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 IO_RSRC_TAG_TABLE_SHIFT 9
104 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
105 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
107 #define IORING_MAX_REG_BUFFERS (1U << 14)
109 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
110 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
113 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
116 u32 head ____cacheline_aligned_in_smp;
117 u32 tail ____cacheline_aligned_in_smp;
121 * This data is shared with the application through the mmap at offsets
122 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
124 * The offsets to the member fields are published through struct
125 * io_sqring_offsets when calling io_uring_setup.
129 * Head and tail offsets into the ring; the offsets need to be
130 * masked to get valid indices.
132 * The kernel controls head of the sq ring and the tail of the cq ring,
133 * and the application controls tail of the sq ring and the head of the
136 struct io_uring sq, cq;
138 * Bitmasks to apply to head and tail offsets (constant, equals
141 u32 sq_ring_mask, cq_ring_mask;
142 /* Ring sizes (constant, power of 2) */
143 u32 sq_ring_entries, cq_ring_entries;
145 * Number of invalid entries dropped by the kernel due to
146 * invalid index stored in array
148 * Written by the kernel, shouldn't be modified by the
149 * application (i.e. get number of "new events" by comparing to
152 * After a new SQ head value was read by the application this
153 * counter includes all submissions that were dropped reaching
154 * the new SQ head (and possibly more).
160 * Written by the kernel, shouldn't be modified by the
163 * The application needs a full memory barrier before checking
164 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
170 * Written by the application, shouldn't be modified by the
175 * Number of completion events lost because the queue was full;
176 * this should be avoided by the application by making sure
177 * there are not more requests pending than there is space in
178 * the completion queue.
180 * Written by the kernel, shouldn't be modified by the
181 * application (i.e. get number of "new events" by comparing to
184 * As completion events come in out of order this counter is not
185 * ordered with any other data.
189 * Ring buffer of completion events.
191 * The kernel writes completion events fresh every time they are
192 * produced, so the application is allowed to modify pending
195 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
198 enum io_uring_cmd_flags {
199 IO_URING_F_NONBLOCK = 1,
200 IO_URING_F_COMPLETE_DEFER = 2,
203 struct io_mapped_ubuf {
206 unsigned int nr_bvecs;
207 unsigned long acct_pages;
208 struct bio_vec bvec[];
213 struct io_overflow_cqe {
214 struct io_uring_cqe cqe;
215 struct list_head list;
218 struct io_fixed_file {
219 /* file * with additional FFS_* flags */
220 unsigned long file_ptr;
224 struct list_head list;
229 struct io_mapped_ubuf *buf;
233 struct io_file_table {
234 /* two level table */
235 struct io_fixed_file **files;
238 struct io_rsrc_node {
239 struct percpu_ref refs;
240 struct list_head node;
241 struct list_head rsrc_list;
242 struct io_rsrc_data *rsrc_data;
243 struct llist_node llist;
247 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
249 struct io_rsrc_data {
250 struct io_ring_ctx *ctx;
256 struct completion done;
261 struct list_head list;
267 struct io_restriction {
268 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
269 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
270 u8 sqe_flags_allowed;
271 u8 sqe_flags_required;
276 IO_SQ_THREAD_SHOULD_STOP = 0,
277 IO_SQ_THREAD_SHOULD_PARK,
282 atomic_t park_pending;
285 /* ctx's that are using this sqd */
286 struct list_head ctx_list;
288 struct task_struct *thread;
289 struct wait_queue_head wait;
291 unsigned sq_thread_idle;
297 struct completion exited;
300 #define IO_IOPOLL_BATCH 8
301 #define IO_COMPL_BATCH 32
302 #define IO_REQ_CACHE_SIZE 32
303 #define IO_REQ_ALLOC_BATCH 8
305 struct io_comp_state {
306 struct io_kiocb *reqs[IO_COMPL_BATCH];
308 /* inline/task_work completion list, under ->uring_lock */
309 struct list_head free_list;
312 struct io_submit_link {
313 struct io_kiocb *head;
314 struct io_kiocb *last;
317 struct io_submit_state {
318 struct blk_plug plug;
319 struct io_submit_link link;
322 * io_kiocb alloc cache
324 void *reqs[IO_REQ_CACHE_SIZE];
325 unsigned int free_reqs;
330 * Batch completion logic
332 struct io_comp_state comp;
335 * File reference cache
339 unsigned int file_refs;
340 unsigned int ios_left;
344 /* const or read-mostly hot data */
346 struct percpu_ref refs;
348 struct io_rings *rings;
350 unsigned int compat: 1;
351 unsigned int drain_next: 1;
352 unsigned int eventfd_async: 1;
353 unsigned int restricted: 1;
354 } ____cacheline_aligned_in_smp;
358 * Ring buffer of indices into array of io_uring_sqe, which is
359 * mmapped by the application using the IORING_OFF_SQES offset.
361 * This indirection could e.g. be used to assign fixed
362 * io_uring_sqe entries to operations and only submit them to
363 * the queue when needed.
365 * The kernel modifies neither the indices array nor the entries
369 struct io_uring_sqe *sq_sqes;
370 unsigned cached_sq_head;
372 unsigned sq_thread_idle;
373 unsigned cached_sq_dropped;
374 unsigned long sq_check_overflow;
376 struct list_head defer_list;
377 struct list_head timeout_list;
378 struct list_head cq_overflow_list;
379 } ____cacheline_aligned_in_smp;
382 struct mutex uring_lock;
383 wait_queue_head_t wait;
384 } ____cacheline_aligned_in_smp;
386 struct io_submit_state submit_state;
387 /* IRQ completion list, under ->completion_lock */
388 struct list_head locked_free_list;
389 unsigned int locked_free_nr;
391 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
392 struct io_sq_data *sq_data; /* if using sq thread polling */
394 struct wait_queue_head sqo_sq_wait;
395 struct list_head sqd_list;
398 * Fixed resources fast path, should be accessed only under uring_lock,
399 * and updated through io_uring_register(2)
401 struct io_rsrc_node *rsrc_node;
403 struct io_file_table file_table;
404 unsigned nr_user_files;
405 unsigned nr_user_bufs;
406 struct io_mapped_ubuf **user_bufs;
408 struct xarray io_buffers;
409 struct xarray personalities;
413 unsigned cached_cq_tail;
415 atomic_t cq_timeouts;
416 unsigned cq_last_tm_flush;
418 unsigned long cq_check_overflow;
419 struct wait_queue_head cq_wait;
420 struct fasync_struct *cq_fasync;
421 struct eventfd_ctx *cq_ev_fd;
422 } ____cacheline_aligned_in_smp;
425 spinlock_t completion_lock;
428 * ->iopoll_list is protected by the ctx->uring_lock for
429 * io_uring instances that don't use IORING_SETUP_SQPOLL.
430 * For SQPOLL, only the single threaded io_sq_thread() will
431 * manipulate the list, hence no extra locking is needed there.
433 struct list_head iopoll_list;
434 struct hlist_head *cancel_hash;
435 unsigned cancel_hash_bits;
436 bool poll_multi_file;
437 } ____cacheline_aligned_in_smp;
439 struct io_restriction restrictions;
441 /* slow path rsrc auxilary data, used by update/register */
443 struct io_rsrc_node *rsrc_backup_node;
444 struct io_mapped_ubuf *dummy_ubuf;
445 struct io_rsrc_data *file_data;
446 struct io_rsrc_data *buf_data;
448 struct delayed_work rsrc_put_work;
449 struct llist_head rsrc_put_llist;
450 struct list_head rsrc_ref_list;
451 spinlock_t rsrc_ref_lock;
454 /* Keep this last, we don't need it for the fast path */
456 #if defined(CONFIG_UNIX)
457 struct socket *ring_sock;
459 /* hashed buffered write serialization */
460 struct io_wq_hash *hash_map;
462 /* Only used for accounting purposes */
463 struct user_struct *user;
464 struct mm_struct *mm_account;
466 /* ctx exit and cancelation */
467 struct callback_head *exit_task_work;
468 struct work_struct exit_work;
469 struct list_head tctx_list;
470 struct completion ref_comp;
474 struct io_uring_task {
475 /* submission side */
478 struct wait_queue_head wait;
479 const struct io_ring_ctx *last;
481 struct percpu_counter inflight;
482 atomic_t inflight_tracked;
485 spinlock_t task_lock;
486 struct io_wq_work_list task_list;
487 unsigned long task_state;
488 struct callback_head task_work;
492 * First field must be the file pointer in all the
493 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
495 struct io_poll_iocb {
497 struct wait_queue_head *head;
501 struct wait_queue_entry wait;
504 struct io_poll_update {
510 bool update_user_data;
518 struct io_timeout_data {
519 struct io_kiocb *req;
520 struct hrtimer timer;
521 struct timespec64 ts;
522 enum hrtimer_mode mode;
527 struct sockaddr __user *addr;
528 int __user *addr_len;
530 unsigned long nofile;
550 struct list_head list;
551 /* head of the link, used by linked timeouts only */
552 struct io_kiocb *head;
555 struct io_timeout_rem {
560 struct timespec64 ts;
565 /* NOTE: kiocb has the file as the first member, so don't do it here */
573 struct sockaddr __user *addr;
580 struct compat_msghdr __user *umsg_compat;
581 struct user_msghdr __user *umsg;
587 struct io_buffer *kbuf;
593 struct filename *filename;
595 unsigned long nofile;
598 struct io_rsrc_update {
624 struct epoll_event event;
628 struct file *file_out;
629 struct file *file_in;
636 struct io_provide_buf {
650 const char __user *filename;
651 struct statx __user *buffer;
663 struct filename *oldpath;
664 struct filename *newpath;
672 struct filename *filename;
675 struct io_completion {
677 struct list_head list;
681 struct io_async_connect {
682 struct sockaddr_storage address;
685 struct io_async_msghdr {
686 struct iovec fast_iov[UIO_FASTIOV];
687 /* points to an allocated iov, if NULL we use fast_iov instead */
688 struct iovec *free_iov;
689 struct sockaddr __user *uaddr;
691 struct sockaddr_storage addr;
695 struct iovec fast_iov[UIO_FASTIOV];
696 const struct iovec *free_iovec;
697 struct iov_iter iter;
699 struct wait_page_queue wpq;
703 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
704 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
705 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
706 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
707 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
708 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
710 /* first byte is taken by user flags, shift it to not overlap */
715 REQ_F_LINK_TIMEOUT_BIT,
716 REQ_F_NEED_CLEANUP_BIT,
718 REQ_F_BUFFER_SELECTED_BIT,
719 REQ_F_LTIMEOUT_ACTIVE_BIT,
720 REQ_F_COMPLETE_INLINE_BIT,
722 REQ_F_DONT_REISSUE_BIT,
723 /* keep async read/write and isreg together and in order */
724 REQ_F_ASYNC_READ_BIT,
725 REQ_F_ASYNC_WRITE_BIT,
728 /* not a real bit, just to check we're not overflowing the space */
734 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
735 /* drain existing IO first */
736 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
738 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
739 /* doesn't sever on completion < 0 */
740 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
742 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
743 /* IOSQE_BUFFER_SELECT */
744 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
746 /* fail rest of links */
747 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
748 /* on inflight list, should be cancelled and waited on exit reliably */
749 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
750 /* read/write uses file position */
751 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
752 /* must not punt to workers */
753 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
754 /* has or had linked timeout */
755 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
757 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
758 /* already went through poll handler */
759 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
760 /* buffer already selected */
761 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
762 /* linked timeout is active, i.e. prepared by link's head */
763 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
764 /* completion is deferred through io_comp_state */
765 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
766 /* caller should reissue async */
767 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
768 /* don't attempt request reissue, see io_rw_reissue() */
769 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
770 /* supports async reads */
771 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
772 /* supports async writes */
773 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
775 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
779 struct io_poll_iocb poll;
780 struct io_poll_iocb *double_poll;
783 struct io_task_work {
784 struct io_wq_work_node node;
785 task_work_func_t func;
789 IORING_RSRC_FILE = 0,
790 IORING_RSRC_BUFFER = 1,
794 * NOTE! Each of the iocb union members has the file pointer
795 * as the first entry in their struct definition. So you can
796 * access the file pointer through any of the sub-structs,
797 * or directly as just 'ki_filp' in this struct.
803 struct io_poll_iocb poll;
804 struct io_poll_update poll_update;
805 struct io_accept accept;
807 struct io_cancel cancel;
808 struct io_timeout timeout;
809 struct io_timeout_rem timeout_rem;
810 struct io_connect connect;
811 struct io_sr_msg sr_msg;
813 struct io_close close;
814 struct io_rsrc_update rsrc_update;
815 struct io_fadvise fadvise;
816 struct io_madvise madvise;
817 struct io_epoll epoll;
818 struct io_splice splice;
819 struct io_provide_buf pbuf;
820 struct io_statx statx;
821 struct io_shutdown shutdown;
822 struct io_rename rename;
823 struct io_unlink unlink;
824 /* use only after cleaning per-op data, see io_clean_op() */
825 struct io_completion compl;
828 /* opcode allocated if it needs to store data for async defer */
831 /* polled IO has completed */
837 struct io_ring_ctx *ctx;
840 struct task_struct *task;
843 struct io_kiocb *link;
844 struct percpu_ref *fixed_rsrc_refs;
846 /* used with ctx->iopoll_list with reads/writes */
847 struct list_head inflight_entry;
849 struct io_task_work io_task_work;
850 struct callback_head task_work;
852 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
853 struct hlist_node hash_node;
854 struct async_poll *apoll;
855 struct io_wq_work work;
856 /* store used ubuf, so we can prevent reloading */
857 struct io_mapped_ubuf *imu;
860 struct io_tctx_node {
861 struct list_head ctx_node;
862 struct task_struct *task;
863 struct io_ring_ctx *ctx;
866 struct io_defer_entry {
867 struct list_head list;
868 struct io_kiocb *req;
873 /* needs req->file assigned */
874 unsigned needs_file : 1;
875 /* hash wq insertion if file is a regular file */
876 unsigned hash_reg_file : 1;
877 /* unbound wq insertion if file is a non-regular file */
878 unsigned unbound_nonreg_file : 1;
879 /* opcode is not supported by this kernel */
880 unsigned not_supported : 1;
881 /* set if opcode supports polled "wait" */
883 unsigned pollout : 1;
884 /* op supports buffer selection */
885 unsigned buffer_select : 1;
886 /* do prep async if is going to be punted */
887 unsigned needs_async_setup : 1;
888 /* should block plug */
890 /* size of async data needed, if any */
891 unsigned short async_size;
894 static const struct io_op_def io_op_defs[] = {
895 [IORING_OP_NOP] = {},
896 [IORING_OP_READV] = {
898 .unbound_nonreg_file = 1,
901 .needs_async_setup = 1,
903 .async_size = sizeof(struct io_async_rw),
905 [IORING_OP_WRITEV] = {
908 .unbound_nonreg_file = 1,
910 .needs_async_setup = 1,
912 .async_size = sizeof(struct io_async_rw),
914 [IORING_OP_FSYNC] = {
917 [IORING_OP_READ_FIXED] = {
919 .unbound_nonreg_file = 1,
922 .async_size = sizeof(struct io_async_rw),
924 [IORING_OP_WRITE_FIXED] = {
927 .unbound_nonreg_file = 1,
930 .async_size = sizeof(struct io_async_rw),
932 [IORING_OP_POLL_ADD] = {
934 .unbound_nonreg_file = 1,
936 [IORING_OP_POLL_REMOVE] = {},
937 [IORING_OP_SYNC_FILE_RANGE] = {
940 [IORING_OP_SENDMSG] = {
942 .unbound_nonreg_file = 1,
944 .needs_async_setup = 1,
945 .async_size = sizeof(struct io_async_msghdr),
947 [IORING_OP_RECVMSG] = {
949 .unbound_nonreg_file = 1,
952 .needs_async_setup = 1,
953 .async_size = sizeof(struct io_async_msghdr),
955 [IORING_OP_TIMEOUT] = {
956 .async_size = sizeof(struct io_timeout_data),
958 [IORING_OP_TIMEOUT_REMOVE] = {
959 /* used by timeout updates' prep() */
961 [IORING_OP_ACCEPT] = {
963 .unbound_nonreg_file = 1,
966 [IORING_OP_ASYNC_CANCEL] = {},
967 [IORING_OP_LINK_TIMEOUT] = {
968 .async_size = sizeof(struct io_timeout_data),
970 [IORING_OP_CONNECT] = {
972 .unbound_nonreg_file = 1,
974 .needs_async_setup = 1,
975 .async_size = sizeof(struct io_async_connect),
977 [IORING_OP_FALLOCATE] = {
980 [IORING_OP_OPENAT] = {},
981 [IORING_OP_CLOSE] = {},
982 [IORING_OP_FILES_UPDATE] = {},
983 [IORING_OP_STATX] = {},
986 .unbound_nonreg_file = 1,
990 .async_size = sizeof(struct io_async_rw),
992 [IORING_OP_WRITE] = {
994 .unbound_nonreg_file = 1,
997 .async_size = sizeof(struct io_async_rw),
999 [IORING_OP_FADVISE] = {
1002 [IORING_OP_MADVISE] = {},
1003 [IORING_OP_SEND] = {
1005 .unbound_nonreg_file = 1,
1008 [IORING_OP_RECV] = {
1010 .unbound_nonreg_file = 1,
1014 [IORING_OP_OPENAT2] = {
1016 [IORING_OP_EPOLL_CTL] = {
1017 .unbound_nonreg_file = 1,
1019 [IORING_OP_SPLICE] = {
1022 .unbound_nonreg_file = 1,
1024 [IORING_OP_PROVIDE_BUFFERS] = {},
1025 [IORING_OP_REMOVE_BUFFERS] = {},
1029 .unbound_nonreg_file = 1,
1031 [IORING_OP_SHUTDOWN] = {
1034 [IORING_OP_RENAMEAT] = {},
1035 [IORING_OP_UNLINKAT] = {},
1038 static bool io_disarm_next(struct io_kiocb *req);
1039 static void io_uring_del_tctx_node(unsigned long index);
1040 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1041 struct task_struct *task,
1043 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1044 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1046 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1047 long res, unsigned int cflags);
1048 static void io_put_req(struct io_kiocb *req);
1049 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1050 static void io_dismantle_req(struct io_kiocb *req);
1051 static void io_put_task(struct task_struct *task, int nr);
1052 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1053 static void io_queue_linked_timeout(struct io_kiocb *req);
1054 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1055 struct io_uring_rsrc_update2 *up,
1057 static void io_clean_op(struct io_kiocb *req);
1058 static struct file *io_file_get(struct io_submit_state *state,
1059 struct io_kiocb *req, int fd, bool fixed);
1060 static void __io_queue_sqe(struct io_kiocb *req);
1061 static void io_rsrc_put_work(struct work_struct *work);
1063 static void io_req_task_queue(struct io_kiocb *req);
1064 static void io_submit_flush_completions(struct io_comp_state *cs,
1065 struct io_ring_ctx *ctx);
1066 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1067 static int io_req_prep_async(struct io_kiocb *req);
1069 static struct kmem_cache *req_cachep;
1071 static const struct file_operations io_uring_fops;
1073 struct sock *io_uring_get_socket(struct file *file)
1075 #if defined(CONFIG_UNIX)
1076 if (file->f_op == &io_uring_fops) {
1077 struct io_ring_ctx *ctx = file->private_data;
1079 return ctx->ring_sock->sk;
1084 EXPORT_SYMBOL(io_uring_get_socket);
1086 #define io_for_each_link(pos, head) \
1087 for (pos = (head); pos; pos = pos->link)
1089 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1091 struct io_ring_ctx *ctx = req->ctx;
1093 if (!req->fixed_rsrc_refs) {
1094 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1095 percpu_ref_get(req->fixed_rsrc_refs);
1099 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1101 bool got = percpu_ref_tryget(ref);
1103 /* already at zero, wait for ->release() */
1105 wait_for_completion(compl);
1106 percpu_ref_resurrect(ref);
1108 percpu_ref_put(ref);
1111 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1114 struct io_kiocb *req;
1116 if (task && head->task != task)
1121 io_for_each_link(req, head) {
1122 if (req->flags & REQ_F_INFLIGHT)
1128 static inline void req_set_fail(struct io_kiocb *req)
1130 req->flags |= REQ_F_FAIL;
1133 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1135 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1137 complete(&ctx->ref_comp);
1140 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1142 return !req->timeout.off;
1145 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1147 struct io_ring_ctx *ctx;
1150 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1155 * Use 5 bits less than the max cq entries, that should give us around
1156 * 32 entries per hash list if totally full and uniformly spread.
1158 hash_bits = ilog2(p->cq_entries);
1162 ctx->cancel_hash_bits = hash_bits;
1163 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1165 if (!ctx->cancel_hash)
1167 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1169 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1170 if (!ctx->dummy_ubuf)
1172 /* set invalid range, so io_import_fixed() fails meeting it */
1173 ctx->dummy_ubuf->ubuf = -1UL;
1175 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1176 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1179 ctx->flags = p->flags;
1180 init_waitqueue_head(&ctx->sqo_sq_wait);
1181 INIT_LIST_HEAD(&ctx->sqd_list);
1182 init_waitqueue_head(&ctx->cq_wait);
1183 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1184 init_completion(&ctx->ref_comp);
1185 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1186 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1187 mutex_init(&ctx->uring_lock);
1188 init_waitqueue_head(&ctx->wait);
1189 spin_lock_init(&ctx->completion_lock);
1190 INIT_LIST_HEAD(&ctx->iopoll_list);
1191 INIT_LIST_HEAD(&ctx->defer_list);
1192 INIT_LIST_HEAD(&ctx->timeout_list);
1193 spin_lock_init(&ctx->rsrc_ref_lock);
1194 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1195 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1196 init_llist_head(&ctx->rsrc_put_llist);
1197 INIT_LIST_HEAD(&ctx->tctx_list);
1198 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1199 INIT_LIST_HEAD(&ctx->locked_free_list);
1202 kfree(ctx->dummy_ubuf);
1203 kfree(ctx->cancel_hash);
1208 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1210 struct io_rings *r = ctx->rings;
1212 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1216 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1218 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1219 struct io_ring_ctx *ctx = req->ctx;
1221 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1227 static void io_req_track_inflight(struct io_kiocb *req)
1229 if (!(req->flags & REQ_F_INFLIGHT)) {
1230 req->flags |= REQ_F_INFLIGHT;
1231 atomic_inc(¤t->io_uring->inflight_tracked);
1235 static void io_prep_async_work(struct io_kiocb *req)
1237 const struct io_op_def *def = &io_op_defs[req->opcode];
1238 struct io_ring_ctx *ctx = req->ctx;
1240 if (!req->work.creds)
1241 req->work.creds = get_current_cred();
1243 req->work.list.next = NULL;
1244 req->work.flags = 0;
1245 if (req->flags & REQ_F_FORCE_ASYNC)
1246 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1248 if (req->flags & REQ_F_ISREG) {
1249 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1250 io_wq_hash_work(&req->work, file_inode(req->file));
1251 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1252 if (def->unbound_nonreg_file)
1253 req->work.flags |= IO_WQ_WORK_UNBOUND;
1256 switch (req->opcode) {
1257 case IORING_OP_SPLICE:
1259 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1260 req->work.flags |= IO_WQ_WORK_UNBOUND;
1265 static void io_prep_async_link(struct io_kiocb *req)
1267 struct io_kiocb *cur;
1269 io_for_each_link(cur, req)
1270 io_prep_async_work(cur);
1273 static void io_queue_async_work(struct io_kiocb *req)
1275 struct io_ring_ctx *ctx = req->ctx;
1276 struct io_kiocb *link = io_prep_linked_timeout(req);
1277 struct io_uring_task *tctx = req->task->io_uring;
1280 BUG_ON(!tctx->io_wq);
1282 /* init ->work of the whole link before punting */
1283 io_prep_async_link(req);
1284 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1285 &req->work, req->flags);
1286 io_wq_enqueue(tctx->io_wq, &req->work);
1288 io_queue_linked_timeout(link);
1291 static void io_kill_timeout(struct io_kiocb *req, int status)
1292 __must_hold(&req->ctx->completion_lock)
1294 struct io_timeout_data *io = req->async_data;
1296 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1297 atomic_set(&req->ctx->cq_timeouts,
1298 atomic_read(&req->ctx->cq_timeouts) + 1);
1299 list_del_init(&req->timeout.list);
1300 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1301 io_put_req_deferred(req, 1);
1305 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1308 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1309 struct io_defer_entry, list);
1311 if (req_need_defer(de->req, de->seq))
1313 list_del_init(&de->list);
1314 io_req_task_queue(de->req);
1316 } while (!list_empty(&ctx->defer_list));
1319 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1323 if (list_empty(&ctx->timeout_list))
1326 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1329 u32 events_needed, events_got;
1330 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1331 struct io_kiocb, timeout.list);
1333 if (io_is_timeout_noseq(req))
1337 * Since seq can easily wrap around over time, subtract
1338 * the last seq at which timeouts were flushed before comparing.
1339 * Assuming not more than 2^31-1 events have happened since,
1340 * these subtractions won't have wrapped, so we can check if
1341 * target is in [last_seq, current_seq] by comparing the two.
1343 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1344 events_got = seq - ctx->cq_last_tm_flush;
1345 if (events_got < events_needed)
1348 list_del_init(&req->timeout.list);
1349 io_kill_timeout(req, 0);
1350 } while (!list_empty(&ctx->timeout_list));
1352 ctx->cq_last_tm_flush = seq;
1355 static void io_commit_cqring(struct io_ring_ctx *ctx)
1357 io_flush_timeouts(ctx);
1359 /* order cqe stores with ring update */
1360 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1362 if (unlikely(!list_empty(&ctx->defer_list)))
1363 __io_queue_deferred(ctx);
1366 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1368 struct io_rings *r = ctx->rings;
1370 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1373 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1375 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1378 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1380 struct io_rings *rings = ctx->rings;
1381 unsigned tail, mask = ctx->cq_entries - 1;
1384 * writes to the cq entry need to come after reading head; the
1385 * control dependency is enough as we're using WRITE_ONCE to
1388 if (__io_cqring_events(ctx) == ctx->cq_entries)
1391 tail = ctx->cached_cq_tail++;
1392 return &rings->cqes[tail & mask];
1395 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1397 if (likely(!ctx->cq_ev_fd))
1399 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1401 return !ctx->eventfd_async || io_wq_current_is_worker();
1404 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1406 /* see waitqueue_active() comment */
1409 if (waitqueue_active(&ctx->wait))
1410 wake_up(&ctx->wait);
1411 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1412 wake_up(&ctx->sq_data->wait);
1413 if (io_should_trigger_evfd(ctx))
1414 eventfd_signal(ctx->cq_ev_fd, 1);
1415 if (waitqueue_active(&ctx->cq_wait)) {
1416 wake_up_interruptible(&ctx->cq_wait);
1417 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1421 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1423 /* see waitqueue_active() comment */
1426 if (ctx->flags & IORING_SETUP_SQPOLL) {
1427 if (waitqueue_active(&ctx->wait))
1428 wake_up(&ctx->wait);
1430 if (io_should_trigger_evfd(ctx))
1431 eventfd_signal(ctx->cq_ev_fd, 1);
1432 if (waitqueue_active(&ctx->cq_wait)) {
1433 wake_up_interruptible(&ctx->cq_wait);
1434 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1438 /* Returns true if there are no backlogged entries after the flush */
1439 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1441 unsigned long flags;
1442 bool all_flushed, posted;
1444 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1448 spin_lock_irqsave(&ctx->completion_lock, flags);
1449 while (!list_empty(&ctx->cq_overflow_list)) {
1450 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1451 struct io_overflow_cqe *ocqe;
1455 ocqe = list_first_entry(&ctx->cq_overflow_list,
1456 struct io_overflow_cqe, list);
1458 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1460 io_account_cq_overflow(ctx);
1463 list_del(&ocqe->list);
1467 all_flushed = list_empty(&ctx->cq_overflow_list);
1469 clear_bit(0, &ctx->sq_check_overflow);
1470 clear_bit(0, &ctx->cq_check_overflow);
1471 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1475 io_commit_cqring(ctx);
1476 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1478 io_cqring_ev_posted(ctx);
1482 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1486 if (test_bit(0, &ctx->cq_check_overflow)) {
1487 /* iopoll syncs against uring_lock, not completion_lock */
1488 if (ctx->flags & IORING_SETUP_IOPOLL)
1489 mutex_lock(&ctx->uring_lock);
1490 ret = __io_cqring_overflow_flush(ctx, force);
1491 if (ctx->flags & IORING_SETUP_IOPOLL)
1492 mutex_unlock(&ctx->uring_lock);
1499 * Shamelessly stolen from the mm implementation of page reference checking,
1500 * see commit f958d7b528b1 for details.
1502 #define req_ref_zero_or_close_to_overflow(req) \
1503 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1505 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1507 return atomic_inc_not_zero(&req->refs);
1510 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1512 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1513 return atomic_sub_and_test(refs, &req->refs);
1516 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1518 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1519 return atomic_dec_and_test(&req->refs);
1522 static inline void req_ref_put(struct io_kiocb *req)
1524 WARN_ON_ONCE(req_ref_put_and_test(req));
1527 static inline void req_ref_get(struct io_kiocb *req)
1529 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1530 atomic_inc(&req->refs);
1533 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1534 long res, unsigned int cflags)
1536 struct io_overflow_cqe *ocqe;
1538 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1541 * If we're in ring overflow flush mode, or in task cancel mode,
1542 * or cannot allocate an overflow entry, then we need to drop it
1545 io_account_cq_overflow(ctx);
1548 if (list_empty(&ctx->cq_overflow_list)) {
1549 set_bit(0, &ctx->sq_check_overflow);
1550 set_bit(0, &ctx->cq_check_overflow);
1551 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1553 ocqe->cqe.user_data = user_data;
1554 ocqe->cqe.res = res;
1555 ocqe->cqe.flags = cflags;
1556 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1560 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1561 long res, unsigned int cflags)
1563 struct io_uring_cqe *cqe;
1565 trace_io_uring_complete(ctx, user_data, res, cflags);
1568 * If we can't get a cq entry, userspace overflowed the
1569 * submission (by quite a lot). Increment the overflow count in
1572 cqe = io_get_cqe(ctx);
1574 WRITE_ONCE(cqe->user_data, user_data);
1575 WRITE_ONCE(cqe->res, res);
1576 WRITE_ONCE(cqe->flags, cflags);
1579 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1582 /* not as hot to bloat with inlining */
1583 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1584 long res, unsigned int cflags)
1586 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1589 static void io_req_complete_post(struct io_kiocb *req, long res,
1590 unsigned int cflags)
1592 struct io_ring_ctx *ctx = req->ctx;
1593 unsigned long flags;
1595 spin_lock_irqsave(&ctx->completion_lock, flags);
1596 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1598 * If we're the last reference to this request, add to our locked
1601 if (req_ref_put_and_test(req)) {
1602 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1603 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1604 io_disarm_next(req);
1606 io_req_task_queue(req->link);
1610 io_dismantle_req(req);
1611 io_put_task(req->task, 1);
1612 list_add(&req->compl.list, &ctx->locked_free_list);
1613 ctx->locked_free_nr++;
1615 if (!percpu_ref_tryget(&ctx->refs))
1618 io_commit_cqring(ctx);
1619 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1622 io_cqring_ev_posted(ctx);
1623 percpu_ref_put(&ctx->refs);
1627 static inline bool io_req_needs_clean(struct io_kiocb *req)
1629 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1630 REQ_F_POLLED | REQ_F_INFLIGHT);
1633 static void io_req_complete_state(struct io_kiocb *req, long res,
1634 unsigned int cflags)
1636 if (io_req_needs_clean(req))
1639 req->compl.cflags = cflags;
1640 req->flags |= REQ_F_COMPLETE_INLINE;
1643 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1644 long res, unsigned cflags)
1646 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1647 io_req_complete_state(req, res, cflags);
1649 io_req_complete_post(req, res, cflags);
1652 static inline void io_req_complete(struct io_kiocb *req, long res)
1654 __io_req_complete(req, 0, res, 0);
1657 static void io_req_complete_failed(struct io_kiocb *req, long res)
1661 io_req_complete_post(req, res, 0);
1664 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1665 struct io_comp_state *cs)
1667 spin_lock_irq(&ctx->completion_lock);
1668 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1669 ctx->locked_free_nr = 0;
1670 spin_unlock_irq(&ctx->completion_lock);
1673 /* Returns true IFF there are requests in the cache */
1674 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1676 struct io_submit_state *state = &ctx->submit_state;
1677 struct io_comp_state *cs = &state->comp;
1681 * If we have more than a batch's worth of requests in our IRQ side
1682 * locked cache, grab the lock and move them over to our submission
1685 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1686 io_flush_cached_locked_reqs(ctx, cs);
1688 nr = state->free_reqs;
1689 while (!list_empty(&cs->free_list)) {
1690 struct io_kiocb *req = list_first_entry(&cs->free_list,
1691 struct io_kiocb, compl.list);
1693 list_del(&req->compl.list);
1694 state->reqs[nr++] = req;
1695 if (nr == ARRAY_SIZE(state->reqs))
1699 state->free_reqs = nr;
1703 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1705 struct io_submit_state *state = &ctx->submit_state;
1707 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1709 if (!state->free_reqs) {
1710 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1713 if (io_flush_cached_reqs(ctx))
1716 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1720 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1721 * retry single alloc to be on the safe side.
1723 if (unlikely(ret <= 0)) {
1724 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1725 if (!state->reqs[0])
1729 state->free_reqs = ret;
1733 return state->reqs[state->free_reqs];
1736 static inline void io_put_file(struct file *file)
1742 static void io_dismantle_req(struct io_kiocb *req)
1744 unsigned int flags = req->flags;
1746 if (io_req_needs_clean(req))
1748 if (!(flags & REQ_F_FIXED_FILE))
1749 io_put_file(req->file);
1750 if (req->fixed_rsrc_refs)
1751 percpu_ref_put(req->fixed_rsrc_refs);
1752 if (req->async_data)
1753 kfree(req->async_data);
1754 if (req->work.creds) {
1755 put_cred(req->work.creds);
1756 req->work.creds = NULL;
1760 /* must to be called somewhat shortly after putting a request */
1761 static inline void io_put_task(struct task_struct *task, int nr)
1763 struct io_uring_task *tctx = task->io_uring;
1765 percpu_counter_sub(&tctx->inflight, nr);
1766 if (unlikely(atomic_read(&tctx->in_idle)))
1767 wake_up(&tctx->wait);
1768 put_task_struct_many(task, nr);
1771 static void __io_free_req(struct io_kiocb *req)
1773 struct io_ring_ctx *ctx = req->ctx;
1775 io_dismantle_req(req);
1776 io_put_task(req->task, 1);
1778 kmem_cache_free(req_cachep, req);
1779 percpu_ref_put(&ctx->refs);
1782 static inline void io_remove_next_linked(struct io_kiocb *req)
1784 struct io_kiocb *nxt = req->link;
1786 req->link = nxt->link;
1790 static bool io_kill_linked_timeout(struct io_kiocb *req)
1791 __must_hold(&req->ctx->completion_lock)
1793 struct io_kiocb *link = req->link;
1796 * Can happen if a linked timeout fired and link had been like
1797 * req -> link t-out -> link t-out [-> ...]
1799 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1800 struct io_timeout_data *io = link->async_data;
1802 io_remove_next_linked(req);
1803 link->timeout.head = NULL;
1804 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1805 io_cqring_fill_event(link->ctx, link->user_data,
1807 io_put_req_deferred(link, 1);
1814 static void io_fail_links(struct io_kiocb *req)
1815 __must_hold(&req->ctx->completion_lock)
1817 struct io_kiocb *nxt, *link = req->link;
1824 trace_io_uring_fail_link(req, link);
1825 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1826 io_put_req_deferred(link, 2);
1831 static bool io_disarm_next(struct io_kiocb *req)
1832 __must_hold(&req->ctx->completion_lock)
1834 bool posted = false;
1836 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1837 posted = io_kill_linked_timeout(req);
1838 if (unlikely((req->flags & REQ_F_FAIL) &&
1839 !(req->flags & REQ_F_HARDLINK))) {
1840 posted |= (req->link != NULL);
1846 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1848 struct io_kiocb *nxt;
1851 * If LINK is set, we have dependent requests in this chain. If we
1852 * didn't fail this request, queue the first one up, moving any other
1853 * dependencies to the next request. In case of failure, fail the rest
1856 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1857 struct io_ring_ctx *ctx = req->ctx;
1858 unsigned long flags;
1861 spin_lock_irqsave(&ctx->completion_lock, flags);
1862 posted = io_disarm_next(req);
1864 io_commit_cqring(req->ctx);
1865 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1867 io_cqring_ev_posted(ctx);
1874 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1876 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1878 return __io_req_find_next(req);
1881 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1885 if (ctx->submit_state.comp.nr) {
1886 mutex_lock(&ctx->uring_lock);
1887 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1888 mutex_unlock(&ctx->uring_lock);
1890 percpu_ref_put(&ctx->refs);
1893 static bool __tctx_task_work(struct io_uring_task *tctx)
1895 struct io_ring_ctx *ctx = NULL;
1896 struct io_wq_work_list list;
1897 struct io_wq_work_node *node;
1899 if (wq_list_empty(&tctx->task_list))
1902 spin_lock_irq(&tctx->task_lock);
1903 list = tctx->task_list;
1904 INIT_WQ_LIST(&tctx->task_list);
1905 spin_unlock_irq(&tctx->task_lock);
1909 struct io_wq_work_node *next = node->next;
1910 struct io_kiocb *req;
1912 req = container_of(node, struct io_kiocb, io_task_work.node);
1913 if (req->ctx != ctx) {
1914 ctx_flush_and_put(ctx);
1916 percpu_ref_get(&ctx->refs);
1919 req->task_work.func(&req->task_work);
1923 ctx_flush_and_put(ctx);
1924 return list.first != NULL;
1927 static void tctx_task_work(struct callback_head *cb)
1929 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1931 clear_bit(0, &tctx->task_state);
1933 while (__tctx_task_work(tctx))
1937 static int io_req_task_work_add(struct io_kiocb *req)
1939 struct task_struct *tsk = req->task;
1940 struct io_uring_task *tctx = tsk->io_uring;
1941 enum task_work_notify_mode notify;
1942 struct io_wq_work_node *node, *prev;
1943 unsigned long flags;
1946 if (unlikely(tsk->flags & PF_EXITING))
1949 WARN_ON_ONCE(!tctx);
1951 spin_lock_irqsave(&tctx->task_lock, flags);
1952 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1953 spin_unlock_irqrestore(&tctx->task_lock, flags);
1955 /* task_work already pending, we're done */
1956 if (test_bit(0, &tctx->task_state) ||
1957 test_and_set_bit(0, &tctx->task_state))
1961 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1962 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1963 * processing task_work. There's no reliable way to tell if TWA_RESUME
1966 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1968 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1969 wake_up_process(tsk);
1974 * Slow path - we failed, find and delete work. if the work is not
1975 * in the list, it got run and we're fine.
1977 spin_lock_irqsave(&tctx->task_lock, flags);
1978 wq_list_for_each(node, prev, &tctx->task_list) {
1979 if (&req->io_task_work.node == node) {
1980 wq_list_del(&tctx->task_list, node, prev);
1985 spin_unlock_irqrestore(&tctx->task_lock, flags);
1986 clear_bit(0, &tctx->task_state);
1990 static bool io_run_task_work_head(struct callback_head **work_head)
1992 struct callback_head *work, *next;
1993 bool executed = false;
1996 work = xchg(work_head, NULL);
2012 static void io_task_work_add_head(struct callback_head **work_head,
2013 struct callback_head *task_work)
2015 struct callback_head *head;
2018 head = READ_ONCE(*work_head);
2019 task_work->next = head;
2020 } while (cmpxchg(work_head, head, task_work) != head);
2023 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2024 task_work_func_t cb)
2026 init_task_work(&req->task_work, cb);
2027 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2030 static void io_req_task_cancel(struct callback_head *cb)
2032 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2033 struct io_ring_ctx *ctx = req->ctx;
2035 /* ctx is guaranteed to stay alive while we hold uring_lock */
2036 mutex_lock(&ctx->uring_lock);
2037 io_req_complete_failed(req, req->result);
2038 mutex_unlock(&ctx->uring_lock);
2041 static void __io_req_task_submit(struct io_kiocb *req)
2043 struct io_ring_ctx *ctx = req->ctx;
2045 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2046 mutex_lock(&ctx->uring_lock);
2047 if (!(current->flags & PF_EXITING) && !current->in_execve)
2048 __io_queue_sqe(req);
2050 io_req_complete_failed(req, -EFAULT);
2051 mutex_unlock(&ctx->uring_lock);
2054 static void io_req_task_submit(struct callback_head *cb)
2056 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2058 __io_req_task_submit(req);
2061 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2064 req->task_work.func = io_req_task_cancel;
2066 if (unlikely(io_req_task_work_add(req)))
2067 io_req_task_work_add_fallback(req, io_req_task_cancel);
2070 static void io_req_task_queue(struct io_kiocb *req)
2072 req->task_work.func = io_req_task_submit;
2074 if (unlikely(io_req_task_work_add(req)))
2075 io_req_task_queue_fail(req, -ECANCELED);
2078 static inline void io_queue_next(struct io_kiocb *req)
2080 struct io_kiocb *nxt = io_req_find_next(req);
2083 io_req_task_queue(nxt);
2086 static void io_free_req(struct io_kiocb *req)
2093 struct task_struct *task;
2098 static inline void io_init_req_batch(struct req_batch *rb)
2105 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2106 struct req_batch *rb)
2109 io_put_task(rb->task, rb->task_refs);
2111 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2114 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2115 struct io_submit_state *state)
2118 io_dismantle_req(req);
2120 if (req->task != rb->task) {
2122 io_put_task(rb->task, rb->task_refs);
2123 rb->task = req->task;
2129 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2130 state->reqs[state->free_reqs++] = req;
2132 list_add(&req->compl.list, &state->comp.free_list);
2135 static void io_submit_flush_completions(struct io_comp_state *cs,
2136 struct io_ring_ctx *ctx)
2139 struct io_kiocb *req;
2140 struct req_batch rb;
2142 io_init_req_batch(&rb);
2143 spin_lock_irq(&ctx->completion_lock);
2144 for (i = 0; i < nr; i++) {
2146 __io_cqring_fill_event(ctx, req->user_data, req->result,
2149 io_commit_cqring(ctx);
2150 spin_unlock_irq(&ctx->completion_lock);
2152 io_cqring_ev_posted(ctx);
2153 for (i = 0; i < nr; i++) {
2156 /* submission and completion refs */
2157 if (req_ref_sub_and_test(req, 2))
2158 io_req_free_batch(&rb, req, &ctx->submit_state);
2161 io_req_free_batch_finish(ctx, &rb);
2166 * Drop reference to request, return next in chain (if there is one) if this
2167 * was the last reference to this request.
2169 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2171 struct io_kiocb *nxt = NULL;
2173 if (req_ref_put_and_test(req)) {
2174 nxt = io_req_find_next(req);
2180 static inline void io_put_req(struct io_kiocb *req)
2182 if (req_ref_put_and_test(req))
2186 static void io_put_req_deferred_cb(struct callback_head *cb)
2188 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2193 static void io_free_req_deferred(struct io_kiocb *req)
2195 req->task_work.func = io_put_req_deferred_cb;
2196 if (unlikely(io_req_task_work_add(req)))
2197 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2200 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2202 if (req_ref_sub_and_test(req, refs))
2203 io_free_req_deferred(req);
2206 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2208 /* See comment at the top of this file */
2210 return __io_cqring_events(ctx);
2213 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2215 struct io_rings *rings = ctx->rings;
2217 /* make sure SQ entry isn't read before tail */
2218 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2221 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2223 unsigned int cflags;
2225 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2226 cflags |= IORING_CQE_F_BUFFER;
2227 req->flags &= ~REQ_F_BUFFER_SELECTED;
2232 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2234 struct io_buffer *kbuf;
2236 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2237 return io_put_kbuf(req, kbuf);
2240 static inline bool io_run_task_work(void)
2243 * Not safe to run on exiting task, and the task_work handling will
2244 * not add work to such a task.
2246 if (unlikely(current->flags & PF_EXITING))
2248 if (current->task_works) {
2249 __set_current_state(TASK_RUNNING);
2258 * Find and free completed poll iocbs
2260 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2261 struct list_head *done)
2263 struct req_batch rb;
2264 struct io_kiocb *req;
2266 /* order with ->result store in io_complete_rw_iopoll() */
2269 io_init_req_batch(&rb);
2270 while (!list_empty(done)) {
2273 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2274 list_del(&req->inflight_entry);
2276 if (READ_ONCE(req->result) == -EAGAIN &&
2277 !(req->flags & REQ_F_DONT_REISSUE)) {
2278 req->iopoll_completed = 0;
2280 io_queue_async_work(req);
2284 if (req->flags & REQ_F_BUFFER_SELECTED)
2285 cflags = io_put_rw_kbuf(req);
2287 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2290 if (req_ref_put_and_test(req))
2291 io_req_free_batch(&rb, req, &ctx->submit_state);
2294 io_commit_cqring(ctx);
2295 io_cqring_ev_posted_iopoll(ctx);
2296 io_req_free_batch_finish(ctx, &rb);
2299 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2302 struct io_kiocb *req, *tmp;
2308 * Only spin for completions if we don't have multiple devices hanging
2309 * off our complete list, and we're under the requested amount.
2311 spin = !ctx->poll_multi_file && *nr_events < min;
2314 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2315 struct kiocb *kiocb = &req->rw.kiocb;
2318 * Move completed and retryable entries to our local lists.
2319 * If we find a request that requires polling, break out
2320 * and complete those lists first, if we have entries there.
2322 if (READ_ONCE(req->iopoll_completed)) {
2323 list_move_tail(&req->inflight_entry, &done);
2326 if (!list_empty(&done))
2329 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2333 /* iopoll may have completed current req */
2334 if (READ_ONCE(req->iopoll_completed))
2335 list_move_tail(&req->inflight_entry, &done);
2342 if (!list_empty(&done))
2343 io_iopoll_complete(ctx, nr_events, &done);
2349 * We can't just wait for polled events to come to us, we have to actively
2350 * find and complete them.
2352 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2354 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2357 mutex_lock(&ctx->uring_lock);
2358 while (!list_empty(&ctx->iopoll_list)) {
2359 unsigned int nr_events = 0;
2361 io_do_iopoll(ctx, &nr_events, 0);
2363 /* let it sleep and repeat later if can't complete a request */
2367 * Ensure we allow local-to-the-cpu processing to take place,
2368 * in this case we need to ensure that we reap all events.
2369 * Also let task_work, etc. to progress by releasing the mutex
2371 if (need_resched()) {
2372 mutex_unlock(&ctx->uring_lock);
2374 mutex_lock(&ctx->uring_lock);
2377 mutex_unlock(&ctx->uring_lock);
2380 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2382 unsigned int nr_events = 0;
2386 * We disallow the app entering submit/complete with polling, but we
2387 * still need to lock the ring to prevent racing with polled issue
2388 * that got punted to a workqueue.
2390 mutex_lock(&ctx->uring_lock);
2392 * Don't enter poll loop if we already have events pending.
2393 * If we do, we can potentially be spinning for commands that
2394 * already triggered a CQE (eg in error).
2396 if (test_bit(0, &ctx->cq_check_overflow))
2397 __io_cqring_overflow_flush(ctx, false);
2398 if (io_cqring_events(ctx))
2402 * If a submit got punted to a workqueue, we can have the
2403 * application entering polling for a command before it gets
2404 * issued. That app will hold the uring_lock for the duration
2405 * of the poll right here, so we need to take a breather every
2406 * now and then to ensure that the issue has a chance to add
2407 * the poll to the issued list. Otherwise we can spin here
2408 * forever, while the workqueue is stuck trying to acquire the
2411 if (list_empty(&ctx->iopoll_list)) {
2412 mutex_unlock(&ctx->uring_lock);
2414 mutex_lock(&ctx->uring_lock);
2416 if (list_empty(&ctx->iopoll_list))
2419 ret = io_do_iopoll(ctx, &nr_events, min);
2420 } while (!ret && nr_events < min && !need_resched());
2422 mutex_unlock(&ctx->uring_lock);
2426 static void kiocb_end_write(struct io_kiocb *req)
2429 * Tell lockdep we inherited freeze protection from submission
2432 if (req->flags & REQ_F_ISREG) {
2433 struct super_block *sb = file_inode(req->file)->i_sb;
2435 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2441 static bool io_resubmit_prep(struct io_kiocb *req)
2443 struct io_async_rw *rw = req->async_data;
2446 return !io_req_prep_async(req);
2447 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2448 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2452 static bool io_rw_should_reissue(struct io_kiocb *req)
2454 umode_t mode = file_inode(req->file)->i_mode;
2455 struct io_ring_ctx *ctx = req->ctx;
2457 if (!S_ISBLK(mode) && !S_ISREG(mode))
2459 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2460 !(ctx->flags & IORING_SETUP_IOPOLL)))
2463 * If ref is dying, we might be running poll reap from the exit work.
2464 * Don't attempt to reissue from that path, just let it fail with
2467 if (percpu_ref_is_dying(&ctx->refs))
2472 static bool io_resubmit_prep(struct io_kiocb *req)
2476 static bool io_rw_should_reissue(struct io_kiocb *req)
2482 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2483 unsigned int issue_flags)
2487 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2488 kiocb_end_write(req);
2489 if (res != req->result) {
2490 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2491 io_rw_should_reissue(req)) {
2492 req->flags |= REQ_F_REISSUE;
2497 if (req->flags & REQ_F_BUFFER_SELECTED)
2498 cflags = io_put_rw_kbuf(req);
2499 __io_req_complete(req, issue_flags, res, cflags);
2502 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2504 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2506 __io_complete_rw(req, res, res2, 0);
2509 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2511 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2513 if (kiocb->ki_flags & IOCB_WRITE)
2514 kiocb_end_write(req);
2515 if (unlikely(res != req->result)) {
2516 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2517 io_resubmit_prep(req))) {
2519 req->flags |= REQ_F_DONT_REISSUE;
2523 WRITE_ONCE(req->result, res);
2524 /* order with io_iopoll_complete() checking ->result */
2526 WRITE_ONCE(req->iopoll_completed, 1);
2530 * After the iocb has been issued, it's safe to be found on the poll list.
2531 * Adding the kiocb to the list AFTER submission ensures that we don't
2532 * find it from a io_do_iopoll() thread before the issuer is done
2533 * accessing the kiocb cookie.
2535 static void io_iopoll_req_issued(struct io_kiocb *req)
2537 struct io_ring_ctx *ctx = req->ctx;
2538 const bool in_async = io_wq_current_is_worker();
2540 /* workqueue context doesn't hold uring_lock, grab it now */
2541 if (unlikely(in_async))
2542 mutex_lock(&ctx->uring_lock);
2545 * Track whether we have multiple files in our lists. This will impact
2546 * how we do polling eventually, not spinning if we're on potentially
2547 * different devices.
2549 if (list_empty(&ctx->iopoll_list)) {
2550 ctx->poll_multi_file = false;
2551 } else if (!ctx->poll_multi_file) {
2552 struct io_kiocb *list_req;
2554 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2556 if (list_req->file != req->file)
2557 ctx->poll_multi_file = true;
2561 * For fast devices, IO may have already completed. If it has, add
2562 * it to the front so we find it first.
2564 if (READ_ONCE(req->iopoll_completed))
2565 list_add(&req->inflight_entry, &ctx->iopoll_list);
2567 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2569 if (unlikely(in_async)) {
2571 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2572 * in sq thread task context or in io worker task context. If
2573 * current task context is sq thread, we don't need to check
2574 * whether should wake up sq thread.
2576 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2577 wq_has_sleeper(&ctx->sq_data->wait))
2578 wake_up(&ctx->sq_data->wait);
2580 mutex_unlock(&ctx->uring_lock);
2584 static inline void io_state_file_put(struct io_submit_state *state)
2586 if (state->file_refs) {
2587 fput_many(state->file, state->file_refs);
2588 state->file_refs = 0;
2593 * Get as many references to a file as we have IOs left in this submission,
2594 * assuming most submissions are for one file, or at least that each file
2595 * has more than one submission.
2597 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2602 if (state->file_refs) {
2603 if (state->fd == fd) {
2607 io_state_file_put(state);
2609 state->file = fget_many(fd, state->ios_left);
2610 if (unlikely(!state->file))
2614 state->file_refs = state->ios_left - 1;
2618 static bool io_bdev_nowait(struct block_device *bdev)
2620 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2624 * If we tracked the file through the SCM inflight mechanism, we could support
2625 * any file. For now, just ensure that anything potentially problematic is done
2628 static bool __io_file_supports_async(struct file *file, int rw)
2630 umode_t mode = file_inode(file)->i_mode;
2632 if (S_ISBLK(mode)) {
2633 if (IS_ENABLED(CONFIG_BLOCK) &&
2634 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2640 if (S_ISREG(mode)) {
2641 if (IS_ENABLED(CONFIG_BLOCK) &&
2642 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2643 file->f_op != &io_uring_fops)
2648 /* any ->read/write should understand O_NONBLOCK */
2649 if (file->f_flags & O_NONBLOCK)
2652 if (!(file->f_mode & FMODE_NOWAIT))
2656 return file->f_op->read_iter != NULL;
2658 return file->f_op->write_iter != NULL;
2661 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2663 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2665 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2668 return __io_file_supports_async(req->file, rw);
2671 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2673 struct io_ring_ctx *ctx = req->ctx;
2674 struct kiocb *kiocb = &req->rw.kiocb;
2675 struct file *file = req->file;
2679 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2680 req->flags |= REQ_F_ISREG;
2682 kiocb->ki_pos = READ_ONCE(sqe->off);
2683 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2684 req->flags |= REQ_F_CUR_POS;
2685 kiocb->ki_pos = file->f_pos;
2687 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2688 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2689 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2693 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2694 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2695 req->flags |= REQ_F_NOWAIT;
2697 ioprio = READ_ONCE(sqe->ioprio);
2699 ret = ioprio_check_cap(ioprio);
2703 kiocb->ki_ioprio = ioprio;
2705 kiocb->ki_ioprio = get_current_ioprio();
2707 if (ctx->flags & IORING_SETUP_IOPOLL) {
2708 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2709 !kiocb->ki_filp->f_op->iopoll)
2712 kiocb->ki_flags |= IOCB_HIPRI;
2713 kiocb->ki_complete = io_complete_rw_iopoll;
2714 req->iopoll_completed = 0;
2716 if (kiocb->ki_flags & IOCB_HIPRI)
2718 kiocb->ki_complete = io_complete_rw;
2721 if (req->opcode == IORING_OP_READ_FIXED ||
2722 req->opcode == IORING_OP_WRITE_FIXED) {
2724 io_req_set_rsrc_node(req);
2727 req->rw.addr = READ_ONCE(sqe->addr);
2728 req->rw.len = READ_ONCE(sqe->len);
2729 req->buf_index = READ_ONCE(sqe->buf_index);
2733 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2739 case -ERESTARTNOINTR:
2740 case -ERESTARTNOHAND:
2741 case -ERESTART_RESTARTBLOCK:
2743 * We can't just restart the syscall, since previously
2744 * submitted sqes may already be in progress. Just fail this
2750 kiocb->ki_complete(kiocb, ret, 0);
2754 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2755 unsigned int issue_flags)
2757 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2758 struct io_async_rw *io = req->async_data;
2759 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2761 /* add previously done IO, if any */
2762 if (io && io->bytes_done > 0) {
2764 ret = io->bytes_done;
2766 ret += io->bytes_done;
2769 if (req->flags & REQ_F_CUR_POS)
2770 req->file->f_pos = kiocb->ki_pos;
2771 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2772 __io_complete_rw(req, ret, 0, issue_flags);
2774 io_rw_done(kiocb, ret);
2776 if (check_reissue && req->flags & REQ_F_REISSUE) {
2777 req->flags &= ~REQ_F_REISSUE;
2778 if (io_resubmit_prep(req)) {
2780 io_queue_async_work(req);
2785 if (req->flags & REQ_F_BUFFER_SELECTED)
2786 cflags = io_put_rw_kbuf(req);
2787 __io_req_complete(req, issue_flags, ret, cflags);
2792 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2793 struct io_mapped_ubuf *imu)
2795 size_t len = req->rw.len;
2796 u64 buf_end, buf_addr = req->rw.addr;
2799 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2801 /* not inside the mapped region */
2802 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2806 * May not be a start of buffer, set size appropriately
2807 * and advance us to the beginning.
2809 offset = buf_addr - imu->ubuf;
2810 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2814 * Don't use iov_iter_advance() here, as it's really slow for
2815 * using the latter parts of a big fixed buffer - it iterates
2816 * over each segment manually. We can cheat a bit here, because
2819 * 1) it's a BVEC iter, we set it up
2820 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2821 * first and last bvec
2823 * So just find our index, and adjust the iterator afterwards.
2824 * If the offset is within the first bvec (or the whole first
2825 * bvec, just use iov_iter_advance(). This makes it easier
2826 * since we can just skip the first segment, which may not
2827 * be PAGE_SIZE aligned.
2829 const struct bio_vec *bvec = imu->bvec;
2831 if (offset <= bvec->bv_len) {
2832 iov_iter_advance(iter, offset);
2834 unsigned long seg_skip;
2836 /* skip first vec */
2837 offset -= bvec->bv_len;
2838 seg_skip = 1 + (offset >> PAGE_SHIFT);
2840 iter->bvec = bvec + seg_skip;
2841 iter->nr_segs -= seg_skip;
2842 iter->count -= bvec->bv_len + offset;
2843 iter->iov_offset = offset & ~PAGE_MASK;
2850 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2852 struct io_ring_ctx *ctx = req->ctx;
2853 struct io_mapped_ubuf *imu = req->imu;
2854 u16 index, buf_index = req->buf_index;
2857 if (unlikely(buf_index >= ctx->nr_user_bufs))
2859 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2860 imu = READ_ONCE(ctx->user_bufs[index]);
2863 return __io_import_fixed(req, rw, iter, imu);
2866 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2869 mutex_unlock(&ctx->uring_lock);
2872 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2875 * "Normal" inline submissions always hold the uring_lock, since we
2876 * grab it from the system call. Same is true for the SQPOLL offload.
2877 * The only exception is when we've detached the request and issue it
2878 * from an async worker thread, grab the lock for that case.
2881 mutex_lock(&ctx->uring_lock);
2884 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2885 int bgid, struct io_buffer *kbuf,
2888 struct io_buffer *head;
2890 if (req->flags & REQ_F_BUFFER_SELECTED)
2893 io_ring_submit_lock(req->ctx, needs_lock);
2895 lockdep_assert_held(&req->ctx->uring_lock);
2897 head = xa_load(&req->ctx->io_buffers, bgid);
2899 if (!list_empty(&head->list)) {
2900 kbuf = list_last_entry(&head->list, struct io_buffer,
2902 list_del(&kbuf->list);
2905 xa_erase(&req->ctx->io_buffers, bgid);
2907 if (*len > kbuf->len)
2910 kbuf = ERR_PTR(-ENOBUFS);
2913 io_ring_submit_unlock(req->ctx, needs_lock);
2918 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2921 struct io_buffer *kbuf;
2924 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2925 bgid = req->buf_index;
2926 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2929 req->rw.addr = (u64) (unsigned long) kbuf;
2930 req->flags |= REQ_F_BUFFER_SELECTED;
2931 return u64_to_user_ptr(kbuf->addr);
2934 #ifdef CONFIG_COMPAT
2935 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2938 struct compat_iovec __user *uiov;
2939 compat_ssize_t clen;
2943 uiov = u64_to_user_ptr(req->rw.addr);
2944 if (!access_ok(uiov, sizeof(*uiov)))
2946 if (__get_user(clen, &uiov->iov_len))
2952 buf = io_rw_buffer_select(req, &len, needs_lock);
2954 return PTR_ERR(buf);
2955 iov[0].iov_base = buf;
2956 iov[0].iov_len = (compat_size_t) len;
2961 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2964 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2968 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2971 len = iov[0].iov_len;
2974 buf = io_rw_buffer_select(req, &len, needs_lock);
2976 return PTR_ERR(buf);
2977 iov[0].iov_base = buf;
2978 iov[0].iov_len = len;
2982 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2985 if (req->flags & REQ_F_BUFFER_SELECTED) {
2986 struct io_buffer *kbuf;
2988 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2989 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2990 iov[0].iov_len = kbuf->len;
2993 if (req->rw.len != 1)
2996 #ifdef CONFIG_COMPAT
2997 if (req->ctx->compat)
2998 return io_compat_import(req, iov, needs_lock);
3001 return __io_iov_buffer_select(req, iov, needs_lock);
3004 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3005 struct iov_iter *iter, bool needs_lock)
3007 void __user *buf = u64_to_user_ptr(req->rw.addr);
3008 size_t sqe_len = req->rw.len;
3009 u8 opcode = req->opcode;
3012 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3014 return io_import_fixed(req, rw, iter);
3017 /* buffer index only valid with fixed read/write, or buffer select */
3018 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3021 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3022 if (req->flags & REQ_F_BUFFER_SELECT) {
3023 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3025 return PTR_ERR(buf);
3026 req->rw.len = sqe_len;
3029 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3034 if (req->flags & REQ_F_BUFFER_SELECT) {
3035 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3037 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3042 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3046 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3048 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3052 * For files that don't have ->read_iter() and ->write_iter(), handle them
3053 * by looping over ->read() or ->write() manually.
3055 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3057 struct kiocb *kiocb = &req->rw.kiocb;
3058 struct file *file = req->file;
3062 * Don't support polled IO through this interface, and we can't
3063 * support non-blocking either. For the latter, this just causes
3064 * the kiocb to be handled from an async context.
3066 if (kiocb->ki_flags & IOCB_HIPRI)
3068 if (kiocb->ki_flags & IOCB_NOWAIT)
3071 while (iov_iter_count(iter)) {
3075 if (!iov_iter_is_bvec(iter)) {
3076 iovec = iov_iter_iovec(iter);
3078 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3079 iovec.iov_len = req->rw.len;
3083 nr = file->f_op->read(file, iovec.iov_base,
3084 iovec.iov_len, io_kiocb_ppos(kiocb));
3086 nr = file->f_op->write(file, iovec.iov_base,
3087 iovec.iov_len, io_kiocb_ppos(kiocb));
3096 if (nr != iovec.iov_len)
3100 iov_iter_advance(iter, nr);
3106 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3107 const struct iovec *fast_iov, struct iov_iter *iter)
3109 struct io_async_rw *rw = req->async_data;
3111 memcpy(&rw->iter, iter, sizeof(*iter));
3112 rw->free_iovec = iovec;
3114 /* can only be fixed buffers, no need to do anything */
3115 if (iov_iter_is_bvec(iter))
3118 unsigned iov_off = 0;
3120 rw->iter.iov = rw->fast_iov;
3121 if (iter->iov != fast_iov) {
3122 iov_off = iter->iov - fast_iov;
3123 rw->iter.iov += iov_off;
3125 if (rw->fast_iov != fast_iov)
3126 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3127 sizeof(struct iovec) * iter->nr_segs);
3129 req->flags |= REQ_F_NEED_CLEANUP;
3133 static inline int io_alloc_async_data(struct io_kiocb *req)
3135 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3136 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3137 return req->async_data == NULL;
3140 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3141 const struct iovec *fast_iov,
3142 struct iov_iter *iter, bool force)
3144 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3146 if (!req->async_data) {
3147 if (io_alloc_async_data(req)) {
3152 io_req_map_rw(req, iovec, fast_iov, iter);
3157 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3159 struct io_async_rw *iorw = req->async_data;
3160 struct iovec *iov = iorw->fast_iov;
3163 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3164 if (unlikely(ret < 0))
3167 iorw->bytes_done = 0;
3168 iorw->free_iovec = iov;
3170 req->flags |= REQ_F_NEED_CLEANUP;
3174 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3176 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3178 return io_prep_rw(req, sqe);
3182 * This is our waitqueue callback handler, registered through lock_page_async()
3183 * when we initially tried to do the IO with the iocb armed our waitqueue.
3184 * This gets called when the page is unlocked, and we generally expect that to
3185 * happen when the page IO is completed and the page is now uptodate. This will
3186 * queue a task_work based retry of the operation, attempting to copy the data
3187 * again. If the latter fails because the page was NOT uptodate, then we will
3188 * do a thread based blocking retry of the operation. That's the unexpected
3191 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3192 int sync, void *arg)
3194 struct wait_page_queue *wpq;
3195 struct io_kiocb *req = wait->private;
3196 struct wait_page_key *key = arg;
3198 wpq = container_of(wait, struct wait_page_queue, wait);
3200 if (!wake_page_match(wpq, key))
3203 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3204 list_del_init(&wait->entry);
3206 /* submit ref gets dropped, acquire a new one */
3208 io_req_task_queue(req);
3213 * This controls whether a given IO request should be armed for async page
3214 * based retry. If we return false here, the request is handed to the async
3215 * worker threads for retry. If we're doing buffered reads on a regular file,
3216 * we prepare a private wait_page_queue entry and retry the operation. This
3217 * will either succeed because the page is now uptodate and unlocked, or it
3218 * will register a callback when the page is unlocked at IO completion. Through
3219 * that callback, io_uring uses task_work to setup a retry of the operation.
3220 * That retry will attempt the buffered read again. The retry will generally
3221 * succeed, or in rare cases where it fails, we then fall back to using the
3222 * async worker threads for a blocking retry.
3224 static bool io_rw_should_retry(struct io_kiocb *req)
3226 struct io_async_rw *rw = req->async_data;
3227 struct wait_page_queue *wait = &rw->wpq;
3228 struct kiocb *kiocb = &req->rw.kiocb;
3230 /* never retry for NOWAIT, we just complete with -EAGAIN */
3231 if (req->flags & REQ_F_NOWAIT)
3234 /* Only for buffered IO */
3235 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3239 * just use poll if we can, and don't attempt if the fs doesn't
3240 * support callback based unlocks
3242 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3245 wait->wait.func = io_async_buf_func;
3246 wait->wait.private = req;
3247 wait->wait.flags = 0;
3248 INIT_LIST_HEAD(&wait->wait.entry);
3249 kiocb->ki_flags |= IOCB_WAITQ;
3250 kiocb->ki_flags &= ~IOCB_NOWAIT;
3251 kiocb->ki_waitq = wait;
3255 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3257 if (req->file->f_op->read_iter)
3258 return call_read_iter(req->file, &req->rw.kiocb, iter);
3259 else if (req->file->f_op->read)
3260 return loop_rw_iter(READ, req, iter);
3265 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3267 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3268 struct kiocb *kiocb = &req->rw.kiocb;
3269 struct iov_iter __iter, *iter = &__iter;
3270 struct io_async_rw *rw = req->async_data;
3271 ssize_t io_size, ret, ret2;
3272 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3278 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3282 io_size = iov_iter_count(iter);
3283 req->result = io_size;
3285 /* Ensure we clear previously set non-block flag */
3286 if (!force_nonblock)
3287 kiocb->ki_flags &= ~IOCB_NOWAIT;
3289 kiocb->ki_flags |= IOCB_NOWAIT;
3291 /* If the file doesn't support async, just async punt */
3292 if (force_nonblock && !io_file_supports_async(req, READ)) {
3293 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3294 return ret ?: -EAGAIN;
3297 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3298 if (unlikely(ret)) {
3303 ret = io_iter_do_read(req, iter);
3305 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3306 req->flags &= ~REQ_F_REISSUE;
3307 /* IOPOLL retry should happen for io-wq threads */
3308 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3310 /* no retry on NONBLOCK nor RWF_NOWAIT */
3311 if (req->flags & REQ_F_NOWAIT)
3313 /* some cases will consume bytes even on error returns */
3314 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3316 } else if (ret == -EIOCBQUEUED) {
3318 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3319 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3320 /* read all, failed, already did sync or don't want to retry */
3324 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3329 rw = req->async_data;
3330 /* now use our persistent iterator, if we aren't already */
3335 rw->bytes_done += ret;
3336 /* if we can retry, do so with the callbacks armed */
3337 if (!io_rw_should_retry(req)) {
3338 kiocb->ki_flags &= ~IOCB_WAITQ;
3343 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3344 * we get -EIOCBQUEUED, then we'll get a notification when the
3345 * desired page gets unlocked. We can also get a partial read
3346 * here, and if we do, then just retry at the new offset.
3348 ret = io_iter_do_read(req, iter);
3349 if (ret == -EIOCBQUEUED)
3351 /* we got some bytes, but not all. retry. */
3352 kiocb->ki_flags &= ~IOCB_WAITQ;
3353 } while (ret > 0 && ret < io_size);
3355 kiocb_done(kiocb, ret, issue_flags);
3357 /* it's faster to check here then delegate to kfree */
3363 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3365 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3367 return io_prep_rw(req, sqe);
3370 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3372 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3373 struct kiocb *kiocb = &req->rw.kiocb;
3374 struct iov_iter __iter, *iter = &__iter;
3375 struct io_async_rw *rw = req->async_data;
3376 ssize_t ret, ret2, io_size;
3377 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3383 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3387 io_size = iov_iter_count(iter);
3388 req->result = io_size;
3390 /* Ensure we clear previously set non-block flag */
3391 if (!force_nonblock)
3392 kiocb->ki_flags &= ~IOCB_NOWAIT;
3394 kiocb->ki_flags |= IOCB_NOWAIT;
3396 /* If the file doesn't support async, just async punt */
3397 if (force_nonblock && !io_file_supports_async(req, WRITE))
3400 /* file path doesn't support NOWAIT for non-direct_IO */
3401 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3402 (req->flags & REQ_F_ISREG))
3405 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3410 * Open-code file_start_write here to grab freeze protection,
3411 * which will be released by another thread in
3412 * io_complete_rw(). Fool lockdep by telling it the lock got
3413 * released so that it doesn't complain about the held lock when
3414 * we return to userspace.
3416 if (req->flags & REQ_F_ISREG) {
3417 sb_start_write(file_inode(req->file)->i_sb);
3418 __sb_writers_release(file_inode(req->file)->i_sb,
3421 kiocb->ki_flags |= IOCB_WRITE;
3423 if (req->file->f_op->write_iter)
3424 ret2 = call_write_iter(req->file, kiocb, iter);
3425 else if (req->file->f_op->write)
3426 ret2 = loop_rw_iter(WRITE, req, iter);
3430 if (req->flags & REQ_F_REISSUE) {
3431 req->flags &= ~REQ_F_REISSUE;
3436 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3437 * retry them without IOCB_NOWAIT.
3439 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3441 /* no retry on NONBLOCK nor RWF_NOWAIT */
3442 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3444 if (!force_nonblock || ret2 != -EAGAIN) {
3445 /* IOPOLL retry should happen for io-wq threads */
3446 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3449 kiocb_done(kiocb, ret2, issue_flags);
3452 /* some cases will consume bytes even on error returns */
3453 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3454 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3455 return ret ?: -EAGAIN;
3458 /* it's reportedly faster than delegating the null check to kfree() */
3464 static int io_renameat_prep(struct io_kiocb *req,
3465 const struct io_uring_sqe *sqe)
3467 struct io_rename *ren = &req->rename;
3468 const char __user *oldf, *newf;
3470 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3473 ren->old_dfd = READ_ONCE(sqe->fd);
3474 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3475 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3476 ren->new_dfd = READ_ONCE(sqe->len);
3477 ren->flags = READ_ONCE(sqe->rename_flags);
3479 ren->oldpath = getname(oldf);
3480 if (IS_ERR(ren->oldpath))
3481 return PTR_ERR(ren->oldpath);
3483 ren->newpath = getname(newf);
3484 if (IS_ERR(ren->newpath)) {
3485 putname(ren->oldpath);
3486 return PTR_ERR(ren->newpath);
3489 req->flags |= REQ_F_NEED_CLEANUP;
3493 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3495 struct io_rename *ren = &req->rename;
3498 if (issue_flags & IO_URING_F_NONBLOCK)
3501 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3502 ren->newpath, ren->flags);
3504 req->flags &= ~REQ_F_NEED_CLEANUP;
3507 io_req_complete(req, ret);
3511 static int io_unlinkat_prep(struct io_kiocb *req,
3512 const struct io_uring_sqe *sqe)
3514 struct io_unlink *un = &req->unlink;
3515 const char __user *fname;
3517 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3520 un->dfd = READ_ONCE(sqe->fd);
3522 un->flags = READ_ONCE(sqe->unlink_flags);
3523 if (un->flags & ~AT_REMOVEDIR)
3526 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3527 un->filename = getname(fname);
3528 if (IS_ERR(un->filename))
3529 return PTR_ERR(un->filename);
3531 req->flags |= REQ_F_NEED_CLEANUP;
3535 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3537 struct io_unlink *un = &req->unlink;
3540 if (issue_flags & IO_URING_F_NONBLOCK)
3543 if (un->flags & AT_REMOVEDIR)
3544 ret = do_rmdir(un->dfd, un->filename);
3546 ret = do_unlinkat(un->dfd, un->filename);
3548 req->flags &= ~REQ_F_NEED_CLEANUP;
3551 io_req_complete(req, ret);
3555 static int io_shutdown_prep(struct io_kiocb *req,
3556 const struct io_uring_sqe *sqe)
3558 #if defined(CONFIG_NET)
3559 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3561 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3565 req->shutdown.how = READ_ONCE(sqe->len);
3572 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3574 #if defined(CONFIG_NET)
3575 struct socket *sock;
3578 if (issue_flags & IO_URING_F_NONBLOCK)
3581 sock = sock_from_file(req->file);
3582 if (unlikely(!sock))
3585 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3588 io_req_complete(req, ret);
3595 static int __io_splice_prep(struct io_kiocb *req,
3596 const struct io_uring_sqe *sqe)
3598 struct io_splice* sp = &req->splice;
3599 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3601 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3605 sp->len = READ_ONCE(sqe->len);
3606 sp->flags = READ_ONCE(sqe->splice_flags);
3608 if (unlikely(sp->flags & ~valid_flags))
3611 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3612 (sp->flags & SPLICE_F_FD_IN_FIXED));
3615 req->flags |= REQ_F_NEED_CLEANUP;
3619 static int io_tee_prep(struct io_kiocb *req,
3620 const struct io_uring_sqe *sqe)
3622 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3624 return __io_splice_prep(req, sqe);
3627 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3629 struct io_splice *sp = &req->splice;
3630 struct file *in = sp->file_in;
3631 struct file *out = sp->file_out;
3632 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3635 if (issue_flags & IO_URING_F_NONBLOCK)
3638 ret = do_tee(in, out, sp->len, flags);
3640 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3642 req->flags &= ~REQ_F_NEED_CLEANUP;
3646 io_req_complete(req, ret);
3650 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3652 struct io_splice* sp = &req->splice;
3654 sp->off_in = READ_ONCE(sqe->splice_off_in);
3655 sp->off_out = READ_ONCE(sqe->off);
3656 return __io_splice_prep(req, sqe);
3659 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3661 struct io_splice *sp = &req->splice;
3662 struct file *in = sp->file_in;
3663 struct file *out = sp->file_out;
3664 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3665 loff_t *poff_in, *poff_out;
3668 if (issue_flags & IO_URING_F_NONBLOCK)
3671 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3672 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3675 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3677 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3679 req->flags &= ~REQ_F_NEED_CLEANUP;
3683 io_req_complete(req, ret);
3688 * IORING_OP_NOP just posts a completion event, nothing else.
3690 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3692 struct io_ring_ctx *ctx = req->ctx;
3694 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3697 __io_req_complete(req, issue_flags, 0, 0);
3701 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3703 struct io_ring_ctx *ctx = req->ctx;
3708 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3710 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3713 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3714 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3717 req->sync.off = READ_ONCE(sqe->off);
3718 req->sync.len = READ_ONCE(sqe->len);
3722 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3724 loff_t end = req->sync.off + req->sync.len;
3727 /* fsync always requires a blocking context */
3728 if (issue_flags & IO_URING_F_NONBLOCK)
3731 ret = vfs_fsync_range(req->file, req->sync.off,
3732 end > 0 ? end : LLONG_MAX,
3733 req->sync.flags & IORING_FSYNC_DATASYNC);
3736 io_req_complete(req, ret);
3740 static int io_fallocate_prep(struct io_kiocb *req,
3741 const struct io_uring_sqe *sqe)
3743 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3745 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3748 req->sync.off = READ_ONCE(sqe->off);
3749 req->sync.len = READ_ONCE(sqe->addr);
3750 req->sync.mode = READ_ONCE(sqe->len);
3754 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3758 /* fallocate always requiring blocking context */
3759 if (issue_flags & IO_URING_F_NONBLOCK)
3761 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3765 io_req_complete(req, ret);
3769 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3771 const char __user *fname;
3774 if (unlikely(sqe->ioprio || sqe->buf_index))
3776 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3779 /* open.how should be already initialised */
3780 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3781 req->open.how.flags |= O_LARGEFILE;
3783 req->open.dfd = READ_ONCE(sqe->fd);
3784 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3785 req->open.filename = getname(fname);
3786 if (IS_ERR(req->open.filename)) {
3787 ret = PTR_ERR(req->open.filename);
3788 req->open.filename = NULL;
3791 req->open.nofile = rlimit(RLIMIT_NOFILE);
3792 req->flags |= REQ_F_NEED_CLEANUP;
3796 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3800 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3802 mode = READ_ONCE(sqe->len);
3803 flags = READ_ONCE(sqe->open_flags);
3804 req->open.how = build_open_how(flags, mode);
3805 return __io_openat_prep(req, sqe);
3808 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3810 struct open_how __user *how;
3814 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3816 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3817 len = READ_ONCE(sqe->len);
3818 if (len < OPEN_HOW_SIZE_VER0)
3821 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3826 return __io_openat_prep(req, sqe);
3829 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3831 struct open_flags op;
3834 bool resolve_nonblock;
3837 ret = build_open_flags(&req->open.how, &op);
3840 nonblock_set = op.open_flag & O_NONBLOCK;
3841 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3842 if (issue_flags & IO_URING_F_NONBLOCK) {
3844 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3845 * it'll always -EAGAIN
3847 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3849 op.lookup_flags |= LOOKUP_CACHED;
3850 op.open_flag |= O_NONBLOCK;
3853 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3857 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3858 /* only retry if RESOLVE_CACHED wasn't already set by application */
3859 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3860 file == ERR_PTR(-EAGAIN)) {
3862 * We could hang on to this 'fd', but seems like marginal
3863 * gain for something that is now known to be a slower path.
3864 * So just put it, and we'll get a new one when we retry.
3872 ret = PTR_ERR(file);
3874 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3875 file->f_flags &= ~O_NONBLOCK;
3876 fsnotify_open(file);
3877 fd_install(ret, file);
3880 putname(req->open.filename);
3881 req->flags &= ~REQ_F_NEED_CLEANUP;
3884 __io_req_complete(req, issue_flags, ret, 0);
3888 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3890 return io_openat2(req, issue_flags);
3893 static int io_remove_buffers_prep(struct io_kiocb *req,
3894 const struct io_uring_sqe *sqe)
3896 struct io_provide_buf *p = &req->pbuf;
3899 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3902 tmp = READ_ONCE(sqe->fd);
3903 if (!tmp || tmp > USHRT_MAX)
3906 memset(p, 0, sizeof(*p));
3908 p->bgid = READ_ONCE(sqe->buf_group);
3912 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3913 int bgid, unsigned nbufs)
3917 /* shouldn't happen */
3921 /* the head kbuf is the list itself */
3922 while (!list_empty(&buf->list)) {
3923 struct io_buffer *nxt;
3925 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3926 list_del(&nxt->list);
3933 xa_erase(&ctx->io_buffers, bgid);
3938 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3940 struct io_provide_buf *p = &req->pbuf;
3941 struct io_ring_ctx *ctx = req->ctx;
3942 struct io_buffer *head;
3944 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3946 io_ring_submit_lock(ctx, !force_nonblock);
3948 lockdep_assert_held(&ctx->uring_lock);
3951 head = xa_load(&ctx->io_buffers, p->bgid);
3953 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3957 /* complete before unlock, IOPOLL may need the lock */
3958 __io_req_complete(req, issue_flags, ret, 0);
3959 io_ring_submit_unlock(ctx, !force_nonblock);
3963 static int io_provide_buffers_prep(struct io_kiocb *req,
3964 const struct io_uring_sqe *sqe)
3966 unsigned long size, tmp_check;
3967 struct io_provide_buf *p = &req->pbuf;
3970 if (sqe->ioprio || sqe->rw_flags)
3973 tmp = READ_ONCE(sqe->fd);
3974 if (!tmp || tmp > USHRT_MAX)
3977 p->addr = READ_ONCE(sqe->addr);
3978 p->len = READ_ONCE(sqe->len);
3980 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3983 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3986 size = (unsigned long)p->len * p->nbufs;
3987 if (!access_ok(u64_to_user_ptr(p->addr), size))
3990 p->bgid = READ_ONCE(sqe->buf_group);
3991 tmp = READ_ONCE(sqe->off);
3992 if (tmp > USHRT_MAX)
3998 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4000 struct io_buffer *buf;
4001 u64 addr = pbuf->addr;
4002 int i, bid = pbuf->bid;
4004 for (i = 0; i < pbuf->nbufs; i++) {
4005 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4010 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4015 INIT_LIST_HEAD(&buf->list);
4018 list_add_tail(&buf->list, &(*head)->list);
4022 return i ? i : -ENOMEM;
4025 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4027 struct io_provide_buf *p = &req->pbuf;
4028 struct io_ring_ctx *ctx = req->ctx;
4029 struct io_buffer *head, *list;
4031 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4033 io_ring_submit_lock(ctx, !force_nonblock);
4035 lockdep_assert_held(&ctx->uring_lock);
4037 list = head = xa_load(&ctx->io_buffers, p->bgid);
4039 ret = io_add_buffers(p, &head);
4040 if (ret >= 0 && !list) {
4041 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4043 __io_remove_buffers(ctx, head, p->bgid, -1U);
4047 /* complete before unlock, IOPOLL may need the lock */
4048 __io_req_complete(req, issue_flags, ret, 0);
4049 io_ring_submit_unlock(ctx, !force_nonblock);
4053 static int io_epoll_ctl_prep(struct io_kiocb *req,
4054 const struct io_uring_sqe *sqe)
4056 #if defined(CONFIG_EPOLL)
4057 if (sqe->ioprio || sqe->buf_index)
4059 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4062 req->epoll.epfd = READ_ONCE(sqe->fd);
4063 req->epoll.op = READ_ONCE(sqe->len);
4064 req->epoll.fd = READ_ONCE(sqe->off);
4066 if (ep_op_has_event(req->epoll.op)) {
4067 struct epoll_event __user *ev;
4069 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4070 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4080 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4082 #if defined(CONFIG_EPOLL)
4083 struct io_epoll *ie = &req->epoll;
4085 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4087 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4088 if (force_nonblock && ret == -EAGAIN)
4093 __io_req_complete(req, issue_flags, ret, 0);
4100 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4102 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4103 if (sqe->ioprio || sqe->buf_index || sqe->off)
4105 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4108 req->madvise.addr = READ_ONCE(sqe->addr);
4109 req->madvise.len = READ_ONCE(sqe->len);
4110 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4117 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4119 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4120 struct io_madvise *ma = &req->madvise;
4123 if (issue_flags & IO_URING_F_NONBLOCK)
4126 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4129 io_req_complete(req, ret);
4136 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4138 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4140 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4143 req->fadvise.offset = READ_ONCE(sqe->off);
4144 req->fadvise.len = READ_ONCE(sqe->len);
4145 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4149 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4151 struct io_fadvise *fa = &req->fadvise;
4154 if (issue_flags & IO_URING_F_NONBLOCK) {
4155 switch (fa->advice) {
4156 case POSIX_FADV_NORMAL:
4157 case POSIX_FADV_RANDOM:
4158 case POSIX_FADV_SEQUENTIAL:
4165 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4168 __io_req_complete(req, issue_flags, ret, 0);
4172 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4174 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4176 if (sqe->ioprio || sqe->buf_index)
4178 if (req->flags & REQ_F_FIXED_FILE)
4181 req->statx.dfd = READ_ONCE(sqe->fd);
4182 req->statx.mask = READ_ONCE(sqe->len);
4183 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4184 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4185 req->statx.flags = READ_ONCE(sqe->statx_flags);
4190 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4192 struct io_statx *ctx = &req->statx;
4195 if (issue_flags & IO_URING_F_NONBLOCK)
4198 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4203 io_req_complete(req, ret);
4207 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4209 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4211 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4212 sqe->rw_flags || sqe->buf_index)
4214 if (req->flags & REQ_F_FIXED_FILE)
4217 req->close.fd = READ_ONCE(sqe->fd);
4221 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4223 struct files_struct *files = current->files;
4224 struct io_close *close = &req->close;
4225 struct fdtable *fdt;
4226 struct file *file = NULL;
4229 spin_lock(&files->file_lock);
4230 fdt = files_fdtable(files);
4231 if (close->fd >= fdt->max_fds) {
4232 spin_unlock(&files->file_lock);
4235 file = fdt->fd[close->fd];
4236 if (!file || file->f_op == &io_uring_fops) {
4237 spin_unlock(&files->file_lock);
4242 /* if the file has a flush method, be safe and punt to async */
4243 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4244 spin_unlock(&files->file_lock);
4248 ret = __close_fd_get_file(close->fd, &file);
4249 spin_unlock(&files->file_lock);
4256 /* No ->flush() or already async, safely close from here */
4257 ret = filp_close(file, current->files);
4263 __io_req_complete(req, issue_flags, ret, 0);
4267 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4269 struct io_ring_ctx *ctx = req->ctx;
4271 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4273 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4276 req->sync.off = READ_ONCE(sqe->off);
4277 req->sync.len = READ_ONCE(sqe->len);
4278 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4282 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4286 /* sync_file_range always requires a blocking context */
4287 if (issue_flags & IO_URING_F_NONBLOCK)
4290 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4294 io_req_complete(req, ret);
4298 #if defined(CONFIG_NET)
4299 static int io_setup_async_msg(struct io_kiocb *req,
4300 struct io_async_msghdr *kmsg)
4302 struct io_async_msghdr *async_msg = req->async_data;
4306 if (io_alloc_async_data(req)) {
4307 kfree(kmsg->free_iov);
4310 async_msg = req->async_data;
4311 req->flags |= REQ_F_NEED_CLEANUP;
4312 memcpy(async_msg, kmsg, sizeof(*kmsg));
4313 async_msg->msg.msg_name = &async_msg->addr;
4314 /* if were using fast_iov, set it to the new one */
4315 if (!async_msg->free_iov)
4316 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4321 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4322 struct io_async_msghdr *iomsg)
4324 iomsg->msg.msg_name = &iomsg->addr;
4325 iomsg->free_iov = iomsg->fast_iov;
4326 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4327 req->sr_msg.msg_flags, &iomsg->free_iov);
4330 static int io_sendmsg_prep_async(struct io_kiocb *req)
4334 ret = io_sendmsg_copy_hdr(req, req->async_data);
4336 req->flags |= REQ_F_NEED_CLEANUP;
4340 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4342 struct io_sr_msg *sr = &req->sr_msg;
4344 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4347 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4348 sr->len = READ_ONCE(sqe->len);
4349 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4350 if (sr->msg_flags & MSG_DONTWAIT)
4351 req->flags |= REQ_F_NOWAIT;
4353 #ifdef CONFIG_COMPAT
4354 if (req->ctx->compat)
4355 sr->msg_flags |= MSG_CMSG_COMPAT;
4360 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4362 struct io_async_msghdr iomsg, *kmsg;
4363 struct socket *sock;
4368 sock = sock_from_file(req->file);
4369 if (unlikely(!sock))
4372 kmsg = req->async_data;
4374 ret = io_sendmsg_copy_hdr(req, &iomsg);
4380 flags = req->sr_msg.msg_flags;
4381 if (issue_flags & IO_URING_F_NONBLOCK)
4382 flags |= MSG_DONTWAIT;
4383 if (flags & MSG_WAITALL)
4384 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4386 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4387 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4388 return io_setup_async_msg(req, kmsg);
4389 if (ret == -ERESTARTSYS)
4392 /* fast path, check for non-NULL to avoid function call */
4394 kfree(kmsg->free_iov);
4395 req->flags &= ~REQ_F_NEED_CLEANUP;
4398 __io_req_complete(req, issue_flags, ret, 0);
4402 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4404 struct io_sr_msg *sr = &req->sr_msg;
4407 struct socket *sock;
4412 sock = sock_from_file(req->file);
4413 if (unlikely(!sock))
4416 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4420 msg.msg_name = NULL;
4421 msg.msg_control = NULL;
4422 msg.msg_controllen = 0;
4423 msg.msg_namelen = 0;
4425 flags = req->sr_msg.msg_flags;
4426 if (issue_flags & IO_URING_F_NONBLOCK)
4427 flags |= MSG_DONTWAIT;
4428 if (flags & MSG_WAITALL)
4429 min_ret = iov_iter_count(&msg.msg_iter);
4431 msg.msg_flags = flags;
4432 ret = sock_sendmsg(sock, &msg);
4433 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4435 if (ret == -ERESTARTSYS)
4440 __io_req_complete(req, issue_flags, ret, 0);
4444 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4445 struct io_async_msghdr *iomsg)
4447 struct io_sr_msg *sr = &req->sr_msg;
4448 struct iovec __user *uiov;
4452 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4453 &iomsg->uaddr, &uiov, &iov_len);
4457 if (req->flags & REQ_F_BUFFER_SELECT) {
4460 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4462 sr->len = iomsg->fast_iov[0].iov_len;
4463 iomsg->free_iov = NULL;
4465 iomsg->free_iov = iomsg->fast_iov;
4466 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4467 &iomsg->free_iov, &iomsg->msg.msg_iter,
4476 #ifdef CONFIG_COMPAT
4477 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4478 struct io_async_msghdr *iomsg)
4480 struct io_sr_msg *sr = &req->sr_msg;
4481 struct compat_iovec __user *uiov;
4486 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4491 uiov = compat_ptr(ptr);
4492 if (req->flags & REQ_F_BUFFER_SELECT) {
4493 compat_ssize_t clen;
4497 if (!access_ok(uiov, sizeof(*uiov)))
4499 if (__get_user(clen, &uiov->iov_len))
4504 iomsg->free_iov = NULL;
4506 iomsg->free_iov = iomsg->fast_iov;
4507 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4508 UIO_FASTIOV, &iomsg->free_iov,
4509 &iomsg->msg.msg_iter, true);
4518 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4519 struct io_async_msghdr *iomsg)
4521 iomsg->msg.msg_name = &iomsg->addr;
4523 #ifdef CONFIG_COMPAT
4524 if (req->ctx->compat)
4525 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4528 return __io_recvmsg_copy_hdr(req, iomsg);
4531 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4534 struct io_sr_msg *sr = &req->sr_msg;
4535 struct io_buffer *kbuf;
4537 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4542 req->flags |= REQ_F_BUFFER_SELECTED;
4546 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4548 return io_put_kbuf(req, req->sr_msg.kbuf);
4551 static int io_recvmsg_prep_async(struct io_kiocb *req)
4555 ret = io_recvmsg_copy_hdr(req, req->async_data);
4557 req->flags |= REQ_F_NEED_CLEANUP;
4561 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4563 struct io_sr_msg *sr = &req->sr_msg;
4565 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4568 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4569 sr->len = READ_ONCE(sqe->len);
4570 sr->bgid = READ_ONCE(sqe->buf_group);
4571 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4572 if (sr->msg_flags & MSG_DONTWAIT)
4573 req->flags |= REQ_F_NOWAIT;
4575 #ifdef CONFIG_COMPAT
4576 if (req->ctx->compat)
4577 sr->msg_flags |= MSG_CMSG_COMPAT;
4582 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4584 struct io_async_msghdr iomsg, *kmsg;
4585 struct socket *sock;
4586 struct io_buffer *kbuf;
4589 int ret, cflags = 0;
4590 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4592 sock = sock_from_file(req->file);
4593 if (unlikely(!sock))
4596 kmsg = req->async_data;
4598 ret = io_recvmsg_copy_hdr(req, &iomsg);
4604 if (req->flags & REQ_F_BUFFER_SELECT) {
4605 kbuf = io_recv_buffer_select(req, !force_nonblock);
4607 return PTR_ERR(kbuf);
4608 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4609 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4610 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4611 1, req->sr_msg.len);
4614 flags = req->sr_msg.msg_flags;
4616 flags |= MSG_DONTWAIT;
4617 if (flags & MSG_WAITALL)
4618 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4620 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4621 kmsg->uaddr, flags);
4622 if (force_nonblock && ret == -EAGAIN)
4623 return io_setup_async_msg(req, kmsg);
4624 if (ret == -ERESTARTSYS)
4627 if (req->flags & REQ_F_BUFFER_SELECTED)
4628 cflags = io_put_recv_kbuf(req);
4629 /* fast path, check for non-NULL to avoid function call */
4631 kfree(kmsg->free_iov);
4632 req->flags &= ~REQ_F_NEED_CLEANUP;
4633 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4635 __io_req_complete(req, issue_flags, ret, cflags);
4639 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4641 struct io_buffer *kbuf;
4642 struct io_sr_msg *sr = &req->sr_msg;
4644 void __user *buf = sr->buf;
4645 struct socket *sock;
4649 int ret, cflags = 0;
4650 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4652 sock = sock_from_file(req->file);
4653 if (unlikely(!sock))
4656 if (req->flags & REQ_F_BUFFER_SELECT) {
4657 kbuf = io_recv_buffer_select(req, !force_nonblock);
4659 return PTR_ERR(kbuf);
4660 buf = u64_to_user_ptr(kbuf->addr);
4663 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4667 msg.msg_name = NULL;
4668 msg.msg_control = NULL;
4669 msg.msg_controllen = 0;
4670 msg.msg_namelen = 0;
4671 msg.msg_iocb = NULL;
4674 flags = req->sr_msg.msg_flags;
4676 flags |= MSG_DONTWAIT;
4677 if (flags & MSG_WAITALL)
4678 min_ret = iov_iter_count(&msg.msg_iter);
4680 ret = sock_recvmsg(sock, &msg, flags);
4681 if (force_nonblock && ret == -EAGAIN)
4683 if (ret == -ERESTARTSYS)
4686 if (req->flags & REQ_F_BUFFER_SELECTED)
4687 cflags = io_put_recv_kbuf(req);
4688 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4690 __io_req_complete(req, issue_flags, ret, cflags);
4694 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4696 struct io_accept *accept = &req->accept;
4698 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4700 if (sqe->ioprio || sqe->len || sqe->buf_index)
4703 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4704 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4705 accept->flags = READ_ONCE(sqe->accept_flags);
4706 accept->nofile = rlimit(RLIMIT_NOFILE);
4710 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4712 struct io_accept *accept = &req->accept;
4713 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4714 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4717 if (req->file->f_flags & O_NONBLOCK)
4718 req->flags |= REQ_F_NOWAIT;
4720 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4721 accept->addr_len, accept->flags,
4723 if (ret == -EAGAIN && force_nonblock)
4726 if (ret == -ERESTARTSYS)
4730 __io_req_complete(req, issue_flags, ret, 0);
4734 static int io_connect_prep_async(struct io_kiocb *req)
4736 struct io_async_connect *io = req->async_data;
4737 struct io_connect *conn = &req->connect;
4739 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4742 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4744 struct io_connect *conn = &req->connect;
4746 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4748 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4751 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4752 conn->addr_len = READ_ONCE(sqe->addr2);
4756 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4758 struct io_async_connect __io, *io;
4759 unsigned file_flags;
4761 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4763 if (req->async_data) {
4764 io = req->async_data;
4766 ret = move_addr_to_kernel(req->connect.addr,
4767 req->connect.addr_len,
4774 file_flags = force_nonblock ? O_NONBLOCK : 0;
4776 ret = __sys_connect_file(req->file, &io->address,
4777 req->connect.addr_len, file_flags);
4778 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4779 if (req->async_data)
4781 if (io_alloc_async_data(req)) {
4785 memcpy(req->async_data, &__io, sizeof(__io));
4788 if (ret == -ERESTARTSYS)
4793 __io_req_complete(req, issue_flags, ret, 0);
4796 #else /* !CONFIG_NET */
4797 #define IO_NETOP_FN(op) \
4798 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4800 return -EOPNOTSUPP; \
4803 #define IO_NETOP_PREP(op) \
4805 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4807 return -EOPNOTSUPP; \
4810 #define IO_NETOP_PREP_ASYNC(op) \
4812 static int io_##op##_prep_async(struct io_kiocb *req) \
4814 return -EOPNOTSUPP; \
4817 IO_NETOP_PREP_ASYNC(sendmsg);
4818 IO_NETOP_PREP_ASYNC(recvmsg);
4819 IO_NETOP_PREP_ASYNC(connect);
4820 IO_NETOP_PREP(accept);
4823 #endif /* CONFIG_NET */
4825 struct io_poll_table {
4826 struct poll_table_struct pt;
4827 struct io_kiocb *req;
4831 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4832 __poll_t mask, task_work_func_t func)
4836 /* for instances that support it check for an event match first: */
4837 if (mask && !(mask & poll->events))
4840 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4842 list_del_init(&poll->wait.entry);
4845 req->task_work.func = func;
4848 * If this fails, then the task is exiting. When a task exits, the
4849 * work gets canceled, so just cancel this request as well instead
4850 * of executing it. We can't safely execute it anyway, as we may not
4851 * have the needed state needed for it anyway.
4853 ret = io_req_task_work_add(req);
4854 if (unlikely(ret)) {
4855 WRITE_ONCE(poll->canceled, true);
4856 io_req_task_work_add_fallback(req, func);
4861 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4862 __acquires(&req->ctx->completion_lock)
4864 struct io_ring_ctx *ctx = req->ctx;
4866 if (!req->result && !READ_ONCE(poll->canceled)) {
4867 struct poll_table_struct pt = { ._key = poll->events };
4869 req->result = vfs_poll(req->file, &pt) & poll->events;
4872 spin_lock_irq(&ctx->completion_lock);
4873 if (!req->result && !READ_ONCE(poll->canceled)) {
4874 add_wait_queue(poll->head, &poll->wait);
4881 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4883 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4884 if (req->opcode == IORING_OP_POLL_ADD)
4885 return req->async_data;
4886 return req->apoll->double_poll;
4889 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4891 if (req->opcode == IORING_OP_POLL_ADD)
4893 return &req->apoll->poll;
4896 static void io_poll_remove_double(struct io_kiocb *req)
4897 __must_hold(&req->ctx->completion_lock)
4899 struct io_poll_iocb *poll = io_poll_get_double(req);
4901 lockdep_assert_held(&req->ctx->completion_lock);
4903 if (poll && poll->head) {
4904 struct wait_queue_head *head = poll->head;
4906 spin_lock(&head->lock);
4907 list_del_init(&poll->wait.entry);
4908 if (poll->wait.private)
4911 spin_unlock(&head->lock);
4915 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4916 __must_hold(&req->ctx->completion_lock)
4918 struct io_ring_ctx *ctx = req->ctx;
4919 unsigned flags = IORING_CQE_F_MORE;
4922 if (READ_ONCE(req->poll.canceled)) {
4924 req->poll.events |= EPOLLONESHOT;
4926 error = mangle_poll(mask);
4928 if (req->poll.events & EPOLLONESHOT)
4930 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4931 io_poll_remove_waitqs(req);
4932 req->poll.done = true;
4935 if (flags & IORING_CQE_F_MORE)
4938 io_commit_cqring(ctx);
4939 return !(flags & IORING_CQE_F_MORE);
4942 static void io_poll_task_func(struct callback_head *cb)
4944 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4945 struct io_ring_ctx *ctx = req->ctx;
4946 struct io_kiocb *nxt;
4948 if (io_poll_rewait(req, &req->poll)) {
4949 spin_unlock_irq(&ctx->completion_lock);
4953 done = io_poll_complete(req, req->result);
4955 hash_del(&req->hash_node);
4958 add_wait_queue(req->poll.head, &req->poll.wait);
4960 spin_unlock_irq(&ctx->completion_lock);
4961 io_cqring_ev_posted(ctx);
4964 nxt = io_put_req_find_next(req);
4966 __io_req_task_submit(nxt);
4971 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4972 int sync, void *key)
4974 struct io_kiocb *req = wait->private;
4975 struct io_poll_iocb *poll = io_poll_get_single(req);
4976 __poll_t mask = key_to_poll(key);
4978 /* for instances that support it check for an event match first: */
4979 if (mask && !(mask & poll->events))
4981 if (!(poll->events & EPOLLONESHOT))
4982 return poll->wait.func(&poll->wait, mode, sync, key);
4984 list_del_init(&wait->entry);
4986 if (poll && poll->head) {
4989 spin_lock(&poll->head->lock);
4990 done = list_empty(&poll->wait.entry);
4992 list_del_init(&poll->wait.entry);
4993 /* make sure double remove sees this as being gone */
4994 wait->private = NULL;
4995 spin_unlock(&poll->head->lock);
4997 /* use wait func handler, so it matches the rq type */
4998 poll->wait.func(&poll->wait, mode, sync, key);
5005 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5006 wait_queue_func_t wake_func)
5010 poll->canceled = false;
5011 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5012 /* mask in events that we always want/need */
5013 poll->events = events | IO_POLL_UNMASK;
5014 INIT_LIST_HEAD(&poll->wait.entry);
5015 init_waitqueue_func_entry(&poll->wait, wake_func);
5018 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5019 struct wait_queue_head *head,
5020 struct io_poll_iocb **poll_ptr)
5022 struct io_kiocb *req = pt->req;
5025 * If poll->head is already set, it's because the file being polled
5026 * uses multiple waitqueues for poll handling (eg one for read, one
5027 * for write). Setup a separate io_poll_iocb if this happens.
5029 if (unlikely(poll->head)) {
5030 struct io_poll_iocb *poll_one = poll;
5032 /* already have a 2nd entry, fail a third attempt */
5034 pt->error = -EINVAL;
5038 * Can't handle multishot for double wait for now, turn it
5039 * into one-shot mode.
5041 if (!(poll_one->events & EPOLLONESHOT))
5042 poll_one->events |= EPOLLONESHOT;
5043 /* double add on the same waitqueue head, ignore */
5044 if (poll_one->head == head)
5046 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5048 pt->error = -ENOMEM;
5051 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5053 poll->wait.private = req;
5060 if (poll->events & EPOLLEXCLUSIVE)
5061 add_wait_queue_exclusive(head, &poll->wait);
5063 add_wait_queue(head, &poll->wait);
5066 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5067 struct poll_table_struct *p)
5069 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5070 struct async_poll *apoll = pt->req->apoll;
5072 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5075 static void io_async_task_func(struct callback_head *cb)
5077 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5078 struct async_poll *apoll = req->apoll;
5079 struct io_ring_ctx *ctx = req->ctx;
5081 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5083 if (io_poll_rewait(req, &apoll->poll)) {
5084 spin_unlock_irq(&ctx->completion_lock);
5088 hash_del(&req->hash_node);
5089 io_poll_remove_double(req);
5090 spin_unlock_irq(&ctx->completion_lock);
5092 if (!READ_ONCE(apoll->poll.canceled))
5093 __io_req_task_submit(req);
5095 io_req_complete_failed(req, -ECANCELED);
5098 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5101 struct io_kiocb *req = wait->private;
5102 struct io_poll_iocb *poll = &req->apoll->poll;
5104 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5107 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5110 static void io_poll_req_insert(struct io_kiocb *req)
5112 struct io_ring_ctx *ctx = req->ctx;
5113 struct hlist_head *list;
5115 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5116 hlist_add_head(&req->hash_node, list);
5119 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5120 struct io_poll_iocb *poll,
5121 struct io_poll_table *ipt, __poll_t mask,
5122 wait_queue_func_t wake_func)
5123 __acquires(&ctx->completion_lock)
5125 struct io_ring_ctx *ctx = req->ctx;
5126 bool cancel = false;
5128 INIT_HLIST_NODE(&req->hash_node);
5129 io_init_poll_iocb(poll, mask, wake_func);
5130 poll->file = req->file;
5131 poll->wait.private = req;
5133 ipt->pt._key = mask;
5135 ipt->error = -EINVAL;
5137 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5139 spin_lock_irq(&ctx->completion_lock);
5140 if (likely(poll->head)) {
5141 spin_lock(&poll->head->lock);
5142 if (unlikely(list_empty(&poll->wait.entry))) {
5148 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5149 list_del_init(&poll->wait.entry);
5151 WRITE_ONCE(poll->canceled, true);
5152 else if (!poll->done) /* actually waiting for an event */
5153 io_poll_req_insert(req);
5154 spin_unlock(&poll->head->lock);
5160 static bool io_arm_poll_handler(struct io_kiocb *req)
5162 const struct io_op_def *def = &io_op_defs[req->opcode];
5163 struct io_ring_ctx *ctx = req->ctx;
5164 struct async_poll *apoll;
5165 struct io_poll_table ipt;
5169 if (!req->file || !file_can_poll(req->file))
5171 if (req->flags & REQ_F_POLLED)
5175 else if (def->pollout)
5179 /* if we can't nonblock try, then no point in arming a poll handler */
5180 if (!io_file_supports_async(req, rw))
5183 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5184 if (unlikely(!apoll))
5186 apoll->double_poll = NULL;
5188 req->flags |= REQ_F_POLLED;
5191 mask = EPOLLONESHOT;
5193 mask |= POLLIN | POLLRDNORM;
5195 mask |= POLLOUT | POLLWRNORM;
5197 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5198 if ((req->opcode == IORING_OP_RECVMSG) &&
5199 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5202 mask |= POLLERR | POLLPRI;
5204 ipt.pt._qproc = io_async_queue_proc;
5206 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5208 if (ret || ipt.error) {
5209 io_poll_remove_double(req);
5210 spin_unlock_irq(&ctx->completion_lock);
5213 spin_unlock_irq(&ctx->completion_lock);
5214 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5215 apoll->poll.events);
5219 static bool __io_poll_remove_one(struct io_kiocb *req,
5220 struct io_poll_iocb *poll, bool do_cancel)
5221 __must_hold(&req->ctx->completion_lock)
5223 bool do_complete = false;
5227 spin_lock(&poll->head->lock);
5229 WRITE_ONCE(poll->canceled, true);
5230 if (!list_empty(&poll->wait.entry)) {
5231 list_del_init(&poll->wait.entry);
5234 spin_unlock(&poll->head->lock);
5235 hash_del(&req->hash_node);
5239 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5240 __must_hold(&req->ctx->completion_lock)
5244 io_poll_remove_double(req);
5245 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5247 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5248 /* non-poll requests have submit ref still */
5254 static bool io_poll_remove_one(struct io_kiocb *req)
5255 __must_hold(&req->ctx->completion_lock)
5259 do_complete = io_poll_remove_waitqs(req);
5261 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5262 io_commit_cqring(req->ctx);
5264 io_put_req_deferred(req, 1);
5271 * Returns true if we found and killed one or more poll requests
5273 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5276 struct hlist_node *tmp;
5277 struct io_kiocb *req;
5280 spin_lock_irq(&ctx->completion_lock);
5281 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5282 struct hlist_head *list;
5284 list = &ctx->cancel_hash[i];
5285 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5286 if (io_match_task(req, tsk, cancel_all))
5287 posted += io_poll_remove_one(req);
5290 spin_unlock_irq(&ctx->completion_lock);
5293 io_cqring_ev_posted(ctx);
5298 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5300 __must_hold(&ctx->completion_lock)
5302 struct hlist_head *list;
5303 struct io_kiocb *req;
5305 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5306 hlist_for_each_entry(req, list, hash_node) {
5307 if (sqe_addr != req->user_data)
5309 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5316 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5318 __must_hold(&ctx->completion_lock)
5320 struct io_kiocb *req;
5322 req = io_poll_find(ctx, sqe_addr, poll_only);
5325 if (io_poll_remove_one(req))
5331 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5336 events = READ_ONCE(sqe->poll32_events);
5338 events = swahw32(events);
5340 if (!(flags & IORING_POLL_ADD_MULTI))
5341 events |= EPOLLONESHOT;
5342 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5345 static int io_poll_update_prep(struct io_kiocb *req,
5346 const struct io_uring_sqe *sqe)
5348 struct io_poll_update *upd = &req->poll_update;
5351 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5353 if (sqe->ioprio || sqe->buf_index)
5355 flags = READ_ONCE(sqe->len);
5356 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5357 IORING_POLL_ADD_MULTI))
5359 /* meaningless without update */
5360 if (flags == IORING_POLL_ADD_MULTI)
5363 upd->old_user_data = READ_ONCE(sqe->addr);
5364 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5365 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5367 upd->new_user_data = READ_ONCE(sqe->off);
5368 if (!upd->update_user_data && upd->new_user_data)
5370 if (upd->update_events)
5371 upd->events = io_poll_parse_events(sqe, flags);
5372 else if (sqe->poll32_events)
5378 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5381 struct io_kiocb *req = wait->private;
5382 struct io_poll_iocb *poll = &req->poll;
5384 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5387 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5388 struct poll_table_struct *p)
5390 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5392 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5395 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5397 struct io_poll_iocb *poll = &req->poll;
5400 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5402 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5404 flags = READ_ONCE(sqe->len);
5405 if (flags & ~IORING_POLL_ADD_MULTI)
5408 poll->events = io_poll_parse_events(sqe, flags);
5412 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5414 struct io_poll_iocb *poll = &req->poll;
5415 struct io_ring_ctx *ctx = req->ctx;
5416 struct io_poll_table ipt;
5419 ipt.pt._qproc = io_poll_queue_proc;
5421 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5424 if (mask) { /* no async, we'd stolen it */
5426 io_poll_complete(req, mask);
5428 spin_unlock_irq(&ctx->completion_lock);
5431 io_cqring_ev_posted(ctx);
5432 if (poll->events & EPOLLONESHOT)
5438 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5440 struct io_ring_ctx *ctx = req->ctx;
5441 struct io_kiocb *preq;
5445 spin_lock_irq(&ctx->completion_lock);
5446 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5452 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5454 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5459 * Don't allow racy completion with singleshot, as we cannot safely
5460 * update those. For multishot, if we're racing with completion, just
5461 * let completion re-add it.
5463 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5464 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5468 /* we now have a detached poll request. reissue. */
5472 spin_unlock_irq(&ctx->completion_lock);
5474 io_req_complete(req, ret);
5477 /* only mask one event flags, keep behavior flags */
5478 if (req->poll_update.update_events) {
5479 preq->poll.events &= ~0xffff;
5480 preq->poll.events |= req->poll_update.events & 0xffff;
5481 preq->poll.events |= IO_POLL_UNMASK;
5483 if (req->poll_update.update_user_data)
5484 preq->user_data = req->poll_update.new_user_data;
5485 spin_unlock_irq(&ctx->completion_lock);
5487 /* complete update request, we're done with it */
5488 io_req_complete(req, ret);
5491 ret = io_poll_add(preq, issue_flags);
5494 io_req_complete(preq, ret);
5500 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5502 struct io_timeout_data *data = container_of(timer,
5503 struct io_timeout_data, timer);
5504 struct io_kiocb *req = data->req;
5505 struct io_ring_ctx *ctx = req->ctx;
5506 unsigned long flags;
5508 spin_lock_irqsave(&ctx->completion_lock, flags);
5509 list_del_init(&req->timeout.list);
5510 atomic_set(&req->ctx->cq_timeouts,
5511 atomic_read(&req->ctx->cq_timeouts) + 1);
5513 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5514 io_commit_cqring(ctx);
5515 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5517 io_cqring_ev_posted(ctx);
5520 return HRTIMER_NORESTART;
5523 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5525 __must_hold(&ctx->completion_lock)
5527 struct io_timeout_data *io;
5528 struct io_kiocb *req;
5531 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5532 found = user_data == req->user_data;
5537 return ERR_PTR(-ENOENT);
5539 io = req->async_data;
5540 if (hrtimer_try_to_cancel(&io->timer) == -1)
5541 return ERR_PTR(-EALREADY);
5542 list_del_init(&req->timeout.list);
5546 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5547 __must_hold(&ctx->completion_lock)
5549 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5552 return PTR_ERR(req);
5555 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5556 io_put_req_deferred(req, 1);
5560 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5561 struct timespec64 *ts, enum hrtimer_mode mode)
5562 __must_hold(&ctx->completion_lock)
5564 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5565 struct io_timeout_data *data;
5568 return PTR_ERR(req);
5570 req->timeout.off = 0; /* noseq */
5571 data = req->async_data;
5572 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5573 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5574 data->timer.function = io_timeout_fn;
5575 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5579 static int io_timeout_remove_prep(struct io_kiocb *req,
5580 const struct io_uring_sqe *sqe)
5582 struct io_timeout_rem *tr = &req->timeout_rem;
5584 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5586 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5588 if (sqe->ioprio || sqe->buf_index || sqe->len)
5591 tr->addr = READ_ONCE(sqe->addr);
5592 tr->flags = READ_ONCE(sqe->timeout_flags);
5593 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5594 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5596 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5598 } else if (tr->flags) {
5599 /* timeout removal doesn't support flags */
5606 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5608 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5613 * Remove or update an existing timeout command
5615 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5617 struct io_timeout_rem *tr = &req->timeout_rem;
5618 struct io_ring_ctx *ctx = req->ctx;
5621 spin_lock_irq(&ctx->completion_lock);
5622 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5623 ret = io_timeout_cancel(ctx, tr->addr);
5625 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5626 io_translate_timeout_mode(tr->flags));
5628 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5629 io_commit_cqring(ctx);
5630 spin_unlock_irq(&ctx->completion_lock);
5631 io_cqring_ev_posted(ctx);
5638 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5639 bool is_timeout_link)
5641 struct io_timeout_data *data;
5643 u32 off = READ_ONCE(sqe->off);
5645 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5647 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5649 if (off && is_timeout_link)
5651 flags = READ_ONCE(sqe->timeout_flags);
5652 if (flags & ~IORING_TIMEOUT_ABS)
5655 req->timeout.off = off;
5657 if (!req->async_data && io_alloc_async_data(req))
5660 data = req->async_data;
5663 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5666 data->mode = io_translate_timeout_mode(flags);
5667 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5668 if (is_timeout_link)
5669 io_req_track_inflight(req);
5673 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5675 struct io_ring_ctx *ctx = req->ctx;
5676 struct io_timeout_data *data = req->async_data;
5677 struct list_head *entry;
5678 u32 tail, off = req->timeout.off;
5680 spin_lock_irq(&ctx->completion_lock);
5683 * sqe->off holds how many events that need to occur for this
5684 * timeout event to be satisfied. If it isn't set, then this is
5685 * a pure timeout request, sequence isn't used.
5687 if (io_is_timeout_noseq(req)) {
5688 entry = ctx->timeout_list.prev;
5692 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5693 req->timeout.target_seq = tail + off;
5695 /* Update the last seq here in case io_flush_timeouts() hasn't.
5696 * This is safe because ->completion_lock is held, and submissions
5697 * and completions are never mixed in the same ->completion_lock section.
5699 ctx->cq_last_tm_flush = tail;
5702 * Insertion sort, ensuring the first entry in the list is always
5703 * the one we need first.
5705 list_for_each_prev(entry, &ctx->timeout_list) {
5706 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5709 if (io_is_timeout_noseq(nxt))
5711 /* nxt.seq is behind @tail, otherwise would've been completed */
5712 if (off >= nxt->timeout.target_seq - tail)
5716 list_add(&req->timeout.list, entry);
5717 data->timer.function = io_timeout_fn;
5718 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5719 spin_unlock_irq(&ctx->completion_lock);
5723 struct io_cancel_data {
5724 struct io_ring_ctx *ctx;
5728 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5730 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5731 struct io_cancel_data *cd = data;
5733 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5736 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5737 struct io_ring_ctx *ctx)
5739 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5740 enum io_wq_cancel cancel_ret;
5743 if (!tctx || !tctx->io_wq)
5746 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5747 switch (cancel_ret) {
5748 case IO_WQ_CANCEL_OK:
5751 case IO_WQ_CANCEL_RUNNING:
5754 case IO_WQ_CANCEL_NOTFOUND:
5762 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5763 struct io_kiocb *req, __u64 sqe_addr,
5766 unsigned long flags;
5769 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5770 spin_lock_irqsave(&ctx->completion_lock, flags);
5773 ret = io_timeout_cancel(ctx, sqe_addr);
5776 ret = io_poll_cancel(ctx, sqe_addr, false);
5780 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5781 io_commit_cqring(ctx);
5782 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5783 io_cqring_ev_posted(ctx);
5789 static int io_async_cancel_prep(struct io_kiocb *req,
5790 const struct io_uring_sqe *sqe)
5792 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5794 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5796 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5799 req->cancel.addr = READ_ONCE(sqe->addr);
5803 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5805 struct io_ring_ctx *ctx = req->ctx;
5806 u64 sqe_addr = req->cancel.addr;
5807 struct io_tctx_node *node;
5810 /* tasks should wait for their io-wq threads, so safe w/o sync */
5811 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5812 spin_lock_irq(&ctx->completion_lock);
5815 ret = io_timeout_cancel(ctx, sqe_addr);
5818 ret = io_poll_cancel(ctx, sqe_addr, false);
5821 spin_unlock_irq(&ctx->completion_lock);
5823 /* slow path, try all io-wq's */
5824 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5826 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5827 struct io_uring_task *tctx = node->task->io_uring;
5829 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5833 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5835 spin_lock_irq(&ctx->completion_lock);
5837 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5838 io_commit_cqring(ctx);
5839 spin_unlock_irq(&ctx->completion_lock);
5840 io_cqring_ev_posted(ctx);
5848 static int io_rsrc_update_prep(struct io_kiocb *req,
5849 const struct io_uring_sqe *sqe)
5851 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5853 if (sqe->ioprio || sqe->rw_flags)
5856 req->rsrc_update.offset = READ_ONCE(sqe->off);
5857 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5858 if (!req->rsrc_update.nr_args)
5860 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5864 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5866 struct io_ring_ctx *ctx = req->ctx;
5867 struct io_uring_rsrc_update2 up;
5870 if (issue_flags & IO_URING_F_NONBLOCK)
5873 up.offset = req->rsrc_update.offset;
5874 up.data = req->rsrc_update.arg;
5879 mutex_lock(&ctx->uring_lock);
5880 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5881 &up, req->rsrc_update.nr_args);
5882 mutex_unlock(&ctx->uring_lock);
5886 __io_req_complete(req, issue_flags, ret, 0);
5890 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5892 switch (req->opcode) {
5895 case IORING_OP_READV:
5896 case IORING_OP_READ_FIXED:
5897 case IORING_OP_READ:
5898 return io_read_prep(req, sqe);
5899 case IORING_OP_WRITEV:
5900 case IORING_OP_WRITE_FIXED:
5901 case IORING_OP_WRITE:
5902 return io_write_prep(req, sqe);
5903 case IORING_OP_POLL_ADD:
5904 return io_poll_add_prep(req, sqe);
5905 case IORING_OP_POLL_REMOVE:
5906 return io_poll_update_prep(req, sqe);
5907 case IORING_OP_FSYNC:
5908 return io_fsync_prep(req, sqe);
5909 case IORING_OP_SYNC_FILE_RANGE:
5910 return io_sfr_prep(req, sqe);
5911 case IORING_OP_SENDMSG:
5912 case IORING_OP_SEND:
5913 return io_sendmsg_prep(req, sqe);
5914 case IORING_OP_RECVMSG:
5915 case IORING_OP_RECV:
5916 return io_recvmsg_prep(req, sqe);
5917 case IORING_OP_CONNECT:
5918 return io_connect_prep(req, sqe);
5919 case IORING_OP_TIMEOUT:
5920 return io_timeout_prep(req, sqe, false);
5921 case IORING_OP_TIMEOUT_REMOVE:
5922 return io_timeout_remove_prep(req, sqe);
5923 case IORING_OP_ASYNC_CANCEL:
5924 return io_async_cancel_prep(req, sqe);
5925 case IORING_OP_LINK_TIMEOUT:
5926 return io_timeout_prep(req, sqe, true);
5927 case IORING_OP_ACCEPT:
5928 return io_accept_prep(req, sqe);
5929 case IORING_OP_FALLOCATE:
5930 return io_fallocate_prep(req, sqe);
5931 case IORING_OP_OPENAT:
5932 return io_openat_prep(req, sqe);
5933 case IORING_OP_CLOSE:
5934 return io_close_prep(req, sqe);
5935 case IORING_OP_FILES_UPDATE:
5936 return io_rsrc_update_prep(req, sqe);
5937 case IORING_OP_STATX:
5938 return io_statx_prep(req, sqe);
5939 case IORING_OP_FADVISE:
5940 return io_fadvise_prep(req, sqe);
5941 case IORING_OP_MADVISE:
5942 return io_madvise_prep(req, sqe);
5943 case IORING_OP_OPENAT2:
5944 return io_openat2_prep(req, sqe);
5945 case IORING_OP_EPOLL_CTL:
5946 return io_epoll_ctl_prep(req, sqe);
5947 case IORING_OP_SPLICE:
5948 return io_splice_prep(req, sqe);
5949 case IORING_OP_PROVIDE_BUFFERS:
5950 return io_provide_buffers_prep(req, sqe);
5951 case IORING_OP_REMOVE_BUFFERS:
5952 return io_remove_buffers_prep(req, sqe);
5954 return io_tee_prep(req, sqe);
5955 case IORING_OP_SHUTDOWN:
5956 return io_shutdown_prep(req, sqe);
5957 case IORING_OP_RENAMEAT:
5958 return io_renameat_prep(req, sqe);
5959 case IORING_OP_UNLINKAT:
5960 return io_unlinkat_prep(req, sqe);
5963 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5968 static int io_req_prep_async(struct io_kiocb *req)
5970 if (!io_op_defs[req->opcode].needs_async_setup)
5972 if (WARN_ON_ONCE(req->async_data))
5974 if (io_alloc_async_data(req))
5977 switch (req->opcode) {
5978 case IORING_OP_READV:
5979 return io_rw_prep_async(req, READ);
5980 case IORING_OP_WRITEV:
5981 return io_rw_prep_async(req, WRITE);
5982 case IORING_OP_SENDMSG:
5983 return io_sendmsg_prep_async(req);
5984 case IORING_OP_RECVMSG:
5985 return io_recvmsg_prep_async(req);
5986 case IORING_OP_CONNECT:
5987 return io_connect_prep_async(req);
5989 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5994 static u32 io_get_sequence(struct io_kiocb *req)
5996 struct io_kiocb *pos;
5997 struct io_ring_ctx *ctx = req->ctx;
5998 u32 total_submitted, nr_reqs = 0;
6000 io_for_each_link(pos, req)
6003 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6004 return total_submitted - nr_reqs;
6007 static int io_req_defer(struct io_kiocb *req)
6009 struct io_ring_ctx *ctx = req->ctx;
6010 struct io_defer_entry *de;
6014 /* Still need defer if there is pending req in defer list. */
6015 if (likely(list_empty_careful(&ctx->defer_list) &&
6016 !(req->flags & REQ_F_IO_DRAIN)))
6019 seq = io_get_sequence(req);
6020 /* Still a chance to pass the sequence check */
6021 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6024 ret = io_req_prep_async(req);
6027 io_prep_async_link(req);
6028 de = kmalloc(sizeof(*de), GFP_KERNEL);
6032 spin_lock_irq(&ctx->completion_lock);
6033 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6034 spin_unlock_irq(&ctx->completion_lock);
6036 io_queue_async_work(req);
6037 return -EIOCBQUEUED;
6040 trace_io_uring_defer(ctx, req, req->user_data);
6043 list_add_tail(&de->list, &ctx->defer_list);
6044 spin_unlock_irq(&ctx->completion_lock);
6045 return -EIOCBQUEUED;
6048 static void io_clean_op(struct io_kiocb *req)
6050 if (req->flags & REQ_F_BUFFER_SELECTED) {
6051 switch (req->opcode) {
6052 case IORING_OP_READV:
6053 case IORING_OP_READ_FIXED:
6054 case IORING_OP_READ:
6055 kfree((void *)(unsigned long)req->rw.addr);
6057 case IORING_OP_RECVMSG:
6058 case IORING_OP_RECV:
6059 kfree(req->sr_msg.kbuf);
6062 req->flags &= ~REQ_F_BUFFER_SELECTED;
6065 if (req->flags & REQ_F_NEED_CLEANUP) {
6066 switch (req->opcode) {
6067 case IORING_OP_READV:
6068 case IORING_OP_READ_FIXED:
6069 case IORING_OP_READ:
6070 case IORING_OP_WRITEV:
6071 case IORING_OP_WRITE_FIXED:
6072 case IORING_OP_WRITE: {
6073 struct io_async_rw *io = req->async_data;
6075 kfree(io->free_iovec);
6078 case IORING_OP_RECVMSG:
6079 case IORING_OP_SENDMSG: {
6080 struct io_async_msghdr *io = req->async_data;
6082 kfree(io->free_iov);
6085 case IORING_OP_SPLICE:
6087 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6088 io_put_file(req->splice.file_in);
6090 case IORING_OP_OPENAT:
6091 case IORING_OP_OPENAT2:
6092 if (req->open.filename)
6093 putname(req->open.filename);
6095 case IORING_OP_RENAMEAT:
6096 putname(req->rename.oldpath);
6097 putname(req->rename.newpath);
6099 case IORING_OP_UNLINKAT:
6100 putname(req->unlink.filename);
6103 req->flags &= ~REQ_F_NEED_CLEANUP;
6105 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6106 kfree(req->apoll->double_poll);
6110 if (req->flags & REQ_F_INFLIGHT) {
6111 struct io_uring_task *tctx = req->task->io_uring;
6113 atomic_dec(&tctx->inflight_tracked);
6114 req->flags &= ~REQ_F_INFLIGHT;
6118 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6120 struct io_ring_ctx *ctx = req->ctx;
6121 const struct cred *creds = NULL;
6124 if (req->work.creds && req->work.creds != current_cred())
6125 creds = override_creds(req->work.creds);
6127 switch (req->opcode) {
6129 ret = io_nop(req, issue_flags);
6131 case IORING_OP_READV:
6132 case IORING_OP_READ_FIXED:
6133 case IORING_OP_READ:
6134 ret = io_read(req, issue_flags);
6136 case IORING_OP_WRITEV:
6137 case IORING_OP_WRITE_FIXED:
6138 case IORING_OP_WRITE:
6139 ret = io_write(req, issue_flags);
6141 case IORING_OP_FSYNC:
6142 ret = io_fsync(req, issue_flags);
6144 case IORING_OP_POLL_ADD:
6145 ret = io_poll_add(req, issue_flags);
6147 case IORING_OP_POLL_REMOVE:
6148 ret = io_poll_update(req, issue_flags);
6150 case IORING_OP_SYNC_FILE_RANGE:
6151 ret = io_sync_file_range(req, issue_flags);
6153 case IORING_OP_SENDMSG:
6154 ret = io_sendmsg(req, issue_flags);
6156 case IORING_OP_SEND:
6157 ret = io_send(req, issue_flags);
6159 case IORING_OP_RECVMSG:
6160 ret = io_recvmsg(req, issue_flags);
6162 case IORING_OP_RECV:
6163 ret = io_recv(req, issue_flags);
6165 case IORING_OP_TIMEOUT:
6166 ret = io_timeout(req, issue_flags);
6168 case IORING_OP_TIMEOUT_REMOVE:
6169 ret = io_timeout_remove(req, issue_flags);
6171 case IORING_OP_ACCEPT:
6172 ret = io_accept(req, issue_flags);
6174 case IORING_OP_CONNECT:
6175 ret = io_connect(req, issue_flags);
6177 case IORING_OP_ASYNC_CANCEL:
6178 ret = io_async_cancel(req, issue_flags);
6180 case IORING_OP_FALLOCATE:
6181 ret = io_fallocate(req, issue_flags);
6183 case IORING_OP_OPENAT:
6184 ret = io_openat(req, issue_flags);
6186 case IORING_OP_CLOSE:
6187 ret = io_close(req, issue_flags);
6189 case IORING_OP_FILES_UPDATE:
6190 ret = io_files_update(req, issue_flags);
6192 case IORING_OP_STATX:
6193 ret = io_statx(req, issue_flags);
6195 case IORING_OP_FADVISE:
6196 ret = io_fadvise(req, issue_flags);
6198 case IORING_OP_MADVISE:
6199 ret = io_madvise(req, issue_flags);
6201 case IORING_OP_OPENAT2:
6202 ret = io_openat2(req, issue_flags);
6204 case IORING_OP_EPOLL_CTL:
6205 ret = io_epoll_ctl(req, issue_flags);
6207 case IORING_OP_SPLICE:
6208 ret = io_splice(req, issue_flags);
6210 case IORING_OP_PROVIDE_BUFFERS:
6211 ret = io_provide_buffers(req, issue_flags);
6213 case IORING_OP_REMOVE_BUFFERS:
6214 ret = io_remove_buffers(req, issue_flags);
6217 ret = io_tee(req, issue_flags);
6219 case IORING_OP_SHUTDOWN:
6220 ret = io_shutdown(req, issue_flags);
6222 case IORING_OP_RENAMEAT:
6223 ret = io_renameat(req, issue_flags);
6225 case IORING_OP_UNLINKAT:
6226 ret = io_unlinkat(req, issue_flags);
6234 revert_creds(creds);
6237 /* If the op doesn't have a file, we're not polling for it */
6238 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6239 io_iopoll_req_issued(req);
6244 static void io_wq_submit_work(struct io_wq_work *work)
6246 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6247 struct io_kiocb *timeout;
6250 timeout = io_prep_linked_timeout(req);
6252 io_queue_linked_timeout(timeout);
6254 if (work->flags & IO_WQ_WORK_CANCEL)
6259 ret = io_issue_sqe(req, 0);
6261 * We can get EAGAIN for polled IO even though we're
6262 * forcing a sync submission from here, since we can't
6263 * wait for request slots on the block side.
6271 /* avoid locking problems by failing it from a clean context */
6273 /* io-wq is going to take one down */
6275 io_req_task_queue_fail(req, ret);
6279 #define FFS_ASYNC_READ 0x1UL
6280 #define FFS_ASYNC_WRITE 0x2UL
6282 #define FFS_ISREG 0x4UL
6284 #define FFS_ISREG 0x0UL
6286 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6288 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6291 struct io_fixed_file *table_l2;
6293 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6294 return &table_l2[i & IORING_FILE_TABLE_MASK];
6297 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6300 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6302 return (struct file *) (slot->file_ptr & FFS_MASK);
6305 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6307 unsigned long file_ptr = (unsigned long) file;
6309 if (__io_file_supports_async(file, READ))
6310 file_ptr |= FFS_ASYNC_READ;
6311 if (__io_file_supports_async(file, WRITE))
6312 file_ptr |= FFS_ASYNC_WRITE;
6313 if (S_ISREG(file_inode(file)->i_mode))
6314 file_ptr |= FFS_ISREG;
6315 file_slot->file_ptr = file_ptr;
6318 static struct file *io_file_get(struct io_submit_state *state,
6319 struct io_kiocb *req, int fd, bool fixed)
6321 struct io_ring_ctx *ctx = req->ctx;
6325 unsigned long file_ptr;
6327 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6329 fd = array_index_nospec(fd, ctx->nr_user_files);
6330 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6331 file = (struct file *) (file_ptr & FFS_MASK);
6332 file_ptr &= ~FFS_MASK;
6333 /* mask in overlapping REQ_F and FFS bits */
6334 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6335 io_req_set_rsrc_node(req);
6337 trace_io_uring_file_get(ctx, fd);
6338 file = __io_file_get(state, fd);
6340 /* we don't allow fixed io_uring files */
6341 if (file && unlikely(file->f_op == &io_uring_fops))
6342 io_req_track_inflight(req);
6348 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6350 struct io_timeout_data *data = container_of(timer,
6351 struct io_timeout_data, timer);
6352 struct io_kiocb *prev, *req = data->req;
6353 struct io_ring_ctx *ctx = req->ctx;
6354 unsigned long flags;
6356 spin_lock_irqsave(&ctx->completion_lock, flags);
6357 prev = req->timeout.head;
6358 req->timeout.head = NULL;
6361 * We don't expect the list to be empty, that will only happen if we
6362 * race with the completion of the linked work.
6365 io_remove_next_linked(prev);
6366 if (!req_ref_inc_not_zero(prev))
6369 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6372 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6373 io_put_req_deferred(prev, 1);
6374 io_put_req_deferred(req, 1);
6376 io_req_complete_post(req, -ETIME, 0);
6378 return HRTIMER_NORESTART;
6381 static void io_queue_linked_timeout(struct io_kiocb *req)
6383 struct io_ring_ctx *ctx = req->ctx;
6385 spin_lock_irq(&ctx->completion_lock);
6387 * If the back reference is NULL, then our linked request finished
6388 * before we got a chance to setup the timer
6390 if (req->timeout.head) {
6391 struct io_timeout_data *data = req->async_data;
6393 data->timer.function = io_link_timeout_fn;
6394 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6397 spin_unlock_irq(&ctx->completion_lock);
6398 /* drop submission reference */
6402 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6404 struct io_kiocb *nxt = req->link;
6406 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6407 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6410 nxt->timeout.head = req;
6411 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6412 req->flags |= REQ_F_LINK_TIMEOUT;
6416 static void __io_queue_sqe(struct io_kiocb *req)
6418 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6421 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6424 * We async punt it if the file wasn't marked NOWAIT, or if the file
6425 * doesn't support non-blocking read/write attempts
6428 /* drop submission reference */
6429 if (req->flags & REQ_F_COMPLETE_INLINE) {
6430 struct io_ring_ctx *ctx = req->ctx;
6431 struct io_comp_state *cs = &ctx->submit_state.comp;
6433 cs->reqs[cs->nr++] = req;
6434 if (cs->nr == ARRAY_SIZE(cs->reqs))
6435 io_submit_flush_completions(cs, ctx);
6439 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6440 if (!io_arm_poll_handler(req)) {
6442 * Queued up for async execution, worker will release
6443 * submit reference when the iocb is actually submitted.
6445 io_queue_async_work(req);
6448 io_req_complete_failed(req, ret);
6451 io_queue_linked_timeout(linked_timeout);
6454 static void io_queue_sqe(struct io_kiocb *req)
6458 ret = io_req_defer(req);
6460 if (ret != -EIOCBQUEUED) {
6462 io_req_complete_failed(req, ret);
6464 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6465 ret = io_req_prep_async(req);
6468 io_queue_async_work(req);
6470 __io_queue_sqe(req);
6475 * Check SQE restrictions (opcode and flags).
6477 * Returns 'true' if SQE is allowed, 'false' otherwise.
6479 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6480 struct io_kiocb *req,
6481 unsigned int sqe_flags)
6483 if (!ctx->restricted)
6486 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6489 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6490 ctx->restrictions.sqe_flags_required)
6493 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6494 ctx->restrictions.sqe_flags_required))
6500 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6501 const struct io_uring_sqe *sqe)
6503 struct io_submit_state *state;
6504 unsigned int sqe_flags;
6505 int personality, ret = 0;
6507 req->opcode = READ_ONCE(sqe->opcode);
6508 /* same numerical values with corresponding REQ_F_*, safe to copy */
6509 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6510 req->user_data = READ_ONCE(sqe->user_data);
6511 req->async_data = NULL;
6515 req->fixed_rsrc_refs = NULL;
6516 /* one is dropped after submission, the other at completion */
6517 atomic_set(&req->refs, 2);
6518 req->task = current;
6520 req->work.creds = NULL;
6522 /* enforce forwards compatibility on users */
6523 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6525 if (unlikely(req->opcode >= IORING_OP_LAST))
6527 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6530 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6531 !io_op_defs[req->opcode].buffer_select)
6534 personality = READ_ONCE(sqe->personality);
6536 req->work.creds = xa_load(&ctx->personalities, personality);
6537 if (!req->work.creds)
6539 get_cred(req->work.creds);
6541 state = &ctx->submit_state;
6544 * Plug now if we have more than 1 IO left after this, and the target
6545 * is potentially a read/write to block based storage.
6547 if (!state->plug_started && state->ios_left > 1 &&
6548 io_op_defs[req->opcode].plug) {
6549 blk_start_plug(&state->plug);
6550 state->plug_started = true;
6553 if (io_op_defs[req->opcode].needs_file) {
6554 bool fixed = req->flags & REQ_F_FIXED_FILE;
6556 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6557 if (unlikely(!req->file))
6565 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6566 const struct io_uring_sqe *sqe)
6568 struct io_submit_link *link = &ctx->submit_state.link;
6571 ret = io_init_req(ctx, req, sqe);
6572 if (unlikely(ret)) {
6575 /* fail even hard links since we don't submit */
6576 req_set_fail(link->head);
6577 io_req_complete_failed(link->head, -ECANCELED);
6580 io_req_complete_failed(req, ret);
6583 ret = io_req_prep(req, sqe);
6587 /* don't need @sqe from now on */
6588 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6589 true, ctx->flags & IORING_SETUP_SQPOLL);
6592 * If we already have a head request, queue this one for async
6593 * submittal once the head completes. If we don't have a head but
6594 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6595 * submitted sync once the chain is complete. If none of those
6596 * conditions are true (normal request), then just queue it.
6599 struct io_kiocb *head = link->head;
6602 * Taking sequential execution of a link, draining both sides
6603 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6604 * requests in the link. So, it drains the head and the
6605 * next after the link request. The last one is done via
6606 * drain_next flag to persist the effect across calls.
6608 if (req->flags & REQ_F_IO_DRAIN) {
6609 head->flags |= REQ_F_IO_DRAIN;
6610 ctx->drain_next = 1;
6612 ret = io_req_prep_async(req);
6615 trace_io_uring_link(ctx, req, head);
6616 link->last->link = req;
6619 /* last request of a link, enqueue the link */
6620 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6625 if (unlikely(ctx->drain_next)) {
6626 req->flags |= REQ_F_IO_DRAIN;
6627 ctx->drain_next = 0;
6629 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6641 * Batched submission is done, ensure local IO is flushed out.
6643 static void io_submit_state_end(struct io_submit_state *state,
6644 struct io_ring_ctx *ctx)
6646 if (state->link.head)
6647 io_queue_sqe(state->link.head);
6649 io_submit_flush_completions(&state->comp, ctx);
6650 if (state->plug_started)
6651 blk_finish_plug(&state->plug);
6652 io_state_file_put(state);
6656 * Start submission side cache.
6658 static void io_submit_state_start(struct io_submit_state *state,
6659 unsigned int max_ios)
6661 state->plug_started = false;
6662 state->ios_left = max_ios;
6663 /* set only head, no need to init link_last in advance */
6664 state->link.head = NULL;
6667 static void io_commit_sqring(struct io_ring_ctx *ctx)
6669 struct io_rings *rings = ctx->rings;
6672 * Ensure any loads from the SQEs are done at this point,
6673 * since once we write the new head, the application could
6674 * write new data to them.
6676 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6680 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6681 * that is mapped by userspace. This means that care needs to be taken to
6682 * ensure that reads are stable, as we cannot rely on userspace always
6683 * being a good citizen. If members of the sqe are validated and then later
6684 * used, it's important that those reads are done through READ_ONCE() to
6685 * prevent a re-load down the line.
6687 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6689 u32 *sq_array = ctx->sq_array;
6690 unsigned head, mask = ctx->sq_entries - 1;
6693 * The cached sq head (or cq tail) serves two purposes:
6695 * 1) allows us to batch the cost of updating the user visible
6697 * 2) allows the kernel side to track the head on its own, even
6698 * though the application is the one updating it.
6700 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & mask]);
6701 if (likely(head < ctx->sq_entries))
6702 return &ctx->sq_sqes[head];
6704 /* drop invalid entries */
6705 ctx->cached_sq_dropped++;
6706 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6710 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6712 struct io_uring_task *tctx;
6715 /* make sure SQ entry isn't read before tail */
6716 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6717 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6720 tctx = current->io_uring;
6721 tctx->cached_refs -= nr;
6722 if (unlikely(tctx->cached_refs < 0)) {
6723 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6725 percpu_counter_add(&tctx->inflight, refill);
6726 refcount_add(refill, ¤t->usage);
6727 tctx->cached_refs += refill;
6729 io_submit_state_start(&ctx->submit_state, nr);
6731 while (submitted < nr) {
6732 const struct io_uring_sqe *sqe;
6733 struct io_kiocb *req;
6735 req = io_alloc_req(ctx);
6736 if (unlikely(!req)) {
6738 submitted = -EAGAIN;
6741 sqe = io_get_sqe(ctx);
6742 if (unlikely(!sqe)) {
6743 kmem_cache_free(req_cachep, req);
6746 /* will complete beyond this point, count as submitted */
6748 if (io_submit_sqe(ctx, req, sqe))
6752 if (unlikely(submitted != nr)) {
6753 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6754 int unused = nr - ref_used;
6756 current->io_uring->cached_refs += unused;
6757 percpu_ref_put_many(&ctx->refs, unused);
6760 io_submit_state_end(&ctx->submit_state, ctx);
6761 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6762 io_commit_sqring(ctx);
6767 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6769 return READ_ONCE(sqd->state);
6772 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6774 /* Tell userspace we may need a wakeup call */
6775 spin_lock_irq(&ctx->completion_lock);
6776 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6777 spin_unlock_irq(&ctx->completion_lock);
6780 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6782 spin_lock_irq(&ctx->completion_lock);
6783 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6784 spin_unlock_irq(&ctx->completion_lock);
6787 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6789 unsigned int to_submit;
6792 to_submit = io_sqring_entries(ctx);
6793 /* if we're handling multiple rings, cap submit size for fairness */
6794 if (cap_entries && to_submit > 8)
6797 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6798 unsigned nr_events = 0;
6800 mutex_lock(&ctx->uring_lock);
6801 if (!list_empty(&ctx->iopoll_list))
6802 io_do_iopoll(ctx, &nr_events, 0);
6805 * Don't submit if refs are dying, good for io_uring_register(),
6806 * but also it is relied upon by io_ring_exit_work()
6808 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6809 !(ctx->flags & IORING_SETUP_R_DISABLED))
6810 ret = io_submit_sqes(ctx, to_submit);
6811 mutex_unlock(&ctx->uring_lock);
6813 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6814 wake_up(&ctx->sqo_sq_wait);
6820 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6822 struct io_ring_ctx *ctx;
6823 unsigned sq_thread_idle = 0;
6825 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6826 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6827 sqd->sq_thread_idle = sq_thread_idle;
6830 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6832 bool did_sig = false;
6833 struct ksignal ksig;
6835 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6836 signal_pending(current)) {
6837 mutex_unlock(&sqd->lock);
6838 if (signal_pending(current))
6839 did_sig = get_signal(&ksig);
6841 mutex_lock(&sqd->lock);
6844 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6847 static int io_sq_thread(void *data)
6849 struct io_sq_data *sqd = data;
6850 struct io_ring_ctx *ctx;
6851 unsigned long timeout = 0;
6852 char buf[TASK_COMM_LEN];
6855 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6856 set_task_comm(current, buf);
6858 if (sqd->sq_cpu != -1)
6859 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6861 set_cpus_allowed_ptr(current, cpu_online_mask);
6862 current->flags |= PF_NO_SETAFFINITY;
6864 mutex_lock(&sqd->lock);
6867 bool cap_entries, sqt_spin, needs_sched;
6869 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6870 if (io_sqd_handle_event(sqd))
6872 timeout = jiffies + sqd->sq_thread_idle;
6877 cap_entries = !list_is_singular(&sqd->ctx_list);
6878 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6879 const struct cred *creds = NULL;
6881 if (ctx->sq_creds != current_cred())
6882 creds = override_creds(ctx->sq_creds);
6883 ret = __io_sq_thread(ctx, cap_entries);
6885 revert_creds(creds);
6886 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6890 if (sqt_spin || !time_after(jiffies, timeout)) {
6894 timeout = jiffies + sqd->sq_thread_idle;
6898 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6899 if (!io_sqd_events_pending(sqd)) {
6901 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6902 io_ring_set_wakeup_flag(ctx);
6904 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6905 !list_empty_careful(&ctx->iopoll_list)) {
6906 needs_sched = false;
6909 if (io_sqring_entries(ctx)) {
6910 needs_sched = false;
6916 mutex_unlock(&sqd->lock);
6918 mutex_lock(&sqd->lock);
6920 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6921 io_ring_clear_wakeup_flag(ctx);
6924 finish_wait(&sqd->wait, &wait);
6925 timeout = jiffies + sqd->sq_thread_idle;
6928 io_uring_cancel_generic(true, sqd);
6930 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6931 io_ring_set_wakeup_flag(ctx);
6933 mutex_unlock(&sqd->lock);
6935 complete(&sqd->exited);
6939 struct io_wait_queue {
6940 struct wait_queue_entry wq;
6941 struct io_ring_ctx *ctx;
6943 unsigned nr_timeouts;
6946 static inline bool io_should_wake(struct io_wait_queue *iowq)
6948 struct io_ring_ctx *ctx = iowq->ctx;
6951 * Wake up if we have enough events, or if a timeout occurred since we
6952 * started waiting. For timeouts, we always want to return to userspace,
6953 * regardless of event count.
6955 return io_cqring_events(ctx) >= iowq->to_wait ||
6956 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6959 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6960 int wake_flags, void *key)
6962 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6966 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6967 * the task, and the next invocation will do it.
6969 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6970 return autoremove_wake_function(curr, mode, wake_flags, key);
6974 static int io_run_task_work_sig(void)
6976 if (io_run_task_work())
6978 if (!signal_pending(current))
6980 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6981 return -ERESTARTSYS;
6985 /* when returns >0, the caller should retry */
6986 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6987 struct io_wait_queue *iowq,
6988 signed long *timeout)
6992 /* make sure we run task_work before checking for signals */
6993 ret = io_run_task_work_sig();
6994 if (ret || io_should_wake(iowq))
6996 /* let the caller flush overflows, retry */
6997 if (test_bit(0, &ctx->cq_check_overflow))
7000 *timeout = schedule_timeout(*timeout);
7001 return !*timeout ? -ETIME : 1;
7005 * Wait until events become available, if we don't already have some. The
7006 * application must reap them itself, as they reside on the shared cq ring.
7008 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7009 const sigset_t __user *sig, size_t sigsz,
7010 struct __kernel_timespec __user *uts)
7012 struct io_wait_queue iowq = {
7015 .func = io_wake_function,
7016 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7019 .to_wait = min_events,
7021 struct io_rings *rings = ctx->rings;
7022 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7026 io_cqring_overflow_flush(ctx, false);
7027 if (io_cqring_events(ctx) >= min_events)
7029 if (!io_run_task_work())
7034 #ifdef CONFIG_COMPAT
7035 if (in_compat_syscall())
7036 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7040 ret = set_user_sigmask(sig, sigsz);
7047 struct timespec64 ts;
7049 if (get_timespec64(&ts, uts))
7051 timeout = timespec64_to_jiffies(&ts);
7054 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7055 trace_io_uring_cqring_wait(ctx, min_events);
7057 /* if we can't even flush overflow, don't wait for more */
7058 if (!io_cqring_overflow_flush(ctx, false)) {
7062 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7063 TASK_INTERRUPTIBLE);
7064 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7065 finish_wait(&ctx->wait, &iowq.wq);
7069 restore_saved_sigmask_unless(ret == -EINTR);
7071 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7074 static void io_free_page_table(void **table, size_t size)
7076 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7078 for (i = 0; i < nr_tables; i++)
7083 static void **io_alloc_page_table(size_t size)
7085 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7086 size_t init_size = size;
7089 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7093 for (i = 0; i < nr_tables; i++) {
7094 unsigned int this_size = min(size, PAGE_SIZE);
7096 table[i] = kzalloc(this_size, GFP_KERNEL);
7098 io_free_page_table(table, init_size);
7106 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7108 spin_lock_bh(&ctx->rsrc_ref_lock);
7111 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7113 spin_unlock_bh(&ctx->rsrc_ref_lock);
7116 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7118 percpu_ref_exit(&ref_node->refs);
7122 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7123 struct io_rsrc_data *data_to_kill)
7125 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7126 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7129 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7131 rsrc_node->rsrc_data = data_to_kill;
7132 io_rsrc_ref_lock(ctx);
7133 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7134 io_rsrc_ref_unlock(ctx);
7136 atomic_inc(&data_to_kill->refs);
7137 percpu_ref_kill(&rsrc_node->refs);
7138 ctx->rsrc_node = NULL;
7141 if (!ctx->rsrc_node) {
7142 ctx->rsrc_node = ctx->rsrc_backup_node;
7143 ctx->rsrc_backup_node = NULL;
7147 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7149 if (ctx->rsrc_backup_node)
7151 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7152 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7155 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7159 /* As we may drop ->uring_lock, other task may have started quiesce */
7163 data->quiesce = true;
7165 ret = io_rsrc_node_switch_start(ctx);
7168 io_rsrc_node_switch(ctx, data);
7170 /* kill initial ref, already quiesced if zero */
7171 if (atomic_dec_and_test(&data->refs))
7173 flush_delayed_work(&ctx->rsrc_put_work);
7174 ret = wait_for_completion_interruptible(&data->done);
7178 atomic_inc(&data->refs);
7179 /* wait for all works potentially completing data->done */
7180 flush_delayed_work(&ctx->rsrc_put_work);
7181 reinit_completion(&data->done);
7183 mutex_unlock(&ctx->uring_lock);
7184 ret = io_run_task_work_sig();
7185 mutex_lock(&ctx->uring_lock);
7187 data->quiesce = false;
7192 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7194 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7195 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7197 return &data->tags[table_idx][off];
7200 static void io_rsrc_data_free(struct io_rsrc_data *data)
7202 size_t size = data->nr * sizeof(data->tags[0][0]);
7205 io_free_page_table((void **)data->tags, size);
7209 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7210 u64 __user *utags, unsigned nr,
7211 struct io_rsrc_data **pdata)
7213 struct io_rsrc_data *data;
7217 data = kzalloc(sizeof(*data), GFP_KERNEL);
7220 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7228 data->do_put = do_put;
7231 for (i = 0; i < nr; i++) {
7232 u64 *tag_slot = io_get_tag_slot(data, i);
7234 if (copy_from_user(tag_slot, &utags[i],
7240 atomic_set(&data->refs, 1);
7241 init_completion(&data->done);
7245 io_rsrc_data_free(data);
7249 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7251 size_t size = nr_files * sizeof(struct io_fixed_file);
7253 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7254 return !!table->files;
7257 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7259 size_t size = nr_files * sizeof(struct io_fixed_file);
7261 io_free_page_table((void **)table->files, size);
7262 table->files = NULL;
7265 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7267 #if defined(CONFIG_UNIX)
7268 if (ctx->ring_sock) {
7269 struct sock *sock = ctx->ring_sock->sk;
7270 struct sk_buff *skb;
7272 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7278 for (i = 0; i < ctx->nr_user_files; i++) {
7281 file = io_file_from_index(ctx, i);
7286 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7287 io_rsrc_data_free(ctx->file_data);
7288 ctx->file_data = NULL;
7289 ctx->nr_user_files = 0;
7292 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7296 if (!ctx->file_data)
7298 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7300 __io_sqe_files_unregister(ctx);
7304 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7305 __releases(&sqd->lock)
7307 WARN_ON_ONCE(sqd->thread == current);
7310 * Do the dance but not conditional clear_bit() because it'd race with
7311 * other threads incrementing park_pending and setting the bit.
7313 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7314 if (atomic_dec_return(&sqd->park_pending))
7315 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7316 mutex_unlock(&sqd->lock);
7319 static void io_sq_thread_park(struct io_sq_data *sqd)
7320 __acquires(&sqd->lock)
7322 WARN_ON_ONCE(sqd->thread == current);
7324 atomic_inc(&sqd->park_pending);
7325 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7326 mutex_lock(&sqd->lock);
7328 wake_up_process(sqd->thread);
7331 static void io_sq_thread_stop(struct io_sq_data *sqd)
7333 WARN_ON_ONCE(sqd->thread == current);
7334 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7336 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7337 mutex_lock(&sqd->lock);
7339 wake_up_process(sqd->thread);
7340 mutex_unlock(&sqd->lock);
7341 wait_for_completion(&sqd->exited);
7344 static void io_put_sq_data(struct io_sq_data *sqd)
7346 if (refcount_dec_and_test(&sqd->refs)) {
7347 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7349 io_sq_thread_stop(sqd);
7354 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7356 struct io_sq_data *sqd = ctx->sq_data;
7359 io_sq_thread_park(sqd);
7360 list_del_init(&ctx->sqd_list);
7361 io_sqd_update_thread_idle(sqd);
7362 io_sq_thread_unpark(sqd);
7364 io_put_sq_data(sqd);
7365 ctx->sq_data = NULL;
7369 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7371 struct io_ring_ctx *ctx_attach;
7372 struct io_sq_data *sqd;
7375 f = fdget(p->wq_fd);
7377 return ERR_PTR(-ENXIO);
7378 if (f.file->f_op != &io_uring_fops) {
7380 return ERR_PTR(-EINVAL);
7383 ctx_attach = f.file->private_data;
7384 sqd = ctx_attach->sq_data;
7387 return ERR_PTR(-EINVAL);
7389 if (sqd->task_tgid != current->tgid) {
7391 return ERR_PTR(-EPERM);
7394 refcount_inc(&sqd->refs);
7399 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7402 struct io_sq_data *sqd;
7405 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7406 sqd = io_attach_sq_data(p);
7411 /* fall through for EPERM case, setup new sqd/task */
7412 if (PTR_ERR(sqd) != -EPERM)
7416 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7418 return ERR_PTR(-ENOMEM);
7420 atomic_set(&sqd->park_pending, 0);
7421 refcount_set(&sqd->refs, 1);
7422 INIT_LIST_HEAD(&sqd->ctx_list);
7423 mutex_init(&sqd->lock);
7424 init_waitqueue_head(&sqd->wait);
7425 init_completion(&sqd->exited);
7429 #if defined(CONFIG_UNIX)
7431 * Ensure the UNIX gc is aware of our file set, so we are certain that
7432 * the io_uring can be safely unregistered on process exit, even if we have
7433 * loops in the file referencing.
7435 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7437 struct sock *sk = ctx->ring_sock->sk;
7438 struct scm_fp_list *fpl;
7439 struct sk_buff *skb;
7442 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7446 skb = alloc_skb(0, GFP_KERNEL);
7455 fpl->user = get_uid(current_user());
7456 for (i = 0; i < nr; i++) {
7457 struct file *file = io_file_from_index(ctx, i + offset);
7461 fpl->fp[nr_files] = get_file(file);
7462 unix_inflight(fpl->user, fpl->fp[nr_files]);
7467 fpl->max = SCM_MAX_FD;
7468 fpl->count = nr_files;
7469 UNIXCB(skb).fp = fpl;
7470 skb->destructor = unix_destruct_scm;
7471 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7472 skb_queue_head(&sk->sk_receive_queue, skb);
7474 for (i = 0; i < nr_files; i++)
7485 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7486 * causes regular reference counting to break down. We rely on the UNIX
7487 * garbage collection to take care of this problem for us.
7489 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7491 unsigned left, total;
7495 left = ctx->nr_user_files;
7497 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7499 ret = __io_sqe_files_scm(ctx, this_files, total);
7503 total += this_files;
7509 while (total < ctx->nr_user_files) {
7510 struct file *file = io_file_from_index(ctx, total);
7520 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7526 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7528 struct file *file = prsrc->file;
7529 #if defined(CONFIG_UNIX)
7530 struct sock *sock = ctx->ring_sock->sk;
7531 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7532 struct sk_buff *skb;
7535 __skb_queue_head_init(&list);
7538 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7539 * remove this entry and rearrange the file array.
7541 skb = skb_dequeue(head);
7543 struct scm_fp_list *fp;
7545 fp = UNIXCB(skb).fp;
7546 for (i = 0; i < fp->count; i++) {
7549 if (fp->fp[i] != file)
7552 unix_notinflight(fp->user, fp->fp[i]);
7553 left = fp->count - 1 - i;
7555 memmove(&fp->fp[i], &fp->fp[i + 1],
7556 left * sizeof(struct file *));
7563 __skb_queue_tail(&list, skb);
7573 __skb_queue_tail(&list, skb);
7575 skb = skb_dequeue(head);
7578 if (skb_peek(&list)) {
7579 spin_lock_irq(&head->lock);
7580 while ((skb = __skb_dequeue(&list)) != NULL)
7581 __skb_queue_tail(head, skb);
7582 spin_unlock_irq(&head->lock);
7589 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7591 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7592 struct io_ring_ctx *ctx = rsrc_data->ctx;
7593 struct io_rsrc_put *prsrc, *tmp;
7595 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7596 list_del(&prsrc->list);
7599 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7601 io_ring_submit_lock(ctx, lock_ring);
7602 spin_lock_irq(&ctx->completion_lock);
7603 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7605 io_commit_cqring(ctx);
7606 spin_unlock_irq(&ctx->completion_lock);
7607 io_cqring_ev_posted(ctx);
7608 io_ring_submit_unlock(ctx, lock_ring);
7611 rsrc_data->do_put(ctx, prsrc);
7615 io_rsrc_node_destroy(ref_node);
7616 if (atomic_dec_and_test(&rsrc_data->refs))
7617 complete(&rsrc_data->done);
7620 static void io_rsrc_put_work(struct work_struct *work)
7622 struct io_ring_ctx *ctx;
7623 struct llist_node *node;
7625 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7626 node = llist_del_all(&ctx->rsrc_put_llist);
7629 struct io_rsrc_node *ref_node;
7630 struct llist_node *next = node->next;
7632 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7633 __io_rsrc_put_work(ref_node);
7638 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7640 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7641 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7642 bool first_add = false;
7644 io_rsrc_ref_lock(ctx);
7647 while (!list_empty(&ctx->rsrc_ref_list)) {
7648 node = list_first_entry(&ctx->rsrc_ref_list,
7649 struct io_rsrc_node, node);
7650 /* recycle ref nodes in order */
7653 list_del(&node->node);
7654 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7656 io_rsrc_ref_unlock(ctx);
7659 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7662 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7664 struct io_rsrc_node *ref_node;
7666 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7670 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7675 INIT_LIST_HEAD(&ref_node->node);
7676 INIT_LIST_HEAD(&ref_node->rsrc_list);
7677 ref_node->done = false;
7681 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7682 unsigned nr_args, u64 __user *tags)
7684 __s32 __user *fds = (__s32 __user *) arg;
7693 if (nr_args > IORING_MAX_FIXED_FILES)
7695 ret = io_rsrc_node_switch_start(ctx);
7698 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7704 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7707 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7708 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7712 /* allow sparse sets */
7715 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7722 if (unlikely(!file))
7726 * Don't allow io_uring instances to be registered. If UNIX
7727 * isn't enabled, then this causes a reference cycle and this
7728 * instance can never get freed. If UNIX is enabled we'll
7729 * handle it just fine, but there's still no point in allowing
7730 * a ring fd as it doesn't support regular read/write anyway.
7732 if (file->f_op == &io_uring_fops) {
7736 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7739 ret = io_sqe_files_scm(ctx);
7741 __io_sqe_files_unregister(ctx);
7745 io_rsrc_node_switch(ctx, NULL);
7748 for (i = 0; i < ctx->nr_user_files; i++) {
7749 file = io_file_from_index(ctx, i);
7753 io_free_file_tables(&ctx->file_table, nr_args);
7754 ctx->nr_user_files = 0;
7756 io_rsrc_data_free(ctx->file_data);
7757 ctx->file_data = NULL;
7761 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7764 #if defined(CONFIG_UNIX)
7765 struct sock *sock = ctx->ring_sock->sk;
7766 struct sk_buff_head *head = &sock->sk_receive_queue;
7767 struct sk_buff *skb;
7770 * See if we can merge this file into an existing skb SCM_RIGHTS
7771 * file set. If there's no room, fall back to allocating a new skb
7772 * and filling it in.
7774 spin_lock_irq(&head->lock);
7775 skb = skb_peek(head);
7777 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7779 if (fpl->count < SCM_MAX_FD) {
7780 __skb_unlink(skb, head);
7781 spin_unlock_irq(&head->lock);
7782 fpl->fp[fpl->count] = get_file(file);
7783 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7785 spin_lock_irq(&head->lock);
7786 __skb_queue_head(head, skb);
7791 spin_unlock_irq(&head->lock);
7798 return __io_sqe_files_scm(ctx, 1, index);
7804 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7805 struct io_rsrc_node *node, void *rsrc)
7807 struct io_rsrc_put *prsrc;
7809 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7813 prsrc->tag = *io_get_tag_slot(data, idx);
7815 list_add(&prsrc->list, &node->rsrc_list);
7819 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7820 struct io_uring_rsrc_update2 *up,
7823 u64 __user *tags = u64_to_user_ptr(up->tags);
7824 __s32 __user *fds = u64_to_user_ptr(up->data);
7825 struct io_rsrc_data *data = ctx->file_data;
7826 struct io_fixed_file *file_slot;
7830 bool needs_switch = false;
7832 if (!ctx->file_data)
7834 if (up->offset + nr_args > ctx->nr_user_files)
7837 for (done = 0; done < nr_args; done++) {
7840 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7841 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7845 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7849 if (fd == IORING_REGISTER_FILES_SKIP)
7852 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7853 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7855 if (file_slot->file_ptr) {
7856 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7857 err = io_queue_rsrc_removal(data, up->offset + done,
7858 ctx->rsrc_node, file);
7861 file_slot->file_ptr = 0;
7862 needs_switch = true;
7871 * Don't allow io_uring instances to be registered. If
7872 * UNIX isn't enabled, then this causes a reference
7873 * cycle and this instance can never get freed. If UNIX
7874 * is enabled we'll handle it just fine, but there's
7875 * still no point in allowing a ring fd as it doesn't
7876 * support regular read/write anyway.
7878 if (file->f_op == &io_uring_fops) {
7883 *io_get_tag_slot(data, up->offset + done) = tag;
7884 io_fixed_file_set(file_slot, file);
7885 err = io_sqe_file_register(ctx, file, i);
7887 file_slot->file_ptr = 0;
7895 io_rsrc_node_switch(ctx, data);
7896 return done ? done : err;
7899 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7901 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7903 req = io_put_req_find_next(req);
7904 return req ? &req->work : NULL;
7907 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7908 struct task_struct *task)
7910 struct io_wq_hash *hash;
7911 struct io_wq_data data;
7912 unsigned int concurrency;
7914 hash = ctx->hash_map;
7916 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7918 return ERR_PTR(-ENOMEM);
7919 refcount_set(&hash->refs, 1);
7920 init_waitqueue_head(&hash->wait);
7921 ctx->hash_map = hash;
7926 data.free_work = io_free_work;
7927 data.do_work = io_wq_submit_work;
7929 /* Do QD, or 4 * CPUS, whatever is smallest */
7930 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7932 return io_wq_create(concurrency, &data);
7935 static int io_uring_alloc_task_context(struct task_struct *task,
7936 struct io_ring_ctx *ctx)
7938 struct io_uring_task *tctx;
7941 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7942 if (unlikely(!tctx))
7945 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7946 if (unlikely(ret)) {
7951 tctx->io_wq = io_init_wq_offload(ctx, task);
7952 if (IS_ERR(tctx->io_wq)) {
7953 ret = PTR_ERR(tctx->io_wq);
7954 percpu_counter_destroy(&tctx->inflight);
7960 init_waitqueue_head(&tctx->wait);
7961 atomic_set(&tctx->in_idle, 0);
7962 atomic_set(&tctx->inflight_tracked, 0);
7963 task->io_uring = tctx;
7964 spin_lock_init(&tctx->task_lock);
7965 INIT_WQ_LIST(&tctx->task_list);
7966 init_task_work(&tctx->task_work, tctx_task_work);
7970 void __io_uring_free(struct task_struct *tsk)
7972 struct io_uring_task *tctx = tsk->io_uring;
7974 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7975 WARN_ON_ONCE(tctx->io_wq);
7976 WARN_ON_ONCE(tctx->cached_refs);
7978 percpu_counter_destroy(&tctx->inflight);
7980 tsk->io_uring = NULL;
7983 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7984 struct io_uring_params *p)
7988 /* Retain compatibility with failing for an invalid attach attempt */
7989 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7990 IORING_SETUP_ATTACH_WQ) {
7993 f = fdget(p->wq_fd);
7997 if (f.file->f_op != &io_uring_fops)
8000 if (ctx->flags & IORING_SETUP_SQPOLL) {
8001 struct task_struct *tsk;
8002 struct io_sq_data *sqd;
8005 sqd = io_get_sq_data(p, &attached);
8011 ctx->sq_creds = get_current_cred();
8013 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8014 if (!ctx->sq_thread_idle)
8015 ctx->sq_thread_idle = HZ;
8017 io_sq_thread_park(sqd);
8018 list_add(&ctx->sqd_list, &sqd->ctx_list);
8019 io_sqd_update_thread_idle(sqd);
8020 /* don't attach to a dying SQPOLL thread, would be racy */
8021 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8022 io_sq_thread_unpark(sqd);
8029 if (p->flags & IORING_SETUP_SQ_AFF) {
8030 int cpu = p->sq_thread_cpu;
8033 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8040 sqd->task_pid = current->pid;
8041 sqd->task_tgid = current->tgid;
8042 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8049 ret = io_uring_alloc_task_context(tsk, ctx);
8050 wake_up_new_task(tsk);
8053 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8054 /* Can't have SQ_AFF without SQPOLL */
8061 complete(&ctx->sq_data->exited);
8063 io_sq_thread_finish(ctx);
8067 static inline void __io_unaccount_mem(struct user_struct *user,
8068 unsigned long nr_pages)
8070 atomic_long_sub(nr_pages, &user->locked_vm);
8073 static inline int __io_account_mem(struct user_struct *user,
8074 unsigned long nr_pages)
8076 unsigned long page_limit, cur_pages, new_pages;
8078 /* Don't allow more pages than we can safely lock */
8079 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8082 cur_pages = atomic_long_read(&user->locked_vm);
8083 new_pages = cur_pages + nr_pages;
8084 if (new_pages > page_limit)
8086 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8087 new_pages) != cur_pages);
8092 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8095 __io_unaccount_mem(ctx->user, nr_pages);
8097 if (ctx->mm_account)
8098 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8101 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8106 ret = __io_account_mem(ctx->user, nr_pages);
8111 if (ctx->mm_account)
8112 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8117 static void io_mem_free(void *ptr)
8124 page = virt_to_head_page(ptr);
8125 if (put_page_testzero(page))
8126 free_compound_page(page);
8129 static void *io_mem_alloc(size_t size)
8131 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8132 __GFP_NORETRY | __GFP_ACCOUNT;
8134 return (void *) __get_free_pages(gfp_flags, get_order(size));
8137 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8140 struct io_rings *rings;
8141 size_t off, sq_array_size;
8143 off = struct_size(rings, cqes, cq_entries);
8144 if (off == SIZE_MAX)
8148 off = ALIGN(off, SMP_CACHE_BYTES);
8156 sq_array_size = array_size(sizeof(u32), sq_entries);
8157 if (sq_array_size == SIZE_MAX)
8160 if (check_add_overflow(off, sq_array_size, &off))
8166 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8168 struct io_mapped_ubuf *imu = *slot;
8171 if (imu != ctx->dummy_ubuf) {
8172 for (i = 0; i < imu->nr_bvecs; i++)
8173 unpin_user_page(imu->bvec[i].bv_page);
8174 if (imu->acct_pages)
8175 io_unaccount_mem(ctx, imu->acct_pages);
8181 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8183 io_buffer_unmap(ctx, &prsrc->buf);
8187 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8191 for (i = 0; i < ctx->nr_user_bufs; i++)
8192 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8193 kfree(ctx->user_bufs);
8194 io_rsrc_data_free(ctx->buf_data);
8195 ctx->user_bufs = NULL;
8196 ctx->buf_data = NULL;
8197 ctx->nr_user_bufs = 0;
8200 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8207 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8209 __io_sqe_buffers_unregister(ctx);
8213 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8214 void __user *arg, unsigned index)
8216 struct iovec __user *src;
8218 #ifdef CONFIG_COMPAT
8220 struct compat_iovec __user *ciovs;
8221 struct compat_iovec ciov;
8223 ciovs = (struct compat_iovec __user *) arg;
8224 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8227 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8228 dst->iov_len = ciov.iov_len;
8232 src = (struct iovec __user *) arg;
8233 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8239 * Not super efficient, but this is just a registration time. And we do cache
8240 * the last compound head, so generally we'll only do a full search if we don't
8243 * We check if the given compound head page has already been accounted, to
8244 * avoid double accounting it. This allows us to account the full size of the
8245 * page, not just the constituent pages of a huge page.
8247 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8248 int nr_pages, struct page *hpage)
8252 /* check current page array */
8253 for (i = 0; i < nr_pages; i++) {
8254 if (!PageCompound(pages[i]))
8256 if (compound_head(pages[i]) == hpage)
8260 /* check previously registered pages */
8261 for (i = 0; i < ctx->nr_user_bufs; i++) {
8262 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8264 for (j = 0; j < imu->nr_bvecs; j++) {
8265 if (!PageCompound(imu->bvec[j].bv_page))
8267 if (compound_head(imu->bvec[j].bv_page) == hpage)
8275 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8276 int nr_pages, struct io_mapped_ubuf *imu,
8277 struct page **last_hpage)
8281 imu->acct_pages = 0;
8282 for (i = 0; i < nr_pages; i++) {
8283 if (!PageCompound(pages[i])) {
8288 hpage = compound_head(pages[i]);
8289 if (hpage == *last_hpage)
8291 *last_hpage = hpage;
8292 if (headpage_already_acct(ctx, pages, i, hpage))
8294 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8298 if (!imu->acct_pages)
8301 ret = io_account_mem(ctx, imu->acct_pages);
8303 imu->acct_pages = 0;
8307 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8308 struct io_mapped_ubuf **pimu,
8309 struct page **last_hpage)
8311 struct io_mapped_ubuf *imu = NULL;
8312 struct vm_area_struct **vmas = NULL;
8313 struct page **pages = NULL;
8314 unsigned long off, start, end, ubuf;
8316 int ret, pret, nr_pages, i;
8318 if (!iov->iov_base) {
8319 *pimu = ctx->dummy_ubuf;
8323 ubuf = (unsigned long) iov->iov_base;
8324 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8325 start = ubuf >> PAGE_SHIFT;
8326 nr_pages = end - start;
8331 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8335 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8340 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8345 mmap_read_lock(current->mm);
8346 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8348 if (pret == nr_pages) {
8349 /* don't support file backed memory */
8350 for (i = 0; i < nr_pages; i++) {
8351 struct vm_area_struct *vma = vmas[i];
8353 if (vma_is_shmem(vma))
8356 !is_file_hugepages(vma->vm_file)) {
8362 ret = pret < 0 ? pret : -EFAULT;
8364 mmap_read_unlock(current->mm);
8367 * if we did partial map, or found file backed vmas,
8368 * release any pages we did get
8371 unpin_user_pages(pages, pret);
8375 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8377 unpin_user_pages(pages, pret);
8381 off = ubuf & ~PAGE_MASK;
8382 size = iov->iov_len;
8383 for (i = 0; i < nr_pages; i++) {
8386 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8387 imu->bvec[i].bv_page = pages[i];
8388 imu->bvec[i].bv_len = vec_len;
8389 imu->bvec[i].bv_offset = off;
8393 /* store original address for later verification */
8395 imu->ubuf_end = ubuf + iov->iov_len;
8396 imu->nr_bvecs = nr_pages;
8407 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8409 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8410 return ctx->user_bufs ? 0 : -ENOMEM;
8413 static int io_buffer_validate(struct iovec *iov)
8415 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8418 * Don't impose further limits on the size and buffer
8419 * constraints here, we'll -EINVAL later when IO is
8420 * submitted if they are wrong.
8423 return iov->iov_len ? -EFAULT : 0;
8427 /* arbitrary limit, but we need something */
8428 if (iov->iov_len > SZ_1G)
8431 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8437 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8438 unsigned int nr_args, u64 __user *tags)
8440 struct page *last_hpage = NULL;
8441 struct io_rsrc_data *data;
8447 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8449 ret = io_rsrc_node_switch_start(ctx);
8452 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8455 ret = io_buffers_map_alloc(ctx, nr_args);
8457 io_rsrc_data_free(data);
8461 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8462 ret = io_copy_iov(ctx, &iov, arg, i);
8465 ret = io_buffer_validate(&iov);
8468 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8473 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8479 WARN_ON_ONCE(ctx->buf_data);
8481 ctx->buf_data = data;
8483 __io_sqe_buffers_unregister(ctx);
8485 io_rsrc_node_switch(ctx, NULL);
8489 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8490 struct io_uring_rsrc_update2 *up,
8491 unsigned int nr_args)
8493 u64 __user *tags = u64_to_user_ptr(up->tags);
8494 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8495 struct page *last_hpage = NULL;
8496 bool needs_switch = false;
8502 if (up->offset + nr_args > ctx->nr_user_bufs)
8505 for (done = 0; done < nr_args; done++) {
8506 struct io_mapped_ubuf *imu;
8507 int offset = up->offset + done;
8510 err = io_copy_iov(ctx, &iov, iovs, done);
8513 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8517 err = io_buffer_validate(&iov);
8520 if (!iov.iov_base && tag) {
8524 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8528 i = array_index_nospec(offset, ctx->nr_user_bufs);
8529 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8530 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8531 ctx->rsrc_node, ctx->user_bufs[i]);
8532 if (unlikely(err)) {
8533 io_buffer_unmap(ctx, &imu);
8536 ctx->user_bufs[i] = NULL;
8537 needs_switch = true;
8540 ctx->user_bufs[i] = imu;
8541 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8545 io_rsrc_node_switch(ctx, ctx->buf_data);
8546 return done ? done : err;
8549 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8551 __s32 __user *fds = arg;
8557 if (copy_from_user(&fd, fds, sizeof(*fds)))
8560 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8561 if (IS_ERR(ctx->cq_ev_fd)) {
8562 int ret = PTR_ERR(ctx->cq_ev_fd);
8563 ctx->cq_ev_fd = NULL;
8570 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8572 if (ctx->cq_ev_fd) {
8573 eventfd_ctx_put(ctx->cq_ev_fd);
8574 ctx->cq_ev_fd = NULL;
8581 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8583 struct io_buffer *buf;
8584 unsigned long index;
8586 xa_for_each(&ctx->io_buffers, index, buf)
8587 __io_remove_buffers(ctx, buf, index, -1U);
8590 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8592 struct io_kiocb *req, *nxt;
8594 list_for_each_entry_safe(req, nxt, list, compl.list) {
8595 if (tsk && req->task != tsk)
8597 list_del(&req->compl.list);
8598 kmem_cache_free(req_cachep, req);
8602 static void io_req_caches_free(struct io_ring_ctx *ctx)
8604 struct io_submit_state *submit_state = &ctx->submit_state;
8605 struct io_comp_state *cs = &ctx->submit_state.comp;
8607 mutex_lock(&ctx->uring_lock);
8609 if (submit_state->free_reqs) {
8610 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8611 submit_state->reqs);
8612 submit_state->free_reqs = 0;
8615 io_flush_cached_locked_reqs(ctx, cs);
8616 io_req_cache_free(&cs->free_list, NULL);
8617 mutex_unlock(&ctx->uring_lock);
8620 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8624 if (!atomic_dec_and_test(&data->refs))
8625 wait_for_completion(&data->done);
8629 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8631 io_sq_thread_finish(ctx);
8633 if (ctx->mm_account) {
8634 mmdrop(ctx->mm_account);
8635 ctx->mm_account = NULL;
8638 mutex_lock(&ctx->uring_lock);
8639 if (io_wait_rsrc_data(ctx->buf_data))
8640 __io_sqe_buffers_unregister(ctx);
8641 if (io_wait_rsrc_data(ctx->file_data))
8642 __io_sqe_files_unregister(ctx);
8644 __io_cqring_overflow_flush(ctx, true);
8645 mutex_unlock(&ctx->uring_lock);
8646 io_eventfd_unregister(ctx);
8647 io_destroy_buffers(ctx);
8649 put_cred(ctx->sq_creds);
8651 /* there are no registered resources left, nobody uses it */
8653 io_rsrc_node_destroy(ctx->rsrc_node);
8654 if (ctx->rsrc_backup_node)
8655 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8656 flush_delayed_work(&ctx->rsrc_put_work);
8658 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8659 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8661 #if defined(CONFIG_UNIX)
8662 if (ctx->ring_sock) {
8663 ctx->ring_sock->file = NULL; /* so that iput() is called */
8664 sock_release(ctx->ring_sock);
8668 io_mem_free(ctx->rings);
8669 io_mem_free(ctx->sq_sqes);
8671 percpu_ref_exit(&ctx->refs);
8672 free_uid(ctx->user);
8673 io_req_caches_free(ctx);
8675 io_wq_put_hash(ctx->hash_map);
8676 kfree(ctx->cancel_hash);
8677 kfree(ctx->dummy_ubuf);
8681 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8683 struct io_ring_ctx *ctx = file->private_data;
8686 poll_wait(file, &ctx->cq_wait, wait);
8688 * synchronizes with barrier from wq_has_sleeper call in
8692 if (!io_sqring_full(ctx))
8693 mask |= EPOLLOUT | EPOLLWRNORM;
8696 * Don't flush cqring overflow list here, just do a simple check.
8697 * Otherwise there could possible be ABBA deadlock:
8700 * lock(&ctx->uring_lock);
8702 * lock(&ctx->uring_lock);
8705 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8706 * pushs them to do the flush.
8708 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8709 mask |= EPOLLIN | EPOLLRDNORM;
8714 static int io_uring_fasync(int fd, struct file *file, int on)
8716 struct io_ring_ctx *ctx = file->private_data;
8718 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8721 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8723 const struct cred *creds;
8725 creds = xa_erase(&ctx->personalities, id);
8734 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8736 return io_run_task_work_head(&ctx->exit_task_work);
8739 struct io_tctx_exit {
8740 struct callback_head task_work;
8741 struct completion completion;
8742 struct io_ring_ctx *ctx;
8745 static void io_tctx_exit_cb(struct callback_head *cb)
8747 struct io_uring_task *tctx = current->io_uring;
8748 struct io_tctx_exit *work;
8750 work = container_of(cb, struct io_tctx_exit, task_work);
8752 * When @in_idle, we're in cancellation and it's racy to remove the
8753 * node. It'll be removed by the end of cancellation, just ignore it.
8755 if (!atomic_read(&tctx->in_idle))
8756 io_uring_del_tctx_node((unsigned long)work->ctx);
8757 complete(&work->completion);
8760 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8762 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8764 return req->ctx == data;
8767 static void io_ring_exit_work(struct work_struct *work)
8769 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8770 unsigned long timeout = jiffies + HZ * 60 * 5;
8771 struct io_tctx_exit exit;
8772 struct io_tctx_node *node;
8776 * If we're doing polled IO and end up having requests being
8777 * submitted async (out-of-line), then completions can come in while
8778 * we're waiting for refs to drop. We need to reap these manually,
8779 * as nobody else will be looking for them.
8782 io_uring_try_cancel_requests(ctx, NULL, true);
8784 struct io_sq_data *sqd = ctx->sq_data;
8785 struct task_struct *tsk;
8787 io_sq_thread_park(sqd);
8789 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8790 io_wq_cancel_cb(tsk->io_uring->io_wq,
8791 io_cancel_ctx_cb, ctx, true);
8792 io_sq_thread_unpark(sqd);
8795 WARN_ON_ONCE(time_after(jiffies, timeout));
8796 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8798 init_completion(&exit.completion);
8799 init_task_work(&exit.task_work, io_tctx_exit_cb);
8802 * Some may use context even when all refs and requests have been put,
8803 * and they are free to do so while still holding uring_lock or
8804 * completion_lock, see __io_req_task_submit(). Apart from other work,
8805 * this lock/unlock section also waits them to finish.
8807 mutex_lock(&ctx->uring_lock);
8808 while (!list_empty(&ctx->tctx_list)) {
8809 WARN_ON_ONCE(time_after(jiffies, timeout));
8811 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8813 /* don't spin on a single task if cancellation failed */
8814 list_rotate_left(&ctx->tctx_list);
8815 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8816 if (WARN_ON_ONCE(ret))
8818 wake_up_process(node->task);
8820 mutex_unlock(&ctx->uring_lock);
8821 wait_for_completion(&exit.completion);
8822 mutex_lock(&ctx->uring_lock);
8824 mutex_unlock(&ctx->uring_lock);
8825 spin_lock_irq(&ctx->completion_lock);
8826 spin_unlock_irq(&ctx->completion_lock);
8828 io_ring_ctx_free(ctx);
8831 /* Returns true if we found and killed one or more timeouts */
8832 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8835 struct io_kiocb *req, *tmp;
8838 spin_lock_irq(&ctx->completion_lock);
8839 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8840 if (io_match_task(req, tsk, cancel_all)) {
8841 io_kill_timeout(req, -ECANCELED);
8846 io_commit_cqring(ctx);
8847 spin_unlock_irq(&ctx->completion_lock);
8849 io_cqring_ev_posted(ctx);
8850 return canceled != 0;
8853 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8855 unsigned long index;
8856 struct creds *creds;
8858 mutex_lock(&ctx->uring_lock);
8859 percpu_ref_kill(&ctx->refs);
8861 __io_cqring_overflow_flush(ctx, true);
8862 xa_for_each(&ctx->personalities, index, creds)
8863 io_unregister_personality(ctx, index);
8864 mutex_unlock(&ctx->uring_lock);
8866 io_kill_timeouts(ctx, NULL, true);
8867 io_poll_remove_all(ctx, NULL, true);
8869 /* if we failed setting up the ctx, we might not have any rings */
8870 io_iopoll_try_reap_events(ctx);
8872 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8874 * Use system_unbound_wq to avoid spawning tons of event kworkers
8875 * if we're exiting a ton of rings at the same time. It just adds
8876 * noise and overhead, there's no discernable change in runtime
8877 * over using system_wq.
8879 queue_work(system_unbound_wq, &ctx->exit_work);
8882 static int io_uring_release(struct inode *inode, struct file *file)
8884 struct io_ring_ctx *ctx = file->private_data;
8886 file->private_data = NULL;
8887 io_ring_ctx_wait_and_kill(ctx);
8891 struct io_task_cancel {
8892 struct task_struct *task;
8896 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8898 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8899 struct io_task_cancel *cancel = data;
8902 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8903 unsigned long flags;
8904 struct io_ring_ctx *ctx = req->ctx;
8906 /* protect against races with linked timeouts */
8907 spin_lock_irqsave(&ctx->completion_lock, flags);
8908 ret = io_match_task(req, cancel->task, cancel->all);
8909 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8911 ret = io_match_task(req, cancel->task, cancel->all);
8916 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8917 struct task_struct *task, bool cancel_all)
8919 struct io_defer_entry *de;
8922 spin_lock_irq(&ctx->completion_lock);
8923 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8924 if (io_match_task(de->req, task, cancel_all)) {
8925 list_cut_position(&list, &ctx->defer_list, &de->list);
8929 spin_unlock_irq(&ctx->completion_lock);
8930 if (list_empty(&list))
8933 while (!list_empty(&list)) {
8934 de = list_first_entry(&list, struct io_defer_entry, list);
8935 list_del_init(&de->list);
8936 io_req_complete_failed(de->req, -ECANCELED);
8942 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8944 struct io_tctx_node *node;
8945 enum io_wq_cancel cret;
8948 mutex_lock(&ctx->uring_lock);
8949 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8950 struct io_uring_task *tctx = node->task->io_uring;
8953 * io_wq will stay alive while we hold uring_lock, because it's
8954 * killed after ctx nodes, which requires to take the lock.
8956 if (!tctx || !tctx->io_wq)
8958 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8959 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8961 mutex_unlock(&ctx->uring_lock);
8966 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8967 struct task_struct *task,
8970 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8971 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8974 enum io_wq_cancel cret;
8978 ret |= io_uring_try_cancel_iowq(ctx);
8979 } else if (tctx && tctx->io_wq) {
8981 * Cancels requests of all rings, not only @ctx, but
8982 * it's fine as the task is in exit/exec.
8984 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8986 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8989 /* SQPOLL thread does its own polling */
8990 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8991 (ctx->sq_data && ctx->sq_data->thread == current)) {
8992 while (!list_empty_careful(&ctx->iopoll_list)) {
8993 io_iopoll_try_reap_events(ctx);
8998 ret |= io_cancel_defer_files(ctx, task, cancel_all);
8999 ret |= io_poll_remove_all(ctx, task, cancel_all);
9000 ret |= io_kill_timeouts(ctx, task, cancel_all);
9001 ret |= io_run_task_work();
9002 ret |= io_run_ctx_fallback(ctx);
9009 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9011 struct io_uring_task *tctx = current->io_uring;
9012 struct io_tctx_node *node;
9015 if (unlikely(!tctx)) {
9016 ret = io_uring_alloc_task_context(current, ctx);
9019 tctx = current->io_uring;
9021 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9022 node = kmalloc(sizeof(*node), GFP_KERNEL);
9026 node->task = current;
9028 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9035 mutex_lock(&ctx->uring_lock);
9036 list_add(&node->ctx_node, &ctx->tctx_list);
9037 mutex_unlock(&ctx->uring_lock);
9044 * Note that this task has used io_uring. We use it for cancelation purposes.
9046 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9048 struct io_uring_task *tctx = current->io_uring;
9050 if (likely(tctx && tctx->last == ctx))
9052 return __io_uring_add_tctx_node(ctx);
9056 * Remove this io_uring_file -> task mapping.
9058 static void io_uring_del_tctx_node(unsigned long index)
9060 struct io_uring_task *tctx = current->io_uring;
9061 struct io_tctx_node *node;
9065 node = xa_erase(&tctx->xa, index);
9069 WARN_ON_ONCE(current != node->task);
9070 WARN_ON_ONCE(list_empty(&node->ctx_node));
9072 mutex_lock(&node->ctx->uring_lock);
9073 list_del(&node->ctx_node);
9074 mutex_unlock(&node->ctx->uring_lock);
9076 if (tctx->last == node->ctx)
9081 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9083 struct io_wq *wq = tctx->io_wq;
9084 struct io_tctx_node *node;
9085 unsigned long index;
9087 xa_for_each(&tctx->xa, index, node)
9088 io_uring_del_tctx_node(index);
9091 * Must be after io_uring_del_task_file() (removes nodes under
9092 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9095 io_wq_put_and_exit(wq);
9099 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9102 return atomic_read(&tctx->inflight_tracked);
9103 return percpu_counter_sum(&tctx->inflight);
9106 static void io_uring_drop_tctx_refs(struct task_struct *task)
9108 struct io_uring_task *tctx = task->io_uring;
9109 unsigned int refs = tctx->cached_refs;
9111 tctx->cached_refs = 0;
9112 percpu_counter_sub(&tctx->inflight, refs);
9113 put_task_struct_many(task, refs);
9117 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9118 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9120 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9122 struct io_uring_task *tctx = current->io_uring;
9123 struct io_ring_ctx *ctx;
9127 WARN_ON_ONCE(sqd && sqd->thread != current);
9129 if (!current->io_uring)
9132 io_wq_exit_start(tctx->io_wq);
9134 io_uring_drop_tctx_refs(current);
9135 atomic_inc(&tctx->in_idle);
9137 /* read completions before cancelations */
9138 inflight = tctx_inflight(tctx, !cancel_all);
9143 struct io_tctx_node *node;
9144 unsigned long index;
9146 xa_for_each(&tctx->xa, index, node) {
9147 /* sqpoll task will cancel all its requests */
9148 if (node->ctx->sq_data)
9150 io_uring_try_cancel_requests(node->ctx, current,
9154 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9155 io_uring_try_cancel_requests(ctx, current,
9159 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9161 * If we've seen completions, retry without waiting. This
9162 * avoids a race where a completion comes in before we did
9163 * prepare_to_wait().
9165 if (inflight == tctx_inflight(tctx, !cancel_all))
9167 finish_wait(&tctx->wait, &wait);
9169 atomic_dec(&tctx->in_idle);
9171 io_uring_clean_tctx(tctx);
9173 /* for exec all current's requests should be gone, kill tctx */
9174 __io_uring_free(current);
9178 void __io_uring_cancel(struct files_struct *files)
9180 io_uring_cancel_generic(!files, NULL);
9183 static void *io_uring_validate_mmap_request(struct file *file,
9184 loff_t pgoff, size_t sz)
9186 struct io_ring_ctx *ctx = file->private_data;
9187 loff_t offset = pgoff << PAGE_SHIFT;
9192 case IORING_OFF_SQ_RING:
9193 case IORING_OFF_CQ_RING:
9196 case IORING_OFF_SQES:
9200 return ERR_PTR(-EINVAL);
9203 page = virt_to_head_page(ptr);
9204 if (sz > page_size(page))
9205 return ERR_PTR(-EINVAL);
9212 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9214 size_t sz = vma->vm_end - vma->vm_start;
9218 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9220 return PTR_ERR(ptr);
9222 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9223 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9226 #else /* !CONFIG_MMU */
9228 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9230 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9233 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9235 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9238 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9239 unsigned long addr, unsigned long len,
9240 unsigned long pgoff, unsigned long flags)
9244 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9246 return PTR_ERR(ptr);
9248 return (unsigned long) ptr;
9251 #endif /* !CONFIG_MMU */
9253 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9258 if (!io_sqring_full(ctx))
9260 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9262 if (!io_sqring_full(ctx))
9265 } while (!signal_pending(current));
9267 finish_wait(&ctx->sqo_sq_wait, &wait);
9271 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9272 struct __kernel_timespec __user **ts,
9273 const sigset_t __user **sig)
9275 struct io_uring_getevents_arg arg;
9278 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9279 * is just a pointer to the sigset_t.
9281 if (!(flags & IORING_ENTER_EXT_ARG)) {
9282 *sig = (const sigset_t __user *) argp;
9288 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9289 * timespec and sigset_t pointers if good.
9291 if (*argsz != sizeof(arg))
9293 if (copy_from_user(&arg, argp, sizeof(arg)))
9295 *sig = u64_to_user_ptr(arg.sigmask);
9296 *argsz = arg.sigmask_sz;
9297 *ts = u64_to_user_ptr(arg.ts);
9301 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9302 u32, min_complete, u32, flags, const void __user *, argp,
9305 struct io_ring_ctx *ctx;
9312 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9313 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9317 if (unlikely(!f.file))
9321 if (unlikely(f.file->f_op != &io_uring_fops))
9325 ctx = f.file->private_data;
9326 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9330 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9334 * For SQ polling, the thread will do all submissions and completions.
9335 * Just return the requested submit count, and wake the thread if
9339 if (ctx->flags & IORING_SETUP_SQPOLL) {
9340 io_cqring_overflow_flush(ctx, false);
9343 if (unlikely(ctx->sq_data->thread == NULL)) {
9346 if (flags & IORING_ENTER_SQ_WAKEUP)
9347 wake_up(&ctx->sq_data->wait);
9348 if (flags & IORING_ENTER_SQ_WAIT) {
9349 ret = io_sqpoll_wait_sq(ctx);
9353 submitted = to_submit;
9354 } else if (to_submit) {
9355 ret = io_uring_add_tctx_node(ctx);
9358 mutex_lock(&ctx->uring_lock);
9359 submitted = io_submit_sqes(ctx, to_submit);
9360 mutex_unlock(&ctx->uring_lock);
9362 if (submitted != to_submit)
9365 if (flags & IORING_ENTER_GETEVENTS) {
9366 const sigset_t __user *sig;
9367 struct __kernel_timespec __user *ts;
9369 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9373 min_complete = min(min_complete, ctx->cq_entries);
9376 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9377 * space applications don't need to do io completion events
9378 * polling again, they can rely on io_sq_thread to do polling
9379 * work, which can reduce cpu usage and uring_lock contention.
9381 if (ctx->flags & IORING_SETUP_IOPOLL &&
9382 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9383 ret = io_iopoll_check(ctx, min_complete);
9385 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9390 percpu_ref_put(&ctx->refs);
9393 return submitted ? submitted : ret;
9396 #ifdef CONFIG_PROC_FS
9397 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9398 const struct cred *cred)
9400 struct user_namespace *uns = seq_user_ns(m);
9401 struct group_info *gi;
9406 seq_printf(m, "%5d\n", id);
9407 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9408 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9409 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9410 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9411 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9412 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9413 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9414 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9415 seq_puts(m, "\n\tGroups:\t");
9416 gi = cred->group_info;
9417 for (g = 0; g < gi->ngroups; g++) {
9418 seq_put_decimal_ull(m, g ? " " : "",
9419 from_kgid_munged(uns, gi->gid[g]));
9421 seq_puts(m, "\n\tCapEff:\t");
9422 cap = cred->cap_effective;
9423 CAP_FOR_EACH_U32(__capi)
9424 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9429 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9431 struct io_sq_data *sq = NULL;
9436 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9437 * since fdinfo case grabs it in the opposite direction of normal use
9438 * cases. If we fail to get the lock, we just don't iterate any
9439 * structures that could be going away outside the io_uring mutex.
9441 has_lock = mutex_trylock(&ctx->uring_lock);
9443 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9449 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9450 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9451 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9452 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9453 struct file *f = io_file_from_index(ctx, i);
9456 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9458 seq_printf(m, "%5u: <none>\n", i);
9460 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9461 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9462 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9463 unsigned int len = buf->ubuf_end - buf->ubuf;
9465 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9467 if (has_lock && !xa_empty(&ctx->personalities)) {
9468 unsigned long index;
9469 const struct cred *cred;
9471 seq_printf(m, "Personalities:\n");
9472 xa_for_each(&ctx->personalities, index, cred)
9473 io_uring_show_cred(m, index, cred);
9475 seq_printf(m, "PollList:\n");
9476 spin_lock_irq(&ctx->completion_lock);
9477 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9478 struct hlist_head *list = &ctx->cancel_hash[i];
9479 struct io_kiocb *req;
9481 hlist_for_each_entry(req, list, hash_node)
9482 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9483 req->task->task_works != NULL);
9485 spin_unlock_irq(&ctx->completion_lock);
9487 mutex_unlock(&ctx->uring_lock);
9490 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9492 struct io_ring_ctx *ctx = f->private_data;
9494 if (percpu_ref_tryget(&ctx->refs)) {
9495 __io_uring_show_fdinfo(ctx, m);
9496 percpu_ref_put(&ctx->refs);
9501 static const struct file_operations io_uring_fops = {
9502 .release = io_uring_release,
9503 .mmap = io_uring_mmap,
9505 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9506 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9508 .poll = io_uring_poll,
9509 .fasync = io_uring_fasync,
9510 #ifdef CONFIG_PROC_FS
9511 .show_fdinfo = io_uring_show_fdinfo,
9515 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9516 struct io_uring_params *p)
9518 struct io_rings *rings;
9519 size_t size, sq_array_offset;
9521 /* make sure these are sane, as we already accounted them */
9522 ctx->sq_entries = p->sq_entries;
9523 ctx->cq_entries = p->cq_entries;
9525 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9526 if (size == SIZE_MAX)
9529 rings = io_mem_alloc(size);
9534 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9535 rings->sq_ring_mask = p->sq_entries - 1;
9536 rings->cq_ring_mask = p->cq_entries - 1;
9537 rings->sq_ring_entries = p->sq_entries;
9538 rings->cq_ring_entries = p->cq_entries;
9540 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9541 if (size == SIZE_MAX) {
9542 io_mem_free(ctx->rings);
9547 ctx->sq_sqes = io_mem_alloc(size);
9548 if (!ctx->sq_sqes) {
9549 io_mem_free(ctx->rings);
9557 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9561 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9565 ret = io_uring_add_tctx_node(ctx);
9570 fd_install(fd, file);
9575 * Allocate an anonymous fd, this is what constitutes the application
9576 * visible backing of an io_uring instance. The application mmaps this
9577 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9578 * we have to tie this fd to a socket for file garbage collection purposes.
9580 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9583 #if defined(CONFIG_UNIX)
9586 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9589 return ERR_PTR(ret);
9592 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9593 O_RDWR | O_CLOEXEC);
9594 #if defined(CONFIG_UNIX)
9596 sock_release(ctx->ring_sock);
9597 ctx->ring_sock = NULL;
9599 ctx->ring_sock->file = file;
9605 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9606 struct io_uring_params __user *params)
9608 struct io_ring_ctx *ctx;
9614 if (entries > IORING_MAX_ENTRIES) {
9615 if (!(p->flags & IORING_SETUP_CLAMP))
9617 entries = IORING_MAX_ENTRIES;
9621 * Use twice as many entries for the CQ ring. It's possible for the
9622 * application to drive a higher depth than the size of the SQ ring,
9623 * since the sqes are only used at submission time. This allows for
9624 * some flexibility in overcommitting a bit. If the application has
9625 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9626 * of CQ ring entries manually.
9628 p->sq_entries = roundup_pow_of_two(entries);
9629 if (p->flags & IORING_SETUP_CQSIZE) {
9631 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9632 * to a power-of-two, if it isn't already. We do NOT impose
9633 * any cq vs sq ring sizing.
9637 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9638 if (!(p->flags & IORING_SETUP_CLAMP))
9640 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9642 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9643 if (p->cq_entries < p->sq_entries)
9646 p->cq_entries = 2 * p->sq_entries;
9649 ctx = io_ring_ctx_alloc(p);
9652 ctx->compat = in_compat_syscall();
9653 if (!capable(CAP_IPC_LOCK))
9654 ctx->user = get_uid(current_user());
9657 * This is just grabbed for accounting purposes. When a process exits,
9658 * the mm is exited and dropped before the files, hence we need to hang
9659 * on to this mm purely for the purposes of being able to unaccount
9660 * memory (locked/pinned vm). It's not used for anything else.
9662 mmgrab(current->mm);
9663 ctx->mm_account = current->mm;
9665 ret = io_allocate_scq_urings(ctx, p);
9669 ret = io_sq_offload_create(ctx, p);
9672 /* always set a rsrc node */
9673 ret = io_rsrc_node_switch_start(ctx);
9676 io_rsrc_node_switch(ctx, NULL);
9678 memset(&p->sq_off, 0, sizeof(p->sq_off));
9679 p->sq_off.head = offsetof(struct io_rings, sq.head);
9680 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9681 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9682 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9683 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9684 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9685 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9687 memset(&p->cq_off, 0, sizeof(p->cq_off));
9688 p->cq_off.head = offsetof(struct io_rings, cq.head);
9689 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9690 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9691 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9692 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9693 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9694 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9696 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9697 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9698 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9699 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9700 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9701 IORING_FEAT_RSRC_TAGS;
9703 if (copy_to_user(params, p, sizeof(*p))) {
9708 file = io_uring_get_file(ctx);
9710 ret = PTR_ERR(file);
9715 * Install ring fd as the very last thing, so we don't risk someone
9716 * having closed it before we finish setup
9718 ret = io_uring_install_fd(ctx, file);
9720 /* fput will clean it up */
9725 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9728 io_ring_ctx_wait_and_kill(ctx);
9733 * Sets up an aio uring context, and returns the fd. Applications asks for a
9734 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9735 * params structure passed in.
9737 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9739 struct io_uring_params p;
9742 if (copy_from_user(&p, params, sizeof(p)))
9744 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9749 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9750 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9751 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9752 IORING_SETUP_R_DISABLED))
9755 return io_uring_create(entries, &p, params);
9758 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9759 struct io_uring_params __user *, params)
9761 return io_uring_setup(entries, params);
9764 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9766 struct io_uring_probe *p;
9770 size = struct_size(p, ops, nr_args);
9771 if (size == SIZE_MAX)
9773 p = kzalloc(size, GFP_KERNEL);
9778 if (copy_from_user(p, arg, size))
9781 if (memchr_inv(p, 0, size))
9784 p->last_op = IORING_OP_LAST - 1;
9785 if (nr_args > IORING_OP_LAST)
9786 nr_args = IORING_OP_LAST;
9788 for (i = 0; i < nr_args; i++) {
9790 if (!io_op_defs[i].not_supported)
9791 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9796 if (copy_to_user(arg, p, size))
9803 static int io_register_personality(struct io_ring_ctx *ctx)
9805 const struct cred *creds;
9809 creds = get_current_cred();
9811 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9812 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9819 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9820 unsigned int nr_args)
9822 struct io_uring_restriction *res;
9826 /* Restrictions allowed only if rings started disabled */
9827 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9830 /* We allow only a single restrictions registration */
9831 if (ctx->restrictions.registered)
9834 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9837 size = array_size(nr_args, sizeof(*res));
9838 if (size == SIZE_MAX)
9841 res = memdup_user(arg, size);
9843 return PTR_ERR(res);
9847 for (i = 0; i < nr_args; i++) {
9848 switch (res[i].opcode) {
9849 case IORING_RESTRICTION_REGISTER_OP:
9850 if (res[i].register_op >= IORING_REGISTER_LAST) {
9855 __set_bit(res[i].register_op,
9856 ctx->restrictions.register_op);
9858 case IORING_RESTRICTION_SQE_OP:
9859 if (res[i].sqe_op >= IORING_OP_LAST) {
9864 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9866 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9867 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9869 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9870 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9879 /* Reset all restrictions if an error happened */
9881 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9883 ctx->restrictions.registered = true;
9889 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9891 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9894 if (ctx->restrictions.registered)
9895 ctx->restricted = 1;
9897 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9898 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9899 wake_up(&ctx->sq_data->wait);
9903 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9904 struct io_uring_rsrc_update2 *up,
9912 if (check_add_overflow(up->offset, nr_args, &tmp))
9914 err = io_rsrc_node_switch_start(ctx);
9919 case IORING_RSRC_FILE:
9920 return __io_sqe_files_update(ctx, up, nr_args);
9921 case IORING_RSRC_BUFFER:
9922 return __io_sqe_buffers_update(ctx, up, nr_args);
9927 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9930 struct io_uring_rsrc_update2 up;
9934 memset(&up, 0, sizeof(up));
9935 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9937 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9940 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9941 unsigned size, unsigned type)
9943 struct io_uring_rsrc_update2 up;
9945 if (size != sizeof(up))
9947 if (copy_from_user(&up, arg, sizeof(up)))
9949 if (!up.nr || up.resv)
9951 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9954 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9955 unsigned int size, unsigned int type)
9957 struct io_uring_rsrc_register rr;
9959 /* keep it extendible */
9960 if (size != sizeof(rr))
9963 memset(&rr, 0, sizeof(rr));
9964 if (copy_from_user(&rr, arg, size))
9966 if (!rr.nr || rr.resv || rr.resv2)
9970 case IORING_RSRC_FILE:
9971 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9972 rr.nr, u64_to_user_ptr(rr.tags));
9973 case IORING_RSRC_BUFFER:
9974 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9975 rr.nr, u64_to_user_ptr(rr.tags));
9980 static bool io_register_op_must_quiesce(int op)
9983 case IORING_REGISTER_BUFFERS:
9984 case IORING_UNREGISTER_BUFFERS:
9985 case IORING_REGISTER_FILES:
9986 case IORING_UNREGISTER_FILES:
9987 case IORING_REGISTER_FILES_UPDATE:
9988 case IORING_REGISTER_PROBE:
9989 case IORING_REGISTER_PERSONALITY:
9990 case IORING_UNREGISTER_PERSONALITY:
9991 case IORING_REGISTER_FILES2:
9992 case IORING_REGISTER_FILES_UPDATE2:
9993 case IORING_REGISTER_BUFFERS2:
9994 case IORING_REGISTER_BUFFERS_UPDATE:
10001 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10002 void __user *arg, unsigned nr_args)
10003 __releases(ctx->uring_lock)
10004 __acquires(ctx->uring_lock)
10009 * We're inside the ring mutex, if the ref is already dying, then
10010 * someone else killed the ctx or is already going through
10011 * io_uring_register().
10013 if (percpu_ref_is_dying(&ctx->refs))
10016 if (ctx->restricted) {
10017 if (opcode >= IORING_REGISTER_LAST)
10019 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10020 if (!test_bit(opcode, ctx->restrictions.register_op))
10024 if (io_register_op_must_quiesce(opcode)) {
10025 percpu_ref_kill(&ctx->refs);
10028 * Drop uring mutex before waiting for references to exit. If
10029 * another thread is currently inside io_uring_enter() it might
10030 * need to grab the uring_lock to make progress. If we hold it
10031 * here across the drain wait, then we can deadlock. It's safe
10032 * to drop the mutex here, since no new references will come in
10033 * after we've killed the percpu ref.
10035 mutex_unlock(&ctx->uring_lock);
10037 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10040 ret = io_run_task_work_sig();
10044 mutex_lock(&ctx->uring_lock);
10047 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10053 case IORING_REGISTER_BUFFERS:
10054 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10056 case IORING_UNREGISTER_BUFFERS:
10058 if (arg || nr_args)
10060 ret = io_sqe_buffers_unregister(ctx);
10062 case IORING_REGISTER_FILES:
10063 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10065 case IORING_UNREGISTER_FILES:
10067 if (arg || nr_args)
10069 ret = io_sqe_files_unregister(ctx);
10071 case IORING_REGISTER_FILES_UPDATE:
10072 ret = io_register_files_update(ctx, arg, nr_args);
10074 case IORING_REGISTER_EVENTFD:
10075 case IORING_REGISTER_EVENTFD_ASYNC:
10079 ret = io_eventfd_register(ctx, arg);
10082 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10083 ctx->eventfd_async = 1;
10085 ctx->eventfd_async = 0;
10087 case IORING_UNREGISTER_EVENTFD:
10089 if (arg || nr_args)
10091 ret = io_eventfd_unregister(ctx);
10093 case IORING_REGISTER_PROBE:
10095 if (!arg || nr_args > 256)
10097 ret = io_probe(ctx, arg, nr_args);
10099 case IORING_REGISTER_PERSONALITY:
10101 if (arg || nr_args)
10103 ret = io_register_personality(ctx);
10105 case IORING_UNREGISTER_PERSONALITY:
10109 ret = io_unregister_personality(ctx, nr_args);
10111 case IORING_REGISTER_ENABLE_RINGS:
10113 if (arg || nr_args)
10115 ret = io_register_enable_rings(ctx);
10117 case IORING_REGISTER_RESTRICTIONS:
10118 ret = io_register_restrictions(ctx, arg, nr_args);
10120 case IORING_REGISTER_FILES2:
10121 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10123 case IORING_REGISTER_FILES_UPDATE2:
10124 ret = io_register_rsrc_update(ctx, arg, nr_args,
10127 case IORING_REGISTER_BUFFERS2:
10128 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10130 case IORING_REGISTER_BUFFERS_UPDATE:
10131 ret = io_register_rsrc_update(ctx, arg, nr_args,
10132 IORING_RSRC_BUFFER);
10139 if (io_register_op_must_quiesce(opcode)) {
10140 /* bring the ctx back to life */
10141 percpu_ref_reinit(&ctx->refs);
10142 reinit_completion(&ctx->ref_comp);
10147 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10148 void __user *, arg, unsigned int, nr_args)
10150 struct io_ring_ctx *ctx;
10159 if (f.file->f_op != &io_uring_fops)
10162 ctx = f.file->private_data;
10164 io_run_task_work();
10166 mutex_lock(&ctx->uring_lock);
10167 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10168 mutex_unlock(&ctx->uring_lock);
10169 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10170 ctx->cq_ev_fd != NULL, ret);
10176 static int __init io_uring_init(void)
10178 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10179 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10180 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10183 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10184 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10185 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10186 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10187 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10188 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10189 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10190 BUILD_BUG_SQE_ELEM(8, __u64, off);
10191 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10192 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10193 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10194 BUILD_BUG_SQE_ELEM(24, __u32, len);
10195 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10196 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10197 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10198 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10199 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10200 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10201 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10202 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10203 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10204 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10205 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10206 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10207 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10208 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10209 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10210 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10211 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10212 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10213 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10215 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10216 sizeof(struct io_uring_rsrc_update));
10217 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10218 sizeof(struct io_uring_rsrc_update2));
10219 /* should fit into one byte */
10220 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10222 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10223 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10224 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10228 __initcall(io_uring_init);