIG_UNIX)
8623 if (ctx->ring_sock) {
8624 ctx->ring_sock->file = NULL; /* so that iput() is called */
8625 sock_release(ctx->ring_sock);
8629 io_mem_free(ctx->rings);
8630 io_mem_free(ctx->sq_sqes);
8632 percpu_ref_exit(&ctx->refs);
8633 free_uid(ctx->user);
8634 io_req_caches_free(ctx);
8636 io_wq_put_hash(ctx->hash_map);
8637 kfree(ctx->cancel_hash);
8638 kfree(ctx->dummy_ubuf);
8642 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8644 struct io_ring_ctx *ctx = file->private_data;
8647 poll_wait(file, &ctx->cq_wait, wait);
8649 * synchronizes with barrier from wq_has_sleeper call in
8653 if (!io_sqring_full(ctx))
8654 mask |= EPOLLOUT | EPOLLWRNORM;
8657 * Don't flush cqring overflow list here, just do a simple check.
8658 * Otherwise there could possible be ABBA deadlock:
8661 * lock(&ctx->uring_lock);
8663 * lock(&ctx->uring_lock);
8666 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8667 * pushs them to do the flush.
8669 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8670 mask |= EPOLLIN | EPOLLRDNORM;
8675 static int io_uring_fasync(int fd, struct file *file, int on)
8677 struct io_ring_ctx *ctx = file->private_data;
8679 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8682 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8684 const struct cred *creds;
8686 creds = xa_erase(&ctx->personalities, id);
8695 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8697 return io_run_task_work_head(&ctx->exit_task_work);
8700 struct io_tctx_exit {
8701 struct callback_head task_work;
8702 struct completion completion;
8703 struct io_ring_ctx *ctx;
8706 static void io_tctx_exit_cb(struct callback_head *cb)
8708 struct io_uring_task *tctx = current->io_uring;
8709 struct io_tctx_exit *work;
8711 work = container_of(cb, struct io_tctx_exit, task_work);
8713 * When @in_idle, we're in cancellation and it's racy to remove the
8714 * node. It'll be removed by the end of cancellation, just ignore it.
8716 if (!atomic_read(&tctx->in_idle))
8717 io_uring_del_tctx_node((unsigned long)work->ctx);
8718 complete(&work->completion);
8721 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8723 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8725 return req->ctx == data;
8728 static void io_ring_exit_work(struct work_struct *work)
8730 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8731 unsigned long timeout = jiffies + HZ * 60 * 5;
8732 struct io_tctx_exit exit;
8733 struct io_tctx_node *node;
8737 * If we're doing polled IO and end up having requests being
8738 * submitted async (out-of-line), then completions can come in while
8739 * we're waiting for refs to drop. We need to reap these manually,
8740 * as nobody else will be looking for them.
8743 io_uring_try_cancel_requests(ctx, NULL, true);
8745 struct io_sq_data *sqd = ctx->sq_data;
8746 struct task_struct *tsk;
8748 io_sq_thread_park(sqd);
8750 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8751 io_wq_cancel_cb(tsk->io_uring->io_wq,
8752 io_cancel_ctx_cb, ctx, true);
8753 io_sq_thread_unpark(sqd);
8756 WARN_ON_ONCE(time_after(jiffies, timeout));
8757 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8759 init_completion(&exit.completion);
8760 init_task_work(&exit.task_work, io_tctx_exit_cb);
8763 * Some may use context even when all refs and requests have been put,
8764 * and they are free to do so while still holding uring_lock or
8765 * completion_lock, see __io_req_task_submit(). Apart from other work,
8766 * this lock/unlock section also waits them to finish.
8768 mutex_lock(&ctx->uring_lock);
8769 while (!list_empty(&ctx->tctx_list)) {
8770 WARN_ON_ONCE(time_after(jiffies, timeout));
8772 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8774 /* don't spin on a single task if cancellation failed */
8775 list_rotate_left(&ctx->tctx_list);
8776 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8777 if (WARN_ON_ONCE(ret))
8779 wake_up_process(node->task);
8781 mutex_unlock(&ctx->uring_lock);
8782 wait_for_completion(&exit.completion);
8783 mutex_lock(&ctx->uring_lock);
8785 mutex_unlock(&ctx->uring_lock);
8786 spin_lock_irq(&ctx->completion_lock);
8787 spin_unlock_irq(&ctx->completion_lock);
8789 io_ring_ctx_free(ctx);
8792 /* Returns true if we found and killed one or more timeouts */
8793 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8796 struct io_kiocb *req, *tmp;
8799 spin_lock_irq(&ctx->completion_lock);
8800 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8801 if (io_match_task(req, tsk, cancel_all)) {
8802 io_kill_timeout(req, -ECANCELED);
8807 io_commit_cqring(ctx);
8808 spin_unlock_irq(&ctx->completion_lock);
8810 io_cqring_ev_posted(ctx);
8811 return canceled != 0;
8814 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8816 unsigned long index;
8817 struct creds *creds;
8819 mutex_lock(&ctx->uring_lock);
8820 percpu_ref_kill(&ctx->refs);
8822 __io_cqring_overflow_flush(ctx, true);
8823 xa_for_each(&ctx->personalities, index, creds)
8824 io_unregister_personality(ctx, index);
8825 mutex_unlock(&ctx->uring_lock);
8827 io_kill_timeouts(ctx, NULL, true);
8828 io_poll_remove_all(ctx, NULL, true);
8830 /* if we failed setting up the ctx, we might not have any rings */
8831 io_iopoll_try_reap_events(ctx);
8833 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8835 * Use system_unbound_wq to avoid spawning tons of event kworkers
8836 * if we're exiting a ton of rings at the same time. It just adds
8837 * noise and overhead, there's no discernable change in runtime
8838 * over using system_wq.
8840 queue_work(system_unbound_wq, &ctx->exit_work);
8843 static int io_uring_release(struct inode *inode, struct file *file)
8845 struct io_ring_ctx *ctx = file->private_data;
8847 file->private_data = NULL;
8848 io_ring_ctx_wait_and_kill(ctx);
8852 struct io_task_cancel {
8853 struct task_struct *task;
8857 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8859 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8860 struct io_task_cancel *cancel = data;
8863 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8864 unsigned long flags;
8865 struct io_ring_ctx *ctx = req->ctx;
8867 /* protect against races with linked timeouts */
8868 spin_lock_irqsave(&ctx->completion_lock, flags);
8869 ret = io_match_task(req, cancel->task, cancel->all);
8870 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8872 ret = io_match_task(req, cancel->task, cancel->all);
8877 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8878 struct task_struct *task, bool cancel_all)
8880 struct io_defer_entry *de;
8883 spin_lock_irq(&ctx->completion_lock);
8884 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8885 if (io_match_task(de->req, task, cancel_all)) {
8886 list_cut_position(&list, &ctx->defer_list, &de->list);
8890 spin_unlock_irq(&ctx->completion_lock);
8891 if (list_empty(&list))
8894 while (!list_empty(&list)) {
8895 de = list_first_entry(&list, struct io_defer_entry, list);
8896 list_del_init(&de->list);
8897 io_req_complete_failed(de->req, -ECANCELED);
8903 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8905 struct io_tctx_node *node;
8906 enum io_wq_cancel cret;
8909 mutex_lock(&ctx->uring_lock);
8910 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8911 struct io_uring_task *tctx = node->task->io_uring;
8914 * io_wq will stay alive while we hold uring_lock, because it's
8915 * killed after ctx nodes, which requires to take the lock.
8917 if (!tctx || !tctx->io_wq)
8919 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8920 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8922 mutex_unlock(&ctx->uring_lock);
8927 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8928 struct task_struct *task,
8931 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8932 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8935 enum io_wq_cancel cret;
8939 ret |= io_uring_try_cancel_iowq(ctx);
8940 } else if (tctx && tctx->io_wq) {
8942 * Cancels requests of all rings, not only @ctx, but
8943 * it's fine as the task is in exit/exec.
8945 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8947 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8950 /* SQPOLL thread does its own polling */
8951 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8952 (ctx->sq_data && ctx->sq_data->thread == current)) {
8953 while (!list_empty_careful(&ctx->iopoll_list)) {
8954 io_iopoll_try_reap_events(ctx);
8959 ret |= io_cancel_defer_files(ctx, task, cancel_all);
8960 ret |= io_poll_remove_all(ctx, task, cancel_all);
8961 ret |= io_kill_timeouts(ctx, task, cancel_all);
8962 ret |= io_run_task_work();
8963 ret |= io_run_ctx_fallback(ctx);
8970 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
8972 struct io_uring_task *tctx = current->io_uring;
8973 struct io_tctx_node *node;
8976 if (unlikely(!tctx)) {
8977 ret = io_uring_alloc_task_context(current, ctx);
8980 tctx = current->io_uring;
8982 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8983 node = kmalloc(sizeof(*node), GFP_KERNEL);
8987 node->task = current;
8989 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8996 mutex_lock(&ctx->uring_lock);
8997 list_add(&node->ctx_node, &ctx->tctx_list);
8998 mutex_unlock(&ctx->uring_lock);
9005 * Note that this task has used io_uring. We use it for cancelation purposes.
9007 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9009 struct io_uring_task *tctx = current->io_uring;
9011 if (likely(tctx && tctx->last == ctx))
9013 return __io_uring_add_tctx_node(ctx);
9017 * Remove this io_uring_file -> task mapping.
9019 static void io_uring_del_tctx_node(unsigned long index)
9021 struct io_uring_task *tctx = current->io_uring;
9022 struct io_tctx_node *node;
9026 node = xa_erase(&tctx->xa, index);
9030 WARN_ON_ONCE(current != node->task);
9031 WARN_ON_ONCE(list_empty(&node->ctx_node));
9033 mutex_lock(&node->ctx->uring_lock);
9034 list_del(&node->ctx_node);
9035 mutex_unlock(&node->ctx->uring_lock);
9037 if (tctx->last == node->ctx)
9042 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9044 struct io_wq *wq = tctx->io_wq;
9045 struct io_tctx_node *node;
9046 unsigned long index;
9048 xa_for_each(&tctx->xa, index, node)
9049 io_uring_del_tctx_node(index);
9052 * Must be after io_uring_del_task_file() (removes nodes under
9053 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9056 io_wq_put_and_exit(wq);
9060 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9063 return atomic_read(&tctx->inflight_tracked);
9064 return percpu_counter_sum(&tctx->inflight);
9067 static void io_uring_try_cancel(bool cancel_all)
9069 struct io_uring_task *tctx = current->io_uring;
9070 struct io_tctx_node *node;
9071 unsigned long index;
9073 xa_for_each(&tctx->xa, index, node) {
9074 struct io_ring_ctx *ctx = node->ctx;
9076 /* sqpoll task will cancel all its requests */
9078 io_uring_try_cancel_requests(ctx, current, cancel_all);
9082 /* should only be called by SQPOLL task */
9083 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd)
9085 struct io_uring_task *tctx = current->io_uring;
9086 struct io_ring_ctx *ctx;
9090 if (!current->io_uring)
9093 io_wq_exit_start(tctx->io_wq);
9095 WARN_ON_ONCE(!sqd || sqd->thread != current);
9097 atomic_inc(&tctx->in_idle);
9099 /* read completions before cancelations */
9100 inflight = tctx_inflight(tctx, false);
9103 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9104 io_uring_try_cancel_requests(ctx, current, true);
9106 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9108 * If we've seen completions, retry without waiting. This
9109 * avoids a race where a completion comes in before we did
9110 * prepare_to_wait().
9112 if (inflight == tctx_inflight(tctx, false))
9114 finish_wait(&tctx->wait, &wait);
9116 atomic_dec(&tctx->in_idle);
9120 * Find any io_uring fd that this task has registered or done IO on, and cancel
9123 void __io_uring_cancel(struct files_struct *files)
9125 struct io_uring_task *tctx = current->io_uring;
9128 bool cancel_all = !files;
9131 io_wq_exit_start(tctx->io_wq);
9133 /* make sure overflow events are dropped */
9134 atomic_inc(&tctx->in_idle);
9136 /* read completions before cancelations */
9137 inflight = tctx_inflight(tctx, !cancel_all);
9140 io_uring_try_cancel(cancel_all);
9141 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9144 * If we've seen completions, retry without waiting. This
9145 * avoids a race where a completion comes in before we did
9146 * prepare_to_wait().
9148 if (inflight == tctx_inflight(tctx, !cancel_all))
9150 finish_wait(&tctx->wait, &wait);
9152 atomic_dec(&tctx->in_idle);
9154 io_uring_clean_tctx(tctx);
9156 /* for exec all current's requests should be gone, kill tctx */
9157 __io_uring_free(current);
9161 static void *io_uring_validate_mmap_request(struct file *file,
9162 loff_t pgoff, size_t sz)
9164 struct io_ring_ctx *ctx = file->private_data;
9165 loff_t offset = pgoff << PAGE_SHIFT;
9170 case IORING_OFF_SQ_RING:
9171 case IORING_OFF_CQ_RING:
9174 case IORING_OFF_SQES:
9178 return ERR_PTR(-EINVAL);
9181 page = virt_to_head_page(ptr);
9182 if (sz > page_size(page))
9183 return ERR_PTR(-EINVAL);
9190 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9192 size_t sz = vma->vm_end - vma->vm_start;
9196 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9198 return PTR_ERR(ptr);
9200 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9201 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9204 #else /* !CONFIG_MMU */
9206 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9208 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9211 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9213 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9216 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9217 unsigned long addr, unsigned long len,
9218 unsigned long pgoff, unsigned long flags)
9222 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9224 return PTR_ERR(ptr);
9226 return (unsigned long) ptr;
9229 #endif /* !CONFIG_MMU */
9231 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9236 if (!io_sqring_full(ctx))
9238 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9240 if (!io_sqring_full(ctx))
9243 } while (!signal_pending(current));
9245 finish_wait(&ctx->sqo_sq_wait, &wait);
9249 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9250 struct __kernel_timespec __user **ts,
9251 const sigset_t __user **sig)
9253 struct io_uring_getevents_arg arg;
9256 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9257 * is just a pointer to the sigset_t.
9259 if (!(flags & IORING_ENTER_EXT_ARG)) {
9260 *sig = (const sigset_t __user *) argp;
9266 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9267 * timespec and sigset_t pointers if good.
9269 if (*argsz != sizeof(arg))
9271 if (copy_from_user(&arg, argp, sizeof(arg)))
9273 *sig = u64_to_user_ptr(arg.sigmask);
9274 *argsz = arg.sigmask_sz;
9275 *ts = u64_to_user_ptr(arg.ts);
9279 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9280 u32, min_complete, u32, flags, const void __user *, argp,
9283 struct io_ring_ctx *ctx;
9290 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9291 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9295 if (unlikely(!f.file))
9299 if (unlikely(f.file->f_op != &io_uring_fops))
9303 ctx = f.file->private_data;
9304 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9308 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9312 * For SQ polling, the thread will do all submissions and completions.
9313 * Just return the requested submit count, and wake the thread if
9317 if (ctx->flags & IORING_SETUP_SQPOLL) {
9318 io_cqring_overflow_flush(ctx, false);
9321 if (unlikely(ctx->sq_data->thread == NULL)) {
9324 if (flags & IORING_ENTER_SQ_WAKEUP)
9325 wake_up(&ctx->sq_data->wait);
9326 if (flags & IORING_ENTER_SQ_WAIT) {
9327 ret = io_sqpoll_wait_sq(ctx);
9331 submitted = to_submit;
9332 } else if (to_submit) {
9333 ret = io_uring_add_tctx_node(ctx);
9336 mutex_lock(&ctx->uring_lock);
9337 submitted = io_submit_sqes(ctx, to_submit);
9338 mutex_unlock(&ctx->uring_lock);
9340 if (submitted != to_submit)
9343 if (flags & IORING_ENTER_GETEVENTS) {
9344 const sigset_t __user *sig;
9345 struct __kernel_timespec __user *ts;
9347 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9351 min_complete = min(min_complete, ctx->cq_entries);
9354 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9355 * space applications don't need to do io completion events
9356 * polling again, they can rely on io_sq_thread to do polling
9357 * work, which can reduce cpu usage and uring_lock contention.
9359 if (ctx->flags & IORING_SETUP_IOPOLL &&
9360 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9361 ret = io_iopoll_check(ctx, min_complete);
9363 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9368 percpu_ref_put(&ctx->refs);
9371 return submitted ? submitted : ret;
9374 #ifdef CONFIG_PROC_FS
9375 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9376 const struct cred *cred)
9378 struct user_namespace *uns = seq_user_ns(m);
9379 struct group_info *gi;
9384 seq_printf(m, "%5d\n", id);
9385 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9386 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9387 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9388 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9389 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9390 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9391 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9392 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9393 seq_puts(m, "\n\tGroups:\t");
9394 gi = cred->group_info;
9395 for (g = 0; g < gi->ngroups; g++) {
9396 seq_put_decimal_ull(m, g ? " " : "",
9397 from_kgid_munged(uns, gi->gid[g]));
9399 seq_puts(m, "\n\tCapEff:\t");
9400 cap = cred->cap_effective;
9401 CAP_FOR_EACH_U32(__capi)
9402 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9407 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9409 struct io_sq_data *sq = NULL;
9414 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9415 * since fdinfo case grabs it in the opposite direction of normal use
9416 * cases. If we fail to get the lock, we just don't iterate any
9417 * structures that could be going away outside the io_uring mutex.
9419 has_lock = mutex_trylock(&ctx->uring_lock);
9421 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9427 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9428 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9429 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9430 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9431 struct file *f = io_file_from_index(ctx, i);
9434 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9436 seq_printf(m, "%5u: <none>\n", i);
9438 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9439 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9440 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9441 unsigned int len = buf->ubuf_end - buf->ubuf;
9443 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9445 if (has_lock && !xa_empty(&ctx->personalities)) {
9446 unsigned long index;
9447 const struct cred *cred;
9449 seq_printf(m, "Personalities:\n");
9450 xa_for_each(&ctx->personalities, index, cred)
9451 io_uring_show_cred(m, index, cred);
9453 seq_printf(m, "PollList:\n");
9454 spin_lock_irq(&ctx->completion_lock);
9455 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9456 struct hlist_head *list = &ctx->cancel_hash[i];
9457 struct io_kiocb *req;
9459 hlist_for_each_entry(req, list, hash_node)
9460 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9461 req->task->task_works != NULL);
9463 spin_unlock_irq(&ctx->completion_lock);
9465 mutex_unlock(&ctx->uring_lock);
9468 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9470 struct io_ring_ctx *ctx = f->private_data;
9472 if (percpu_ref_tryget(&ctx->refs)) {
9473 __io_uring_show_fdinfo(ctx, m);
9474 percpu_ref_put(&ctx->refs);
9479 static const struct file_operations io_uring_fops = {
9480 .release = io_uring_release,
9481 .mmap = io_uring_mmap,
9483 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9484 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9486 .poll = io_uring_poll,
9487 .fasync = io_uring_fasync,
9488 #ifdef CONFIG_PROC_FS
9489 .show_fdinfo = io_uring_show_fdinfo,
9493 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9494 struct io_uring_params *p)
9496 struct io_rings *rings;
9497 size_t size, sq_array_offset;
9499 /* make sure these are sane, as we already accounted them */
9500 ctx->sq_entries = p->sq_entries;
9501 ctx->cq_entries = p->cq_entries;
9503 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9504 if (size == SIZE_MAX)
9507 rings = io_mem_alloc(size);
9512 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9513 rings->sq_ring_mask = p->sq_entries - 1;
9514 rings->cq_ring_mask = p->cq_entries - 1;
9515 rings->sq_ring_entries = p->sq_entries;
9516 rings->cq_ring_entries = p->cq_entries;
9518 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9519 if (size == SIZE_MAX) {
9520 io_mem_free(ctx->rings);
9525 ctx->sq_sqes = io_mem_alloc(size);
9526 if (!ctx->sq_sqes) {
9527 io_mem_free(ctx->rings);
9535 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9539 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9543 ret = io_uring_add_tctx_node(ctx);
9548 fd_install(fd, file);
9553 * Allocate an anonymous fd, this is what constitutes the application
9554 * visible backing of an io_uring instance. The application mmaps this
9555 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9556 * we have to tie this fd to a socket for file garbage collection purposes.
9558 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9561 #if defined(CONFIG_UNIX)
9564 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9567 return ERR_PTR(ret);
9570 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9571 O_RDWR | O_CLOEXEC);
9572 #if defined(CONFIG_UNIX)
9574 sock_release(ctx->ring_sock);
9575 ctx->ring_sock = NULL;
9577 ctx->ring_sock->file = file;
9583 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9584 struct io_uring_params __user *params)
9586 struct io_ring_ctx *ctx;
9592 if (entries > IORING_MAX_ENTRIES) {
9593 if (!(p->flags & IORING_SETUP_CLAMP))
9595 entries = IORING_MAX_ENTRIES;
9599 * Use twice as many entries for the CQ ring. It's possible for the
9600 * application to drive a higher depth than the size of the SQ ring,
9601 * since the sqes are only used at submission time. This allows for
9602 * some flexibility in overcommitting a bit. If the application has
9603 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9604 * of CQ ring entries manually.
9606 p->sq_entries = roundup_pow_of_two(entries);
9607 if (p->flags & IORING_SETUP_CQSIZE) {
9609 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9610 * to a power-of-two, if it isn't already. We do NOT impose
9611 * any cq vs sq ring sizing.
9615 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9616 if (!(p->flags & IORING_SETUP_CLAMP))
9618 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9620 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9621 if (p->cq_entries < p->sq_entries)
9624 p->cq_entries = 2 * p->sq_entries;
9627 ctx = io_ring_ctx_alloc(p);
9630 ctx->compat = in_compat_syscall();
9631 if (!capable(CAP_IPC_LOCK))
9632 ctx->user = get_uid(current_user());
9635 * This is just grabbed for accounting purposes. When a process exits,
9636 * the mm is exited and dropped before the files, hence we need to hang
9637 * on to this mm purely for the purposes of being able to unaccount
9638 * memory (locked/pinned vm). It's not used for anything else.
9640 mmgrab(current->mm);
9641 ctx->mm_account = current->mm;
9643 ret = io_allocate_scq_urings(ctx, p);
9647 ret = io_sq_offload_create(ctx, p);
9650 /* always set a rsrc node */
9651 ret = io_rsrc_node_switch_start(ctx);
9654 io_rsrc_node_switch(ctx, NULL);
9656 memset(&p->sq_off, 0, sizeof(p->sq_off));
9657 p->sq_off.head = offsetof(struct io_rings, sq.head);
9658 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9659 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9660 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9661 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9662 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9663 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9665 memset(&p->cq_off, 0, sizeof(p->cq_off));
9666 p->cq_off.head = offsetof(struct io_rings, cq.head);
9667 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9668 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9669 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9670 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9671 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9672 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9674 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9675 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9676 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9677 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9678 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9679 IORING_FEAT_RSRC_TAGS;
9681 if (copy_to_user(params, p, sizeof(*p))) {
9686 file = io_uring_get_file(ctx);
9688 ret = PTR_ERR(file);
9693 * Install ring fd as the very last thing, so we don't risk someone
9694 * having closed it before we finish setup
9696 ret = io_uring_install_fd(ctx, file);
9698 /* fput will clean it up */
9703 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9706 io_ring_ctx_wait_and_kill(ctx);
9711 * Sets up an aio uring context, and returns the fd. Applications asks for a
9712 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9713 * params structure passed in.
9715 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9717 struct io_uring_params p;
9720 if (copy_from_user(&p, params, sizeof(p)))
9722 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9727 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9728 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9729 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9730 IORING_SETUP_R_DISABLED))
9733 return io_uring_create(entries, &p, params);
9736 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9737 struct io_uring_params __user *, params)
9739 return io_uring_setup(entries, params);
9742 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9744 struct io_uring_probe *p;
9748 size = struct_size(p, ops, nr_args);
9749 if (size == SIZE_MAX)
9751 p = kzalloc(size, GFP_KERNEL);
9756 if (copy_from_user(p, arg, size))
9759 if (memchr_inv(p, 0, size))
9762 p->last_op = IORING_OP_LAST - 1;
9763 if (nr_args > IORING_OP_LAST)
9764 nr_args = IORING_OP_LAST;
9766 for (i = 0; i < nr_args; i++) {
9768 if (!io_op_defs[i].not_supported)
9769 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9774 if (copy_to_user(arg, p, size))
9781 static int io_register_personality(struct io_ring_ctx *ctx)
9783 const struct cred *creds;
9787 creds = get_current_cred();
9789 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9790 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9797 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9798 unsigned int nr_args)
9800 struct io_uring_restriction *res;
9804 /* Restrictions allowed only if rings started disabled */
9805 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9808 /* We allow only a single restrictions registration */
9809 if (ctx->restrictions.registered)
9812 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9815 size = array_size(nr_args, sizeof(*res));
9816 if (size == SIZE_MAX)
9819 res = memdup_user(arg, size);
9821 return PTR_ERR(res);
9825 for (i = 0; i < nr_args; i++) {
9826 switch (res[i].opcode) {
9827 case IORING_RESTRICTION_REGISTER_OP:
9828 if (res[i].register_op >= IORING_REGISTER_LAST) {
9833 __set_bit(res[i].register_op,
9834 ctx->restrictions.register_op);
9836 case IORING_RESTRICTION_SQE_OP:
9837 if (res[i].sqe_op >= IORING_OP_LAST) {
9842 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9844 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9845 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9847 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9848 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9857 /* Reset all restrictions if an error happened */
9859 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9861 ctx->restrictions.registered = true;
9867 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9869 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9872 if (ctx->restrictions.registered)
9873 ctx->restricted = 1;
9875 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9876 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9877 wake_up(&ctx->sq_data->wait);
9881 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9882 struct io_uring_rsrc_update2 *up,
9890 if (check_add_overflow(up->offset, nr_args, &tmp))
9892 err = io_rsrc_node_switch_start(ctx);
9897 case IORING_RSRC_FILE:
9898 return __io_sqe_files_update(ctx, up, nr_args);
9899 case IORING_RSRC_BUFFER:
9900 return __io_sqe_buffers_update(ctx, up, nr_args);
9905 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9908 struct io_uring_rsrc_update2 up;
9912 memset(&up, 0, sizeof(up));
9913 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9915 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9918 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9919 unsigned size, unsigned type)
9921 struct io_uring_rsrc_update2 up;
9923 if (size != sizeof(up))
9925 if (copy_from_user(&up, arg, sizeof(up)))
9927 if (!up.nr || up.resv)
9929 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9932 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9933 unsigned int size, unsigned int type)
9935 struct io_uring_rsrc_register rr;
9937 /* keep it extendible */
9938 if (size != sizeof(rr))
9941 memset(&rr, 0, sizeof(rr));
9942 if (copy_from_user(&rr, arg, size))
9944 if (!rr.nr || rr.resv || rr.resv2)
9948 case IORING_RSRC_FILE:
9949 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9950 rr.nr, u64_to_user_ptr(rr.tags));
9951 case IORING_RSRC_BUFFER:
9952 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9953 rr.nr, u64_to_user_ptr(rr.tags));
9958 static bool io_register_op_must_quiesce(int op)
9961 case IORING_REGISTER_BUFFERS:
9962 case IORING_UNREGISTER_BUFFERS:
9963 case IORING_REGISTER_FILES:
9964 case IORING_UNREGISTER_FILES:
9965 case IORING_REGISTER_FILES_UPDATE:
9966 case IORING_REGISTER_PROBE:
9967 case IORING_REGISTER_PERSONALITY:
9968 case IORING_UNREGISTER_PERSONALITY:
9969 case IORING_REGISTER_FILES2:
9970 case IORING_REGISTER_FILES_UPDATE2:
9971 case IORING_REGISTER_BUFFERS2:
9972 case IORING_REGISTER_BUFFERS_UPDATE:
9979 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9980 void __user *arg, unsigned nr_args)
9981 __releases(ctx->uring_lock)
9982 __acquires(ctx->uring_lock)
9987 * We're inside the ring mutex, if the ref is already dying, then
9988 * someone else killed the ctx or is already going through
9989 * io_uring_register().
9991 if (percpu_ref_is_dying(&ctx->refs))
9994 if (ctx->restricted) {
9995 if (opcode >= IORING_REGISTER_LAST)
9997 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
9998 if (!test_bit(opcode, ctx->restrictions.register_op))
10002 if (io_register_op_must_quiesce(opcode)) {
10003 percpu_ref_kill(&ctx->refs);
10006 * Drop uring mutex before waiting for references to exit. If
10007 * another thread is currently inside io_uring_enter() it might
10008 * need to grab the uring_lock to make progress. If we hold it
10009 * here across the drain wait, then we can deadlock. It's safe
10010 * to drop the mutex here, since no new references will come in
10011 * after we've killed the percpu ref.
10013 mutex_unlock(&ctx->uring_lock);
10015 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10018 ret = io_run_task_work_sig();
10022 mutex_lock(&ctx->uring_lock);
10025 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10031 case IORING_REGISTER_BUFFERS:
10032 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10034 case IORING_UNREGISTER_BUFFERS:
10036 if (arg || nr_args)
10038 ret = io_sqe_buffers_unregister(ctx);
10040 case IORING_REGISTER_FILES:
10041 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10043 case IORING_UNREGISTER_FILES:
10045 if (arg || nr_args)
10047 ret = io_sqe_files_unregister(ctx);
10049 case IORING_REGISTER_FILES_UPDATE:
10050 ret = io_register_files_update(ctx, arg, nr_args);
10052 case IORING_REGISTER_EVENTFD:
10053 case IORING_REGISTER_EVENTFD_ASYNC:
10057 ret = io_eventfd_register(ctx, arg);
10060 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10061 ctx->eventfd_async = 1;
10063 ctx->eventfd_async = 0;
10065 case IORING_UNREGISTER_EVENTFD:
10067 if (arg || nr_args)
10069 ret = io_eventfd_unregister(ctx);
10071 case IORING_REGISTER_PROBE:
10073 if (!arg || nr_args > 256)
10075 ret = io_probe(ctx, arg, nr_args);
10077 case IORING_REGISTER_PERSONALITY:
10079 if (arg || nr_args)
10081 ret = io_register_personality(ctx);
10083 case IORING_UNREGISTER_PERSONALITY:
10087 ret = io_unregister_personality(ctx, nr_args);
10089 case IORING_REGISTER_ENABLE_RINGS:
10091 if (arg || nr_args)
10093 ret = io_register_enable_rings(ctx);
10095 case IORING_REGISTER_RESTRICTIONS:
10096 ret = io_register_restrictions(ctx, arg, nr_args);
10098 case IORING_REGISTER_FILES2:
10099 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10101 case IORING_REGISTER_FILES_UPDATE2:
10102 ret = io_register_rsrc_update(ctx, arg, nr_args,
10105 case IORING_REGISTER_BUFFERS2:
10106 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10108 case IORING_REGISTER_BUFFERS_UPDATE:
10109 ret = io_register_rsrc_update(ctx, arg, nr_args,
10110 IORING_RSRC_BUFFER);
10117 if (io_register_op_must_quiesce(opcode)) {
10118 /* bring the ctx back to life */
10119 percpu_ref_reinit(&ctx->refs);
10120 reinit_completion(&ctx->ref_comp);
10125 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10126 void __user *, arg, unsigned int, nr_args)
10128 struct io_ring_ctx *ctx;
10137 if (f.file->f_op != &io_uring_fops)
10140 ctx = f.file->private_data;
10142 io_run_task_work();
10144 mutex_lock(&ctx->uring_lock);
10145 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10146 mutex_unlock(&ctx->uring_lock);
10147 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10148 ctx->cq_ev_fd != NULL, ret);
10154 static int __init io_uring_init(void)
10156 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10157 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10158 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10161 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10162 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10163 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10164 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10165 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10166 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10167 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10168 BUILD_BUG_SQE_ELEM(8, __u64, off);
10169 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10170 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10171 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10172 BUILD_BUG_SQE_ELEM(24, __u32, len);
10173 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10174 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10175 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10176 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10177 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10178 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10179 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10180 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10181 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10182 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10183 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10184 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10185 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10186 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10187 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10188 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10189 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10190 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10191 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10193 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10194 sizeof(struct io_uring_rsrc_update));
10195 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10196 sizeof(struct io_uring_rsrc_update2));
10197 /* should fit into one byte */
10198 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10200 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10201 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10202 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10206 __initcall(io_uring_init);