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 | \
114 u32 head ____cacheline_aligned_in_smp;
115 u32 tail ____cacheline_aligned_in_smp;
119 * This data is shared with the application through the mmap at offsets
120 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
122 * The offsets to the member fields are published through struct
123 * io_sqring_offsets when calling io_uring_setup.
127 * Head and tail offsets into the ring; the offsets need to be
128 * masked to get valid indices.
130 * The kernel controls head of the sq ring and the tail of the cq ring,
131 * and the application controls tail of the sq ring and the head of the
134 struct io_uring sq, cq;
136 * Bitmasks to apply to head and tail offsets (constant, equals
139 u32 sq_ring_mask, cq_ring_mask;
140 /* Ring sizes (constant, power of 2) */
141 u32 sq_ring_entries, cq_ring_entries;
143 * Number of invalid entries dropped by the kernel due to
144 * invalid index stored in array
146 * Written by the kernel, shouldn't be modified by the
147 * application (i.e. get number of "new events" by comparing to
150 * After a new SQ head value was read by the application this
151 * counter includes all submissions that were dropped reaching
152 * the new SQ head (and possibly more).
158 * Written by the kernel, shouldn't be modified by the
161 * The application needs a full memory barrier before checking
162 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
168 * Written by the application, shouldn't be modified by the
173 * Number of completion events lost because the queue was full;
174 * this should be avoided by the application by making sure
175 * there are not more requests pending than there is space in
176 * the completion queue.
178 * Written by the kernel, shouldn't be modified by the
179 * application (i.e. get number of "new events" by comparing to
182 * As completion events come in out of order this counter is not
183 * ordered with any other data.
187 * Ring buffer of completion events.
189 * The kernel writes completion events fresh every time they are
190 * produced, so the application is allowed to modify pending
193 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
196 enum io_uring_cmd_flags {
197 IO_URING_F_NONBLOCK = 1,
198 IO_URING_F_COMPLETE_DEFER = 2,
201 struct io_mapped_ubuf {
204 unsigned int nr_bvecs;
205 unsigned long acct_pages;
206 struct bio_vec bvec[];
211 struct io_overflow_cqe {
212 struct io_uring_cqe cqe;
213 struct list_head list;
216 struct io_fixed_file {
217 /* file * with additional FFS_* flags */
218 unsigned long file_ptr;
222 struct list_head list;
227 struct io_mapped_ubuf *buf;
231 struct io_file_table {
232 /* two level table */
233 struct io_fixed_file **files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_IOPOLL_BATCH 8
299 #define IO_COMPL_BATCH 32
300 #define IO_REQ_CACHE_SIZE 32
301 #define IO_REQ_ALLOC_BATCH 8
303 struct io_comp_state {
304 struct io_kiocb *reqs[IO_COMPL_BATCH];
306 /* inline/task_work completion list, under ->uring_lock */
307 struct list_head free_list;
310 struct io_submit_link {
311 struct io_kiocb *head;
312 struct io_kiocb *last;
315 struct io_submit_state {
316 struct blk_plug plug;
317 struct io_submit_link link;
320 * io_kiocb alloc cache
322 void *reqs[IO_REQ_CACHE_SIZE];
323 unsigned int free_reqs;
328 * Batch completion logic
330 struct io_comp_state comp;
333 * File reference cache
337 unsigned int file_refs;
338 unsigned int ios_left;
343 struct percpu_ref refs;
344 } ____cacheline_aligned_in_smp;
348 unsigned int compat: 1;
349 unsigned int drain_next: 1;
350 unsigned int eventfd_async: 1;
351 unsigned int restricted: 1;
354 * Ring buffer of indices into array of io_uring_sqe, which is
355 * mmapped by the application using the IORING_OFF_SQES offset.
357 * This indirection could e.g. be used to assign fixed
358 * io_uring_sqe entries to operations and only submit them to
359 * the queue when needed.
361 * The kernel modifies neither the indices array nor the entries
365 unsigned cached_sq_head;
367 unsigned sq_thread_idle;
368 unsigned cached_sq_dropped;
369 unsigned long sq_check_overflow;
371 struct list_head defer_list;
372 struct list_head timeout_list;
373 struct list_head cq_overflow_list;
375 struct io_uring_sqe *sq_sqes;
376 } ____cacheline_aligned_in_smp;
379 struct mutex uring_lock;
380 wait_queue_head_t wait;
381 } ____cacheline_aligned_in_smp;
383 struct io_submit_state submit_state;
384 /* IRQ completion list, under ->completion_lock */
385 struct list_head locked_free_list;
386 unsigned int locked_free_nr;
388 struct io_rings *rings;
390 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
397 * Fixed resources fast path, should be accessed only under uring_lock,
398 * and updated through io_uring_register(2)
400 struct io_rsrc_node *rsrc_node;
402 struct io_file_table file_table;
403 unsigned nr_user_files;
404 unsigned nr_user_bufs;
405 struct io_mapped_ubuf **user_bufs;
407 struct xarray io_buffers;
408 struct xarray personalities;
412 unsigned cached_cq_tail;
414 atomic_t cq_timeouts;
415 unsigned cq_last_tm_flush;
417 unsigned long cq_check_overflow;
418 struct wait_queue_head cq_wait;
419 struct fasync_struct *cq_fasync;
420 struct eventfd_ctx *cq_ev_fd;
421 } ____cacheline_aligned_in_smp;
424 spinlock_t completion_lock;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list;
433 struct hlist_head *cancel_hash;
434 unsigned cancel_hash_bits;
435 bool poll_multi_file;
436 } ____cacheline_aligned_in_smp;
438 struct io_restriction restrictions;
440 /* slow path rsrc auxilary data, used by update/register */
442 struct io_rsrc_node *rsrc_backup_node;
443 struct io_mapped_ubuf *dummy_ubuf;
444 struct io_rsrc_data *file_data;
445 struct io_rsrc_data *buf_data;
447 struct delayed_work rsrc_put_work;
448 struct llist_head rsrc_put_llist;
449 struct list_head rsrc_ref_list;
450 spinlock_t rsrc_ref_lock;
453 /* Keep this last, we don't need it for the fast path */
455 #if defined(CONFIG_UNIX)
456 struct socket *ring_sock;
458 /* hashed buffered write serialization */
459 struct io_wq_hash *hash_map;
461 /* Only used for accounting purposes */
462 struct user_struct *user;
463 struct mm_struct *mm_account;
465 /* ctx exit and cancelation */
466 struct callback_head *exit_task_work;
467 struct work_struct exit_work;
468 struct list_head tctx_list;
469 struct completion ref_comp;
473 struct io_uring_task {
474 /* submission side */
476 struct wait_queue_head wait;
477 const struct io_ring_ctx *last;
479 struct percpu_counter inflight;
480 atomic_t inflight_tracked;
483 spinlock_t task_lock;
484 struct io_wq_work_list task_list;
485 unsigned long task_state;
486 struct callback_head task_work;
490 * First field must be the file pointer in all the
491 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
493 struct io_poll_iocb {
495 struct wait_queue_head *head;
499 struct wait_queue_entry wait;
502 struct io_poll_update {
508 bool update_user_data;
516 struct io_timeout_data {
517 struct io_kiocb *req;
518 struct hrtimer timer;
519 struct timespec64 ts;
520 enum hrtimer_mode mode;
525 struct sockaddr __user *addr;
526 int __user *addr_len;
528 unsigned long nofile;
548 struct list_head list;
549 /* head of the link, used by linked timeouts only */
550 struct io_kiocb *head;
553 struct io_timeout_rem {
558 struct timespec64 ts;
563 /* NOTE: kiocb has the file as the first member, so don't do it here */
571 struct sockaddr __user *addr;
578 struct compat_msghdr __user *umsg_compat;
579 struct user_msghdr __user *umsg;
585 struct io_buffer *kbuf;
591 struct filename *filename;
593 unsigned long nofile;
596 struct io_rsrc_update {
622 struct epoll_event event;
626 struct file *file_out;
627 struct file *file_in;
634 struct io_provide_buf {
648 const char __user *filename;
649 struct statx __user *buffer;
661 struct filename *oldpath;
662 struct filename *newpath;
670 struct filename *filename;
673 struct io_completion {
675 struct list_head list;
679 struct io_async_connect {
680 struct sockaddr_storage address;
683 struct io_async_msghdr {
684 struct iovec fast_iov[UIO_FASTIOV];
685 /* points to an allocated iov, if NULL we use fast_iov instead */
686 struct iovec *free_iov;
687 struct sockaddr __user *uaddr;
689 struct sockaddr_storage addr;
693 struct iovec fast_iov[UIO_FASTIOV];
694 const struct iovec *free_iovec;
695 struct iov_iter iter;
697 struct wait_page_queue wpq;
701 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
702 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
703 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
704 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
705 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
706 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
708 /* first byte is taken by user flags, shift it to not overlap */
713 REQ_F_LINK_TIMEOUT_BIT,
714 REQ_F_NEED_CLEANUP_BIT,
716 REQ_F_BUFFER_SELECTED_BIT,
717 REQ_F_LTIMEOUT_ACTIVE_BIT,
718 REQ_F_COMPLETE_INLINE_BIT,
720 REQ_F_DONT_REISSUE_BIT,
721 /* keep async read/write and isreg together and in order */
722 REQ_F_ASYNC_READ_BIT,
723 REQ_F_ASYNC_WRITE_BIT,
726 /* not a real bit, just to check we're not overflowing the space */
732 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
733 /* drain existing IO first */
734 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
736 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
737 /* doesn't sever on completion < 0 */
738 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
740 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
741 /* IOSQE_BUFFER_SELECT */
742 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
744 /* fail rest of links */
745 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
746 /* on inflight list, should be cancelled and waited on exit reliably */
747 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
748 /* read/write uses file position */
749 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
750 /* must not punt to workers */
751 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
752 /* has or had linked timeout */
753 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
755 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
756 /* already went through poll handler */
757 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
758 /* buffer already selected */
759 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
760 /* linked timeout is active, i.e. prepared by link's head */
761 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
762 /* completion is deferred through io_comp_state */
763 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
764 /* caller should reissue async */
765 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
766 /* don't attempt request reissue, see io_rw_reissue() */
767 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
768 /* supports async reads */
769 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
770 /* supports async writes */
771 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
773 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
777 struct io_poll_iocb poll;
778 struct io_poll_iocb *double_poll;
781 struct io_task_work {
782 struct io_wq_work_node node;
783 task_work_func_t func;
787 IORING_RSRC_FILE = 0,
788 IORING_RSRC_BUFFER = 1,
792 * NOTE! Each of the iocb union members has the file pointer
793 * as the first entry in their struct definition. So you can
794 * access the file pointer through any of the sub-structs,
795 * or directly as just 'ki_filp' in this struct.
801 struct io_poll_iocb poll;
802 struct io_poll_update poll_update;
803 struct io_accept accept;
805 struct io_cancel cancel;
806 struct io_timeout timeout;
807 struct io_timeout_rem timeout_rem;
808 struct io_connect connect;
809 struct io_sr_msg sr_msg;
811 struct io_close close;
812 struct io_rsrc_update rsrc_update;
813 struct io_fadvise fadvise;
814 struct io_madvise madvise;
815 struct io_epoll epoll;
816 struct io_splice splice;
817 struct io_provide_buf pbuf;
818 struct io_statx statx;
819 struct io_shutdown shutdown;
820 struct io_rename rename;
821 struct io_unlink unlink;
822 /* use only after cleaning per-op data, see io_clean_op() */
823 struct io_completion compl;
826 /* opcode allocated if it needs to store data for async defer */
829 /* polled IO has completed */
835 struct io_ring_ctx *ctx;
838 struct task_struct *task;
841 struct io_kiocb *link;
842 struct percpu_ref *fixed_rsrc_refs;
844 /* used with ctx->iopoll_list with reads/writes */
845 struct list_head inflight_entry;
847 struct io_task_work io_task_work;
848 struct callback_head task_work;
850 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
851 struct hlist_node hash_node;
852 struct async_poll *apoll;
853 struct io_wq_work work;
854 /* store used ubuf, so we can prevent reloading */
855 struct io_mapped_ubuf *imu;
858 struct io_tctx_node {
859 struct list_head ctx_node;
860 struct task_struct *task;
861 struct io_ring_ctx *ctx;
864 struct io_defer_entry {
865 struct list_head list;
866 struct io_kiocb *req;
871 /* needs req->file assigned */
872 unsigned needs_file : 1;
873 /* hash wq insertion if file is a regular file */
874 unsigned hash_reg_file : 1;
875 /* unbound wq insertion if file is a non-regular file */
876 unsigned unbound_nonreg_file : 1;
877 /* opcode is not supported by this kernel */
878 unsigned not_supported : 1;
879 /* set if opcode supports polled "wait" */
881 unsigned pollout : 1;
882 /* op supports buffer selection */
883 unsigned buffer_select : 1;
884 /* do prep async if is going to be punted */
885 unsigned needs_async_setup : 1;
886 /* should block plug */
888 /* size of async data needed, if any */
889 unsigned short async_size;
892 static const struct io_op_def io_op_defs[] = {
893 [IORING_OP_NOP] = {},
894 [IORING_OP_READV] = {
896 .unbound_nonreg_file = 1,
899 .needs_async_setup = 1,
901 .async_size = sizeof(struct io_async_rw),
903 [IORING_OP_WRITEV] = {
906 .unbound_nonreg_file = 1,
908 .needs_async_setup = 1,
910 .async_size = sizeof(struct io_async_rw),
912 [IORING_OP_FSYNC] = {
915 [IORING_OP_READ_FIXED] = {
917 .unbound_nonreg_file = 1,
920 .async_size = sizeof(struct io_async_rw),
922 [IORING_OP_WRITE_FIXED] = {
925 .unbound_nonreg_file = 1,
928 .async_size = sizeof(struct io_async_rw),
930 [IORING_OP_POLL_ADD] = {
932 .unbound_nonreg_file = 1,
934 [IORING_OP_POLL_REMOVE] = {},
935 [IORING_OP_SYNC_FILE_RANGE] = {
938 [IORING_OP_SENDMSG] = {
940 .unbound_nonreg_file = 1,
942 .needs_async_setup = 1,
943 .async_size = sizeof(struct io_async_msghdr),
945 [IORING_OP_RECVMSG] = {
947 .unbound_nonreg_file = 1,
950 .needs_async_setup = 1,
951 .async_size = sizeof(struct io_async_msghdr),
953 [IORING_OP_TIMEOUT] = {
954 .async_size = sizeof(struct io_timeout_data),
956 [IORING_OP_TIMEOUT_REMOVE] = {
957 /* used by timeout updates' prep() */
959 [IORING_OP_ACCEPT] = {
961 .unbound_nonreg_file = 1,
964 [IORING_OP_ASYNC_CANCEL] = {},
965 [IORING_OP_LINK_TIMEOUT] = {
966 .async_size = sizeof(struct io_timeout_data),
968 [IORING_OP_CONNECT] = {
970 .unbound_nonreg_file = 1,
972 .needs_async_setup = 1,
973 .async_size = sizeof(struct io_async_connect),
975 [IORING_OP_FALLOCATE] = {
978 [IORING_OP_OPENAT] = {},
979 [IORING_OP_CLOSE] = {},
980 [IORING_OP_FILES_UPDATE] = {},
981 [IORING_OP_STATX] = {},
984 .unbound_nonreg_file = 1,
988 .async_size = sizeof(struct io_async_rw),
990 [IORING_OP_WRITE] = {
992 .unbound_nonreg_file = 1,
995 .async_size = sizeof(struct io_async_rw),
997 [IORING_OP_FADVISE] = {
1000 [IORING_OP_MADVISE] = {},
1001 [IORING_OP_SEND] = {
1003 .unbound_nonreg_file = 1,
1006 [IORING_OP_RECV] = {
1008 .unbound_nonreg_file = 1,
1012 [IORING_OP_OPENAT2] = {
1014 [IORING_OP_EPOLL_CTL] = {
1015 .unbound_nonreg_file = 1,
1017 [IORING_OP_SPLICE] = {
1020 .unbound_nonreg_file = 1,
1022 [IORING_OP_PROVIDE_BUFFERS] = {},
1023 [IORING_OP_REMOVE_BUFFERS] = {},
1027 .unbound_nonreg_file = 1,
1029 [IORING_OP_SHUTDOWN] = {
1032 [IORING_OP_RENAMEAT] = {},
1033 [IORING_OP_UNLINKAT] = {},
1036 static bool io_disarm_next(struct io_kiocb *req);
1037 static void io_uring_del_tctx_node(unsigned long index);
1038 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1039 struct task_struct *task,
1041 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd);
1042 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1044 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1045 long res, unsigned int cflags);
1046 static void io_put_req(struct io_kiocb *req);
1047 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1048 static void io_dismantle_req(struct io_kiocb *req);
1049 static void io_put_task(struct task_struct *task, int nr);
1050 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1051 static void io_queue_linked_timeout(struct io_kiocb *req);
1052 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1053 struct io_uring_rsrc_update2 *up,
1055 static void io_clean_op(struct io_kiocb *req);
1056 static struct file *io_file_get(struct io_submit_state *state,
1057 struct io_kiocb *req, int fd, bool fixed);
1058 static void __io_queue_sqe(struct io_kiocb *req);
1059 static void io_rsrc_put_work(struct work_struct *work);
1061 static void io_req_task_queue(struct io_kiocb *req);
1062 static void io_submit_flush_completions(struct io_comp_state *cs,
1063 struct io_ring_ctx *ctx);
1064 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1065 static int io_req_prep_async(struct io_kiocb *req);
1067 static struct kmem_cache *req_cachep;
1069 static const struct file_operations io_uring_fops;
1071 struct sock *io_uring_get_socket(struct file *file)
1073 #if defined(CONFIG_UNIX)
1074 if (file->f_op == &io_uring_fops) {
1075 struct io_ring_ctx *ctx = file->private_data;
1077 return ctx->ring_sock->sk;
1082 EXPORT_SYMBOL(io_uring_get_socket);
1084 #define io_for_each_link(pos, head) \
1085 for (pos = (head); pos; pos = pos->link)
1087 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1089 struct io_ring_ctx *ctx = req->ctx;
1091 if (!req->fixed_rsrc_refs) {
1092 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1093 percpu_ref_get(req->fixed_rsrc_refs);
1097 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1099 bool got = percpu_ref_tryget(ref);
1101 /* already at zero, wait for ->release() */
1103 wait_for_completion(compl);
1104 percpu_ref_resurrect(ref);
1106 percpu_ref_put(ref);
1109 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1112 struct io_kiocb *req;
1114 if (task && head->task != task)
1119 io_for_each_link(req, head) {
1120 if (req->flags & REQ_F_INFLIGHT)
1126 static inline void req_set_fail(struct io_kiocb *req)
1128 req->flags |= REQ_F_FAIL;
1131 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1133 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1135 complete(&ctx->ref_comp);
1138 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1140 return !req->timeout.off;
1143 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1145 struct io_ring_ctx *ctx;
1148 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1153 * Use 5 bits less than the max cq entries, that should give us around
1154 * 32 entries per hash list if totally full and uniformly spread.
1156 hash_bits = ilog2(p->cq_entries);
1160 ctx->cancel_hash_bits = hash_bits;
1161 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1163 if (!ctx->cancel_hash)
1165 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1167 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1168 if (!ctx->dummy_ubuf)
1170 /* set invalid range, so io_import_fixed() fails meeting it */
1171 ctx->dummy_ubuf->ubuf = -1UL;
1173 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1174 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1177 ctx->flags = p->flags;
1178 init_waitqueue_head(&ctx->sqo_sq_wait);
1179 INIT_LIST_HEAD(&ctx->sqd_list);
1180 init_waitqueue_head(&ctx->cq_wait);
1181 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1182 init_completion(&ctx->ref_comp);
1183 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1184 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1185 mutex_init(&ctx->uring_lock);
1186 init_waitqueue_head(&ctx->wait);
1187 spin_lock_init(&ctx->completion_lock);
1188 INIT_LIST_HEAD(&ctx->iopoll_list);
1189 INIT_LIST_HEAD(&ctx->defer_list);
1190 INIT_LIST_HEAD(&ctx->timeout_list);
1191 spin_lock_init(&ctx->rsrc_ref_lock);
1192 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1193 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1194 init_llist_head(&ctx->rsrc_put_llist);
1195 INIT_LIST_HEAD(&ctx->tctx_list);
1196 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1197 INIT_LIST_HEAD(&ctx->locked_free_list);
1200 kfree(ctx->dummy_ubuf);
1201 kfree(ctx->cancel_hash);
1206 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1208 struct io_rings *r = ctx->rings;
1210 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1214 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1216 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1217 struct io_ring_ctx *ctx = req->ctx;
1219 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1225 static void io_req_track_inflight(struct io_kiocb *req)
1227 if (!(req->flags & REQ_F_INFLIGHT)) {
1228 req->flags |= REQ_F_INFLIGHT;
1229 atomic_inc(¤t->io_uring->inflight_tracked);
1233 static void io_prep_async_work(struct io_kiocb *req)
1235 const struct io_op_def *def = &io_op_defs[req->opcode];
1236 struct io_ring_ctx *ctx = req->ctx;
1238 if (!req->work.creds)
1239 req->work.creds = get_current_cred();
1241 req->work.list.next = NULL;
1242 req->work.flags = 0;
1243 if (req->flags & REQ_F_FORCE_ASYNC)
1244 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1246 if (req->flags & REQ_F_ISREG) {
1247 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1248 io_wq_hash_work(&req->work, file_inode(req->file));
1249 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1250 if (def->unbound_nonreg_file)
1251 req->work.flags |= IO_WQ_WORK_UNBOUND;
1254 switch (req->opcode) {
1255 case IORING_OP_SPLICE:
1257 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1258 req->work.flags |= IO_WQ_WORK_UNBOUND;
1263 static void io_prep_async_link(struct io_kiocb *req)
1265 struct io_kiocb *cur;
1267 io_for_each_link(cur, req)
1268 io_prep_async_work(cur);
1271 static void io_queue_async_work(struct io_kiocb *req)
1273 struct io_ring_ctx *ctx = req->ctx;
1274 struct io_kiocb *link = io_prep_linked_timeout(req);
1275 struct io_uring_task *tctx = req->task->io_uring;
1278 BUG_ON(!tctx->io_wq);
1280 /* init ->work of the whole link before punting */
1281 io_prep_async_link(req);
1282 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1283 &req->work, req->flags);
1284 io_wq_enqueue(tctx->io_wq, &req->work);
1286 io_queue_linked_timeout(link);
1289 static void io_kill_timeout(struct io_kiocb *req, int status)
1290 __must_hold(&req->ctx->completion_lock)
1292 struct io_timeout_data *io = req->async_data;
1294 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1295 atomic_set(&req->ctx->cq_timeouts,
1296 atomic_read(&req->ctx->cq_timeouts) + 1);
1297 list_del_init(&req->timeout.list);
1298 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1299 io_put_req_deferred(req, 1);
1303 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1306 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1307 struct io_defer_entry, list);
1309 if (req_need_defer(de->req, de->seq))
1311 list_del_init(&de->list);
1312 io_req_task_queue(de->req);
1314 } while (!list_empty(&ctx->defer_list));
1317 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1321 if (list_empty(&ctx->timeout_list))
1324 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1327 u32 events_needed, events_got;
1328 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1329 struct io_kiocb, timeout.list);
1331 if (io_is_timeout_noseq(req))
1335 * Since seq can easily wrap around over time, subtract
1336 * the last seq at which timeouts were flushed before comparing.
1337 * Assuming not more than 2^31-1 events have happened since,
1338 * these subtractions won't have wrapped, so we can check if
1339 * target is in [last_seq, current_seq] by comparing the two.
1341 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1342 events_got = seq - ctx->cq_last_tm_flush;
1343 if (events_got < events_needed)
1346 list_del_init(&req->timeout.list);
1347 io_kill_timeout(req, 0);
1348 } while (!list_empty(&ctx->timeout_list));
1350 ctx->cq_last_tm_flush = seq;
1353 static void io_commit_cqring(struct io_ring_ctx *ctx)
1355 io_flush_timeouts(ctx);
1357 /* order cqe stores with ring update */
1358 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1360 if (unlikely(!list_empty(&ctx->defer_list)))
1361 __io_queue_deferred(ctx);
1364 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1366 struct io_rings *r = ctx->rings;
1368 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1371 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1373 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1376 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1378 struct io_rings *rings = ctx->rings;
1379 unsigned tail, mask = ctx->cq_entries - 1;
1382 * writes to the cq entry need to come after reading head; the
1383 * control dependency is enough as we're using WRITE_ONCE to
1386 if (__io_cqring_events(ctx) == ctx->cq_entries)
1389 tail = ctx->cached_cq_tail++;
1390 return &rings->cqes[tail & mask];
1393 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1395 if (likely(!ctx->cq_ev_fd))
1397 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1399 return !ctx->eventfd_async || io_wq_current_is_worker();
1402 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1404 /* see waitqueue_active() comment */
1407 if (waitqueue_active(&ctx->wait))
1408 wake_up(&ctx->wait);
1409 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1410 wake_up(&ctx->sq_data->wait);
1411 if (io_should_trigger_evfd(ctx))
1412 eventfd_signal(ctx->cq_ev_fd, 1);
1413 if (waitqueue_active(&ctx->cq_wait)) {
1414 wake_up_interruptible(&ctx->cq_wait);
1415 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1419 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1421 /* see waitqueue_active() comment */
1424 if (ctx->flags & IORING_SETUP_SQPOLL) {
1425 if (waitqueue_active(&ctx->wait))
1426 wake_up(&ctx->wait);
1428 if (io_should_trigger_evfd(ctx))
1429 eventfd_signal(ctx->cq_ev_fd, 1);
1430 if (waitqueue_active(&ctx->cq_wait)) {
1431 wake_up_interruptible(&ctx->cq_wait);
1432 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1436 /* Returns true if there are no backlogged entries after the flush */
1437 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1439 unsigned long flags;
1440 bool all_flushed, posted;
1442 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1446 spin_lock_irqsave(&ctx->completion_lock, flags);
1447 while (!list_empty(&ctx->cq_overflow_list)) {
1448 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1449 struct io_overflow_cqe *ocqe;
1453 ocqe = list_first_entry(&ctx->cq_overflow_list,
1454 struct io_overflow_cqe, list);
1456 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1458 io_account_cq_overflow(ctx);
1461 list_del(&ocqe->list);
1465 all_flushed = list_empty(&ctx->cq_overflow_list);
1467 clear_bit(0, &ctx->sq_check_overflow);
1468 clear_bit(0, &ctx->cq_check_overflow);
1469 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1473 io_commit_cqring(ctx);
1474 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1476 io_cqring_ev_posted(ctx);
1480 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1484 if (test_bit(0, &ctx->cq_check_overflow)) {
1485 /* iopoll syncs against uring_lock, not completion_lock */
1486 if (ctx->flags & IORING_SETUP_IOPOLL)
1487 mutex_lock(&ctx->uring_lock);
1488 ret = __io_cqring_overflow_flush(ctx, force);
1489 if (ctx->flags & IORING_SETUP_IOPOLL)
1490 mutex_unlock(&ctx->uring_lock);
1497 * Shamelessly stolen from the mm implementation of page reference checking,
1498 * see commit f958d7b528b1 for details.
1500 #define req_ref_zero_or_close_to_overflow(req) \
1501 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1503 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1505 return atomic_inc_not_zero(&req->refs);
1508 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1510 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1511 return atomic_sub_and_test(refs, &req->refs);
1514 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1516 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1517 return atomic_dec_and_test(&req->refs);
1520 static inline void req_ref_put(struct io_kiocb *req)
1522 WARN_ON_ONCE(req_ref_put_and_test(req));
1525 static inline void req_ref_get(struct io_kiocb *req)
1527 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1528 atomic_inc(&req->refs);
1531 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1532 long res, unsigned int cflags)
1534 struct io_overflow_cqe *ocqe;
1536 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1539 * If we're in ring overflow flush mode, or in task cancel mode,
1540 * or cannot allocate an overflow entry, then we need to drop it
1543 io_account_cq_overflow(ctx);
1546 if (list_empty(&ctx->cq_overflow_list)) {
1547 set_bit(0, &ctx->sq_check_overflow);
1548 set_bit(0, &ctx->cq_check_overflow);
1549 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1551 ocqe->cqe.user_data = user_data;
1552 ocqe->cqe.res = res;
1553 ocqe->cqe.flags = cflags;
1554 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1558 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1559 long res, unsigned int cflags)
1561 struct io_uring_cqe *cqe;
1563 trace_io_uring_complete(ctx, user_data, res, cflags);
1566 * If we can't get a cq entry, userspace overflowed the
1567 * submission (by quite a lot). Increment the overflow count in
1570 cqe = io_get_cqe(ctx);
1572 WRITE_ONCE(cqe->user_data, user_data);
1573 WRITE_ONCE(cqe->res, res);
1574 WRITE_ONCE(cqe->flags, cflags);
1577 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1580 /* not as hot to bloat with inlining */
1581 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1582 long res, unsigned int cflags)
1584 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1587 static void io_req_complete_post(struct io_kiocb *req, long res,
1588 unsigned int cflags)
1590 struct io_ring_ctx *ctx = req->ctx;
1591 unsigned long flags;
1593 spin_lock_irqsave(&ctx->completion_lock, flags);
1594 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1596 * If we're the last reference to this request, add to our locked
1599 if (req_ref_put_and_test(req)) {
1600 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1601 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1602 io_disarm_next(req);
1604 io_req_task_queue(req->link);
1608 io_dismantle_req(req);
1609 io_put_task(req->task, 1);
1610 list_add(&req->compl.list, &ctx->locked_free_list);
1611 ctx->locked_free_nr++;
1613 if (!percpu_ref_tryget(&ctx->refs))
1616 io_commit_cqring(ctx);
1617 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1620 io_cqring_ev_posted(ctx);
1621 percpu_ref_put(&ctx->refs);
1625 static inline bool io_req_needs_clean(struct io_kiocb *req)
1627 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1628 REQ_F_POLLED | REQ_F_INFLIGHT);
1631 static void io_req_complete_state(struct io_kiocb *req, long res,
1632 unsigned int cflags)
1634 if (io_req_needs_clean(req))
1637 req->compl.cflags = cflags;
1638 req->flags |= REQ_F_COMPLETE_INLINE;
1641 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1642 long res, unsigned cflags)
1644 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1645 io_req_complete_state(req, res, cflags);
1647 io_req_complete_post(req, res, cflags);
1650 static inline void io_req_complete(struct io_kiocb *req, long res)
1652 __io_req_complete(req, 0, res, 0);
1655 static void io_req_complete_failed(struct io_kiocb *req, long res)
1659 io_req_complete_post(req, res, 0);
1662 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1663 struct io_comp_state *cs)
1665 spin_lock_irq(&ctx->completion_lock);
1666 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1667 ctx->locked_free_nr = 0;
1668 spin_unlock_irq(&ctx->completion_lock);
1671 /* Returns true IFF there are requests in the cache */
1672 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1674 struct io_submit_state *state = &ctx->submit_state;
1675 struct io_comp_state *cs = &state->comp;
1679 * If we have more than a batch's worth of requests in our IRQ side
1680 * locked cache, grab the lock and move them over to our submission
1683 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1684 io_flush_cached_locked_reqs(ctx, cs);
1686 nr = state->free_reqs;
1687 while (!list_empty(&cs->free_list)) {
1688 struct io_kiocb *req = list_first_entry(&cs->free_list,
1689 struct io_kiocb, compl.list);
1691 list_del(&req->compl.list);
1692 state->reqs[nr++] = req;
1693 if (nr == ARRAY_SIZE(state->reqs))
1697 state->free_reqs = nr;
1701 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1703 struct io_submit_state *state = &ctx->submit_state;
1705 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1707 if (!state->free_reqs) {
1708 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1711 if (io_flush_cached_reqs(ctx))
1714 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1718 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1719 * retry single alloc to be on the safe side.
1721 if (unlikely(ret <= 0)) {
1722 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1723 if (!state->reqs[0])
1727 state->free_reqs = ret;
1731 return state->reqs[state->free_reqs];
1734 static inline void io_put_file(struct file *file)
1740 static void io_dismantle_req(struct io_kiocb *req)
1742 unsigned int flags = req->flags;
1744 if (io_req_needs_clean(req))
1746 if (!(flags & REQ_F_FIXED_FILE))
1747 io_put_file(req->file);
1748 if (req->fixed_rsrc_refs)
1749 percpu_ref_put(req->fixed_rsrc_refs);
1750 if (req->async_data)
1751 kfree(req->async_data);
1752 if (req->work.creds) {
1753 put_cred(req->work.creds);
1754 req->work.creds = NULL;
1758 /* must to be called somewhat shortly after putting a request */
1759 static inline void io_put_task(struct task_struct *task, int nr)
1761 struct io_uring_task *tctx = task->io_uring;
1763 percpu_counter_sub(&tctx->inflight, nr);
1764 if (unlikely(atomic_read(&tctx->in_idle)))
1765 wake_up(&tctx->wait);
1766 put_task_struct_many(task, nr);
1769 static void __io_free_req(struct io_kiocb *req)
1771 struct io_ring_ctx *ctx = req->ctx;
1773 io_dismantle_req(req);
1774 io_put_task(req->task, 1);
1776 kmem_cache_free(req_cachep, req);
1777 percpu_ref_put(&ctx->refs);
1780 static inline void io_remove_next_linked(struct io_kiocb *req)
1782 struct io_kiocb *nxt = req->link;
1784 req->link = nxt->link;
1788 static bool io_kill_linked_timeout(struct io_kiocb *req)
1789 __must_hold(&req->ctx->completion_lock)
1791 struct io_kiocb *link = req->link;
1794 * Can happen if a linked timeout fired and link had been like
1795 * req -> link t-out -> link t-out [-> ...]
1797 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1798 struct io_timeout_data *io = link->async_data;
1800 io_remove_next_linked(req);
1801 link->timeout.head = NULL;
1802 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1803 io_cqring_fill_event(link->ctx, link->user_data,
1805 io_put_req_deferred(link, 1);
1812 static void io_fail_links(struct io_kiocb *req)
1813 __must_hold(&req->ctx->completion_lock)
1815 struct io_kiocb *nxt, *link = req->link;
1822 trace_io_uring_fail_link(req, link);
1823 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1824 io_put_req_deferred(link, 2);
1829 static bool io_disarm_next(struct io_kiocb *req)
1830 __must_hold(&req->ctx->completion_lock)
1832 bool posted = false;
1834 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1835 posted = io_kill_linked_timeout(req);
1836 if (unlikely((req->flags & REQ_F_FAIL) &&
1837 !(req->flags & REQ_F_HARDLINK))) {
1838 posted |= (req->link != NULL);
1844 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1846 struct io_kiocb *nxt;
1849 * If LINK is set, we have dependent requests in this chain. If we
1850 * didn't fail this request, queue the first one up, moving any other
1851 * dependencies to the next request. In case of failure, fail the rest
1854 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1855 struct io_ring_ctx *ctx = req->ctx;
1856 unsigned long flags;
1859 spin_lock_irqsave(&ctx->completion_lock, flags);
1860 posted = io_disarm_next(req);
1862 io_commit_cqring(req->ctx);
1863 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1865 io_cqring_ev_posted(ctx);
1872 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1874 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1876 return __io_req_find_next(req);
1879 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1883 if (ctx->submit_state.comp.nr) {
1884 mutex_lock(&ctx->uring_lock);
1885 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1886 mutex_unlock(&ctx->uring_lock);
1888 percpu_ref_put(&ctx->refs);
1891 static bool __tctx_task_work(struct io_uring_task *tctx)
1893 struct io_ring_ctx *ctx = NULL;
1894 struct io_wq_work_list list;
1895 struct io_wq_work_node *node;
1897 if (wq_list_empty(&tctx->task_list))
1900 spin_lock_irq(&tctx->task_lock);
1901 list = tctx->task_list;
1902 INIT_WQ_LIST(&tctx->task_list);
1903 spin_unlock_irq(&tctx->task_lock);
1907 struct io_wq_work_node *next = node->next;
1908 struct io_kiocb *req;
1910 req = container_of(node, struct io_kiocb, io_task_work.node);
1911 if (req->ctx != ctx) {
1912 ctx_flush_and_put(ctx);
1914 percpu_ref_get(&ctx->refs);
1917 req->task_work.func(&req->task_work);
1921 ctx_flush_and_put(ctx);
1922 return list.first != NULL;
1925 static void tctx_task_work(struct callback_head *cb)
1927 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1929 clear_bit(0, &tctx->task_state);
1931 while (__tctx_task_work(tctx))
1935 static int io_req_task_work_add(struct io_kiocb *req)
1937 struct task_struct *tsk = req->task;
1938 struct io_uring_task *tctx = tsk->io_uring;
1939 enum task_work_notify_mode notify;
1940 struct io_wq_work_node *node, *prev;
1941 unsigned long flags;
1944 if (unlikely(tsk->flags & PF_EXITING))
1947 WARN_ON_ONCE(!tctx);
1949 spin_lock_irqsave(&tctx->task_lock, flags);
1950 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1951 spin_unlock_irqrestore(&tctx->task_lock, flags);
1953 /* task_work already pending, we're done */
1954 if (test_bit(0, &tctx->task_state) ||
1955 test_and_set_bit(0, &tctx->task_state))
1959 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1960 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1961 * processing task_work. There's no reliable way to tell if TWA_RESUME
1964 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1966 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1967 wake_up_process(tsk);
1972 * Slow path - we failed, find and delete work. if the work is not
1973 * in the list, it got run and we're fine.
1975 spin_lock_irqsave(&tctx->task_lock, flags);
1976 wq_list_for_each(node, prev, &tctx->task_list) {
1977 if (&req->io_task_work.node == node) {
1978 wq_list_del(&tctx->task_list, node, prev);
1983 spin_unlock_irqrestore(&tctx->task_lock, flags);
1984 clear_bit(0, &tctx->task_state);
1988 static bool io_run_task_work_head(struct callback_head **work_head)
1990 struct callback_head *work, *next;
1991 bool executed = false;
1994 work = xchg(work_head, NULL);
2010 static void io_task_work_add_head(struct callback_head **work_head,
2011 struct callback_head *task_work)
2013 struct callback_head *head;
2016 head = READ_ONCE(*work_head);
2017 task_work->next = head;
2018 } while (cmpxchg(work_head, head, task_work) != head);
2021 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2022 task_work_func_t cb)
2024 init_task_work(&req->task_work, cb);
2025 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2028 static void io_req_task_cancel(struct callback_head *cb)
2030 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2031 struct io_ring_ctx *ctx = req->ctx;
2033 /* ctx is guaranteed to stay alive while we hold uring_lock */
2034 mutex_lock(&ctx->uring_lock);
2035 io_req_complete_failed(req, req->result);
2036 mutex_unlock(&ctx->uring_lock);
2039 static void __io_req_task_submit(struct io_kiocb *req)
2041 struct io_ring_ctx *ctx = req->ctx;
2043 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2044 mutex_lock(&ctx->uring_lock);
2045 if (!(current->flags & PF_EXITING) && !current->in_execve)
2046 __io_queue_sqe(req);
2048 io_req_complete_failed(req, -EFAULT);
2049 mutex_unlock(&ctx->uring_lock);
2052 static void io_req_task_submit(struct callback_head *cb)
2054 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2056 __io_req_task_submit(req);
2059 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2062 req->task_work.func = io_req_task_cancel;
2064 if (unlikely(io_req_task_work_add(req)))
2065 io_req_task_work_add_fallback(req, io_req_task_cancel);
2068 static void io_req_task_queue(struct io_kiocb *req)
2070 req->task_work.func = io_req_task_submit;
2072 if (unlikely(io_req_task_work_add(req)))
2073 io_req_task_queue_fail(req, -ECANCELED);
2076 static inline void io_queue_next(struct io_kiocb *req)
2078 struct io_kiocb *nxt = io_req_find_next(req);
2081 io_req_task_queue(nxt);
2084 static void io_free_req(struct io_kiocb *req)
2091 struct task_struct *task;
2096 static inline void io_init_req_batch(struct req_batch *rb)
2103 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2104 struct req_batch *rb)
2107 io_put_task(rb->task, rb->task_refs);
2109 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2112 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2113 struct io_submit_state *state)
2116 io_dismantle_req(req);
2118 if (req->task != rb->task) {
2120 io_put_task(rb->task, rb->task_refs);
2121 rb->task = req->task;
2127 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2128 state->reqs[state->free_reqs++] = req;
2130 list_add(&req->compl.list, &state->comp.free_list);
2133 static void io_submit_flush_completions(struct io_comp_state *cs,
2134 struct io_ring_ctx *ctx)
2137 struct io_kiocb *req;
2138 struct req_batch rb;
2140 io_init_req_batch(&rb);
2141 spin_lock_irq(&ctx->completion_lock);
2142 for (i = 0; i < nr; i++) {
2144 __io_cqring_fill_event(ctx, req->user_data, req->result,
2147 io_commit_cqring(ctx);
2148 spin_unlock_irq(&ctx->completion_lock);
2150 io_cqring_ev_posted(ctx);
2151 for (i = 0; i < nr; i++) {
2154 /* submission and completion refs */
2155 if (req_ref_sub_and_test(req, 2))
2156 io_req_free_batch(&rb, req, &ctx->submit_state);
2159 io_req_free_batch_finish(ctx, &rb);
2164 * Drop reference to request, return next in chain (if there is one) if this
2165 * was the last reference to this request.
2167 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2169 struct io_kiocb *nxt = NULL;
2171 if (req_ref_put_and_test(req)) {
2172 nxt = io_req_find_next(req);
2178 static inline void io_put_req(struct io_kiocb *req)
2180 if (req_ref_put_and_test(req))
2184 static void io_put_req_deferred_cb(struct callback_head *cb)
2186 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2191 static void io_free_req_deferred(struct io_kiocb *req)
2193 req->task_work.func = io_put_req_deferred_cb;
2194 if (unlikely(io_req_task_work_add(req)))
2195 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2198 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2200 if (req_ref_sub_and_test(req, refs))
2201 io_free_req_deferred(req);
2204 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2206 /* See comment at the top of this file */
2208 return __io_cqring_events(ctx);
2211 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2213 struct io_rings *rings = ctx->rings;
2215 /* make sure SQ entry isn't read before tail */
2216 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2219 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2221 unsigned int cflags;
2223 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2224 cflags |= IORING_CQE_F_BUFFER;
2225 req->flags &= ~REQ_F_BUFFER_SELECTED;
2230 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2232 struct io_buffer *kbuf;
2234 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2235 return io_put_kbuf(req, kbuf);
2238 static inline bool io_run_task_work(void)
2241 * Not safe to run on exiting task, and the task_work handling will
2242 * not add work to such a task.
2244 if (unlikely(current->flags & PF_EXITING))
2246 if (current->task_works) {
2247 __set_current_state(TASK_RUNNING);
2256 * Find and free completed poll iocbs
2258 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2259 struct list_head *done)
2261 struct req_batch rb;
2262 struct io_kiocb *req;
2264 /* order with ->result store in io_complete_rw_iopoll() */
2267 io_init_req_batch(&rb);
2268 while (!list_empty(done)) {
2271 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2272 list_del(&req->inflight_entry);
2274 if (READ_ONCE(req->result) == -EAGAIN &&
2275 !(req->flags & REQ_F_DONT_REISSUE)) {
2276 req->iopoll_completed = 0;
2278 io_queue_async_work(req);
2282 if (req->flags & REQ_F_BUFFER_SELECTED)
2283 cflags = io_put_rw_kbuf(req);
2285 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2288 if (req_ref_put_and_test(req))
2289 io_req_free_batch(&rb, req, &ctx->submit_state);
2292 io_commit_cqring(ctx);
2293 io_cqring_ev_posted_iopoll(ctx);
2294 io_req_free_batch_finish(ctx, &rb);
2297 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2300 struct io_kiocb *req, *tmp;
2306 * Only spin for completions if we don't have multiple devices hanging
2307 * off our complete list, and we're under the requested amount.
2309 spin = !ctx->poll_multi_file && *nr_events < min;
2312 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2313 struct kiocb *kiocb = &req->rw.kiocb;
2316 * Move completed and retryable entries to our local lists.
2317 * If we find a request that requires polling, break out
2318 * and complete those lists first, if we have entries there.
2320 if (READ_ONCE(req->iopoll_completed)) {
2321 list_move_tail(&req->inflight_entry, &done);
2324 if (!list_empty(&done))
2327 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2331 /* iopoll may have completed current req */
2332 if (READ_ONCE(req->iopoll_completed))
2333 list_move_tail(&req->inflight_entry, &done);
2340 if (!list_empty(&done))
2341 io_iopoll_complete(ctx, nr_events, &done);
2347 * We can't just wait for polled events to come to us, we have to actively
2348 * find and complete them.
2350 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2352 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2355 mutex_lock(&ctx->uring_lock);
2356 while (!list_empty(&ctx->iopoll_list)) {
2357 unsigned int nr_events = 0;
2359 io_do_iopoll(ctx, &nr_events, 0);
2361 /* let it sleep and repeat later if can't complete a request */
2365 * Ensure we allow local-to-the-cpu processing to take place,
2366 * in this case we need to ensure that we reap all events.
2367 * Also let task_work, etc. to progress by releasing the mutex
2369 if (need_resched()) {
2370 mutex_unlock(&ctx->uring_lock);
2372 mutex_lock(&ctx->uring_lock);
2375 mutex_unlock(&ctx->uring_lock);
2378 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2380 unsigned int nr_events = 0;
2384 * We disallow the app entering submit/complete with polling, but we
2385 * still need to lock the ring to prevent racing with polled issue
2386 * that got punted to a workqueue.
2388 mutex_lock(&ctx->uring_lock);
2390 * Don't enter poll loop if we already have events pending.
2391 * If we do, we can potentially be spinning for commands that
2392 * already triggered a CQE (eg in error).
2394 if (test_bit(0, &ctx->cq_check_overflow))
2395 __io_cqring_overflow_flush(ctx, false);
2396 if (io_cqring_events(ctx))
2400 * If a submit got punted to a workqueue, we can have the
2401 * application entering polling for a command before it gets
2402 * issued. That app will hold the uring_lock for the duration
2403 * of the poll right here, so we need to take a breather every
2404 * now and then to ensure that the issue has a chance to add
2405 * the poll to the issued list. Otherwise we can spin here
2406 * forever, while the workqueue is stuck trying to acquire the
2409 if (list_empty(&ctx->iopoll_list)) {
2410 mutex_unlock(&ctx->uring_lock);
2412 mutex_lock(&ctx->uring_lock);
2414 if (list_empty(&ctx->iopoll_list))
2417 ret = io_do_iopoll(ctx, &nr_events, min);
2418 } while (!ret && nr_events < min && !need_resched());
2420 mutex_unlock(&ctx->uring_lock);
2424 static void kiocb_end_write(struct io_kiocb *req)
2427 * Tell lockdep we inherited freeze protection from submission
2430 if (req->flags & REQ_F_ISREG) {
2431 struct super_block *sb = file_inode(req->file)->i_sb;
2433 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2439 static bool io_resubmit_prep(struct io_kiocb *req)
2441 struct io_async_rw *rw = req->async_data;
2444 return !io_req_prep_async(req);
2445 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2446 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2450 static bool io_rw_should_reissue(struct io_kiocb *req)
2452 umode_t mode = file_inode(req->file)->i_mode;
2453 struct io_ring_ctx *ctx = req->ctx;
2455 if (!S_ISBLK(mode) && !S_ISREG(mode))
2457 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2458 !(ctx->flags & IORING_SETUP_IOPOLL)))
2461 * If ref is dying, we might be running poll reap from the exit work.
2462 * Don't attempt to reissue from that path, just let it fail with
2465 if (percpu_ref_is_dying(&ctx->refs))
2470 static bool io_resubmit_prep(struct io_kiocb *req)
2474 static bool io_rw_should_reissue(struct io_kiocb *req)
2480 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2481 unsigned int issue_flags)
2485 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2486 kiocb_end_write(req);
2487 if (res != req->result) {
2488 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2489 io_rw_should_reissue(req)) {
2490 req->flags |= REQ_F_REISSUE;
2495 if (req->flags & REQ_F_BUFFER_SELECTED)
2496 cflags = io_put_rw_kbuf(req);
2497 __io_req_complete(req, issue_flags, res, cflags);
2500 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2502 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2504 __io_complete_rw(req, res, res2, 0);
2507 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2509 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2511 if (kiocb->ki_flags & IOCB_WRITE)
2512 kiocb_end_write(req);
2513 if (unlikely(res != req->result)) {
2514 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2515 io_resubmit_prep(req))) {
2517 req->flags |= REQ_F_DONT_REISSUE;
2521 WRITE_ONCE(req->result, res);
2522 /* order with io_iopoll_complete() checking ->result */
2524 WRITE_ONCE(req->iopoll_completed, 1);
2528 * After the iocb has been issued, it's safe to be found on the poll list.
2529 * Adding the kiocb to the list AFTER submission ensures that we don't
2530 * find it from a io_do_iopoll() thread before the issuer is done
2531 * accessing the kiocb cookie.
2533 static void io_iopoll_req_issued(struct io_kiocb *req)
2535 struct io_ring_ctx *ctx = req->ctx;
2536 const bool in_async = io_wq_current_is_worker();
2538 /* workqueue context doesn't hold uring_lock, grab it now */
2539 if (unlikely(in_async))
2540 mutex_lock(&ctx->uring_lock);
2543 * Track whether we have multiple files in our lists. This will impact
2544 * how we do polling eventually, not spinning if we're on potentially
2545 * different devices.
2547 if (list_empty(&ctx->iopoll_list)) {
2548 ctx->poll_multi_file = false;
2549 } else if (!ctx->poll_multi_file) {
2550 struct io_kiocb *list_req;
2552 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2554 if (list_req->file != req->file)
2555 ctx->poll_multi_file = true;
2559 * For fast devices, IO may have already completed. If it has, add
2560 * it to the front so we find it first.
2562 if (READ_ONCE(req->iopoll_completed))
2563 list_add(&req->inflight_entry, &ctx->iopoll_list);
2565 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2567 if (unlikely(in_async)) {
2569 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2570 * in sq thread task context or in io worker task context. If
2571 * current task context is sq thread, we don't need to check
2572 * whether should wake up sq thread.
2574 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2575 wq_has_sleeper(&ctx->sq_data->wait))
2576 wake_up(&ctx->sq_data->wait);
2578 mutex_unlock(&ctx->uring_lock);
2582 static inline void io_state_file_put(struct io_submit_state *state)
2584 if (state->file_refs) {
2585 fput_many(state->file, state->file_refs);
2586 state->file_refs = 0;
2591 * Get as many references to a file as we have IOs left in this submission,
2592 * assuming most submissions are for one file, or at least that each file
2593 * has more than one submission.
2595 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2600 if (state->file_refs) {
2601 if (state->fd == fd) {
2605 io_state_file_put(state);
2607 state->file = fget_many(fd, state->ios_left);
2608 if (unlikely(!state->file))
2612 state->file_refs = state->ios_left - 1;
2616 static bool io_bdev_nowait(struct block_device *bdev)
2618 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2622 * If we tracked the file through the SCM inflight mechanism, we could support
2623 * any file. For now, just ensure that anything potentially problematic is done
2626 static bool __io_file_supports_async(struct file *file, int rw)
2628 umode_t mode = file_inode(file)->i_mode;
2630 if (S_ISBLK(mode)) {
2631 if (IS_ENABLED(CONFIG_BLOCK) &&
2632 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2638 if (S_ISREG(mode)) {
2639 if (IS_ENABLED(CONFIG_BLOCK) &&
2640 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2641 file->f_op != &io_uring_fops)
2646 /* any ->read/write should understand O_NONBLOCK */
2647 if (file->f_flags & O_NONBLOCK)
2650 if (!(file->f_mode & FMODE_NOWAIT))
2654 return file->f_op->read_iter != NULL;
2656 return file->f_op->write_iter != NULL;
2659 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2661 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2663 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2666 return __io_file_supports_async(req->file, rw);
2669 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2671 struct io_ring_ctx *ctx = req->ctx;
2672 struct kiocb *kiocb = &req->rw.kiocb;
2673 struct file *file = req->file;
2677 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2678 req->flags |= REQ_F_ISREG;
2680 kiocb->ki_pos = READ_ONCE(sqe->off);
2681 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2682 req->flags |= REQ_F_CUR_POS;
2683 kiocb->ki_pos = file->f_pos;
2685 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2686 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2687 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2691 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2692 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2693 req->flags |= REQ_F_NOWAIT;
2695 ioprio = READ_ONCE(sqe->ioprio);
2697 ret = ioprio_check_cap(ioprio);
2701 kiocb->ki_ioprio = ioprio;
2703 kiocb->ki_ioprio = get_current_ioprio();
2705 if (ctx->flags & IORING_SETUP_IOPOLL) {
2706 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2707 !kiocb->ki_filp->f_op->iopoll)
2710 kiocb->ki_flags |= IOCB_HIPRI;
2711 kiocb->ki_complete = io_complete_rw_iopoll;
2712 req->iopoll_completed = 0;
2714 if (kiocb->ki_flags & IOCB_HIPRI)
2716 kiocb->ki_complete = io_complete_rw;
2719 if (req->opcode == IORING_OP_READ_FIXED ||
2720 req->opcode == IORING_OP_WRITE_FIXED) {
2722 io_req_set_rsrc_node(req);
2725 req->rw.addr = READ_ONCE(sqe->addr);
2726 req->rw.len = READ_ONCE(sqe->len);
2727 req->buf_index = READ_ONCE(sqe->buf_index);
2731 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2737 case -ERESTARTNOINTR:
2738 case -ERESTARTNOHAND:
2739 case -ERESTART_RESTARTBLOCK:
2741 * We can't just restart the syscall, since previously
2742 * submitted sqes may already be in progress. Just fail this
2748 kiocb->ki_complete(kiocb, ret, 0);
2752 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2753 unsigned int issue_flags)
2755 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2756 struct io_async_rw *io = req->async_data;
2757 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2759 /* add previously done IO, if any */
2760 if (io && io->bytes_done > 0) {
2762 ret = io->bytes_done;
2764 ret += io->bytes_done;
2767 if (req->flags & REQ_F_CUR_POS)
2768 req->file->f_pos = kiocb->ki_pos;
2769 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2770 __io_complete_rw(req, ret, 0, issue_flags);
2772 io_rw_done(kiocb, ret);
2774 if (check_reissue && req->flags & REQ_F_REISSUE) {
2775 req->flags &= ~REQ_F_REISSUE;
2776 if (io_resubmit_prep(req)) {
2778 io_queue_async_work(req);
2783 if (req->flags & REQ_F_BUFFER_SELECTED)
2784 cflags = io_put_rw_kbuf(req);
2785 __io_req_complete(req, issue_flags, ret, cflags);
2790 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2791 struct io_mapped_ubuf *imu)
2793 size_t len = req->rw.len;
2794 u64 buf_end, buf_addr = req->rw.addr;
2797 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2799 /* not inside the mapped region */
2800 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2804 * May not be a start of buffer, set size appropriately
2805 * and advance us to the beginning.
2807 offset = buf_addr - imu->ubuf;
2808 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2812 * Don't use iov_iter_advance() here, as it's really slow for
2813 * using the latter parts of a big fixed buffer - it iterates
2814 * over each segment manually. We can cheat a bit here, because
2817 * 1) it's a BVEC iter, we set it up
2818 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2819 * first and last bvec
2821 * So just find our index, and adjust the iterator afterwards.
2822 * If the offset is within the first bvec (or the whole first
2823 * bvec, just use iov_iter_advance(). This makes it easier
2824 * since we can just skip the first segment, which may not
2825 * be PAGE_SIZE aligned.
2827 const struct bio_vec *bvec = imu->bvec;
2829 if (offset <= bvec->bv_len) {
2830 iov_iter_advance(iter, offset);
2832 unsigned long seg_skip;
2834 /* skip first vec */
2835 offset -= bvec->bv_len;
2836 seg_skip = 1 + (offset >> PAGE_SHIFT);
2838 iter->bvec = bvec + seg_skip;
2839 iter->nr_segs -= seg_skip;
2840 iter->count -= bvec->bv_len + offset;
2841 iter->iov_offset = offset & ~PAGE_MASK;
2848 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2850 struct io_ring_ctx *ctx = req->ctx;
2851 struct io_mapped_ubuf *imu = req->imu;
2852 u16 index, buf_index = req->buf_index;
2855 if (unlikely(buf_index >= ctx->nr_user_bufs))
2857 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2858 imu = READ_ONCE(ctx->user_bufs[index]);
2861 return __io_import_fixed(req, rw, iter, imu);
2864 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2867 mutex_unlock(&ctx->uring_lock);
2870 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2873 * "Normal" inline submissions always hold the uring_lock, since we
2874 * grab it from the system call. Same is true for the SQPOLL offload.
2875 * The only exception is when we've detached the request and issue it
2876 * from an async worker thread, grab the lock for that case.
2879 mutex_lock(&ctx->uring_lock);
2882 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2883 int bgid, struct io_buffer *kbuf,
2886 struct io_buffer *head;
2888 if (req->flags & REQ_F_BUFFER_SELECTED)
2891 io_ring_submit_lock(req->ctx, needs_lock);
2893 lockdep_assert_held(&req->ctx->uring_lock);
2895 head = xa_load(&req->ctx->io_buffers, bgid);
2897 if (!list_empty(&head->list)) {
2898 kbuf = list_last_entry(&head->list, struct io_buffer,
2900 list_del(&kbuf->list);
2903 xa_erase(&req->ctx->io_buffers, bgid);
2905 if (*len > kbuf->len)
2908 kbuf = ERR_PTR(-ENOBUFS);
2911 io_ring_submit_unlock(req->ctx, needs_lock);
2916 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2919 struct io_buffer *kbuf;
2922 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2923 bgid = req->buf_index;
2924 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2927 req->rw.addr = (u64) (unsigned long) kbuf;
2928 req->flags |= REQ_F_BUFFER_SELECTED;
2929 return u64_to_user_ptr(kbuf->addr);
2932 #ifdef CONFIG_COMPAT
2933 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2936 struct compat_iovec __user *uiov;
2937 compat_ssize_t clen;
2941 uiov = u64_to_user_ptr(req->rw.addr);
2942 if (!access_ok(uiov, sizeof(*uiov)))
2944 if (__get_user(clen, &uiov->iov_len))
2950 buf = io_rw_buffer_select(req, &len, needs_lock);
2952 return PTR_ERR(buf);
2953 iov[0].iov_base = buf;
2954 iov[0].iov_len = (compat_size_t) len;
2959 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2962 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2966 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2969 len = iov[0].iov_len;
2972 buf = io_rw_buffer_select(req, &len, needs_lock);
2974 return PTR_ERR(buf);
2975 iov[0].iov_base = buf;
2976 iov[0].iov_len = len;
2980 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2983 if (req->flags & REQ_F_BUFFER_SELECTED) {
2984 struct io_buffer *kbuf;
2986 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2987 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2988 iov[0].iov_len = kbuf->len;
2991 if (req->rw.len != 1)
2994 #ifdef CONFIG_COMPAT
2995 if (req->ctx->compat)
2996 return io_compat_import(req, iov, needs_lock);
2999 return __io_iov_buffer_select(req, iov, needs_lock);
3002 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3003 struct iov_iter *iter, bool needs_lock)
3005 void __user *buf = u64_to_user_ptr(req->rw.addr);
3006 size_t sqe_len = req->rw.len;
3007 u8 opcode = req->opcode;
3010 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3012 return io_import_fixed(req, rw, iter);
3015 /* buffer index only valid with fixed read/write, or buffer select */
3016 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3019 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3020 if (req->flags & REQ_F_BUFFER_SELECT) {
3021 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3023 return PTR_ERR(buf);
3024 req->rw.len = sqe_len;
3027 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3032 if (req->flags & REQ_F_BUFFER_SELECT) {
3033 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3035 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3040 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3044 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3046 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3050 * For files that don't have ->read_iter() and ->write_iter(), handle them
3051 * by looping over ->read() or ->write() manually.
3053 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3055 struct kiocb *kiocb = &req->rw.kiocb;
3056 struct file *file = req->file;
3060 * Don't support polled IO through this interface, and we can't
3061 * support non-blocking either. For the latter, this just causes
3062 * the kiocb to be handled from an async context.
3064 if (kiocb->ki_flags & IOCB_HIPRI)
3066 if (kiocb->ki_flags & IOCB_NOWAIT)
3069 while (iov_iter_count(iter)) {
3073 if (!iov_iter_is_bvec(iter)) {
3074 iovec = iov_iter_iovec(iter);
3076 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3077 iovec.iov_len = req->rw.len;
3081 nr = file->f_op->read(file, iovec.iov_base,
3082 iovec.iov_len, io_kiocb_ppos(kiocb));
3084 nr = file->f_op->write(file, iovec.iov_base,
3085 iovec.iov_len, io_kiocb_ppos(kiocb));
3094 if (nr != iovec.iov_len)
3098 iov_iter_advance(iter, nr);
3104 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3105 const struct iovec *fast_iov, struct iov_iter *iter)
3107 struct io_async_rw *rw = req->async_data;
3109 memcpy(&rw->iter, iter, sizeof(*iter));
3110 rw->free_iovec = iovec;
3112 /* can only be fixed buffers, no need to do anything */
3113 if (iov_iter_is_bvec(iter))
3116 unsigned iov_off = 0;
3118 rw->iter.iov = rw->fast_iov;
3119 if (iter->iov != fast_iov) {
3120 iov_off = iter->iov - fast_iov;
3121 rw->iter.iov += iov_off;
3123 if (rw->fast_iov != fast_iov)
3124 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3125 sizeof(struct iovec) * iter->nr_segs);
3127 req->flags |= REQ_F_NEED_CLEANUP;
3131 static inline int io_alloc_async_data(struct io_kiocb *req)
3133 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3134 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3135 return req->async_data == NULL;
3138 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3139 const struct iovec *fast_iov,
3140 struct iov_iter *iter, bool force)
3142 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3144 if (!req->async_data) {
3145 if (io_alloc_async_data(req)) {
3150 io_req_map_rw(req, iovec, fast_iov, iter);
3155 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3157 struct io_async_rw *iorw = req->async_data;
3158 struct iovec *iov = iorw->fast_iov;
3161 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3162 if (unlikely(ret < 0))
3165 iorw->bytes_done = 0;
3166 iorw->free_iovec = iov;
3168 req->flags |= REQ_F_NEED_CLEANUP;
3172 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3174 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3176 return io_prep_rw(req, sqe);
3180 * This is our waitqueue callback handler, registered through lock_page_async()
3181 * when we initially tried to do the IO with the iocb armed our waitqueue.
3182 * This gets called when the page is unlocked, and we generally expect that to
3183 * happen when the page IO is completed and the page is now uptodate. This will
3184 * queue a task_work based retry of the operation, attempting to copy the data
3185 * again. If the latter fails because the page was NOT uptodate, then we will
3186 * do a thread based blocking retry of the operation. That's the unexpected
3189 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3190 int sync, void *arg)
3192 struct wait_page_queue *wpq;
3193 struct io_kiocb *req = wait->private;
3194 struct wait_page_key *key = arg;
3196 wpq = container_of(wait, struct wait_page_queue, wait);
3198 if (!wake_page_match(wpq, key))
3201 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3202 list_del_init(&wait->entry);
3204 /* submit ref gets dropped, acquire a new one */
3206 io_req_task_queue(req);
3211 * This controls whether a given IO request should be armed for async page
3212 * based retry. If we return false here, the request is handed to the async
3213 * worker threads for retry. If we're doing buffered reads on a regular file,
3214 * we prepare a private wait_page_queue entry and retry the operation. This
3215 * will either succeed because the page is now uptodate and unlocked, or it
3216 * will register a callback when the page is unlocked at IO completion. Through
3217 * that callback, io_uring uses task_work to setup a retry of the operation.
3218 * That retry will attempt the buffered read again. The retry will generally
3219 * succeed, or in rare cases where it fails, we then fall back to using the
3220 * async worker threads for a blocking retry.
3222 static bool io_rw_should_retry(struct io_kiocb *req)
3224 struct io_async_rw *rw = req->async_data;
3225 struct wait_page_queue *wait = &rw->wpq;
3226 struct kiocb *kiocb = &req->rw.kiocb;
3228 /* never retry for NOWAIT, we just complete with -EAGAIN */
3229 if (req->flags & REQ_F_NOWAIT)
3232 /* Only for buffered IO */
3233 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3237 * just use poll if we can, and don't attempt if the fs doesn't
3238 * support callback based unlocks
3240 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3243 wait->wait.func = io_async_buf_func;
3244 wait->wait.private = req;
3245 wait->wait.flags = 0;
3246 INIT_LIST_HEAD(&wait->wait.entry);
3247 kiocb->ki_flags |= IOCB_WAITQ;
3248 kiocb->ki_flags &= ~IOCB_NOWAIT;
3249 kiocb->ki_waitq = wait;
3253 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3255 if (req->file->f_op->read_iter)
3256 return call_read_iter(req->file, &req->rw.kiocb, iter);
3257 else if (req->file->f_op->read)
3258 return loop_rw_iter(READ, req, iter);
3263 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3265 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3266 struct kiocb *kiocb = &req->rw.kiocb;
3267 struct iov_iter __iter, *iter = &__iter;
3268 struct io_async_rw *rw = req->async_data;
3269 ssize_t io_size, ret, ret2;
3270 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3276 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3280 io_size = iov_iter_count(iter);
3281 req->result = io_size;
3283 /* Ensure we clear previously set non-block flag */
3284 if (!force_nonblock)
3285 kiocb->ki_flags &= ~IOCB_NOWAIT;
3287 kiocb->ki_flags |= IOCB_NOWAIT;
3289 /* If the file doesn't support async, just async punt */
3290 if (force_nonblock && !io_file_supports_async(req, READ)) {
3291 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3292 return ret ?: -EAGAIN;
3295 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3296 if (unlikely(ret)) {
3301 ret = io_iter_do_read(req, iter);
3303 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3304 req->flags &= ~REQ_F_REISSUE;
3305 /* IOPOLL retry should happen for io-wq threads */
3306 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3308 /* no retry on NONBLOCK nor RWF_NOWAIT */
3309 if (req->flags & REQ_F_NOWAIT)
3311 /* some cases will consume bytes even on error returns */
3312 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3314 } else if (ret == -EIOCBQUEUED) {
3316 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3317 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3318 /* read all, failed, already did sync or don't want to retry */
3322 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3327 rw = req->async_data;
3328 /* now use our persistent iterator, if we aren't already */
3333 rw->bytes_done += ret;
3334 /* if we can retry, do so with the callbacks armed */
3335 if (!io_rw_should_retry(req)) {
3336 kiocb->ki_flags &= ~IOCB_WAITQ;
3341 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3342 * we get -EIOCBQUEUED, then we'll get a notification when the
3343 * desired page gets unlocked. We can also get a partial read
3344 * here, and if we do, then just retry at the new offset.
3346 ret = io_iter_do_read(req, iter);
3347 if (ret == -EIOCBQUEUED)
3349 /* we got some bytes, but not all. retry. */
3350 kiocb->ki_flags &= ~IOCB_WAITQ;
3351 } while (ret > 0 && ret < io_size);
3353 kiocb_done(kiocb, ret, issue_flags);
3355 /* it's faster to check here then delegate to kfree */
3361 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3363 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3365 return io_prep_rw(req, sqe);
3368 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3370 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3371 struct kiocb *kiocb = &req->rw.kiocb;
3372 struct iov_iter __iter, *iter = &__iter;
3373 struct io_async_rw *rw = req->async_data;
3374 ssize_t ret, ret2, io_size;
3375 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3381 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3385 io_size = iov_iter_count(iter);
3386 req->result = io_size;
3388 /* Ensure we clear previously set non-block flag */
3389 if (!force_nonblock)
3390 kiocb->ki_flags &= ~IOCB_NOWAIT;
3392 kiocb->ki_flags |= IOCB_NOWAIT;
3394 /* If the file doesn't support async, just async punt */
3395 if (force_nonblock && !io_file_supports_async(req, WRITE))
3398 /* file path doesn't support NOWAIT for non-direct_IO */
3399 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3400 (req->flags & REQ_F_ISREG))
3403 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3408 * Open-code file_start_write here to grab freeze protection,
3409 * which will be released by another thread in
3410 * io_complete_rw(). Fool lockdep by telling it the lock got
3411 * released so that it doesn't complain about the held lock when
3412 * we return to userspace.
3414 if (req->flags & REQ_F_ISREG) {
3415 sb_start_write(file_inode(req->file)->i_sb);
3416 __sb_writers_release(file_inode(req->file)->i_sb,
3419 kiocb->ki_flags |= IOCB_WRITE;
3421 if (req->file->f_op->write_iter)
3422 ret2 = call_write_iter(req->file, kiocb, iter);
3423 else if (req->file->f_op->write)
3424 ret2 = loop_rw_iter(WRITE, req, iter);
3428 if (req->flags & REQ_F_REISSUE) {
3429 req->flags &= ~REQ_F_REISSUE;
3434 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3435 * retry them without IOCB_NOWAIT.
3437 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3439 /* no retry on NONBLOCK nor RWF_NOWAIT */
3440 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3442 if (!force_nonblock || ret2 != -EAGAIN) {
3443 /* IOPOLL retry should happen for io-wq threads */
3444 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3447 kiocb_done(kiocb, ret2, issue_flags);
3450 /* some cases will consume bytes even on error returns */
3451 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3452 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3453 return ret ?: -EAGAIN;
3456 /* it's reportedly faster than delegating the null check to kfree() */
3462 static int io_renameat_prep(struct io_kiocb *req,
3463 const struct io_uring_sqe *sqe)
3465 struct io_rename *ren = &req->rename;
3466 const char __user *oldf, *newf;
3468 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3471 ren->old_dfd = READ_ONCE(sqe->fd);
3472 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3473 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3474 ren->new_dfd = READ_ONCE(sqe->len);
3475 ren->flags = READ_ONCE(sqe->rename_flags);
3477 ren->oldpath = getname(oldf);
3478 if (IS_ERR(ren->oldpath))
3479 return PTR_ERR(ren->oldpath);
3481 ren->newpath = getname(newf);
3482 if (IS_ERR(ren->newpath)) {
3483 putname(ren->oldpath);
3484 return PTR_ERR(ren->newpath);
3487 req->flags |= REQ_F_NEED_CLEANUP;
3491 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3493 struct io_rename *ren = &req->rename;
3496 if (issue_flags & IO_URING_F_NONBLOCK)
3499 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3500 ren->newpath, ren->flags);
3502 req->flags &= ~REQ_F_NEED_CLEANUP;
3505 io_req_complete(req, ret);
3509 static int io_unlinkat_prep(struct io_kiocb *req,
3510 const struct io_uring_sqe *sqe)
3512 struct io_unlink *un = &req->unlink;
3513 const char __user *fname;
3515 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3518 un->dfd = READ_ONCE(sqe->fd);
3520 un->flags = READ_ONCE(sqe->unlink_flags);
3521 if (un->flags & ~AT_REMOVEDIR)
3524 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3525 un->filename = getname(fname);
3526 if (IS_ERR(un->filename))
3527 return PTR_ERR(un->filename);
3529 req->flags |= REQ_F_NEED_CLEANUP;
3533 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3535 struct io_unlink *un = &req->unlink;
3538 if (issue_flags & IO_URING_F_NONBLOCK)
3541 if (un->flags & AT_REMOVEDIR)
3542 ret = do_rmdir(un->dfd, un->filename);
3544 ret = do_unlinkat(un->dfd, un->filename);
3546 req->flags &= ~REQ_F_NEED_CLEANUP;
3549 io_req_complete(req, ret);
3553 static int io_shutdown_prep(struct io_kiocb *req,
3554 const struct io_uring_sqe *sqe)
3556 #if defined(CONFIG_NET)
3557 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3559 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3563 req->shutdown.how = READ_ONCE(sqe->len);
3570 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3572 #if defined(CONFIG_NET)
3573 struct socket *sock;
3576 if (issue_flags & IO_URING_F_NONBLOCK)
3579 sock = sock_from_file(req->file);
3580 if (unlikely(!sock))
3583 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3586 io_req_complete(req, ret);
3593 static int __io_splice_prep(struct io_kiocb *req,
3594 const struct io_uring_sqe *sqe)
3596 struct io_splice* sp = &req->splice;
3597 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3599 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3603 sp->len = READ_ONCE(sqe->len);
3604 sp->flags = READ_ONCE(sqe->splice_flags);
3606 if (unlikely(sp->flags & ~valid_flags))
3609 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3610 (sp->flags & SPLICE_F_FD_IN_FIXED));
3613 req->flags |= REQ_F_NEED_CLEANUP;
3617 static int io_tee_prep(struct io_kiocb *req,
3618 const struct io_uring_sqe *sqe)
3620 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3622 return __io_splice_prep(req, sqe);
3625 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3627 struct io_splice *sp = &req->splice;
3628 struct file *in = sp->file_in;
3629 struct file *out = sp->file_out;
3630 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3633 if (issue_flags & IO_URING_F_NONBLOCK)
3636 ret = do_tee(in, out, sp->len, flags);
3638 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3640 req->flags &= ~REQ_F_NEED_CLEANUP;
3644 io_req_complete(req, ret);
3648 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3650 struct io_splice* sp = &req->splice;
3652 sp->off_in = READ_ONCE(sqe->splice_off_in);
3653 sp->off_out = READ_ONCE(sqe->off);
3654 return __io_splice_prep(req, sqe);
3657 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3659 struct io_splice *sp = &req->splice;
3660 struct file *in = sp->file_in;
3661 struct file *out = sp->file_out;
3662 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3663 loff_t *poff_in, *poff_out;
3666 if (issue_flags & IO_URING_F_NONBLOCK)
3669 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3670 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3673 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3675 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3677 req->flags &= ~REQ_F_NEED_CLEANUP;
3681 io_req_complete(req, ret);
3686 * IORING_OP_NOP just posts a completion event, nothing else.
3688 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3690 struct io_ring_ctx *ctx = req->ctx;
3692 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3695 __io_req_complete(req, issue_flags, 0, 0);
3699 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3701 struct io_ring_ctx *ctx = req->ctx;
3706 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3708 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3711 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3712 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3715 req->sync.off = READ_ONCE(sqe->off);
3716 req->sync.len = READ_ONCE(sqe->len);
3720 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3722 loff_t end = req->sync.off + req->sync.len;
3725 /* fsync always requires a blocking context */
3726 if (issue_flags & IO_URING_F_NONBLOCK)
3729 ret = vfs_fsync_range(req->file, req->sync.off,
3730 end > 0 ? end : LLONG_MAX,
3731 req->sync.flags & IORING_FSYNC_DATASYNC);
3734 io_req_complete(req, ret);
3738 static int io_fallocate_prep(struct io_kiocb *req,
3739 const struct io_uring_sqe *sqe)
3741 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3743 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3746 req->sync.off = READ_ONCE(sqe->off);
3747 req->sync.len = READ_ONCE(sqe->addr);
3748 req->sync.mode = READ_ONCE(sqe->len);
3752 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3756 /* fallocate always requiring blocking context */
3757 if (issue_flags & IO_URING_F_NONBLOCK)
3759 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3763 io_req_complete(req, ret);
3767 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3769 const char __user *fname;
3772 if (unlikely(sqe->ioprio || sqe->buf_index))
3774 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3777 /* open.how should be already initialised */
3778 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3779 req->open.how.flags |= O_LARGEFILE;
3781 req->open.dfd = READ_ONCE(sqe->fd);
3782 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3783 req->open.filename = getname(fname);
3784 if (IS_ERR(req->open.filename)) {
3785 ret = PTR_ERR(req->open.filename);
3786 req->open.filename = NULL;
3789 req->open.nofile = rlimit(RLIMIT_NOFILE);
3790 req->flags |= REQ_F_NEED_CLEANUP;
3794 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3798 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3800 mode = READ_ONCE(sqe->len);
3801 flags = READ_ONCE(sqe->open_flags);
3802 req->open.how = build_open_how(flags, mode);
3803 return __io_openat_prep(req, sqe);
3806 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3808 struct open_how __user *how;
3812 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3814 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3815 len = READ_ONCE(sqe->len);
3816 if (len < OPEN_HOW_SIZE_VER0)
3819 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3824 return __io_openat_prep(req, sqe);
3827 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3829 struct open_flags op;
3832 bool resolve_nonblock;
3835 ret = build_open_flags(&req->open.how, &op);
3838 nonblock_set = op.open_flag & O_NONBLOCK;
3839 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3840 if (issue_flags & IO_URING_F_NONBLOCK) {
3842 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3843 * it'll always -EAGAIN
3845 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3847 op.lookup_flags |= LOOKUP_CACHED;
3848 op.open_flag |= O_NONBLOCK;
3851 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3855 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3856 /* only retry if RESOLVE_CACHED wasn't already set by application */
3857 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3858 file == ERR_PTR(-EAGAIN)) {
3860 * We could hang on to this 'fd', but seems like marginal
3861 * gain for something that is now known to be a slower path.
3862 * So just put it, and we'll get a new one when we retry.
3870 ret = PTR_ERR(file);
3872 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3873 file->f_flags &= ~O_NONBLOCK;
3874 fsnotify_open(file);
3875 fd_install(ret, file);
3878 putname(req->open.filename);
3879 req->flags &= ~REQ_F_NEED_CLEANUP;
3882 __io_req_complete(req, issue_flags, ret, 0);
3886 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3888 return io_openat2(req, issue_flags);
3891 static int io_remove_buffers_prep(struct io_kiocb *req,
3892 const struct io_uring_sqe *sqe)
3894 struct io_provide_buf *p = &req->pbuf;
3897 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3900 tmp = READ_ONCE(sqe->fd);
3901 if (!tmp || tmp > USHRT_MAX)
3904 memset(p, 0, sizeof(*p));
3906 p->bgid = READ_ONCE(sqe->buf_group);
3910 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3911 int bgid, unsigned nbufs)
3915 /* shouldn't happen */
3919 /* the head kbuf is the list itself */
3920 while (!list_empty(&buf->list)) {
3921 struct io_buffer *nxt;
3923 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3924 list_del(&nxt->list);
3931 xa_erase(&ctx->io_buffers, bgid);
3936 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3938 struct io_provide_buf *p = &req->pbuf;
3939 struct io_ring_ctx *ctx = req->ctx;
3940 struct io_buffer *head;
3942 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3944 io_ring_submit_lock(ctx, !force_nonblock);
3946 lockdep_assert_held(&ctx->uring_lock);
3949 head = xa_load(&ctx->io_buffers, p->bgid);
3951 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3955 /* complete before unlock, IOPOLL may need the lock */
3956 __io_req_complete(req, issue_flags, ret, 0);
3957 io_ring_submit_unlock(ctx, !force_nonblock);
3961 static int io_provide_buffers_prep(struct io_kiocb *req,
3962 const struct io_uring_sqe *sqe)
3964 unsigned long size, tmp_check;
3965 struct io_provide_buf *p = &req->pbuf;
3968 if (sqe->ioprio || sqe->rw_flags)
3971 tmp = READ_ONCE(sqe->fd);
3972 if (!tmp || tmp > USHRT_MAX)
3975 p->addr = READ_ONCE(sqe->addr);
3976 p->len = READ_ONCE(sqe->len);
3978 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3981 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3984 size = (unsigned long)p->len * p->nbufs;
3985 if (!access_ok(u64_to_user_ptr(p->addr), size))
3988 p->bgid = READ_ONCE(sqe->buf_group);
3989 tmp = READ_ONCE(sqe->off);
3990 if (tmp > USHRT_MAX)
3996 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3998 struct io_buffer *buf;
3999 u64 addr = pbuf->addr;
4000 int i, bid = pbuf->bid;
4002 for (i = 0; i < pbuf->nbufs; i++) {
4003 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4008 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4013 INIT_LIST_HEAD(&buf->list);
4016 list_add_tail(&buf->list, &(*head)->list);
4020 return i ? i : -ENOMEM;
4023 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4025 struct io_provide_buf *p = &req->pbuf;
4026 struct io_ring_ctx *ctx = req->ctx;
4027 struct io_buffer *head, *list;
4029 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4031 io_ring_submit_lock(ctx, !force_nonblock);
4033 lockdep_assert_held(&ctx->uring_lock);
4035 list = head = xa_load(&ctx->io_buffers, p->bgid);
4037 ret = io_add_buffers(p, &head);
4038 if (ret >= 0 && !list) {
4039 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4041 __io_remove_buffers(ctx, head, p->bgid, -1U);
4045 /* complete before unlock, IOPOLL may need the lock */
4046 __io_req_complete(req, issue_flags, ret, 0);
4047 io_ring_submit_unlock(ctx, !force_nonblock);
4051 static int io_epoll_ctl_prep(struct io_kiocb *req,
4052 const struct io_uring_sqe *sqe)
4054 #if defined(CONFIG_EPOLL)
4055 if (sqe->ioprio || sqe->buf_index)
4057 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4060 req->epoll.epfd = READ_ONCE(sqe->fd);
4061 req->epoll.op = READ_ONCE(sqe->len);
4062 req->epoll.fd = READ_ONCE(sqe->off);
4064 if (ep_op_has_event(req->epoll.op)) {
4065 struct epoll_event __user *ev;
4067 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4068 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4078 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4080 #if defined(CONFIG_EPOLL)
4081 struct io_epoll *ie = &req->epoll;
4083 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4085 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4086 if (force_nonblock && ret == -EAGAIN)
4091 __io_req_complete(req, issue_flags, ret, 0);
4098 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4100 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4101 if (sqe->ioprio || sqe->buf_index || sqe->off)
4103 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4106 req->madvise.addr = READ_ONCE(sqe->addr);
4107 req->madvise.len = READ_ONCE(sqe->len);
4108 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4115 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4117 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4118 struct io_madvise *ma = &req->madvise;
4121 if (issue_flags & IO_URING_F_NONBLOCK)
4124 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4127 io_req_complete(req, ret);
4134 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4136 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4138 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4141 req->fadvise.offset = READ_ONCE(sqe->off);
4142 req->fadvise.len = READ_ONCE(sqe->len);
4143 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4147 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4149 struct io_fadvise *fa = &req->fadvise;
4152 if (issue_flags & IO_URING_F_NONBLOCK) {
4153 switch (fa->advice) {
4154 case POSIX_FADV_NORMAL:
4155 case POSIX_FADV_RANDOM:
4156 case POSIX_FADV_SEQUENTIAL:
4163 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4166 __io_req_complete(req, issue_flags, ret, 0);
4170 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4172 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4174 if (sqe->ioprio || sqe->buf_index)
4176 if (req->flags & REQ_F_FIXED_FILE)
4179 req->statx.dfd = READ_ONCE(sqe->fd);
4180 req->statx.mask = READ_ONCE(sqe->len);
4181 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4182 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4183 req->statx.flags = READ_ONCE(sqe->statx_flags);
4188 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4190 struct io_statx *ctx = &req->statx;
4193 if (issue_flags & IO_URING_F_NONBLOCK)
4196 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4201 io_req_complete(req, ret);
4205 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4207 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4209 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4210 sqe->rw_flags || sqe->buf_index)
4212 if (req->flags & REQ_F_FIXED_FILE)
4215 req->close.fd = READ_ONCE(sqe->fd);
4219 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4221 struct files_struct *files = current->files;
4222 struct io_close *close = &req->close;
4223 struct fdtable *fdt;
4224 struct file *file = NULL;
4227 spin_lock(&files->file_lock);
4228 fdt = files_fdtable(files);
4229 if (close->fd >= fdt->max_fds) {
4230 spin_unlock(&files->file_lock);
4233 file = fdt->fd[close->fd];
4234 if (!file || file->f_op == &io_uring_fops) {
4235 spin_unlock(&files->file_lock);
4240 /* if the file has a flush method, be safe and punt to async */
4241 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4242 spin_unlock(&files->file_lock);
4246 ret = __close_fd_get_file(close->fd, &file);
4247 spin_unlock(&files->file_lock);
4254 /* No ->flush() or already async, safely close from here */
4255 ret = filp_close(file, current->files);
4261 __io_req_complete(req, issue_flags, ret, 0);
4265 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4267 struct io_ring_ctx *ctx = req->ctx;
4269 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4271 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4274 req->sync.off = READ_ONCE(sqe->off);
4275 req->sync.len = READ_ONCE(sqe->len);
4276 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4280 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4284 /* sync_file_range always requires a blocking context */
4285 if (issue_flags & IO_URING_F_NONBLOCK)
4288 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4292 io_req_complete(req, ret);
4296 #if defined(CONFIG_NET)
4297 static int io_setup_async_msg(struct io_kiocb *req,
4298 struct io_async_msghdr *kmsg)
4300 struct io_async_msghdr *async_msg = req->async_data;
4304 if (io_alloc_async_data(req)) {
4305 kfree(kmsg->free_iov);
4308 async_msg = req->async_data;
4309 req->flags |= REQ_F_NEED_CLEANUP;
4310 memcpy(async_msg, kmsg, sizeof(*kmsg));
4311 async_msg->msg.msg_name = &async_msg->addr;
4312 /* if were using fast_iov, set it to the new one */
4313 if (!async_msg->free_iov)
4314 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4319 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4320 struct io_async_msghdr *iomsg)
4322 iomsg->msg.msg_name = &iomsg->addr;
4323 iomsg->free_iov = iomsg->fast_iov;
4324 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4325 req->sr_msg.msg_flags, &iomsg->free_iov);
4328 static int io_sendmsg_prep_async(struct io_kiocb *req)
4332 ret = io_sendmsg_copy_hdr(req, req->async_data);
4334 req->flags |= REQ_F_NEED_CLEANUP;
4338 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4340 struct io_sr_msg *sr = &req->sr_msg;
4342 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4345 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4346 sr->len = READ_ONCE(sqe->len);
4347 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4348 if (sr->msg_flags & MSG_DONTWAIT)
4349 req->flags |= REQ_F_NOWAIT;
4351 #ifdef CONFIG_COMPAT
4352 if (req->ctx->compat)
4353 sr->msg_flags |= MSG_CMSG_COMPAT;
4358 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4360 struct io_async_msghdr iomsg, *kmsg;
4361 struct socket *sock;
4366 sock = sock_from_file(req->file);
4367 if (unlikely(!sock))
4370 kmsg = req->async_data;
4372 ret = io_sendmsg_copy_hdr(req, &iomsg);
4378 flags = req->sr_msg.msg_flags;
4379 if (issue_flags & IO_URING_F_NONBLOCK)
4380 flags |= MSG_DONTWAIT;
4381 if (flags & MSG_WAITALL)
4382 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4384 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4385 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4386 return io_setup_async_msg(req, kmsg);
4387 if (ret == -ERESTARTSYS)
4390 /* fast path, check for non-NULL to avoid function call */
4392 kfree(kmsg->free_iov);
4393 req->flags &= ~REQ_F_NEED_CLEANUP;
4396 __io_req_complete(req, issue_flags, ret, 0);
4400 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4402 struct io_sr_msg *sr = &req->sr_msg;
4405 struct socket *sock;
4410 sock = sock_from_file(req->file);
4411 if (unlikely(!sock))
4414 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4418 msg.msg_name = NULL;
4419 msg.msg_control = NULL;
4420 msg.msg_controllen = 0;
4421 msg.msg_namelen = 0;
4423 flags = req->sr_msg.msg_flags;
4424 if (issue_flags & IO_URING_F_NONBLOCK)
4425 flags |= MSG_DONTWAIT;
4426 if (flags & MSG_WAITALL)
4427 min_ret = iov_iter_count(&msg.msg_iter);
4429 msg.msg_flags = flags;
4430 ret = sock_sendmsg(sock, &msg);
4431 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4433 if (ret == -ERESTARTSYS)
4438 __io_req_complete(req, issue_flags, ret, 0);
4442 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4443 struct io_async_msghdr *iomsg)
4445 struct io_sr_msg *sr = &req->sr_msg;
4446 struct iovec __user *uiov;
4450 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4451 &iomsg->uaddr, &uiov, &iov_len);
4455 if (req->flags & REQ_F_BUFFER_SELECT) {
4458 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4460 sr->len = iomsg->fast_iov[0].iov_len;
4461 iomsg->free_iov = NULL;
4463 iomsg->free_iov = iomsg->fast_iov;
4464 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4465 &iomsg->free_iov, &iomsg->msg.msg_iter,
4474 #ifdef CONFIG_COMPAT
4475 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4476 struct io_async_msghdr *iomsg)
4478 struct io_sr_msg *sr = &req->sr_msg;
4479 struct compat_iovec __user *uiov;
4484 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4489 uiov = compat_ptr(ptr);
4490 if (req->flags & REQ_F_BUFFER_SELECT) {
4491 compat_ssize_t clen;
4495 if (!access_ok(uiov, sizeof(*uiov)))
4497 if (__get_user(clen, &uiov->iov_len))
4502 iomsg->free_iov = NULL;
4504 iomsg->free_iov = iomsg->fast_iov;
4505 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4506 UIO_FASTIOV, &iomsg->free_iov,
4507 &iomsg->msg.msg_iter, true);
4516 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4517 struct io_async_msghdr *iomsg)
4519 iomsg->msg.msg_name = &iomsg->addr;
4521 #ifdef CONFIG_COMPAT
4522 if (req->ctx->compat)
4523 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4526 return __io_recvmsg_copy_hdr(req, iomsg);
4529 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4532 struct io_sr_msg *sr = &req->sr_msg;
4533 struct io_buffer *kbuf;
4535 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4540 req->flags |= REQ_F_BUFFER_SELECTED;
4544 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4546 return io_put_kbuf(req, req->sr_msg.kbuf);
4549 static int io_recvmsg_prep_async(struct io_kiocb *req)
4553 ret = io_recvmsg_copy_hdr(req, req->async_data);
4555 req->flags |= REQ_F_NEED_CLEANUP;
4559 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4561 struct io_sr_msg *sr = &req->sr_msg;
4563 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4566 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4567 sr->len = READ_ONCE(sqe->len);
4568 sr->bgid = READ_ONCE(sqe->buf_group);
4569 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4570 if (sr->msg_flags & MSG_DONTWAIT)
4571 req->flags |= REQ_F_NOWAIT;
4573 #ifdef CONFIG_COMPAT
4574 if (req->ctx->compat)
4575 sr->msg_flags |= MSG_CMSG_COMPAT;
4580 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4582 struct io_async_msghdr iomsg, *kmsg;
4583 struct socket *sock;
4584 struct io_buffer *kbuf;
4587 int ret, cflags = 0;
4588 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4590 sock = sock_from_file(req->file);
4591 if (unlikely(!sock))
4594 kmsg = req->async_data;
4596 ret = io_recvmsg_copy_hdr(req, &iomsg);
4602 if (req->flags & REQ_F_BUFFER_SELECT) {
4603 kbuf = io_recv_buffer_select(req, !force_nonblock);
4605 return PTR_ERR(kbuf);
4606 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4607 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4608 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4609 1, req->sr_msg.len);
4612 flags = req->sr_msg.msg_flags;
4614 flags |= MSG_DONTWAIT;
4615 if (flags & MSG_WAITALL)
4616 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4618 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4619 kmsg->uaddr, flags);
4620 if (force_nonblock && ret == -EAGAIN)
4621 return io_setup_async_msg(req, kmsg);
4622 if (ret == -ERESTARTSYS)
4625 if (req->flags & REQ_F_BUFFER_SELECTED)
4626 cflags = io_put_recv_kbuf(req);
4627 /* fast path, check for non-NULL to avoid function call */
4629 kfree(kmsg->free_iov);
4630 req->flags &= ~REQ_F_NEED_CLEANUP;
4631 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4633 __io_req_complete(req, issue_flags, ret, cflags);
4637 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4639 struct io_buffer *kbuf;
4640 struct io_sr_msg *sr = &req->sr_msg;
4642 void __user *buf = sr->buf;
4643 struct socket *sock;
4647 int ret, cflags = 0;
4648 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4650 sock = sock_from_file(req->file);
4651 if (unlikely(!sock))
4654 if (req->flags & REQ_F_BUFFER_SELECT) {
4655 kbuf = io_recv_buffer_select(req, !force_nonblock);
4657 return PTR_ERR(kbuf);
4658 buf = u64_to_user_ptr(kbuf->addr);
4661 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4665 msg.msg_name = NULL;
4666 msg.msg_control = NULL;
4667 msg.msg_controllen = 0;
4668 msg.msg_namelen = 0;
4669 msg.msg_iocb = NULL;
4672 flags = req->sr_msg.msg_flags;
4674 flags |= MSG_DONTWAIT;
4675 if (flags & MSG_WAITALL)
4676 min_ret = iov_iter_count(&msg.msg_iter);
4678 ret = sock_recvmsg(sock, &msg, flags);
4679 if (force_nonblock && ret == -EAGAIN)
4681 if (ret == -ERESTARTSYS)
4684 if (req->flags & REQ_F_BUFFER_SELECTED)
4685 cflags = io_put_recv_kbuf(req);
4686 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4688 __io_req_complete(req, issue_flags, ret, cflags);
4692 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4694 struct io_accept *accept = &req->accept;
4696 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4698 if (sqe->ioprio || sqe->len || sqe->buf_index)
4701 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4702 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4703 accept->flags = READ_ONCE(sqe->accept_flags);
4704 accept->nofile = rlimit(RLIMIT_NOFILE);
4708 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4710 struct io_accept *accept = &req->accept;
4711 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4712 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4715 if (req->file->f_flags & O_NONBLOCK)
4716 req->flags |= REQ_F_NOWAIT;
4718 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4719 accept->addr_len, accept->flags,
4721 if (ret == -EAGAIN && force_nonblock)
4724 if (ret == -ERESTARTSYS)
4728 __io_req_complete(req, issue_flags, ret, 0);
4732 static int io_connect_prep_async(struct io_kiocb *req)
4734 struct io_async_connect *io = req->async_data;
4735 struct io_connect *conn = &req->connect;
4737 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4740 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4742 struct io_connect *conn = &req->connect;
4744 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4746 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4749 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4750 conn->addr_len = READ_ONCE(sqe->addr2);
4754 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4756 struct io_async_connect __io, *io;
4757 unsigned file_flags;
4759 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4761 if (req->async_data) {
4762 io = req->async_data;
4764 ret = move_addr_to_kernel(req->connect.addr,
4765 req->connect.addr_len,
4772 file_flags = force_nonblock ? O_NONBLOCK : 0;
4774 ret = __sys_connect_file(req->file, &io->address,
4775 req->connect.addr_len, file_flags);
4776 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4777 if (req->async_data)
4779 if (io_alloc_async_data(req)) {
4783 memcpy(req->async_data, &__io, sizeof(__io));
4786 if (ret == -ERESTARTSYS)
4791 __io_req_complete(req, issue_flags, ret, 0);
4794 #else /* !CONFIG_NET */
4795 #define IO_NETOP_FN(op) \
4796 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4798 return -EOPNOTSUPP; \
4801 #define IO_NETOP_PREP(op) \
4803 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4805 return -EOPNOTSUPP; \
4808 #define IO_NETOP_PREP_ASYNC(op) \
4810 static int io_##op##_prep_async(struct io_kiocb *req) \
4812 return -EOPNOTSUPP; \
4815 IO_NETOP_PREP_ASYNC(sendmsg);
4816 IO_NETOP_PREP_ASYNC(recvmsg);
4817 IO_NETOP_PREP_ASYNC(connect);
4818 IO_NETOP_PREP(accept);
4821 #endif /* CONFIG_NET */
4823 struct io_poll_table {
4824 struct poll_table_struct pt;
4825 struct io_kiocb *req;
4829 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4830 __poll_t mask, task_work_func_t func)
4834 /* for instances that support it check for an event match first: */
4835 if (mask && !(mask & poll->events))
4838 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4840 list_del_init(&poll->wait.entry);
4843 req->task_work.func = func;
4846 * If this fails, then the task is exiting. When a task exits, the
4847 * work gets canceled, so just cancel this request as well instead
4848 * of executing it. We can't safely execute it anyway, as we may not
4849 * have the needed state needed for it anyway.
4851 ret = io_req_task_work_add(req);
4852 if (unlikely(ret)) {
4853 WRITE_ONCE(poll->canceled, true);
4854 io_req_task_work_add_fallback(req, func);
4859 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4860 __acquires(&req->ctx->completion_lock)
4862 struct io_ring_ctx *ctx = req->ctx;
4864 if (!req->result && !READ_ONCE(poll->canceled)) {
4865 struct poll_table_struct pt = { ._key = poll->events };
4867 req->result = vfs_poll(req->file, &pt) & poll->events;
4870 spin_lock_irq(&ctx->completion_lock);
4871 if (!req->result && !READ_ONCE(poll->canceled)) {
4872 add_wait_queue(poll->head, &poll->wait);
4879 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4881 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4882 if (req->opcode == IORING_OP_POLL_ADD)
4883 return req->async_data;
4884 return req->apoll->double_poll;
4887 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4889 if (req->opcode == IORING_OP_POLL_ADD)
4891 return &req->apoll->poll;
4894 static void io_poll_remove_double(struct io_kiocb *req)
4895 __must_hold(&req->ctx->completion_lock)
4897 struct io_poll_iocb *poll = io_poll_get_double(req);
4899 lockdep_assert_held(&req->ctx->completion_lock);
4901 if (poll && poll->head) {
4902 struct wait_queue_head *head = poll->head;
4904 spin_lock(&head->lock);
4905 list_del_init(&poll->wait.entry);
4906 if (poll->wait.private)
4909 spin_unlock(&head->lock);
4913 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4914 __must_hold(&req->ctx->completion_lock)
4916 struct io_ring_ctx *ctx = req->ctx;
4917 unsigned flags = IORING_CQE_F_MORE;
4920 if (READ_ONCE(req->poll.canceled)) {
4922 req->poll.events |= EPOLLONESHOT;
4924 error = mangle_poll(mask);
4926 if (req->poll.events & EPOLLONESHOT)
4928 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4929 io_poll_remove_waitqs(req);
4930 req->poll.done = true;
4933 if (flags & IORING_CQE_F_MORE)
4936 io_commit_cqring(ctx);
4937 return !(flags & IORING_CQE_F_MORE);
4940 static void io_poll_task_func(struct callback_head *cb)
4942 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4943 struct io_ring_ctx *ctx = req->ctx;
4944 struct io_kiocb *nxt;
4946 if (io_poll_rewait(req, &req->poll)) {
4947 spin_unlock_irq(&ctx->completion_lock);
4951 done = io_poll_complete(req, req->result);
4953 hash_del(&req->hash_node);
4956 add_wait_queue(req->poll.head, &req->poll.wait);
4958 spin_unlock_irq(&ctx->completion_lock);
4959 io_cqring_ev_posted(ctx);
4962 nxt = io_put_req_find_next(req);
4964 __io_req_task_submit(nxt);
4969 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4970 int sync, void *key)
4972 struct io_kiocb *req = wait->private;
4973 struct io_poll_iocb *poll = io_poll_get_single(req);
4974 __poll_t mask = key_to_poll(key);
4976 /* for instances that support it check for an event match first: */
4977 if (mask && !(mask & poll->events))
4979 if (!(poll->events & EPOLLONESHOT))
4980 return poll->wait.func(&poll->wait, mode, sync, key);
4982 list_del_init(&wait->entry);
4984 if (poll && poll->head) {
4987 spin_lock(&poll->head->lock);
4988 done = list_empty(&poll->wait.entry);
4990 list_del_init(&poll->wait.entry);
4991 /* make sure double remove sees this as being gone */
4992 wait->private = NULL;
4993 spin_unlock(&poll->head->lock);
4995 /* use wait func handler, so it matches the rq type */
4996 poll->wait.func(&poll->wait, mode, sync, key);
5003 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5004 wait_queue_func_t wake_func)
5008 poll->canceled = false;
5009 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5010 /* mask in events that we always want/need */
5011 poll->events = events | IO_POLL_UNMASK;
5012 INIT_LIST_HEAD(&poll->wait.entry);
5013 init_waitqueue_func_entry(&poll->wait, wake_func);
5016 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5017 struct wait_queue_head *head,
5018 struct io_poll_iocb **poll_ptr)
5020 struct io_kiocb *req = pt->req;
5023 * If poll->head is already set, it's because the file being polled
5024 * uses multiple waitqueues for poll handling (eg one for read, one
5025 * for write). Setup a separate io_poll_iocb if this happens.
5027 if (unlikely(poll->head)) {
5028 struct io_poll_iocb *poll_one = poll;
5030 /* already have a 2nd entry, fail a third attempt */
5032 pt->error = -EINVAL;
5036 * Can't handle multishot for double wait for now, turn it
5037 * into one-shot mode.
5039 if (!(poll_one->events & EPOLLONESHOT))
5040 poll_one->events |= EPOLLONESHOT;
5041 /* double add on the same waitqueue head, ignore */
5042 if (poll_one->head == head)
5044 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5046 pt->error = -ENOMEM;
5049 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5051 poll->wait.private = req;
5058 if (poll->events & EPOLLEXCLUSIVE)
5059 add_wait_queue_exclusive(head, &poll->wait);
5061 add_wait_queue(head, &poll->wait);
5064 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5065 struct poll_table_struct *p)
5067 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5068 struct async_poll *apoll = pt->req->apoll;
5070 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5073 static void io_async_task_func(struct callback_head *cb)
5075 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5076 struct async_poll *apoll = req->apoll;
5077 struct io_ring_ctx *ctx = req->ctx;
5079 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5081 if (io_poll_rewait(req, &apoll->poll)) {
5082 spin_unlock_irq(&ctx->completion_lock);
5086 hash_del(&req->hash_node);
5087 io_poll_remove_double(req);
5088 spin_unlock_irq(&ctx->completion_lock);
5090 if (!READ_ONCE(apoll->poll.canceled))
5091 __io_req_task_submit(req);
5093 io_req_complete_failed(req, -ECANCELED);
5096 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5099 struct io_kiocb *req = wait->private;
5100 struct io_poll_iocb *poll = &req->apoll->poll;
5102 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5105 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5108 static void io_poll_req_insert(struct io_kiocb *req)
5110 struct io_ring_ctx *ctx = req->ctx;
5111 struct hlist_head *list;
5113 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5114 hlist_add_head(&req->hash_node, list);
5117 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5118 struct io_poll_iocb *poll,
5119 struct io_poll_table *ipt, __poll_t mask,
5120 wait_queue_func_t wake_func)
5121 __acquires(&ctx->completion_lock)
5123 struct io_ring_ctx *ctx = req->ctx;
5124 bool cancel = false;
5126 INIT_HLIST_NODE(&req->hash_node);
5127 io_init_poll_iocb(poll, mask, wake_func);
5128 poll->file = req->file;
5129 poll->wait.private = req;
5131 ipt->pt._key = mask;
5133 ipt->error = -EINVAL;
5135 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5137 spin_lock_irq(&ctx->completion_lock);
5138 if (likely(poll->head)) {
5139 spin_lock(&poll->head->lock);
5140 if (unlikely(list_empty(&poll->wait.entry))) {
5146 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5147 list_del_init(&poll->wait.entry);
5149 WRITE_ONCE(poll->canceled, true);
5150 else if (!poll->done) /* actually waiting for an event */
5151 io_poll_req_insert(req);
5152 spin_unlock(&poll->head->lock);
5158 static bool io_arm_poll_handler(struct io_kiocb *req)
5160 const struct io_op_def *def = &io_op_defs[req->opcode];
5161 struct io_ring_ctx *ctx = req->ctx;
5162 struct async_poll *apoll;
5163 struct io_poll_table ipt;
5167 if (!req->file || !file_can_poll(req->file))
5169 if (req->flags & REQ_F_POLLED)
5173 else if (def->pollout)
5177 /* if we can't nonblock try, then no point in arming a poll handler */
5178 if (!io_file_supports_async(req, rw))
5181 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5182 if (unlikely(!apoll))
5184 apoll->double_poll = NULL;
5186 req->flags |= REQ_F_POLLED;
5189 mask = EPOLLONESHOT;
5191 mask |= POLLIN | POLLRDNORM;
5193 mask |= POLLOUT | POLLWRNORM;
5195 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5196 if ((req->opcode == IORING_OP_RECVMSG) &&
5197 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5200 mask |= POLLERR | POLLPRI;
5202 ipt.pt._qproc = io_async_queue_proc;
5204 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5206 if (ret || ipt.error) {
5207 io_poll_remove_double(req);
5208 spin_unlock_irq(&ctx->completion_lock);
5211 spin_unlock_irq(&ctx->completion_lock);
5212 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5213 apoll->poll.events);
5217 static bool __io_poll_remove_one(struct io_kiocb *req,
5218 struct io_poll_iocb *poll, bool do_cancel)
5219 __must_hold(&req->ctx->completion_lock)
5221 bool do_complete = false;
5225 spin_lock(&poll->head->lock);
5227 WRITE_ONCE(poll->canceled, true);
5228 if (!list_empty(&poll->wait.entry)) {
5229 list_del_init(&poll->wait.entry);
5232 spin_unlock(&poll->head->lock);
5233 hash_del(&req->hash_node);
5237 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5238 __must_hold(&req->ctx->completion_lock)
5242 io_poll_remove_double(req);
5243 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5245 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5246 /* non-poll requests have submit ref still */
5252 static bool io_poll_remove_one(struct io_kiocb *req)
5253 __must_hold(&req->ctx->completion_lock)
5257 do_complete = io_poll_remove_waitqs(req);
5259 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5260 io_commit_cqring(req->ctx);
5262 io_put_req_deferred(req, 1);
5269 * Returns true if we found and killed one or more poll requests
5271 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5274 struct hlist_node *tmp;
5275 struct io_kiocb *req;
5278 spin_lock_irq(&ctx->completion_lock);
5279 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5280 struct hlist_head *list;
5282 list = &ctx->cancel_hash[i];
5283 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5284 if (io_match_task(req, tsk, cancel_all))
5285 posted += io_poll_remove_one(req);
5288 spin_unlock_irq(&ctx->completion_lock);
5291 io_cqring_ev_posted(ctx);
5296 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5298 __must_hold(&ctx->completion_lock)
5300 struct hlist_head *list;
5301 struct io_kiocb *req;
5303 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5304 hlist_for_each_entry(req, list, hash_node) {
5305 if (sqe_addr != req->user_data)
5307 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5314 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5316 __must_hold(&ctx->completion_lock)
5318 struct io_kiocb *req;
5320 req = io_poll_find(ctx, sqe_addr, poll_only);
5323 if (io_poll_remove_one(req))
5329 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5334 events = READ_ONCE(sqe->poll32_events);
5336 events = swahw32(events);
5338 if (!(flags & IORING_POLL_ADD_MULTI))
5339 events |= EPOLLONESHOT;
5340 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5343 static int io_poll_update_prep(struct io_kiocb *req,
5344 const struct io_uring_sqe *sqe)
5346 struct io_poll_update *upd = &req->poll_update;
5349 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5351 if (sqe->ioprio || sqe->buf_index)
5353 flags = READ_ONCE(sqe->len);
5354 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5355 IORING_POLL_ADD_MULTI))
5357 /* meaningless without update */
5358 if (flags == IORING_POLL_ADD_MULTI)
5361 upd->old_user_data = READ_ONCE(sqe->addr);
5362 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5363 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5365 upd->new_user_data = READ_ONCE(sqe->off);
5366 if (!upd->update_user_data && upd->new_user_data)
5368 if (upd->update_events)
5369 upd->events = io_poll_parse_events(sqe, flags);
5370 else if (sqe->poll32_events)
5376 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5379 struct io_kiocb *req = wait->private;
5380 struct io_poll_iocb *poll = &req->poll;
5382 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5385 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5386 struct poll_table_struct *p)
5388 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5390 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5393 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5395 struct io_poll_iocb *poll = &req->poll;
5398 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5400 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5402 flags = READ_ONCE(sqe->len);
5403 if (flags & ~IORING_POLL_ADD_MULTI)
5406 poll->events = io_poll_parse_events(sqe, flags);
5410 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5412 struct io_poll_iocb *poll = &req->poll;
5413 struct io_ring_ctx *ctx = req->ctx;
5414 struct io_poll_table ipt;
5417 ipt.pt._qproc = io_poll_queue_proc;
5419 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5422 if (mask) { /* no async, we'd stolen it */
5424 io_poll_complete(req, mask);
5426 spin_unlock_irq(&ctx->completion_lock);
5429 io_cqring_ev_posted(ctx);
5430 if (poll->events & EPOLLONESHOT)
5436 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5438 struct io_ring_ctx *ctx = req->ctx;
5439 struct io_kiocb *preq;
5443 spin_lock_irq(&ctx->completion_lock);
5444 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5450 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5452 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5457 * Don't allow racy completion with singleshot, as we cannot safely
5458 * update those. For multishot, if we're racing with completion, just
5459 * let completion re-add it.
5461 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5462 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5466 /* we now have a detached poll request. reissue. */
5470 spin_unlock_irq(&ctx->completion_lock);
5472 io_req_complete(req, ret);
5475 /* only mask one event flags, keep behavior flags */
5476 if (req->poll_update.update_events) {
5477 preq->poll.events &= ~0xffff;
5478 preq->poll.events |= req->poll_update.events & 0xffff;
5479 preq->poll.events |= IO_POLL_UNMASK;
5481 if (req->poll_update.update_user_data)
5482 preq->user_data = req->poll_update.new_user_data;
5483 spin_unlock_irq(&ctx->completion_lock);
5485 /* complete update request, we're done with it */
5486 io_req_complete(req, ret);
5489 ret = io_poll_add(preq, issue_flags);
5492 io_req_complete(preq, ret);
5498 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5500 struct io_timeout_data *data = container_of(timer,
5501 struct io_timeout_data, timer);
5502 struct io_kiocb *req = data->req;
5503 struct io_ring_ctx *ctx = req->ctx;
5504 unsigned long flags;
5506 spin_lock_irqsave(&ctx->completion_lock, flags);
5507 list_del_init(&req->timeout.list);
5508 atomic_set(&req->ctx->cq_timeouts,
5509 atomic_read(&req->ctx->cq_timeouts) + 1);
5511 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5512 io_commit_cqring(ctx);
5513 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5515 io_cqring_ev_posted(ctx);
5518 return HRTIMER_NORESTART;
5521 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5523 __must_hold(&ctx->completion_lock)
5525 struct io_timeout_data *io;
5526 struct io_kiocb *req;
5529 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5530 found = user_data == req->user_data;
5535 return ERR_PTR(-ENOENT);
5537 io = req->async_data;
5538 if (hrtimer_try_to_cancel(&io->timer) == -1)
5539 return ERR_PTR(-EALREADY);
5540 list_del_init(&req->timeout.list);
5544 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5545 __must_hold(&ctx->completion_lock)
5547 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5550 return PTR_ERR(req);
5553 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5554 io_put_req_deferred(req, 1);
5558 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5559 struct timespec64 *ts, enum hrtimer_mode mode)
5560 __must_hold(&ctx->completion_lock)
5562 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5563 struct io_timeout_data *data;
5566 return PTR_ERR(req);
5568 req->timeout.off = 0; /* noseq */
5569 data = req->async_data;
5570 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5571 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5572 data->timer.function = io_timeout_fn;
5573 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5577 static int io_timeout_remove_prep(struct io_kiocb *req,
5578 const struct io_uring_sqe *sqe)
5580 struct io_timeout_rem *tr = &req->timeout_rem;
5582 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5584 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5586 if (sqe->ioprio || sqe->buf_index || sqe->len)
5589 tr->addr = READ_ONCE(sqe->addr);
5590 tr->flags = READ_ONCE(sqe->timeout_flags);
5591 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5592 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5594 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5596 } else if (tr->flags) {
5597 /* timeout removal doesn't support flags */
5604 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5606 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5611 * Remove or update an existing timeout command
5613 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5615 struct io_timeout_rem *tr = &req->timeout_rem;
5616 struct io_ring_ctx *ctx = req->ctx;
5619 spin_lock_irq(&ctx->completion_lock);
5620 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5621 ret = io_timeout_cancel(ctx, tr->addr);
5623 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5624 io_translate_timeout_mode(tr->flags));
5626 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5627 io_commit_cqring(ctx);
5628 spin_unlock_irq(&ctx->completion_lock);
5629 io_cqring_ev_posted(ctx);
5636 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5637 bool is_timeout_link)
5639 struct io_timeout_data *data;
5641 u32 off = READ_ONCE(sqe->off);
5643 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5645 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5647 if (off && is_timeout_link)
5649 flags = READ_ONCE(sqe->timeout_flags);
5650 if (flags & ~IORING_TIMEOUT_ABS)
5653 req->timeout.off = off;
5655 if (!req->async_data && io_alloc_async_data(req))
5658 data = req->async_data;
5661 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5664 data->mode = io_translate_timeout_mode(flags);
5665 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5666 if (is_timeout_link)
5667 io_req_track_inflight(req);
5671 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5673 struct io_ring_ctx *ctx = req->ctx;
5674 struct io_timeout_data *data = req->async_data;
5675 struct list_head *entry;
5676 u32 tail, off = req->timeout.off;
5678 spin_lock_irq(&ctx->completion_lock);
5681 * sqe->off holds how many events that need to occur for this
5682 * timeout event to be satisfied. If it isn't set, then this is
5683 * a pure timeout request, sequence isn't used.
5685 if (io_is_timeout_noseq(req)) {
5686 entry = ctx->timeout_list.prev;
5690 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5691 req->timeout.target_seq = tail + off;
5693 /* Update the last seq here in case io_flush_timeouts() hasn't.
5694 * This is safe because ->completion_lock is held, and submissions
5695 * and completions are never mixed in the same ->completion_lock section.
5697 ctx->cq_last_tm_flush = tail;
5700 * Insertion sort, ensuring the first entry in the list is always
5701 * the one we need first.
5703 list_for_each_prev(entry, &ctx->timeout_list) {
5704 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5707 if (io_is_timeout_noseq(nxt))
5709 /* nxt.seq is behind @tail, otherwise would've been completed */
5710 if (off >= nxt->timeout.target_seq - tail)
5714 list_add(&req->timeout.list, entry);
5715 data->timer.function = io_timeout_fn;
5716 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5717 spin_unlock_irq(&ctx->completion_lock);
5721 struct io_cancel_data {
5722 struct io_ring_ctx *ctx;
5726 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5728 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5729 struct io_cancel_data *cd = data;
5731 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5734 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5735 struct io_ring_ctx *ctx)
5737 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5738 enum io_wq_cancel cancel_ret;
5741 if (!tctx || !tctx->io_wq)
5744 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5745 switch (cancel_ret) {
5746 case IO_WQ_CANCEL_OK:
5749 case IO_WQ_CANCEL_RUNNING:
5752 case IO_WQ_CANCEL_NOTFOUND:
5760 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5761 struct io_kiocb *req, __u64 sqe_addr,
5764 unsigned long flags;
5767 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5768 spin_lock_irqsave(&ctx->completion_lock, flags);
5771 ret = io_timeout_cancel(ctx, sqe_addr);
5774 ret = io_poll_cancel(ctx, sqe_addr, false);
5778 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5779 io_commit_cqring(ctx);
5780 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5781 io_cqring_ev_posted(ctx);
5787 static int io_async_cancel_prep(struct io_kiocb *req,
5788 const struct io_uring_sqe *sqe)
5790 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5792 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5794 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5797 req->cancel.addr = READ_ONCE(sqe->addr);
5801 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5803 struct io_ring_ctx *ctx = req->ctx;
5804 u64 sqe_addr = req->cancel.addr;
5805 struct io_tctx_node *node;
5808 /* tasks should wait for their io-wq threads, so safe w/o sync */
5809 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5810 spin_lock_irq(&ctx->completion_lock);
5813 ret = io_timeout_cancel(ctx, sqe_addr);
5816 ret = io_poll_cancel(ctx, sqe_addr, false);
5819 spin_unlock_irq(&ctx->completion_lock);
5821 /* slow path, try all io-wq's */
5822 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5824 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5825 struct io_uring_task *tctx = node->task->io_uring;
5827 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5831 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5833 spin_lock_irq(&ctx->completion_lock);
5835 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5836 io_commit_cqring(ctx);
5837 spin_unlock_irq(&ctx->completion_lock);
5838 io_cqring_ev_posted(ctx);
5846 static int io_rsrc_update_prep(struct io_kiocb *req,
5847 const struct io_uring_sqe *sqe)
5849 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5851 if (sqe->ioprio || sqe->rw_flags)
5854 req->rsrc_update.offset = READ_ONCE(sqe->off);
5855 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5856 if (!req->rsrc_update.nr_args)
5858 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5862 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5864 struct io_ring_ctx *ctx = req->ctx;
5865 struct io_uring_rsrc_update2 up;
5868 if (issue_flags & IO_URING_F_NONBLOCK)
5871 up.offset = req->rsrc_update.offset;
5872 up.data = req->rsrc_update.arg;
5877 mutex_lock(&ctx->uring_lock);
5878 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5879 &up, req->rsrc_update.nr_args);
5880 mutex_unlock(&ctx->uring_lock);
5884 __io_req_complete(req, issue_flags, ret, 0);
5888 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5890 switch (req->opcode) {
5893 case IORING_OP_READV:
5894 case IORING_OP_READ_FIXED:
5895 case IORING_OP_READ:
5896 return io_read_prep(req, sqe);
5897 case IORING_OP_WRITEV:
5898 case IORING_OP_WRITE_FIXED:
5899 case IORING_OP_WRITE:
5900 return io_write_prep(req, sqe);
5901 case IORING_OP_POLL_ADD:
5902 return io_poll_add_prep(req, sqe);
5903 case IORING_OP_POLL_REMOVE:
5904 return io_poll_update_prep(req, sqe);
5905 case IORING_OP_FSYNC:
5906 return io_fsync_prep(req, sqe);
5907 case IORING_OP_SYNC_FILE_RANGE:
5908 return io_sfr_prep(req, sqe);
5909 case IORING_OP_SENDMSG:
5910 case IORING_OP_SEND:
5911 return io_sendmsg_prep(req, sqe);
5912 case IORING_OP_RECVMSG:
5913 case IORING_OP_RECV:
5914 return io_recvmsg_prep(req, sqe);
5915 case IORING_OP_CONNECT:
5916 return io_connect_prep(req, sqe);
5917 case IORING_OP_TIMEOUT:
5918 return io_timeout_prep(req, sqe, false);
5919 case IORING_OP_TIMEOUT_REMOVE:
5920 return io_timeout_remove_prep(req, sqe);
5921 case IORING_OP_ASYNC_CANCEL:
5922 return io_async_cancel_prep(req, sqe);
5923 case IORING_OP_LINK_TIMEOUT:
5924 return io_timeout_prep(req, sqe, true);
5925 case IORING_OP_ACCEPT:
5926 return io_accept_prep(req, sqe);
5927 case IORING_OP_FALLOCATE:
5928 return io_fallocate_prep(req, sqe);
5929 case IORING_OP_OPENAT:
5930 return io_openat_prep(req, sqe);
5931 case IORING_OP_CLOSE:
5932 return io_close_prep(req, sqe);
5933 case IORING_OP_FILES_UPDATE:
5934 return io_rsrc_update_prep(req, sqe);
5935 case IORING_OP_STATX:
5936 return io_statx_prep(req, sqe);
5937 case IORING_OP_FADVISE:
5938 return io_fadvise_prep(req, sqe);
5939 case IORING_OP_MADVISE:
5940 return io_madvise_prep(req, sqe);
5941 case IORING_OP_OPENAT2:
5942 return io_openat2_prep(req, sqe);
5943 case IORING_OP_EPOLL_CTL:
5944 return io_epoll_ctl_prep(req, sqe);
5945 case IORING_OP_SPLICE:
5946 return io_splice_prep(req, sqe);
5947 case IORING_OP_PROVIDE_BUFFERS:
5948 return io_provide_buffers_prep(req, sqe);
5949 case IORING_OP_REMOVE_BUFFERS:
5950 return io_remove_buffers_prep(req, sqe);
5952 return io_tee_prep(req, sqe);
5953 case IORING_OP_SHUTDOWN:
5954 return io_shutdown_prep(req, sqe);
5955 case IORING_OP_RENAMEAT:
5956 return io_renameat_prep(req, sqe);
5957 case IORING_OP_UNLINKAT:
5958 return io_unlinkat_prep(req, sqe);
5961 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5966 static int io_req_prep_async(struct io_kiocb *req)
5968 if (!io_op_defs[req->opcode].needs_async_setup)
5970 if (WARN_ON_ONCE(req->async_data))
5972 if (io_alloc_async_data(req))
5975 switch (req->opcode) {
5976 case IORING_OP_READV:
5977 return io_rw_prep_async(req, READ);
5978 case IORING_OP_WRITEV:
5979 return io_rw_prep_async(req, WRITE);
5980 case IORING_OP_SENDMSG:
5981 return io_sendmsg_prep_async(req);
5982 case IORING_OP_RECVMSG:
5983 return io_recvmsg_prep_async(req);
5984 case IORING_OP_CONNECT:
5985 return io_connect_prep_async(req);
5987 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5992 static u32 io_get_sequence(struct io_kiocb *req)
5994 struct io_kiocb *pos;
5995 struct io_ring_ctx *ctx = req->ctx;
5996 u32 total_submitted, nr_reqs = 0;
5998 io_for_each_link(pos, req)
6001 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6002 return total_submitted - nr_reqs;
6005 static int io_req_defer(struct io_kiocb *req)
6007 struct io_ring_ctx *ctx = req->ctx;
6008 struct io_defer_entry *de;
6012 /* Still need defer if there is pending req in defer list. */
6013 if (likely(list_empty_careful(&ctx->defer_list) &&
6014 !(req->flags & REQ_F_IO_DRAIN)))
6017 seq = io_get_sequence(req);
6018 /* Still a chance to pass the sequence check */
6019 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6022 ret = io_req_prep_async(req);
6025 io_prep_async_link(req);
6026 de = kmalloc(sizeof(*de), GFP_KERNEL);
6030 spin_lock_irq(&ctx->completion_lock);
6031 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6032 spin_unlock_irq(&ctx->completion_lock);
6034 io_queue_async_work(req);
6035 return -EIOCBQUEUED;
6038 trace_io_uring_defer(ctx, req, req->user_data);
6041 list_add_tail(&de->list, &ctx->defer_list);
6042 spin_unlock_irq(&ctx->completion_lock);
6043 return -EIOCBQUEUED;
6046 static void io_clean_op(struct io_kiocb *req)
6048 if (req->flags & REQ_F_BUFFER_SELECTED) {
6049 switch (req->opcode) {
6050 case IORING_OP_READV:
6051 case IORING_OP_READ_FIXED:
6052 case IORING_OP_READ:
6053 kfree((void *)(unsigned long)req->rw.addr);
6055 case IORING_OP_RECVMSG:
6056 case IORING_OP_RECV:
6057 kfree(req->sr_msg.kbuf);
6060 req->flags &= ~REQ_F_BUFFER_SELECTED;
6063 if (req->flags & REQ_F_NEED_CLEANUP) {
6064 switch (req->opcode) {
6065 case IORING_OP_READV:
6066 case IORING_OP_READ_FIXED:
6067 case IORING_OP_READ:
6068 case IORING_OP_WRITEV:
6069 case IORING_OP_WRITE_FIXED:
6070 case IORING_OP_WRITE: {
6071 struct io_async_rw *io = req->async_data;
6073 kfree(io->free_iovec);
6076 case IORING_OP_RECVMSG:
6077 case IORING_OP_SENDMSG: {
6078 struct io_async_msghdr *io = req->async_data;
6080 kfree(io->free_iov);
6083 case IORING_OP_SPLICE:
6085 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6086 io_put_file(req->splice.file_in);
6088 case IORING_OP_OPENAT:
6089 case IORING_OP_OPENAT2:
6090 if (req->open.filename)
6091 putname(req->open.filename);
6093 case IORING_OP_RENAMEAT:
6094 putname(req->rename.oldpath);
6095 putname(req->rename.newpath);
6097 case IORING_OP_UNLINKAT:
6098 putname(req->unlink.filename);
6101 req->flags &= ~REQ_F_NEED_CLEANUP;
6103 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6104 kfree(req->apoll->double_poll);
6108 if (req->flags & REQ_F_INFLIGHT) {
6109 struct io_uring_task *tctx = req->task->io_uring;
6111 atomic_dec(&tctx->inflight_tracked);
6112 req->flags &= ~REQ_F_INFLIGHT;
6116 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6118 struct io_ring_ctx *ctx = req->ctx;
6119 const struct cred *creds = NULL;
6122 if (req->work.creds && req->work.creds != current_cred())
6123 creds = override_creds(req->work.creds);
6125 switch (req->opcode) {
6127 ret = io_nop(req, issue_flags);
6129 case IORING_OP_READV:
6130 case IORING_OP_READ_FIXED:
6131 case IORING_OP_READ:
6132 ret = io_read(req, issue_flags);
6134 case IORING_OP_WRITEV:
6135 case IORING_OP_WRITE_FIXED:
6136 case IORING_OP_WRITE:
6137 ret = io_write(req, issue_flags);
6139 case IORING_OP_FSYNC:
6140 ret = io_fsync(req, issue_flags);
6142 case IORING_OP_POLL_ADD:
6143 ret = io_poll_add(req, issue_flags);
6145 case IORING_OP_POLL_REMOVE:
6146 ret = io_poll_update(req, issue_flags);
6148 case IORING_OP_SYNC_FILE_RANGE:
6149 ret = io_sync_file_range(req, issue_flags);
6151 case IORING_OP_SENDMSG:
6152 ret = io_sendmsg(req, issue_flags);
6154 case IORING_OP_SEND:
6155 ret = io_send(req, issue_flags);
6157 case IORING_OP_RECVMSG:
6158 ret = io_recvmsg(req, issue_flags);
6160 case IORING_OP_RECV:
6161 ret = io_recv(req, issue_flags);
6163 case IORING_OP_TIMEOUT:
6164 ret = io_timeout(req, issue_flags);
6166 case IORING_OP_TIMEOUT_REMOVE:
6167 ret = io_timeout_remove(req, issue_flags);
6169 case IORING_OP_ACCEPT:
6170 ret = io_accept(req, issue_flags);
6172 case IORING_OP_CONNECT:
6173 ret = io_connect(req, issue_flags);
6175 case IORING_OP_ASYNC_CANCEL:
6176 ret = io_async_cancel(req, issue_flags);
6178 case IORING_OP_FALLOCATE:
6179 ret = io_fallocate(req, issue_flags);
6181 case IORING_OP_OPENAT:
6182 ret = io_openat(req, issue_flags);
6184 case IORING_OP_CLOSE:
6185 ret = io_close(req, issue_flags);
6187 case IORING_OP_FILES_UPDATE:
6188 ret = io_files_update(req, issue_flags);
6190 case IORING_OP_STATX:
6191 ret = io_statx(req, issue_flags);
6193 case IORING_OP_FADVISE:
6194 ret = io_fadvise(req, issue_flags);
6196 case IORING_OP_MADVISE:
6197 ret = io_madvise(req, issue_flags);
6199 case IORING_OP_OPENAT2:
6200 ret = io_openat2(req, issue_flags);
6202 case IORING_OP_EPOLL_CTL:
6203 ret = io_epoll_ctl(req, issue_flags);
6205 case IORING_OP_SPLICE:
6206 ret = io_splice(req, issue_flags);
6208 case IORING_OP_PROVIDE_BUFFERS:
6209 ret = io_provide_buffers(req, issue_flags);
6211 case IORING_OP_REMOVE_BUFFERS:
6212 ret = io_remove_buffers(req, issue_flags);
6215 ret = io_tee(req, issue_flags);
6217 case IORING_OP_SHUTDOWN:
6218 ret = io_shutdown(req, issue_flags);
6220 case IORING_OP_RENAMEAT:
6221 ret = io_renameat(req, issue_flags);
6223 case IORING_OP_UNLINKAT:
6224 ret = io_unlinkat(req, issue_flags);
6232 revert_creds(creds);
6235 /* If the op doesn't have a file, we're not polling for it */
6236 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6237 io_iopoll_req_issued(req);
6242 static void io_wq_submit_work(struct io_wq_work *work)
6244 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6245 struct io_kiocb *timeout;
6248 timeout = io_prep_linked_timeout(req);
6250 io_queue_linked_timeout(timeout);
6252 if (work->flags & IO_WQ_WORK_CANCEL)
6257 ret = io_issue_sqe(req, 0);
6259 * We can get EAGAIN for polled IO even though we're
6260 * forcing a sync submission from here, since we can't
6261 * wait for request slots on the block side.
6269 /* avoid locking problems by failing it from a clean context */
6271 /* io-wq is going to take one down */
6273 io_req_task_queue_fail(req, ret);
6277 #define FFS_ASYNC_READ 0x1UL
6278 #define FFS_ASYNC_WRITE 0x2UL
6280 #define FFS_ISREG 0x4UL
6282 #define FFS_ISREG 0x0UL
6284 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6286 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6289 struct io_fixed_file *table_l2;
6291 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6292 return &table_l2[i & IORING_FILE_TABLE_MASK];
6295 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6298 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6300 return (struct file *) (slot->file_ptr & FFS_MASK);
6303 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6305 unsigned long file_ptr = (unsigned long) file;
6307 if (__io_file_supports_async(file, READ))
6308 file_ptr |= FFS_ASYNC_READ;
6309 if (__io_file_supports_async(file, WRITE))
6310 file_ptr |= FFS_ASYNC_WRITE;
6311 if (S_ISREG(file_inode(file)->i_mode))
6312 file_ptr |= FFS_ISREG;
6313 file_slot->file_ptr = file_ptr;
6316 static struct file *io_file_get(struct io_submit_state *state,
6317 struct io_kiocb *req, int fd, bool fixed)
6319 struct io_ring_ctx *ctx = req->ctx;
6323 unsigned long file_ptr;
6325 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6327 fd = array_index_nospec(fd, ctx->nr_user_files);
6328 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6329 file = (struct file *) (file_ptr & FFS_MASK);
6330 file_ptr &= ~FFS_MASK;
6331 /* mask in overlapping REQ_F and FFS bits */
6332 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6333 io_req_set_rsrc_node(req);
6335 trace_io_uring_file_get(ctx, fd);
6336 file = __io_file_get(state, fd);
6338 /* we don't allow fixed io_uring files */
6339 if (file && unlikely(file->f_op == &io_uring_fops))
6340 io_req_track_inflight(req);
6346 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6348 struct io_timeout_data *data = container_of(timer,
6349 struct io_timeout_data, timer);
6350 struct io_kiocb *prev, *req = data->req;
6351 struct io_ring_ctx *ctx = req->ctx;
6352 unsigned long flags;
6354 spin_lock_irqsave(&ctx->completion_lock, flags);
6355 prev = req->timeout.head;
6356 req->timeout.head = NULL;
6359 * We don't expect the list to be empty, that will only happen if we
6360 * race with the completion of the linked work.
6363 io_remove_next_linked(prev);
6364 if (!req_ref_inc_not_zero(prev))
6367 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6370 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6371 io_put_req_deferred(prev, 1);
6372 io_put_req_deferred(req, 1);
6374 io_req_complete_post(req, -ETIME, 0);
6376 return HRTIMER_NORESTART;
6379 static void io_queue_linked_timeout(struct io_kiocb *req)
6381 struct io_ring_ctx *ctx = req->ctx;
6383 spin_lock_irq(&ctx->completion_lock);
6385 * If the back reference is NULL, then our linked request finished
6386 * before we got a chance to setup the timer
6388 if (req->timeout.head) {
6389 struct io_timeout_data *data = req->async_data;
6391 data->timer.function = io_link_timeout_fn;
6392 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6395 spin_unlock_irq(&ctx->completion_lock);
6396 /* drop submission reference */
6400 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6402 struct io_kiocb *nxt = req->link;
6404 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6405 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6408 nxt->timeout.head = req;
6409 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6410 req->flags |= REQ_F_LINK_TIMEOUT;
6414 static void __io_queue_sqe(struct io_kiocb *req)
6416 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6419 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6422 * We async punt it if the file wasn't marked NOWAIT, or if the file
6423 * doesn't support non-blocking read/write attempts
6426 /* drop submission reference */
6427 if (req->flags & REQ_F_COMPLETE_INLINE) {
6428 struct io_ring_ctx *ctx = req->ctx;
6429 struct io_comp_state *cs = &ctx->submit_state.comp;
6431 cs->reqs[cs->nr++] = req;
6432 if (cs->nr == ARRAY_SIZE(cs->reqs))
6433 io_submit_flush_completions(cs, ctx);
6437 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6438 if (!io_arm_poll_handler(req)) {
6440 * Queued up for async execution, worker will release
6441 * submit reference when the iocb is actually submitted.
6443 io_queue_async_work(req);
6446 io_req_complete_failed(req, ret);
6449 io_queue_linked_timeout(linked_timeout);
6452 static void io_queue_sqe(struct io_kiocb *req)
6456 ret = io_req_defer(req);
6458 if (ret != -EIOCBQUEUED) {
6460 io_req_complete_failed(req, ret);
6462 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6463 ret = io_req_prep_async(req);
6466 io_queue_async_work(req);
6468 __io_queue_sqe(req);
6473 * Check SQE restrictions (opcode and flags).
6475 * Returns 'true' if SQE is allowed, 'false' otherwise.
6477 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6478 struct io_kiocb *req,
6479 unsigned int sqe_flags)
6481 if (!ctx->restricted)
6484 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6487 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6488 ctx->restrictions.sqe_flags_required)
6491 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6492 ctx->restrictions.sqe_flags_required))
6498 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6499 const struct io_uring_sqe *sqe)
6501 struct io_submit_state *state;
6502 unsigned int sqe_flags;
6503 int personality, ret = 0;
6505 req->opcode = READ_ONCE(sqe->opcode);
6506 /* same numerical values with corresponding REQ_F_*, safe to copy */
6507 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6508 req->user_data = READ_ONCE(sqe->user_data);
6509 req->async_data = NULL;
6513 req->fixed_rsrc_refs = NULL;
6514 /* one is dropped after submission, the other at completion */
6515 atomic_set(&req->refs, 2);
6516 req->task = current;
6518 req->work.creds = NULL;
6520 /* enforce forwards compatibility on users */
6521 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6523 if (unlikely(req->opcode >= IORING_OP_LAST))
6525 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6528 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6529 !io_op_defs[req->opcode].buffer_select)
6532 personality = READ_ONCE(sqe->personality);
6534 req->work.creds = xa_load(&ctx->personalities, personality);
6535 if (!req->work.creds)
6537 get_cred(req->work.creds);
6539 state = &ctx->submit_state;
6542 * Plug now if we have more than 1 IO left after this, and the target
6543 * is potentially a read/write to block based storage.
6545 if (!state->plug_started && state->ios_left > 1 &&
6546 io_op_defs[req->opcode].plug) {
6547 blk_start_plug(&state->plug);
6548 state->plug_started = true;
6551 if (io_op_defs[req->opcode].needs_file) {
6552 bool fixed = req->flags & REQ_F_FIXED_FILE;
6554 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6555 if (unlikely(!req->file))
6563 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6564 const struct io_uring_sqe *sqe)
6566 struct io_submit_link *link = &ctx->submit_state.link;
6569 ret = io_init_req(ctx, req, sqe);
6570 if (unlikely(ret)) {
6573 /* fail even hard links since we don't submit */
6574 req_set_fail(link->head);
6575 io_req_complete_failed(link->head, -ECANCELED);
6578 io_req_complete_failed(req, ret);
6581 ret = io_req_prep(req, sqe);
6585 /* don't need @sqe from now on */
6586 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6587 true, ctx->flags & IORING_SETUP_SQPOLL);
6590 * If we already have a head request, queue this one for async
6591 * submittal once the head completes. If we don't have a head but
6592 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6593 * submitted sync once the chain is complete. If none of those
6594 * conditions are true (normal request), then just queue it.
6597 struct io_kiocb *head = link->head;
6600 * Taking sequential execution of a link, draining both sides
6601 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6602 * requests in the link. So, it drains the head and the
6603 * next after the link request. The last one is done via
6604 * drain_next flag to persist the effect across calls.
6606 if (req->flags & REQ_F_IO_DRAIN) {
6607 head->flags |= REQ_F_IO_DRAIN;
6608 ctx->drain_next = 1;
6610 ret = io_req_prep_async(req);
6613 trace_io_uring_link(ctx, req, head);
6614 link->last->link = req;
6617 /* last request of a link, enqueue the link */
6618 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6623 if (unlikely(ctx->drain_next)) {
6624 req->flags |= REQ_F_IO_DRAIN;
6625 ctx->drain_next = 0;
6627 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6639 * Batched submission is done, ensure local IO is flushed out.
6641 static void io_submit_state_end(struct io_submit_state *state,
6642 struct io_ring_ctx *ctx)
6644 if (state->link.head)
6645 io_queue_sqe(state->link.head);
6647 io_submit_flush_completions(&state->comp, ctx);
6648 if (state->plug_started)
6649 blk_finish_plug(&state->plug);
6650 io_state_file_put(state);
6654 * Start submission side cache.
6656 static void io_submit_state_start(struct io_submit_state *state,
6657 unsigned int max_ios)
6659 state->plug_started = false;
6660 state->ios_left = max_ios;
6661 /* set only head, no need to init link_last in advance */
6662 state->link.head = NULL;
6665 static void io_commit_sqring(struct io_ring_ctx *ctx)
6667 struct io_rings *rings = ctx->rings;
6670 * Ensure any loads from the SQEs are done at this point,
6671 * since once we write the new head, the application could
6672 * write new data to them.
6674 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6678 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6679 * that is mapped by userspace. This means that care needs to be taken to
6680 * ensure that reads are stable, as we cannot rely on userspace always
6681 * being a good citizen. If members of the sqe are validated and then later
6682 * used, it's important that those reads are done through READ_ONCE() to
6683 * prevent a re-load down the line.
6685 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6687 u32 *sq_array = ctx->sq_array;
6688 unsigned head, mask = ctx->sq_entries - 1;
6691 * The cached sq head (or cq tail) serves two purposes:
6693 * 1) allows us to batch the cost of updating the user visible
6695 * 2) allows the kernel side to track the head on its own, even
6696 * though the application is the one updating it.
6698 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & mask]);
6699 if (likely(head < ctx->sq_entries))
6700 return &ctx->sq_sqes[head];
6702 /* drop invalid entries */
6703 ctx->cached_sq_dropped++;
6704 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6708 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6712 /* make sure SQ entry isn't read before tail */
6713 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6715 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6718 percpu_counter_add(¤t->io_uring->inflight, nr);
6719 refcount_add(nr, ¤t->usage);
6720 io_submit_state_start(&ctx->submit_state, nr);
6722 while (submitted < nr) {
6723 const struct io_uring_sqe *sqe;
6724 struct io_kiocb *req;
6726 req = io_alloc_req(ctx);
6727 if (unlikely(!req)) {
6729 submitted = -EAGAIN;
6732 sqe = io_get_sqe(ctx);
6733 if (unlikely(!sqe)) {
6734 kmem_cache_free(req_cachep, req);
6737 /* will complete beyond this point, count as submitted */
6739 if (io_submit_sqe(ctx, req, sqe))
6743 if (unlikely(submitted != nr)) {
6744 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6745 struct io_uring_task *tctx = current->io_uring;
6746 int unused = nr - ref_used;
6748 percpu_ref_put_many(&ctx->refs, unused);
6749 percpu_counter_sub(&tctx->inflight, unused);
6750 put_task_struct_many(current, unused);
6753 io_submit_state_end(&ctx->submit_state, ctx);
6754 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6755 io_commit_sqring(ctx);
6760 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6762 return READ_ONCE(sqd->state);
6765 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6767 /* Tell userspace we may need a wakeup call */
6768 spin_lock_irq(&ctx->completion_lock);
6769 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6770 spin_unlock_irq(&ctx->completion_lock);
6773 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
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 int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6782 unsigned int to_submit;
6785 to_submit = io_sqring_entries(ctx);
6786 /* if we're handling multiple rings, cap submit size for fairness */
6787 if (cap_entries && to_submit > 8)
6790 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6791 unsigned nr_events = 0;
6793 mutex_lock(&ctx->uring_lock);
6794 if (!list_empty(&ctx->iopoll_list))
6795 io_do_iopoll(ctx, &nr_events, 0);
6798 * Don't submit if refs are dying, good for io_uring_register(),
6799 * but also it is relied upon by io_ring_exit_work()
6801 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6802 !(ctx->flags & IORING_SETUP_R_DISABLED))
6803 ret = io_submit_sqes(ctx, to_submit);
6804 mutex_unlock(&ctx->uring_lock);
6806 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6807 wake_up(&ctx->sqo_sq_wait);
6813 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6815 struct io_ring_ctx *ctx;
6816 unsigned sq_thread_idle = 0;
6818 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6819 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6820 sqd->sq_thread_idle = sq_thread_idle;
6823 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6825 bool did_sig = false;
6826 struct ksignal ksig;
6828 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6829 signal_pending(current)) {
6830 mutex_unlock(&sqd->lock);
6831 if (signal_pending(current))
6832 did_sig = get_signal(&ksig);
6834 mutex_lock(&sqd->lock);
6837 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6840 static int io_sq_thread(void *data)
6842 struct io_sq_data *sqd = data;
6843 struct io_ring_ctx *ctx;
6844 unsigned long timeout = 0;
6845 char buf[TASK_COMM_LEN];
6848 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6849 set_task_comm(current, buf);
6851 if (sqd->sq_cpu != -1)
6852 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6854 set_cpus_allowed_ptr(current, cpu_online_mask);
6855 current->flags |= PF_NO_SETAFFINITY;
6857 mutex_lock(&sqd->lock);
6860 bool cap_entries, sqt_spin, needs_sched;
6862 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6863 if (io_sqd_handle_event(sqd))
6865 timeout = jiffies + sqd->sq_thread_idle;
6870 cap_entries = !list_is_singular(&sqd->ctx_list);
6871 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6872 const struct cred *creds = NULL;
6874 if (ctx->sq_creds != current_cred())
6875 creds = override_creds(ctx->sq_creds);
6876 ret = __io_sq_thread(ctx, cap_entries);
6878 revert_creds(creds);
6879 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6883 if (sqt_spin || !time_after(jiffies, timeout)) {
6887 timeout = jiffies + sqd->sq_thread_idle;
6891 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6892 if (!io_sqd_events_pending(sqd)) {
6894 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6895 io_ring_set_wakeup_flag(ctx);
6897 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6898 !list_empty_careful(&ctx->iopoll_list)) {
6899 needs_sched = false;
6902 if (io_sqring_entries(ctx)) {
6903 needs_sched = false;
6909 mutex_unlock(&sqd->lock);
6911 mutex_lock(&sqd->lock);
6913 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6914 io_ring_clear_wakeup_flag(ctx);
6917 finish_wait(&sqd->wait, &wait);
6918 timeout = jiffies + sqd->sq_thread_idle;
6921 io_uring_cancel_sqpoll(sqd);
6923 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6924 io_ring_set_wakeup_flag(ctx);
6926 mutex_unlock(&sqd->lock);
6928 complete(&sqd->exited);
6932 struct io_wait_queue {
6933 struct wait_queue_entry wq;
6934 struct io_ring_ctx *ctx;
6936 unsigned nr_timeouts;
6939 static inline bool io_should_wake(struct io_wait_queue *iowq)
6941 struct io_ring_ctx *ctx = iowq->ctx;
6944 * Wake up if we have enough events, or if a timeout occurred since we
6945 * started waiting. For timeouts, we always want to return to userspace,
6946 * regardless of event count.
6948 return io_cqring_events(ctx) >= iowq->to_wait ||
6949 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6952 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6953 int wake_flags, void *key)
6955 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6959 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6960 * the task, and the next invocation will do it.
6962 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6963 return autoremove_wake_function(curr, mode, wake_flags, key);
6967 static int io_run_task_work_sig(void)
6969 if (io_run_task_work())
6971 if (!signal_pending(current))
6973 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6974 return -ERESTARTSYS;
6978 /* when returns >0, the caller should retry */
6979 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6980 struct io_wait_queue *iowq,
6981 signed long *timeout)
6985 /* make sure we run task_work before checking for signals */
6986 ret = io_run_task_work_sig();
6987 if (ret || io_should_wake(iowq))
6989 /* let the caller flush overflows, retry */
6990 if (test_bit(0, &ctx->cq_check_overflow))
6993 *timeout = schedule_timeout(*timeout);
6994 return !*timeout ? -ETIME : 1;
6998 * Wait until events become available, if we don't already have some. The
6999 * application must reap them itself, as they reside on the shared cq ring.
7001 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7002 const sigset_t __user *sig, size_t sigsz,
7003 struct __kernel_timespec __user *uts)
7005 struct io_wait_queue iowq = {
7008 .func = io_wake_function,
7009 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7012 .to_wait = min_events,
7014 struct io_rings *rings = ctx->rings;
7015 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7019 io_cqring_overflow_flush(ctx, false);
7020 if (io_cqring_events(ctx) >= min_events)
7022 if (!io_run_task_work())
7027 #ifdef CONFIG_COMPAT
7028 if (in_compat_syscall())
7029 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7033 ret = set_user_sigmask(sig, sigsz);
7040 struct timespec64 ts;
7042 if (get_timespec64(&ts, uts))
7044 timeout = timespec64_to_jiffies(&ts);
7047 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7048 trace_io_uring_cqring_wait(ctx, min_events);
7050 /* if we can't even flush overflow, don't wait for more */
7051 if (!io_cqring_overflow_flush(ctx, false)) {
7055 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7056 TASK_INTERRUPTIBLE);
7057 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7058 finish_wait(&ctx->wait, &iowq.wq);
7062 restore_saved_sigmask_unless(ret == -EINTR);
7064 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7067 static void io_free_page_table(void **table, size_t size)
7069 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7071 for (i = 0; i < nr_tables; i++)
7076 static void **io_alloc_page_table(size_t size)
7078 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7079 size_t init_size = size;
7082 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7086 for (i = 0; i < nr_tables; i++) {
7087 unsigned int this_size = min(size, PAGE_SIZE);
7089 table[i] = kzalloc(this_size, GFP_KERNEL);
7091 io_free_page_table(table, init_size);
7099 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7101 spin_lock_bh(&ctx->rsrc_ref_lock);
7104 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7106 spin_unlock_bh(&ctx->rsrc_ref_lock);
7109 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7111 percpu_ref_exit(&ref_node->refs);
7115 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7116 struct io_rsrc_data *data_to_kill)
7118 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7119 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7122 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7124 rsrc_node->rsrc_data = data_to_kill;
7125 io_rsrc_ref_lock(ctx);
7126 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7127 io_rsrc_ref_unlock(ctx);
7129 atomic_inc(&data_to_kill->refs);
7130 percpu_ref_kill(&rsrc_node->refs);
7131 ctx->rsrc_node = NULL;
7134 if (!ctx->rsrc_node) {
7135 ctx->rsrc_node = ctx->rsrc_backup_node;
7136 ctx->rsrc_backup_node = NULL;
7140 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7142 if (ctx->rsrc_backup_node)
7144 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7145 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7148 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7152 /* As we may drop ->uring_lock, other task may have started quiesce */
7156 data->quiesce = true;
7158 ret = io_rsrc_node_switch_start(ctx);
7161 io_rsrc_node_switch(ctx, data);
7163 /* kill initial ref, already quiesced if zero */
7164 if (atomic_dec_and_test(&data->refs))
7166 flush_delayed_work(&ctx->rsrc_put_work);
7167 ret = wait_for_completion_interruptible(&data->done);
7171 atomic_inc(&data->refs);
7172 /* wait for all works potentially completing data->done */
7173 flush_delayed_work(&ctx->rsrc_put_work);
7174 reinit_completion(&data->done);
7176 mutex_unlock(&ctx->uring_lock);
7177 ret = io_run_task_work_sig();
7178 mutex_lock(&ctx->uring_lock);
7180 data->quiesce = false;
7185 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7187 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7188 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7190 return &data->tags[table_idx][off];
7193 static void io_rsrc_data_free(struct io_rsrc_data *data)
7195 size_t size = data->nr * sizeof(data->tags[0][0]);
7198 io_free_page_table((void **)data->tags, size);
7202 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7203 u64 __user *utags, unsigned nr,
7204 struct io_rsrc_data **pdata)
7206 struct io_rsrc_data *data;
7210 data = kzalloc(sizeof(*data), GFP_KERNEL);
7213 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7221 data->do_put = do_put;
7224 for (i = 0; i < nr; i++) {
7225 if (copy_from_user(io_get_tag_slot(data, i), &utags[i],
7226 sizeof(data->tags[i])))
7231 atomic_set(&data->refs, 1);
7232 init_completion(&data->done);
7236 io_rsrc_data_free(data);
7240 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7242 size_t size = nr_files * sizeof(struct io_fixed_file);
7244 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7245 return !!table->files;
7248 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7250 size_t size = nr_files * sizeof(struct io_fixed_file);
7252 io_free_page_table((void **)table->files, size);
7253 table->files = NULL;
7256 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7258 #if defined(CONFIG_UNIX)
7259 if (ctx->ring_sock) {
7260 struct sock *sock = ctx->ring_sock->sk;
7261 struct sk_buff *skb;
7263 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7269 for (i = 0; i < ctx->nr_user_files; i++) {
7272 file = io_file_from_index(ctx, i);
7277 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7278 io_rsrc_data_free(ctx->file_data);
7279 ctx->file_data = NULL;
7280 ctx->nr_user_files = 0;
7283 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7287 if (!ctx->file_data)
7289 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7291 __io_sqe_files_unregister(ctx);
7295 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7296 __releases(&sqd->lock)
7298 WARN_ON_ONCE(sqd->thread == current);
7301 * Do the dance but not conditional clear_bit() because it'd race with
7302 * other threads incrementing park_pending and setting the bit.
7304 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7305 if (atomic_dec_return(&sqd->park_pending))
7306 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7307 mutex_unlock(&sqd->lock);
7310 static void io_sq_thread_park(struct io_sq_data *sqd)
7311 __acquires(&sqd->lock)
7313 WARN_ON_ONCE(sqd->thread == current);
7315 atomic_inc(&sqd->park_pending);
7316 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7317 mutex_lock(&sqd->lock);
7319 wake_up_process(sqd->thread);
7322 static void io_sq_thread_stop(struct io_sq_data *sqd)
7324 WARN_ON_ONCE(sqd->thread == current);
7325 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7327 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7328 mutex_lock(&sqd->lock);
7330 wake_up_process(sqd->thread);
7331 mutex_unlock(&sqd->lock);
7332 wait_for_completion(&sqd->exited);
7335 static void io_put_sq_data(struct io_sq_data *sqd)
7337 if (refcount_dec_and_test(&sqd->refs)) {
7338 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7340 io_sq_thread_stop(sqd);
7345 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7347 struct io_sq_data *sqd = ctx->sq_data;
7350 io_sq_thread_park(sqd);
7351 list_del_init(&ctx->sqd_list);
7352 io_sqd_update_thread_idle(sqd);
7353 io_sq_thread_unpark(sqd);
7355 io_put_sq_data(sqd);
7356 ctx->sq_data = NULL;
7360 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7362 struct io_ring_ctx *ctx_attach;
7363 struct io_sq_data *sqd;
7366 f = fdget(p->wq_fd);
7368 return ERR_PTR(-ENXIO);
7369 if (f.file->f_op != &io_uring_fops) {
7371 return ERR_PTR(-EINVAL);
7374 ctx_attach = f.file->private_data;
7375 sqd = ctx_attach->sq_data;
7378 return ERR_PTR(-EINVAL);
7380 if (sqd->task_tgid != current->tgid) {
7382 return ERR_PTR(-EPERM);
7385 refcount_inc(&sqd->refs);
7390 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7393 struct io_sq_data *sqd;
7396 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7397 sqd = io_attach_sq_data(p);
7402 /* fall through for EPERM case, setup new sqd/task */
7403 if (PTR_ERR(sqd) != -EPERM)
7407 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7409 return ERR_PTR(-ENOMEM);
7411 atomic_set(&sqd->park_pending, 0);
7412 refcount_set(&sqd->refs, 1);
7413 INIT_LIST_HEAD(&sqd->ctx_list);
7414 mutex_init(&sqd->lock);
7415 init_waitqueue_head(&sqd->wait);
7416 init_completion(&sqd->exited);
7420 #if defined(CONFIG_UNIX)
7422 * Ensure the UNIX gc is aware of our file set, so we are certain that
7423 * the io_uring can be safely unregistered on process exit, even if we have
7424 * loops in the file referencing.
7426 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7428 struct sock *sk = ctx->ring_sock->sk;
7429 struct scm_fp_list *fpl;
7430 struct sk_buff *skb;
7433 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7437 skb = alloc_skb(0, GFP_KERNEL);
7446 fpl->user = get_uid(current_user());
7447 for (i = 0; i < nr; i++) {
7448 struct file *file = io_file_from_index(ctx, i + offset);
7452 fpl->fp[nr_files] = get_file(file);
7453 unix_inflight(fpl->user, fpl->fp[nr_files]);
7458 fpl->max = SCM_MAX_FD;
7459 fpl->count = nr_files;
7460 UNIXCB(skb).fp = fpl;
7461 skb->destructor = unix_destruct_scm;
7462 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7463 skb_queue_head(&sk->sk_receive_queue, skb);
7465 for (i = 0; i < nr_files; i++)
7476 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7477 * causes regular reference counting to break down. We rely on the UNIX
7478 * garbage collection to take care of this problem for us.
7480 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7482 unsigned left, total;
7486 left = ctx->nr_user_files;
7488 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7490 ret = __io_sqe_files_scm(ctx, this_files, total);
7494 total += this_files;
7500 while (total < ctx->nr_user_files) {
7501 struct file *file = io_file_from_index(ctx, total);
7511 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7517 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7519 struct file *file = prsrc->file;
7520 #if defined(CONFIG_UNIX)
7521 struct sock *sock = ctx->ring_sock->sk;
7522 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7523 struct sk_buff *skb;
7526 __skb_queue_head_init(&list);
7529 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7530 * remove this entry and rearrange the file array.
7532 skb = skb_dequeue(head);
7534 struct scm_fp_list *fp;
7536 fp = UNIXCB(skb).fp;
7537 for (i = 0; i < fp->count; i++) {
7540 if (fp->fp[i] != file)
7543 unix_notinflight(fp->user, fp->fp[i]);
7544 left = fp->count - 1 - i;
7546 memmove(&fp->fp[i], &fp->fp[i + 1],
7547 left * sizeof(struct file *));
7554 __skb_queue_tail(&list, skb);
7564 __skb_queue_tail(&list, skb);
7566 skb = skb_dequeue(head);
7569 if (skb_peek(&list)) {
7570 spin_lock_irq(&head->lock);
7571 while ((skb = __skb_dequeue(&list)) != NULL)
7572 __skb_queue_tail(head, skb);
7573 spin_unlock_irq(&head->lock);
7580 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7582 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7583 struct io_ring_ctx *ctx = rsrc_data->ctx;
7584 struct io_rsrc_put *prsrc, *tmp;
7586 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7587 list_del(&prsrc->list);
7590 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7592 io_ring_submit_lock(ctx, lock_ring);
7593 spin_lock_irq(&ctx->completion_lock);
7594 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7596 io_commit_cqring(ctx);
7597 spin_unlock_irq(&ctx->completion_lock);
7598 io_cqring_ev_posted(ctx);
7599 io_ring_submit_unlock(ctx, lock_ring);
7602 rsrc_data->do_put(ctx, prsrc);
7606 io_rsrc_node_destroy(ref_node);
7607 if (atomic_dec_and_test(&rsrc_data->refs))
7608 complete(&rsrc_data->done);
7611 static void io_rsrc_put_work(struct work_struct *work)
7613 struct io_ring_ctx *ctx;
7614 struct llist_node *node;
7616 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7617 node = llist_del_all(&ctx->rsrc_put_llist);
7620 struct io_rsrc_node *ref_node;
7621 struct llist_node *next = node->next;
7623 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7624 __io_rsrc_put_work(ref_node);
7629 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7631 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7632 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7633 bool first_add = false;
7635 io_rsrc_ref_lock(ctx);
7638 while (!list_empty(&ctx->rsrc_ref_list)) {
7639 node = list_first_entry(&ctx->rsrc_ref_list,
7640 struct io_rsrc_node, node);
7641 /* recycle ref nodes in order */
7644 list_del(&node->node);
7645 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7647 io_rsrc_ref_unlock(ctx);
7650 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7653 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7655 struct io_rsrc_node *ref_node;
7657 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7661 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7666 INIT_LIST_HEAD(&ref_node->node);
7667 INIT_LIST_HEAD(&ref_node->rsrc_list);
7668 ref_node->done = false;
7672 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7673 unsigned nr_args, u64 __user *tags)
7675 __s32 __user *fds = (__s32 __user *) arg;
7684 if (nr_args > IORING_MAX_FIXED_FILES)
7686 ret = io_rsrc_node_switch_start(ctx);
7689 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7695 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7698 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7699 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7703 /* allow sparse sets */
7706 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7713 if (unlikely(!file))
7717 * Don't allow io_uring instances to be registered. If UNIX
7718 * isn't enabled, then this causes a reference cycle and this
7719 * instance can never get freed. If UNIX is enabled we'll
7720 * handle it just fine, but there's still no point in allowing
7721 * a ring fd as it doesn't support regular read/write anyway.
7723 if (file->f_op == &io_uring_fops) {
7727 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7730 ret = io_sqe_files_scm(ctx);
7732 __io_sqe_files_unregister(ctx);
7736 io_rsrc_node_switch(ctx, NULL);
7739 for (i = 0; i < ctx->nr_user_files; i++) {
7740 file = io_file_from_index(ctx, i);
7744 io_free_file_tables(&ctx->file_table, nr_args);
7745 ctx->nr_user_files = 0;
7747 io_rsrc_data_free(ctx->file_data);
7748 ctx->file_data = NULL;
7752 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7755 #if defined(CONFIG_UNIX)
7756 struct sock *sock = ctx->ring_sock->sk;
7757 struct sk_buff_head *head = &sock->sk_receive_queue;
7758 struct sk_buff *skb;
7761 * See if we can merge this file into an existing skb SCM_RIGHTS
7762 * file set. If there's no room, fall back to allocating a new skb
7763 * and filling it in.
7765 spin_lock_irq(&head->lock);
7766 skb = skb_peek(head);
7768 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7770 if (fpl->count < SCM_MAX_FD) {
7771 __skb_unlink(skb, head);
7772 spin_unlock_irq(&head->lock);
7773 fpl->fp[fpl->count] = get_file(file);
7774 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7776 spin_lock_irq(&head->lock);
7777 __skb_queue_head(head, skb);
7782 spin_unlock_irq(&head->lock);
7789 return __io_sqe_files_scm(ctx, 1, index);
7795 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7796 struct io_rsrc_node *node, void *rsrc)
7798 struct io_rsrc_put *prsrc;
7800 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7804 prsrc->tag = *io_get_tag_slot(data, idx);
7806 list_add(&prsrc->list, &node->rsrc_list);
7810 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7811 struct io_uring_rsrc_update2 *up,
7814 u64 __user *tags = u64_to_user_ptr(up->tags);
7815 __s32 __user *fds = u64_to_user_ptr(up->data);
7816 struct io_rsrc_data *data = ctx->file_data;
7817 struct io_fixed_file *file_slot;
7821 bool needs_switch = false;
7823 if (!ctx->file_data)
7825 if (up->offset + nr_args > ctx->nr_user_files)
7828 for (done = 0; done < nr_args; done++) {
7831 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7832 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7836 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7840 if (fd == IORING_REGISTER_FILES_SKIP)
7843 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7844 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7846 if (file_slot->file_ptr) {
7847 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7848 err = io_queue_rsrc_removal(data, up->offset + done,
7849 ctx->rsrc_node, file);
7852 file_slot->file_ptr = 0;
7853 needs_switch = true;
7862 * Don't allow io_uring instances to be registered. If
7863 * UNIX isn't enabled, then this causes a reference
7864 * cycle and this instance can never get freed. If UNIX
7865 * is enabled we'll handle it just fine, but there's
7866 * still no point in allowing a ring fd as it doesn't
7867 * support regular read/write anyway.
7869 if (file->f_op == &io_uring_fops) {
7874 *io_get_tag_slot(data, up->offset + done) = tag;
7875 io_fixed_file_set(file_slot, file);
7876 err = io_sqe_file_register(ctx, file, i);
7878 file_slot->file_ptr = 0;
7886 io_rsrc_node_switch(ctx, data);
7887 return done ? done : err;
7890 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7892 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7894 req = io_put_req_find_next(req);
7895 return req ? &req->work : NULL;
7898 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7899 struct task_struct *task)
7901 struct io_wq_hash *hash;
7902 struct io_wq_data data;
7903 unsigned int concurrency;
7905 hash = ctx->hash_map;
7907 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7909 return ERR_PTR(-ENOMEM);
7910 refcount_set(&hash->refs, 1);
7911 init_waitqueue_head(&hash->wait);
7912 ctx->hash_map = hash;
7917 data.free_work = io_free_work;
7918 data.do_work = io_wq_submit_work;
7920 /* Do QD, or 4 * CPUS, whatever is smallest */
7921 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7923 return io_wq_create(concurrency, &data);
7926 static int io_uring_alloc_task_context(struct task_struct *task,
7927 struct io_ring_ctx *ctx)
7929 struct io_uring_task *tctx;
7932 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7933 if (unlikely(!tctx))
7936 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7937 if (unlikely(ret)) {
7942 tctx->io_wq = io_init_wq_offload(ctx, task);
7943 if (IS_ERR(tctx->io_wq)) {
7944 ret = PTR_ERR(tctx->io_wq);
7945 percpu_counter_destroy(&tctx->inflight);
7951 init_waitqueue_head(&tctx->wait);
7953 atomic_set(&tctx->in_idle, 0);
7954 atomic_set(&tctx->inflight_tracked, 0);
7955 task->io_uring = tctx;
7956 spin_lock_init(&tctx->task_lock);
7957 INIT_WQ_LIST(&tctx->task_list);
7958 tctx->task_state = 0;
7959 init_task_work(&tctx->task_work, tctx_task_work);
7963 void __io_uring_free(struct task_struct *tsk)
7965 struct io_uring_task *tctx = tsk->io_uring;
7967 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7968 WARN_ON_ONCE(tctx->io_wq);
7970 percpu_counter_destroy(&tctx->inflight);
7972 tsk->io_uring = NULL;
7975 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7976 struct io_uring_params *p)
7980 /* Retain compatibility with failing for an invalid attach attempt */
7981 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7982 IORING_SETUP_ATTACH_WQ) {
7985 f = fdget(p->wq_fd);
7989 if (f.file->f_op != &io_uring_fops)
7992 if (ctx->flags & IORING_SETUP_SQPOLL) {
7993 struct task_struct *tsk;
7994 struct io_sq_data *sqd;
7997 sqd = io_get_sq_data(p, &attached);
8003 ctx->sq_creds = get_current_cred();
8005 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8006 if (!ctx->sq_thread_idle)
8007 ctx->sq_thread_idle = HZ;
8009 io_sq_thread_park(sqd);
8010 list_add(&ctx->sqd_list, &sqd->ctx_list);
8011 io_sqd_update_thread_idle(sqd);
8012 /* don't attach to a dying SQPOLL thread, would be racy */
8013 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8014 io_sq_thread_unpark(sqd);
8021 if (p->flags & IORING_SETUP_SQ_AFF) {
8022 int cpu = p->sq_thread_cpu;
8025 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8032 sqd->task_pid = current->pid;
8033 sqd->task_tgid = current->tgid;
8034 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8041 ret = io_uring_alloc_task_context(tsk, ctx);
8042 wake_up_new_task(tsk);
8045 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8046 /* Can't have SQ_AFF without SQPOLL */
8053 complete(&ctx->sq_data->exited);
8055 io_sq_thread_finish(ctx);
8059 static inline void __io_unaccount_mem(struct user_struct *user,
8060 unsigned long nr_pages)
8062 atomic_long_sub(nr_pages, &user->locked_vm);
8065 static inline int __io_account_mem(struct user_struct *user,
8066 unsigned long nr_pages)
8068 unsigned long page_limit, cur_pages, new_pages;
8070 /* Don't allow more pages than we can safely lock */
8071 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8074 cur_pages = atomic_long_read(&user->locked_vm);
8075 new_pages = cur_pages + nr_pages;
8076 if (new_pages > page_limit)
8078 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8079 new_pages) != cur_pages);
8084 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8087 __io_unaccount_mem(ctx->user, nr_pages);
8089 if (ctx->mm_account)
8090 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8093 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8098 ret = __io_account_mem(ctx->user, nr_pages);
8103 if (ctx->mm_account)
8104 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8109 static void io_mem_free(void *ptr)
8116 page = virt_to_head_page(ptr);
8117 if (put_page_testzero(page))
8118 free_compound_page(page);
8121 static void *io_mem_alloc(size_t size)
8123 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8124 __GFP_NORETRY | __GFP_ACCOUNT;
8126 return (void *) __get_free_pages(gfp_flags, get_order(size));
8129 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8132 struct io_rings *rings;
8133 size_t off, sq_array_size;
8135 off = struct_size(rings, cqes, cq_entries);
8136 if (off == SIZE_MAX)
8140 off = ALIGN(off, SMP_CACHE_BYTES);
8148 sq_array_size = array_size(sizeof(u32), sq_entries);
8149 if (sq_array_size == SIZE_MAX)
8152 if (check_add_overflow(off, sq_array_size, &off))
8158 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8160 struct io_mapped_ubuf *imu = *slot;
8163 if (imu != ctx->dummy_ubuf) {
8164 for (i = 0; i < imu->nr_bvecs; i++)
8165 unpin_user_page(imu->bvec[i].bv_page);
8166 if (imu->acct_pages)
8167 io_unaccount_mem(ctx, imu->acct_pages);
8173 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8175 io_buffer_unmap(ctx, &prsrc->buf);
8179 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8183 for (i = 0; i < ctx->nr_user_bufs; i++)
8184 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8185 kfree(ctx->user_bufs);
8186 io_rsrc_data_free(ctx->buf_data);
8187 ctx->user_bufs = NULL;
8188 ctx->buf_data = NULL;
8189 ctx->nr_user_bufs = 0;
8192 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8199 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8201 __io_sqe_buffers_unregister(ctx);
8205 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8206 void __user *arg, unsigned index)
8208 struct iovec __user *src;
8210 #ifdef CONFIG_COMPAT
8212 struct compat_iovec __user *ciovs;
8213 struct compat_iovec ciov;
8215 ciovs = (struct compat_iovec __user *) arg;
8216 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8219 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8220 dst->iov_len = ciov.iov_len;
8224 src = (struct iovec __user *) arg;
8225 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8231 * Not super efficient, but this is just a registration time. And we do cache
8232 * the last compound head, so generally we'll only do a full search if we don't
8235 * We check if the given compound head page has already been accounted, to
8236 * avoid double accounting it. This allows us to account the full size of the
8237 * page, not just the constituent pages of a huge page.
8239 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8240 int nr_pages, struct page *hpage)
8244 /* check current page array */
8245 for (i = 0; i < nr_pages; i++) {
8246 if (!PageCompound(pages[i]))
8248 if (compound_head(pages[i]) == hpage)
8252 /* check previously registered pages */
8253 for (i = 0; i < ctx->nr_user_bufs; i++) {
8254 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8256 for (j = 0; j < imu->nr_bvecs; j++) {
8257 if (!PageCompound(imu->bvec[j].bv_page))
8259 if (compound_head(imu->bvec[j].bv_page) == hpage)
8267 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8268 int nr_pages, struct io_mapped_ubuf *imu,
8269 struct page **last_hpage)
8273 imu->acct_pages = 0;
8274 for (i = 0; i < nr_pages; i++) {
8275 if (!PageCompound(pages[i])) {
8280 hpage = compound_head(pages[i]);
8281 if (hpage == *last_hpage)
8283 *last_hpage = hpage;
8284 if (headpage_already_acct(ctx, pages, i, hpage))
8286 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8290 if (!imu->acct_pages)
8293 ret = io_account_mem(ctx, imu->acct_pages);
8295 imu->acct_pages = 0;
8299 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8300 struct io_mapped_ubuf **pimu,
8301 struct page **last_hpage)
8303 struct io_mapped_ubuf *imu = NULL;
8304 struct vm_area_struct **vmas = NULL;
8305 struct page **pages = NULL;
8306 unsigned long off, start, end, ubuf;
8308 int ret, pret, nr_pages, i;
8310 if (!iov->iov_base) {
8311 *pimu = ctx->dummy_ubuf;
8315 ubuf = (unsigned long) iov->iov_base;
8316 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8317 start = ubuf >> PAGE_SHIFT;
8318 nr_pages = end - start;
8323 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8327 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8332 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8337 mmap_read_lock(current->mm);
8338 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8340 if (pret == nr_pages) {
8341 /* don't support file backed memory */
8342 for (i = 0; i < nr_pages; i++) {
8343 struct vm_area_struct *vma = vmas[i];
8345 if (vma_is_shmem(vma))
8348 !is_file_hugepages(vma->vm_file)) {
8354 ret = pret < 0 ? pret : -EFAULT;
8356 mmap_read_unlock(current->mm);
8359 * if we did partial map, or found file backed vmas,
8360 * release any pages we did get
8363 unpin_user_pages(pages, pret);
8367 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8369 unpin_user_pages(pages, pret);
8373 off = ubuf & ~PAGE_MASK;
8374 size = iov->iov_len;
8375 for (i = 0; i < nr_pages; i++) {
8378 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8379 imu->bvec[i].bv_page = pages[i];
8380 imu->bvec[i].bv_len = vec_len;
8381 imu->bvec[i].bv_offset = off;
8385 /* store original address for later verification */
8387 imu->ubuf_end = ubuf + iov->iov_len;
8388 imu->nr_bvecs = nr_pages;
8399 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8401 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8402 return ctx->user_bufs ? 0 : -ENOMEM;
8405 static int io_buffer_validate(struct iovec *iov)
8407 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8410 * Don't impose further limits on the size and buffer
8411 * constraints here, we'll -EINVAL later when IO is
8412 * submitted if they are wrong.
8415 return iov->iov_len ? -EFAULT : 0;
8419 /* arbitrary limit, but we need something */
8420 if (iov->iov_len > SZ_1G)
8423 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8429 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8430 unsigned int nr_args, u64 __user *tags)
8432 struct page *last_hpage = NULL;
8433 struct io_rsrc_data *data;
8439 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8441 ret = io_rsrc_node_switch_start(ctx);
8444 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8447 ret = io_buffers_map_alloc(ctx, nr_args);
8449 io_rsrc_data_free(data);
8453 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8454 ret = io_copy_iov(ctx, &iov, arg, i);
8457 ret = io_buffer_validate(&iov);
8460 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8465 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8471 WARN_ON_ONCE(ctx->buf_data);
8473 ctx->buf_data = data;
8475 __io_sqe_buffers_unregister(ctx);
8477 io_rsrc_node_switch(ctx, NULL);
8481 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8482 struct io_uring_rsrc_update2 *up,
8483 unsigned int nr_args)
8485 u64 __user *tags = u64_to_user_ptr(up->tags);
8486 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8487 struct page *last_hpage = NULL;
8488 bool needs_switch = false;
8494 if (up->offset + nr_args > ctx->nr_user_bufs)
8497 for (done = 0; done < nr_args; done++) {
8498 struct io_mapped_ubuf *imu;
8499 int offset = up->offset + done;
8502 err = io_copy_iov(ctx, &iov, iovs, done);
8505 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8509 err = io_buffer_validate(&iov);
8512 if (!iov.iov_base && tag) {
8516 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8520 i = array_index_nospec(offset, ctx->nr_user_bufs);
8521 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8522 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8523 ctx->rsrc_node, ctx->user_bufs[i]);
8524 if (unlikely(err)) {
8525 io_buffer_unmap(ctx, &imu);
8528 ctx->user_bufs[i] = NULL;
8529 needs_switch = true;
8532 ctx->user_bufs[i] = imu;
8533 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8537 io_rsrc_node_switch(ctx, ctx->buf_data);
8538 return done ? done : err;
8541 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8543 __s32 __user *fds = arg;
8549 if (copy_from_user(&fd, fds, sizeof(*fds)))
8552 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8553 if (IS_ERR(ctx->cq_ev_fd)) {
8554 int ret = PTR_ERR(ctx->cq_ev_fd);
8555 ctx->cq_ev_fd = NULL;
8562 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8564 if (ctx->cq_ev_fd) {
8565 eventfd_ctx_put(ctx->cq_ev_fd);
8566 ctx->cq_ev_fd = NULL;
8573 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8575 struct io_buffer *buf;
8576 unsigned long index;
8578 xa_for_each(&ctx->io_buffers, index, buf)
8579 __io_remove_buffers(ctx, buf, index, -1U);
8582 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8584 struct io_kiocb *req, *nxt;
8586 list_for_each_entry_safe(req, nxt, list, compl.list) {
8587 if (tsk && req->task != tsk)
8589 list_del(&req->compl.list);
8590 kmem_cache_free(req_cachep, req);
8594 static void io_req_caches_free(struct io_ring_ctx *ctx)
8596 struct io_submit_state *submit_state = &ctx->submit_state;
8597 struct io_comp_state *cs = &ctx->submit_state.comp;
8599 mutex_lock(&ctx->uring_lock);
8601 if (submit_state->free_reqs) {
8602 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8603 submit_state->reqs);
8604 submit_state->free_reqs = 0;
8607 io_flush_cached_locked_reqs(ctx, cs);
8608 io_req_cache_free(&cs->free_list, NULL);
8609 mutex_unlock(&ctx->uring_lock);
8612 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8616 if (!atomic_dec_and_test(&data->refs))
8617 wait_for_completion(&data->done);
8621 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8623 io_sq_thread_finish(ctx);
8625 if (ctx->mm_account) {
8626 mmdrop(ctx->mm_account);
8627 ctx->mm_account = NULL;
8630 mutex_lock(&ctx->uring_lock);
8631 if (io_wait_rsrc_data(ctx->buf_data))
8632 __io_sqe_buffers_unregister(ctx);
8633 if (io_wait_rsrc_data(ctx->file_data))
8634 __io_sqe_files_unregister(ctx);
8636 __io_cqring_overflow_flush(ctx, true);
8637 mutex_unlock(&ctx->uring_lock);
8638 io_eventfd_unregister(ctx);
8639 io_destroy_buffers(ctx);
8641 put_cred(ctx->sq_creds);
8643 /* there are no registered resources left, nobody uses it */
8645 io_rsrc_node_destroy(ctx->rsrc_node);
8646 if (ctx->rsrc_backup_node)
8647 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8648 flush_delayed_work(&ctx->rsrc_put_work);
8650 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8651 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8653 #if defined(CONFIG_UNIX)
8654 if (ctx->ring_sock) {
8655 ctx->ring_sock->file = NULL; /* so that iput() is called */
8656 sock_release(ctx->ring_sock);
8660 io_mem_free(ctx->rings);
8661 io_mem_free(ctx->sq_sqes);
8663 percpu_ref_exit(&ctx->refs);
8664 free_uid(ctx->user);
8665 io_req_caches_free(ctx);
8667 io_wq_put_hash(ctx->hash_map);
8668 kfree(ctx->cancel_hash);
8669 kfree(ctx->dummy_ubuf);
8673 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8675 struct io_ring_ctx *ctx = file->private_data;
8678 poll_wait(file, &ctx->cq_wait, wait);
8680 * synchronizes with barrier from wq_has_sleeper call in
8684 if (!io_sqring_full(ctx))
8685 mask |= EPOLLOUT | EPOLLWRNORM;
8688 * Don't flush cqring overflow list here, just do a simple check.
8689 * Otherwise there could possible be ABBA deadlock:
8692 * lock(&ctx->uring_lock);
8694 * lock(&ctx->uring_lock);
8697 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8698 * pushs them to do the flush.
8700 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8701 mask |= EPOLLIN | EPOLLRDNORM;
8706 static int io_uring_fasync(int fd, struct file *file, int on)
8708 struct io_ring_ctx *ctx = file->private_data;
8710 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8713 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8715 const struct cred *creds;
8717 creds = xa_erase(&ctx->personalities, id);
8726 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8728 return io_run_task_work_head(&ctx->exit_task_work);
8731 struct io_tctx_exit {
8732 struct callback_head task_work;
8733 struct completion completion;
8734 struct io_ring_ctx *ctx;
8737 static void io_tctx_exit_cb(struct callback_head *cb)
8739 struct io_uring_task *tctx = current->io_uring;
8740 struct io_tctx_exit *work;
8742 work = container_of(cb, struct io_tctx_exit, task_work);
8744 * When @in_idle, we're in cancellation and it's racy to remove the
8745 * node. It'll be removed by the end of cancellation, just ignore it.
8747 if (!atomic_read(&tctx->in_idle))
8748 io_uring_del_tctx_node((unsigned long)work->ctx);
8749 complete(&work->completion);
8752 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8754 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8756 return req->ctx == data;
8759 static void io_ring_exit_work(struct work_struct *work)
8761 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8762 unsigned long timeout = jiffies + HZ * 60 * 5;
8763 struct io_tctx_exit exit;
8764 struct io_tctx_node *node;
8768 * If we're doing polled IO and end up having requests being
8769 * submitted async (out-of-line), then completions can come in while
8770 * we're waiting for refs to drop. We need to reap these manually,
8771 * as nobody else will be looking for them.
8774 io_uring_try_cancel_requests(ctx, NULL, true);
8776 struct io_sq_data *sqd = ctx->sq_data;
8777 struct task_struct *tsk;
8779 io_sq_thread_park(sqd);
8781 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8782 io_wq_cancel_cb(tsk->io_uring->io_wq,
8783 io_cancel_ctx_cb, ctx, true);
8784 io_sq_thread_unpark(sqd);
8787 WARN_ON_ONCE(time_after(jiffies, timeout));
8788 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8790 init_completion(&exit.completion);
8791 init_task_work(&exit.task_work, io_tctx_exit_cb);
8794 * Some may use context even when all refs and requests have been put,
8795 * and they are free to do so while still holding uring_lock or
8796 * completion_lock, see __io_req_task_submit(). Apart from other work,
8797 * this lock/unlock section also waits them to finish.
8799 mutex_lock(&ctx->uring_lock);
8800 while (!list_empty(&ctx->tctx_list)) {
8801 WARN_ON_ONCE(time_after(jiffies, timeout));
8803 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8805 /* don't spin on a single task if cancellation failed */
8806 list_rotate_left(&ctx->tctx_list);
8807 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8808 if (WARN_ON_ONCE(ret))
8810 wake_up_process(node->task);
8812 mutex_unlock(&ctx->uring_lock);
8813 wait_for_completion(&exit.completion);
8814 mutex_lock(&ctx->uring_lock);
8816 mutex_unlock(&ctx->uring_lock);
8817 spin_lock_irq(&ctx->completion_lock);
8818 spin_unlock_irq(&ctx->completion_lock);
8820 io_ring_ctx_free(ctx);
8823 /* Returns true if we found and killed one or more timeouts */
8824 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8827 struct io_kiocb *req, *tmp;
8830 spin_lock_irq(&ctx->completion_lock);
8831 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8832 if (io_match_task(req, tsk, cancel_all)) {
8833 io_kill_timeout(req, -ECANCELED);
8838 io_commit_cqring(ctx);
8839 spin_unlock_irq(&ctx->completion_lock);
8841 io_cqring_ev_posted(ctx);
8842 return canceled != 0;
8845 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8847 unsigned long index;
8848 struct creds *creds;
8850 mutex_lock(&ctx->uring_lock);
8851 percpu_ref_kill(&ctx->refs);
8853 __io_cqring_overflow_flush(ctx, true);
8854 xa_for_each(&ctx->personalities, index, creds)
8855 io_unregister_personality(ctx, index);
8856 mutex_unlock(&ctx->uring_lock);
8858 io_kill_timeouts(ctx, NULL, true);
8859 io_poll_remove_all(ctx, NULL, true);
8861 /* if we failed setting up the ctx, we might not have any rings */
8862 io_iopoll_try_reap_events(ctx);
8864 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8866 * Use system_unbound_wq to avoid spawning tons of event kworkers
8867 * if we're exiting a ton of rings at the same time. It just adds
8868 * noise and overhead, there's no discernable change in runtime
8869 * over using system_wq.
8871 queue_work(system_unbound_wq, &ctx->exit_work);
8874 static int io_uring_release(struct inode *inode, struct file *file)
8876 struct io_ring_ctx *ctx = file->private_data;
8878 file->private_data = NULL;
8879 io_ring_ctx_wait_and_kill(ctx);
8883 struct io_task_cancel {
8884 struct task_struct *task;
8888 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8890 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8891 struct io_task_cancel *cancel = data;
8894 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8895 unsigned long flags;
8896 struct io_ring_ctx *ctx = req->ctx;
8898 /* protect against races with linked timeouts */
8899 spin_lock_irqsave(&ctx->completion_lock, flags);
8900 ret = io_match_task(req, cancel->task, cancel->all);
8901 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8903 ret = io_match_task(req, cancel->task, cancel->all);
8908 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8909 struct task_struct *task, bool cancel_all)
8911 struct io_defer_entry *de;
8914 spin_lock_irq(&ctx->completion_lock);
8915 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8916 if (io_match_task(de->req, task, cancel_all)) {
8917 list_cut_position(&list, &ctx->defer_list, &de->list);
8921 spin_unlock_irq(&ctx->completion_lock);
8922 if (list_empty(&list))
8925 while (!list_empty(&list)) {
8926 de = list_first_entry(&list, struct io_defer_entry, list);
8927 list_del_init(&de->list);
8928 io_req_complete_failed(de->req, -ECANCELED);
8934 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8936 struct io_tctx_node *node;
8937 enum io_wq_cancel cret;
8940 mutex_lock(&ctx->uring_lock);
8941 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8942 struct io_uring_task *tctx = node->task->io_uring;
8945 * io_wq will stay alive while we hold uring_lock, because it's
8946 * killed after ctx nodes, which requires to take the lock.
8948 if (!tctx || !tctx->io_wq)
8950 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8951 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8953 mutex_unlock(&ctx->uring_lock);
8958 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8959 struct task_struct *task,
8962 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8963 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8966 enum io_wq_cancel cret;
8970 ret |= io_uring_try_cancel_iowq(ctx);
8971 } else if (tctx && tctx->io_wq) {
8973 * Cancels requests of all rings, not only @ctx, but
8974 * it's fine as the task is in exit/exec.
8976 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8978 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8981 /* SQPOLL thread does its own polling */
8982 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8983 (ctx->sq_data && ctx->sq_data->thread == current)) {
8984 while (!list_empty_careful(&ctx->iopoll_list)) {
8985 io_iopoll_try_reap_events(ctx);
8990 ret |= io_cancel_defer_files(ctx, task, cancel_all);
8991 ret |= io_poll_remove_all(ctx, task, cancel_all);
8992 ret |= io_kill_timeouts(ctx, task, cancel_all);
8993 ret |= io_run_task_work();
8994 ret |= io_run_ctx_fallback(ctx);
9001 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9003 struct io_uring_task *tctx = current->io_uring;
9004 struct io_tctx_node *node;
9007 if (unlikely(!tctx)) {
9008 ret = io_uring_alloc_task_context(current, ctx);
9011 tctx = current->io_uring;
9013 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9014 node = kmalloc(sizeof(*node), GFP_KERNEL);
9018 node->task = current;
9020 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9027 mutex_lock(&ctx->uring_lock);
9028 list_add(&node->ctx_node, &ctx->tctx_list);
9029 mutex_unlock(&ctx->uring_lock);
9036 * Note that this task has used io_uring. We use it for cancelation purposes.
9038 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9040 struct io_uring_task *tctx = current->io_uring;
9042 if (likely(tctx && tctx->last == ctx))
9044 return __io_uring_add_tctx_node(ctx);
9048 * Remove this io_uring_file -> task mapping.
9050 static void io_uring_del_tctx_node(unsigned long index)
9052 struct io_uring_task *tctx = current->io_uring;
9053 struct io_tctx_node *node;
9057 node = xa_erase(&tctx->xa, index);
9061 WARN_ON_ONCE(current != node->task);
9062 WARN_ON_ONCE(list_empty(&node->ctx_node));
9064 mutex_lock(&node->ctx->uring_lock);
9065 list_del(&node->ctx_node);
9066 mutex_unlock(&node->ctx->uring_lock);
9068 if (tctx->last == node->ctx)
9073 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9075 struct io_wq *wq = tctx->io_wq;
9076 struct io_tctx_node *node;
9077 unsigned long index;
9079 xa_for_each(&tctx->xa, index, node)
9080 io_uring_del_tctx_node(index);
9083 * Must be after io_uring_del_task_file() (removes nodes under
9084 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9087 io_wq_put_and_exit(wq);
9091 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9094 return atomic_read(&tctx->inflight_tracked);
9095 return percpu_counter_sum(&tctx->inflight);
9098 static void io_uring_try_cancel(bool cancel_all)
9100 struct io_uring_task *tctx = current->io_uring;
9101 struct io_tctx_node *node;
9102 unsigned long index;
9104 xa_for_each(&tctx->xa, index, node) {
9105 struct io_ring_ctx *ctx = node->ctx;
9107 /* sqpoll task will cancel all its requests */
9109 io_uring_try_cancel_requests(ctx, current, cancel_all);
9113 /* should only be called by SQPOLL task */
9114 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd)
9116 struct io_uring_task *tctx = current->io_uring;
9117 struct io_ring_ctx *ctx;
9121 if (!current->io_uring)
9124 io_wq_exit_start(tctx->io_wq);
9126 WARN_ON_ONCE(!sqd || sqd->thread != current);
9128 atomic_inc(&tctx->in_idle);
9130 /* read completions before cancelations */
9131 inflight = tctx_inflight(tctx, false);
9134 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9135 io_uring_try_cancel_requests(ctx, current, true);
9137 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9139 * If we've seen completions, retry without waiting. This
9140 * avoids a race where a completion comes in before we did
9141 * prepare_to_wait().
9143 if (inflight == tctx_inflight(tctx, false))
9145 finish_wait(&tctx->wait, &wait);
9147 atomic_dec(&tctx->in_idle);
9151 * Find any io_uring fd that this task has registered or done IO on, and cancel
9154 void __io_uring_cancel(struct files_struct *files)
9156 struct io_uring_task *tctx = current->io_uring;
9159 bool cancel_all = !files;
9162 io_wq_exit_start(tctx->io_wq);
9164 /* make sure overflow events are dropped */
9165 atomic_inc(&tctx->in_idle);
9167 /* read completions before cancelations */
9168 inflight = tctx_inflight(tctx, !cancel_all);
9171 io_uring_try_cancel(cancel_all);
9172 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9175 * If we've seen completions, retry without waiting. This
9176 * avoids a race where a completion comes in before we did
9177 * prepare_to_wait().
9179 if (inflight == tctx_inflight(tctx, !cancel_all))
9181 finish_wait(&tctx->wait, &wait);
9183 atomic_dec(&tctx->in_idle);
9185 io_uring_clean_tctx(tctx);
9187 /* for exec all current's requests should be gone, kill tctx */
9188 __io_uring_free(current);
9192 static void *io_uring_validate_mmap_request(struct file *file,
9193 loff_t pgoff, size_t sz)
9195 struct io_ring_ctx *ctx = file->private_data;
9196 loff_t offset = pgoff << PAGE_SHIFT;
9201 case IORING_OFF_SQ_RING:
9202 case IORING_OFF_CQ_RING:
9205 case IORING_OFF_SQES:
9209 return ERR_PTR(-EINVAL);
9212 page = virt_to_head_page(ptr);
9213 if (sz > page_size(page))
9214 return ERR_PTR(-EINVAL);
9221 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9223 size_t sz = vma->vm_end - vma->vm_start;
9227 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9229 return PTR_ERR(ptr);
9231 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9232 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9235 #else /* !CONFIG_MMU */
9237 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9239 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9242 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9244 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9247 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9248 unsigned long addr, unsigned long len,
9249 unsigned long pgoff, unsigned long flags)
9253 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9255 return PTR_ERR(ptr);
9257 return (unsigned long) ptr;
9260 #endif /* !CONFIG_MMU */
9262 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9267 if (!io_sqring_full(ctx))
9269 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9271 if (!io_sqring_full(ctx))
9274 } while (!signal_pending(current));
9276 finish_wait(&ctx->sqo_sq_wait, &wait);
9280 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9281 struct __kernel_timespec __user **ts,
9282 const sigset_t __user **sig)
9284 struct io_uring_getevents_arg arg;
9287 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9288 * is just a pointer to the sigset_t.
9290 if (!(flags & IORING_ENTER_EXT_ARG)) {
9291 *sig = (const sigset_t __user *) argp;
9297 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9298 * timespec and sigset_t pointers if good.
9300 if (*argsz != sizeof(arg))
9302 if (copy_from_user(&arg, argp, sizeof(arg)))
9304 *sig = u64_to_user_ptr(arg.sigmask);
9305 *argsz = arg.sigmask_sz;
9306 *ts = u64_to_user_ptr(arg.ts);
9310 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9311 u32, min_complete, u32, flags, const void __user *, argp,
9314 struct io_ring_ctx *ctx;
9321 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9322 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9326 if (unlikely(!f.file))
9330 if (unlikely(f.file->f_op != &io_uring_fops))
9334 ctx = f.file->private_data;
9335 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9339 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9343 * For SQ polling, the thread will do all submissions and completions.
9344 * Just return the requested submit count, and wake the thread if
9348 if (ctx->flags & IORING_SETUP_SQPOLL) {
9349 io_cqring_overflow_flush(ctx, false);
9352 if (unlikely(ctx->sq_data->thread == NULL)) {
9355 if (flags & IORING_ENTER_SQ_WAKEUP)
9356 wake_up(&ctx->sq_data->wait);
9357 if (flags & IORING_ENTER_SQ_WAIT) {
9358 ret = io_sqpoll_wait_sq(ctx);
9362 submitted = to_submit;
9363 } else if (to_submit) {
9364 ret = io_uring_add_tctx_node(ctx);
9367 mutex_lock(&ctx->uring_lock);
9368 submitted = io_submit_sqes(ctx, to_submit);
9369 mutex_unlock(&ctx->uring_lock);
9371 if (submitted != to_submit)
9374 if (flags & IORING_ENTER_GETEVENTS) {
9375 const sigset_t __user *sig;
9376 struct __kernel_timespec __user *ts;
9378 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9382 min_complete = min(min_complete, ctx->cq_entries);
9385 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9386 * space applications don't need to do io completion events
9387 * polling again, they can rely on io_sq_thread to do polling
9388 * work, which can reduce cpu usage and uring_lock contention.
9390 if (ctx->flags & IORING_SETUP_IOPOLL &&
9391 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9392 ret = io_iopoll_check(ctx, min_complete);
9394 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9399 percpu_ref_put(&ctx->refs);
9402 return submitted ? submitted : ret;
9405 #ifdef CONFIG_PROC_FS
9406 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9407 const struct cred *cred)
9409 struct user_namespace *uns = seq_user_ns(m);
9410 struct group_info *gi;
9415 seq_printf(m, "%5d\n", id);
9416 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9417 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9418 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9419 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9420 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9421 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9422 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9423 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9424 seq_puts(m, "\n\tGroups:\t");
9425 gi = cred->group_info;
9426 for (g = 0; g < gi->ngroups; g++) {
9427 seq_put_decimal_ull(m, g ? " " : "",
9428 from_kgid_munged(uns, gi->gid[g]));
9430 seq_puts(m, "\n\tCapEff:\t");
9431 cap = cred->cap_effective;
9432 CAP_FOR_EACH_U32(__capi)
9433 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9438 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9440 struct io_sq_data *sq = NULL;
9445 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9446 * since fdinfo case grabs it in the opposite direction of normal use
9447 * cases. If we fail to get the lock, we just don't iterate any
9448 * structures that could be going away outside the io_uring mutex.
9450 has_lock = mutex_trylock(&ctx->uring_lock);
9452 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9458 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9459 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9460 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9461 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9462 struct file *f = io_file_from_index(ctx, i);
9465 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9467 seq_printf(m, "%5u: <none>\n", i);
9469 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9470 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9471 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9472 unsigned int len = buf->ubuf_end - buf->ubuf;
9474 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9476 if (has_lock && !xa_empty(&ctx->personalities)) {
9477 unsigned long index;
9478 const struct cred *cred;
9480 seq_printf(m, "Personalities:\n");
9481 xa_for_each(&ctx->personalities, index, cred)
9482 io_uring_show_cred(m, index, cred);
9484 seq_printf(m, "PollList:\n");
9485 spin_lock_irq(&ctx->completion_lock);
9486 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9487 struct hlist_head *list = &ctx->cancel_hash[i];
9488 struct io_kiocb *req;
9490 hlist_for_each_entry(req, list, hash_node)
9491 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9492 req->task->task_works != NULL);
9494 spin_unlock_irq(&ctx->completion_lock);
9496 mutex_unlock(&ctx->uring_lock);
9499 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9501 struct io_ring_ctx *ctx = f->private_data;
9503 if (percpu_ref_tryget(&ctx->refs)) {
9504 __io_uring_show_fdinfo(ctx, m);
9505 percpu_ref_put(&ctx->refs);
9510 static const struct file_operations io_uring_fops = {
9511 .release = io_uring_release,
9512 .mmap = io_uring_mmap,
9514 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9515 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9517 .poll = io_uring_poll,
9518 .fasync = io_uring_fasync,
9519 #ifdef CONFIG_PROC_FS
9520 .show_fdinfo = io_uring_show_fdinfo,
9524 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9525 struct io_uring_params *p)
9527 struct io_rings *rings;
9528 size_t size, sq_array_offset;
9530 /* make sure these are sane, as we already accounted them */
9531 ctx->sq_entries = p->sq_entries;
9532 ctx->cq_entries = p->cq_entries;
9534 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9535 if (size == SIZE_MAX)
9538 rings = io_mem_alloc(size);
9543 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9544 rings->sq_ring_mask = p->sq_entries - 1;
9545 rings->cq_ring_mask = p->cq_entries - 1;
9546 rings->sq_ring_entries = p->sq_entries;
9547 rings->cq_ring_entries = p->cq_entries;
9549 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9550 if (size == SIZE_MAX) {
9551 io_mem_free(ctx->rings);
9556 ctx->sq_sqes = io_mem_alloc(size);
9557 if (!ctx->sq_sqes) {
9558 io_mem_free(ctx->rings);
9566 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9570 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9574 ret = io_uring_add_tctx_node(ctx);
9579 fd_install(fd, file);
9584 * Allocate an anonymous fd, this is what constitutes the application
9585 * visible backing of an io_uring instance. The application mmaps this
9586 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9587 * we have to tie this fd to a socket for file garbage collection purposes.
9589 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9592 #if defined(CONFIG_UNIX)
9595 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9598 return ERR_PTR(ret);
9601 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9602 O_RDWR | O_CLOEXEC);
9603 #if defined(CONFIG_UNIX)
9605 sock_release(ctx->ring_sock);
9606 ctx->ring_sock = NULL;
9608 ctx->ring_sock->file = file;
9614 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9615 struct io_uring_params __user *params)
9617 struct io_ring_ctx *ctx;
9623 if (entries > IORING_MAX_ENTRIES) {
9624 if (!(p->flags & IORING_SETUP_CLAMP))
9626 entries = IORING_MAX_ENTRIES;
9630 * Use twice as many entries for the CQ ring. It's possible for the
9631 * application to drive a higher depth than the size of the SQ ring,
9632 * since the sqes are only used at submission time. This allows for
9633 * some flexibility in overcommitting a bit. If the application has
9634 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9635 * of CQ ring entries manually.
9637 p->sq_entries = roundup_pow_of_two(entries);
9638 if (p->flags & IORING_SETUP_CQSIZE) {
9640 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9641 * to a power-of-two, if it isn't already. We do NOT impose
9642 * any cq vs sq ring sizing.
9646 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9647 if (!(p->flags & IORING_SETUP_CLAMP))
9649 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9651 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9652 if (p->cq_entries < p->sq_entries)
9655 p->cq_entries = 2 * p->sq_entries;
9658 ctx = io_ring_ctx_alloc(p);
9661 ctx->compat = in_compat_syscall();
9662 if (!capable(CAP_IPC_LOCK))
9663 ctx->user = get_uid(current_user());
9666 * This is just grabbed for accounting purposes. When a process exits,
9667 * the mm is exited and dropped before the files, hence we need to hang
9668 * on to this mm purely for the purposes of being able to unaccount
9669 * memory (locked/pinned vm). It's not used for anything else.
9671 mmgrab(current->mm);
9672 ctx->mm_account = current->mm;
9674 ret = io_allocate_scq_urings(ctx, p);
9678 ret = io_sq_offload_create(ctx, p);
9681 /* always set a rsrc node */
9682 ret = io_rsrc_node_switch_start(ctx);
9685 io_rsrc_node_switch(ctx, NULL);
9687 memset(&p->sq_off, 0, sizeof(p->sq_off));
9688 p->sq_off.head = offsetof(struct io_rings, sq.head);
9689 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9690 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9691 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9692 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9693 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9694 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9696 memset(&p->cq_off, 0, sizeof(p->cq_off));
9697 p->cq_off.head = offsetof(struct io_rings, cq.head);
9698 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9699 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9700 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9701 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9702 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9703 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9705 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9706 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9707 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9708 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9709 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9710 IORING_FEAT_RSRC_TAGS;
9712 if (copy_to_user(params, p, sizeof(*p))) {
9717 file = io_uring_get_file(ctx);
9719 ret = PTR_ERR(file);
9724 * Install ring fd as the very last thing, so we don't risk someone
9725 * having closed it before we finish setup
9727 ret = io_uring_install_fd(ctx, file);
9729 /* fput will clean it up */
9734 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9737 io_ring_ctx_wait_and_kill(ctx);
9742 * Sets up an aio uring context, and returns the fd. Applications asks for a
9743 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9744 * params structure passed in.
9746 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9748 struct io_uring_params p;
9751 if (copy_from_user(&p, params, sizeof(p)))
9753 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9758 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9759 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9760 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9761 IORING_SETUP_R_DISABLED))
9764 return io_uring_create(entries, &p, params);
9767 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9768 struct io_uring_params __user *, params)
9770 return io_uring_setup(entries, params);
9773 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9775 struct io_uring_probe *p;
9779 size = struct_size(p, ops, nr_args);
9780 if (size == SIZE_MAX)
9782 p = kzalloc(size, GFP_KERNEL);
9787 if (copy_from_user(p, arg, size))
9790 if (memchr_inv(p, 0, size))
9793 p->last_op = IORING_OP_LAST - 1;
9794 if (nr_args > IORING_OP_LAST)
9795 nr_args = IORING_OP_LAST;
9797 for (i = 0; i < nr_args; i++) {
9799 if (!io_op_defs[i].not_supported)
9800 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9805 if (copy_to_user(arg, p, size))
9812 static int io_register_personality(struct io_ring_ctx *ctx)
9814 const struct cred *creds;
9818 creds = get_current_cred();
9820 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9821 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9828 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9829 unsigned int nr_args)
9831 struct io_uring_restriction *res;
9835 /* Restrictions allowed only if rings started disabled */
9836 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9839 /* We allow only a single restrictions registration */
9840 if (ctx->restrictions.registered)
9843 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9846 size = array_size(nr_args, sizeof(*res));
9847 if (size == SIZE_MAX)
9850 res = memdup_user(arg, size);
9852 return PTR_ERR(res);
9856 for (i = 0; i < nr_args; i++) {
9857 switch (res[i].opcode) {
9858 case IORING_RESTRICTION_REGISTER_OP:
9859 if (res[i].register_op >= IORING_REGISTER_LAST) {
9864 __set_bit(res[i].register_op,
9865 ctx->restrictions.register_op);
9867 case IORING_RESTRICTION_SQE_OP:
9868 if (res[i].sqe_op >= IORING_OP_LAST) {
9873 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9875 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9876 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9878 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9879 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9888 /* Reset all restrictions if an error happened */
9890 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9892 ctx->restrictions.registered = true;
9898 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9900 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9903 if (ctx->restrictions.registered)
9904 ctx->restricted = 1;
9906 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9907 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9908 wake_up(&ctx->sq_data->wait);
9912 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9913 struct io_uring_rsrc_update2 *up,
9921 if (check_add_overflow(up->offset, nr_args, &tmp))
9923 err = io_rsrc_node_switch_start(ctx);
9928 case IORING_RSRC_FILE:
9929 return __io_sqe_files_update(ctx, up, nr_args);
9930 case IORING_RSRC_BUFFER:
9931 return __io_sqe_buffers_update(ctx, up, nr_args);
9936 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9939 struct io_uring_rsrc_update2 up;
9943 memset(&up, 0, sizeof(up));
9944 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9946 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9949 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9950 unsigned size, unsigned type)
9952 struct io_uring_rsrc_update2 up;
9954 if (size != sizeof(up))
9956 if (copy_from_user(&up, arg, sizeof(up)))
9958 if (!up.nr || up.resv)
9960 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9963 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9964 unsigned int size, unsigned int type)
9966 struct io_uring_rsrc_register rr;
9968 /* keep it extendible */
9969 if (size != sizeof(rr))
9972 memset(&rr, 0, sizeof(rr));
9973 if (copy_from_user(&rr, arg, size))
9975 if (!rr.nr || rr.resv || rr.resv2)
9979 case IORING_RSRC_FILE:
9980 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9981 rr.nr, u64_to_user_ptr(rr.tags));
9982 case IORING_RSRC_BUFFER:
9983 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9984 rr.nr, u64_to_user_ptr(rr.tags));
9989 static bool io_register_op_must_quiesce(int op)
9992 case IORING_REGISTER_BUFFERS:
9993 case IORING_UNREGISTER_BUFFERS:
9994 case IORING_REGISTER_FILES:
9995 case IORING_UNREGISTER_FILES:
9996 case IORING_REGISTER_FILES_UPDATE:
9997 case IORING_REGISTER_PROBE:
9998 case IORING_REGISTER_PERSONALITY:
9999 case IORING_UNREGISTER_PERSONALITY:
10000 case IORING_REGISTER_FILES2:
10001 case IORING_REGISTER_FILES_UPDATE2:
10002 case IORING_REGISTER_BUFFERS2:
10003 case IORING_REGISTER_BUFFERS_UPDATE:
10010 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10011 void __user *arg, unsigned nr_args)
10012 __releases(ctx->uring_lock)
10013 __acquires(ctx->uring_lock)
10018 * We're inside the ring mutex, if the ref is already dying, then
10019 * someone else killed the ctx or is already going through
10020 * io_uring_register().
10022 if (percpu_ref_is_dying(&ctx->refs))
10025 if (ctx->restricted) {
10026 if (opcode >= IORING_REGISTER_LAST)
10028 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10029 if (!test_bit(opcode, ctx->restrictions.register_op))
10033 if (io_register_op_must_quiesce(opcode)) {
10034 percpu_ref_kill(&ctx->refs);
10037 * Drop uring mutex before waiting for references to exit. If
10038 * another thread is currently inside io_uring_enter() it might
10039 * need to grab the uring_lock to make progress. If we hold it
10040 * here across the drain wait, then we can deadlock. It's safe
10041 * to drop the mutex here, since no new references will come in
10042 * after we've killed the percpu ref.
10044 mutex_unlock(&ctx->uring_lock);
10046 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10049 ret = io_run_task_work_sig();
10053 mutex_lock(&ctx->uring_lock);
10056 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10062 case IORING_REGISTER_BUFFERS:
10063 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10065 case IORING_UNREGISTER_BUFFERS:
10067 if (arg || nr_args)
10069 ret = io_sqe_buffers_unregister(ctx);
10071 case IORING_REGISTER_FILES:
10072 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10074 case IORING_UNREGISTER_FILES:
10076 if (arg || nr_args)
10078 ret = io_sqe_files_unregister(ctx);
10080 case IORING_REGISTER_FILES_UPDATE:
10081 ret = io_register_files_update(ctx, arg, nr_args);
10083 case IORING_REGISTER_EVENTFD:
10084 case IORING_REGISTER_EVENTFD_ASYNC:
10088 ret = io_eventfd_register(ctx, arg);
10091 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10092 ctx->eventfd_async = 1;
10094 ctx->eventfd_async = 0;
10096 case IORING_UNREGISTER_EVENTFD:
10098 if (arg || nr_args)
10100 ret = io_eventfd_unregister(ctx);
10102 case IORING_REGISTER_PROBE:
10104 if (!arg || nr_args > 256)
10106 ret = io_probe(ctx, arg, nr_args);
10108 case IORING_REGISTER_PERSONALITY:
10110 if (arg || nr_args)
10112 ret = io_register_personality(ctx);
10114 case IORING_UNREGISTER_PERSONALITY:
10118 ret = io_unregister_personality(ctx, nr_args);
10120 case IORING_REGISTER_ENABLE_RINGS:
10122 if (arg || nr_args)
10124 ret = io_register_enable_rings(ctx);
10126 case IORING_REGISTER_RESTRICTIONS:
10127 ret = io_register_restrictions(ctx, arg, nr_args);
10129 case IORING_REGISTER_FILES2:
10130 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10132 case IORING_REGISTER_FILES_UPDATE2:
10133 ret = io_register_rsrc_update(ctx, arg, nr_args,
10136 case IORING_REGISTER_BUFFERS2:
10137 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10139 case IORING_REGISTER_BUFFERS_UPDATE:
10140 ret = io_register_rsrc_update(ctx, arg, nr_args,
10141 IORING_RSRC_BUFFER);
10148 if (io_register_op_must_quiesce(opcode)) {
10149 /* bring the ctx back to life */
10150 percpu_ref_reinit(&ctx->refs);
10151 reinit_completion(&ctx->ref_comp);
10156 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10157 void __user *, arg, unsigned int, nr_args)
10159 struct io_ring_ctx *ctx;
10168 if (f.file->f_op != &io_uring_fops)
10171 ctx = f.file->private_data;
10173 io_run_task_work();
10175 mutex_lock(&ctx->uring_lock);
10176 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10177 mutex_unlock(&ctx->uring_lock);
10178 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10179 ctx->cq_ev_fd != NULL, ret);
10185 static int __init io_uring_init(void)
10187 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10188 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10189 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10192 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10193 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10194 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10195 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10196 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10197 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10198 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10199 BUILD_BUG_SQE_ELEM(8, __u64, off);
10200 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10201 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10202 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10203 BUILD_BUG_SQE_ELEM(24, __u32, len);
10204 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10205 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10206 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10207 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10208 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10209 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10210 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10211 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10212 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10213 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10214 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10215 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10216 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10217 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10218 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10219 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10220 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10221 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10222 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10224 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10225 sizeof(struct io_uring_rsrc_update));
10226 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10227 sizeof(struct io_uring_rsrc_update2));
10228 /* should fit into one byte */
10229 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10231 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10232 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10233 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10237 __initcall(io_uring_init);