1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define IO_RSRC_TAG_TABLE_SHIFT 9
104 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
105 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
107 #define IORING_MAX_REG_BUFFERS (1U << 14)
109 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
110 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
113 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
116 u32 head ____cacheline_aligned_in_smp;
117 u32 tail ____cacheline_aligned_in_smp;
121 * This data is shared with the application through the mmap at offsets
122 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
124 * The offsets to the member fields are published through struct
125 * io_sqring_offsets when calling io_uring_setup.
129 * Head and tail offsets into the ring; the offsets need to be
130 * masked to get valid indices.
132 * The kernel controls head of the sq ring and the tail of the cq ring,
133 * and the application controls tail of the sq ring and the head of the
136 struct io_uring sq, cq;
138 * Bitmasks to apply to head and tail offsets (constant, equals
141 u32 sq_ring_mask, cq_ring_mask;
142 /* Ring sizes (constant, power of 2) */
143 u32 sq_ring_entries, cq_ring_entries;
145 * Number of invalid entries dropped by the kernel due to
146 * invalid index stored in array
148 * Written by the kernel, shouldn't be modified by the
149 * application (i.e. get number of "new events" by comparing to
152 * After a new SQ head value was read by the application this
153 * counter includes all submissions that were dropped reaching
154 * the new SQ head (and possibly more).
160 * Written by the kernel, shouldn't be modified by the
163 * The application needs a full memory barrier before checking
164 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
170 * Written by the application, shouldn't be modified by the
175 * Number of completion events lost because the queue was full;
176 * this should be avoided by the application by making sure
177 * there are not more requests pending than there is space in
178 * the completion queue.
180 * Written by the kernel, shouldn't be modified by the
181 * application (i.e. get number of "new events" by comparing to
184 * As completion events come in out of order this counter is not
185 * ordered with any other data.
189 * Ring buffer of completion events.
191 * The kernel writes completion events fresh every time they are
192 * produced, so the application is allowed to modify pending
195 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
198 enum io_uring_cmd_flags {
199 IO_URING_F_NONBLOCK = 1,
200 IO_URING_F_COMPLETE_DEFER = 2,
203 struct io_mapped_ubuf {
206 unsigned int nr_bvecs;
207 unsigned long acct_pages;
208 struct bio_vec bvec[];
213 struct io_overflow_cqe {
214 struct io_uring_cqe cqe;
215 struct list_head list;
218 struct io_fixed_file {
219 /* file * with additional FFS_* flags */
220 unsigned long file_ptr;
224 struct list_head list;
229 struct io_mapped_ubuf *buf;
233 struct io_file_table {
234 /* two level table */
235 struct io_fixed_file **files;
238 struct io_rsrc_node {
239 struct percpu_ref refs;
240 struct list_head node;
241 struct list_head rsrc_list;
242 struct io_rsrc_data *rsrc_data;
243 struct llist_node llist;
247 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
249 struct io_rsrc_data {
250 struct io_ring_ctx *ctx;
256 struct completion done;
261 struct list_head list;
267 struct io_restriction {
268 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
269 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
270 u8 sqe_flags_allowed;
271 u8 sqe_flags_required;
276 IO_SQ_THREAD_SHOULD_STOP = 0,
277 IO_SQ_THREAD_SHOULD_PARK,
282 atomic_t park_pending;
285 /* ctx's that are using this sqd */
286 struct list_head ctx_list;
288 struct task_struct *thread;
289 struct wait_queue_head wait;
291 unsigned sq_thread_idle;
297 struct completion exited;
300 #define IO_IOPOLL_BATCH 8
301 #define IO_COMPL_BATCH 32
302 #define IO_REQ_CACHE_SIZE 32
303 #define IO_REQ_ALLOC_BATCH 8
305 struct io_comp_state {
306 struct io_kiocb *reqs[IO_COMPL_BATCH];
308 /* inline/task_work completion list, under ->uring_lock */
309 struct list_head free_list;
312 struct io_submit_link {
313 struct io_kiocb *head;
314 struct io_kiocb *last;
317 struct io_submit_state {
318 struct blk_plug plug;
319 struct io_submit_link link;
322 * io_kiocb alloc cache
324 void *reqs[IO_REQ_CACHE_SIZE];
325 unsigned int free_reqs;
330 * Batch completion logic
332 struct io_comp_state comp;
335 * File reference cache
339 unsigned int file_refs;
340 unsigned int ios_left;
344 /* const or read-mostly hot data */
346 struct percpu_ref refs;
348 struct io_rings *rings;
350 unsigned int compat: 1;
351 unsigned int drain_next: 1;
352 unsigned int eventfd_async: 1;
353 unsigned int restricted: 1;
354 } ____cacheline_aligned_in_smp;
356 /* submission data */
359 * Ring buffer of indices into array of io_uring_sqe, which is
360 * mmapped by the application using the IORING_OFF_SQES offset.
362 * This indirection could e.g. be used to assign fixed
363 * io_uring_sqe entries to operations and only submit them to
364 * the queue when needed.
366 * The kernel modifies neither the indices array nor the entries
370 struct io_uring_sqe *sq_sqes;
371 unsigned cached_sq_head;
373 unsigned cached_sq_dropped;
374 unsigned long sq_check_overflow;
375 struct list_head defer_list;
378 * Fixed resources fast path, should be accessed only under
379 * uring_lock, and updated through io_uring_register(2)
381 struct io_rsrc_node *rsrc_node;
382 struct io_file_table file_table;
383 unsigned nr_user_files;
384 unsigned nr_user_bufs;
385 struct io_mapped_ubuf **user_bufs;
387 struct io_submit_state submit_state;
388 struct list_head timeout_list;
389 struct list_head cq_overflow_list;
390 struct xarray io_buffers;
391 struct xarray personalities;
393 unsigned sq_thread_idle;
394 } ____cacheline_aligned_in_smp;
397 struct mutex uring_lock;
398 wait_queue_head_t wait;
399 } ____cacheline_aligned_in_smp;
401 /* IRQ completion list, under ->completion_lock */
402 struct list_head locked_free_list;
403 unsigned int locked_free_nr;
405 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
406 struct io_sq_data *sq_data; /* if using sq thread polling */
408 struct wait_queue_head sqo_sq_wait;
409 struct list_head sqd_list;
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 */
477 struct wait_queue_head wait;
478 const struct io_ring_ctx *last;
480 struct percpu_counter inflight;
481 atomic_t inflight_tracked;
484 spinlock_t task_lock;
485 struct io_wq_work_list task_list;
486 unsigned long task_state;
487 struct callback_head task_work;
491 * First field must be the file pointer in all the
492 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
494 struct io_poll_iocb {
496 struct wait_queue_head *head;
500 struct wait_queue_entry wait;
503 struct io_poll_update {
509 bool update_user_data;
517 struct io_timeout_data {
518 struct io_kiocb *req;
519 struct hrtimer timer;
520 struct timespec64 ts;
521 enum hrtimer_mode mode;
526 struct sockaddr __user *addr;
527 int __user *addr_len;
529 unsigned long nofile;
549 struct list_head list;
550 /* head of the link, used by linked timeouts only */
551 struct io_kiocb *head;
554 struct io_timeout_rem {
559 struct timespec64 ts;
564 /* NOTE: kiocb has the file as the first member, so don't do it here */
572 struct sockaddr __user *addr;
579 struct compat_msghdr __user *umsg_compat;
580 struct user_msghdr __user *umsg;
586 struct io_buffer *kbuf;
592 struct filename *filename;
594 unsigned long nofile;
597 struct io_rsrc_update {
623 struct epoll_event event;
627 struct file *file_out;
628 struct file *file_in;
635 struct io_provide_buf {
649 const char __user *filename;
650 struct statx __user *buffer;
662 struct filename *oldpath;
663 struct filename *newpath;
671 struct filename *filename;
674 struct io_completion {
676 struct list_head list;
680 struct io_async_connect {
681 struct sockaddr_storage address;
684 struct io_async_msghdr {
685 struct iovec fast_iov[UIO_FASTIOV];
686 /* points to an allocated iov, if NULL we use fast_iov instead */
687 struct iovec *free_iov;
688 struct sockaddr __user *uaddr;
690 struct sockaddr_storage addr;
694 struct iovec fast_iov[UIO_FASTIOV];
695 const struct iovec *free_iovec;
696 struct iov_iter iter;
698 struct wait_page_queue wpq;
702 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
703 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
704 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
705 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
706 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
707 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
709 /* first byte is taken by user flags, shift it to not overlap */
714 REQ_F_LINK_TIMEOUT_BIT,
715 REQ_F_NEED_CLEANUP_BIT,
717 REQ_F_BUFFER_SELECTED_BIT,
718 REQ_F_LTIMEOUT_ACTIVE_BIT,
719 REQ_F_COMPLETE_INLINE_BIT,
721 REQ_F_DONT_REISSUE_BIT,
722 /* keep async read/write and isreg together and in order */
723 REQ_F_ASYNC_READ_BIT,
724 REQ_F_ASYNC_WRITE_BIT,
727 /* not a real bit, just to check we're not overflowing the space */
733 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
734 /* drain existing IO first */
735 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
737 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
738 /* doesn't sever on completion < 0 */
739 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
741 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
742 /* IOSQE_BUFFER_SELECT */
743 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
745 /* fail rest of links */
746 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
747 /* on inflight list, should be cancelled and waited on exit reliably */
748 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
749 /* read/write uses file position */
750 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
751 /* must not punt to workers */
752 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
753 /* has or had linked timeout */
754 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
756 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
757 /* already went through poll handler */
758 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
759 /* buffer already selected */
760 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
761 /* linked timeout is active, i.e. prepared by link's head */
762 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
763 /* completion is deferred through io_comp_state */
764 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
765 /* caller should reissue async */
766 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
767 /* don't attempt request reissue, see io_rw_reissue() */
768 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
769 /* supports async reads */
770 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
771 /* supports async writes */
772 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
774 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
778 struct io_poll_iocb poll;
779 struct io_poll_iocb *double_poll;
782 struct io_task_work {
783 struct io_wq_work_node node;
784 task_work_func_t func;
788 IORING_RSRC_FILE = 0,
789 IORING_RSRC_BUFFER = 1,
793 * NOTE! Each of the iocb union members has the file pointer
794 * as the first entry in their struct definition. So you can
795 * access the file pointer through any of the sub-structs,
796 * or directly as just 'ki_filp' in this struct.
802 struct io_poll_iocb poll;
803 struct io_poll_update poll_update;
804 struct io_accept accept;
806 struct io_cancel cancel;
807 struct io_timeout timeout;
808 struct io_timeout_rem timeout_rem;
809 struct io_connect connect;
810 struct io_sr_msg sr_msg;
812 struct io_close close;
813 struct io_rsrc_update rsrc_update;
814 struct io_fadvise fadvise;
815 struct io_madvise madvise;
816 struct io_epoll epoll;
817 struct io_splice splice;
818 struct io_provide_buf pbuf;
819 struct io_statx statx;
820 struct io_shutdown shutdown;
821 struct io_rename rename;
822 struct io_unlink unlink;
823 /* use only after cleaning per-op data, see io_clean_op() */
824 struct io_completion compl;
827 /* opcode allocated if it needs to store data for async defer */
830 /* polled IO has completed */
836 struct io_ring_ctx *ctx;
839 struct task_struct *task;
842 struct io_kiocb *link;
843 struct percpu_ref *fixed_rsrc_refs;
845 /* used with ctx->iopoll_list with reads/writes */
846 struct list_head inflight_entry;
848 struct io_task_work io_task_work;
849 struct callback_head task_work;
851 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
852 struct hlist_node hash_node;
853 struct async_poll *apoll;
854 struct io_wq_work work;
855 /* store used ubuf, so we can prevent reloading */
856 struct io_mapped_ubuf *imu;
859 struct io_tctx_node {
860 struct list_head ctx_node;
861 struct task_struct *task;
862 struct io_ring_ctx *ctx;
865 struct io_defer_entry {
866 struct list_head list;
867 struct io_kiocb *req;
872 /* needs req->file assigned */
873 unsigned needs_file : 1;
874 /* hash wq insertion if file is a regular file */
875 unsigned hash_reg_file : 1;
876 /* unbound wq insertion if file is a non-regular file */
877 unsigned unbound_nonreg_file : 1;
878 /* opcode is not supported by this kernel */
879 unsigned not_supported : 1;
880 /* set if opcode supports polled "wait" */
882 unsigned pollout : 1;
883 /* op supports buffer selection */
884 unsigned buffer_select : 1;
885 /* do prep async if is going to be punted */
886 unsigned needs_async_setup : 1;
887 /* should block plug */
889 /* size of async data needed, if any */
890 unsigned short async_size;
893 static const struct io_op_def io_op_defs[] = {
894 [IORING_OP_NOP] = {},
895 [IORING_OP_READV] = {
897 .unbound_nonreg_file = 1,
900 .needs_async_setup = 1,
902 .async_size = sizeof(struct io_async_rw),
904 [IORING_OP_WRITEV] = {
907 .unbound_nonreg_file = 1,
909 .needs_async_setup = 1,
911 .async_size = sizeof(struct io_async_rw),
913 [IORING_OP_FSYNC] = {
916 [IORING_OP_READ_FIXED] = {
918 .unbound_nonreg_file = 1,
921 .async_size = sizeof(struct io_async_rw),
923 [IORING_OP_WRITE_FIXED] = {
926 .unbound_nonreg_file = 1,
929 .async_size = sizeof(struct io_async_rw),
931 [IORING_OP_POLL_ADD] = {
933 .unbound_nonreg_file = 1,
935 [IORING_OP_POLL_REMOVE] = {},
936 [IORING_OP_SYNC_FILE_RANGE] = {
939 [IORING_OP_SENDMSG] = {
941 .unbound_nonreg_file = 1,
943 .needs_async_setup = 1,
944 .async_size = sizeof(struct io_async_msghdr),
946 [IORING_OP_RECVMSG] = {
948 .unbound_nonreg_file = 1,
951 .needs_async_setup = 1,
952 .async_size = sizeof(struct io_async_msghdr),
954 [IORING_OP_TIMEOUT] = {
955 .async_size = sizeof(struct io_timeout_data),
957 [IORING_OP_TIMEOUT_REMOVE] = {
958 /* used by timeout updates' prep() */
960 [IORING_OP_ACCEPT] = {
962 .unbound_nonreg_file = 1,
965 [IORING_OP_ASYNC_CANCEL] = {},
966 [IORING_OP_LINK_TIMEOUT] = {
967 .async_size = sizeof(struct io_timeout_data),
969 [IORING_OP_CONNECT] = {
971 .unbound_nonreg_file = 1,
973 .needs_async_setup = 1,
974 .async_size = sizeof(struct io_async_connect),
976 [IORING_OP_FALLOCATE] = {
979 [IORING_OP_OPENAT] = {},
980 [IORING_OP_CLOSE] = {},
981 [IORING_OP_FILES_UPDATE] = {},
982 [IORING_OP_STATX] = {},
985 .unbound_nonreg_file = 1,
989 .async_size = sizeof(struct io_async_rw),
991 [IORING_OP_WRITE] = {
993 .unbound_nonreg_file = 1,
996 .async_size = sizeof(struct io_async_rw),
998 [IORING_OP_FADVISE] = {
1001 [IORING_OP_MADVISE] = {},
1002 [IORING_OP_SEND] = {
1004 .unbound_nonreg_file = 1,
1007 [IORING_OP_RECV] = {
1009 .unbound_nonreg_file = 1,
1013 [IORING_OP_OPENAT2] = {
1015 [IORING_OP_EPOLL_CTL] = {
1016 .unbound_nonreg_file = 1,
1018 [IORING_OP_SPLICE] = {
1021 .unbound_nonreg_file = 1,
1023 [IORING_OP_PROVIDE_BUFFERS] = {},
1024 [IORING_OP_REMOVE_BUFFERS] = {},
1028 .unbound_nonreg_file = 1,
1030 [IORING_OP_SHUTDOWN] = {
1033 [IORING_OP_RENAMEAT] = {},
1034 [IORING_OP_UNLINKAT] = {},
1037 static bool io_disarm_next(struct io_kiocb *req);
1038 static void io_uring_del_tctx_node(unsigned long index);
1039 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1040 struct task_struct *task,
1042 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1043 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1045 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1046 long res, unsigned int cflags);
1047 static void io_put_req(struct io_kiocb *req);
1048 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1049 static void io_dismantle_req(struct io_kiocb *req);
1050 static void io_put_task(struct task_struct *task, int nr);
1051 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1052 static void io_queue_linked_timeout(struct io_kiocb *req);
1053 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1054 struct io_uring_rsrc_update2 *up,
1056 static void io_clean_op(struct io_kiocb *req);
1057 static struct file *io_file_get(struct io_submit_state *state,
1058 struct io_kiocb *req, int fd, bool fixed);
1059 static void __io_queue_sqe(struct io_kiocb *req);
1060 static void io_rsrc_put_work(struct work_struct *work);
1062 static void io_req_task_queue(struct io_kiocb *req);
1063 static void io_submit_flush_completions(struct io_comp_state *cs,
1064 struct io_ring_ctx *ctx);
1065 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1066 static int io_req_prep_async(struct io_kiocb *req);
1068 static struct kmem_cache *req_cachep;
1070 static const struct file_operations io_uring_fops;
1072 struct sock *io_uring_get_socket(struct file *file)
1074 #if defined(CONFIG_UNIX)
1075 if (file->f_op == &io_uring_fops) {
1076 struct io_ring_ctx *ctx = file->private_data;
1078 return ctx->ring_sock->sk;
1083 EXPORT_SYMBOL(io_uring_get_socket);
1085 #define io_for_each_link(pos, head) \
1086 for (pos = (head); pos; pos = pos->link)
1088 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1090 struct io_ring_ctx *ctx = req->ctx;
1092 if (!req->fixed_rsrc_refs) {
1093 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1094 percpu_ref_get(req->fixed_rsrc_refs);
1098 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1100 bool got = percpu_ref_tryget(ref);
1102 /* already at zero, wait for ->release() */
1104 wait_for_completion(compl);
1105 percpu_ref_resurrect(ref);
1107 percpu_ref_put(ref);
1110 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1113 struct io_kiocb *req;
1115 if (task && head->task != task)
1120 io_for_each_link(req, head) {
1121 if (req->flags & REQ_F_INFLIGHT)
1127 static inline void req_set_fail(struct io_kiocb *req)
1129 req->flags |= REQ_F_FAIL;
1132 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1134 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1136 complete(&ctx->ref_comp);
1139 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1141 return !req->timeout.off;
1144 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1146 struct io_ring_ctx *ctx;
1149 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1154 * Use 5 bits less than the max cq entries, that should give us around
1155 * 32 entries per hash list if totally full and uniformly spread.
1157 hash_bits = ilog2(p->cq_entries);
1161 ctx->cancel_hash_bits = hash_bits;
1162 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1164 if (!ctx->cancel_hash)
1166 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1168 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1169 if (!ctx->dummy_ubuf)
1171 /* set invalid range, so io_import_fixed() fails meeting it */
1172 ctx->dummy_ubuf->ubuf = -1UL;
1174 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1175 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1178 ctx->flags = p->flags;
1179 init_waitqueue_head(&ctx->sqo_sq_wait);
1180 INIT_LIST_HEAD(&ctx->sqd_list);
1181 init_waitqueue_head(&ctx->cq_wait);
1182 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1183 init_completion(&ctx->ref_comp);
1184 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1185 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1186 mutex_init(&ctx->uring_lock);
1187 init_waitqueue_head(&ctx->wait);
1188 spin_lock_init(&ctx->completion_lock);
1189 INIT_LIST_HEAD(&ctx->iopoll_list);
1190 INIT_LIST_HEAD(&ctx->defer_list);
1191 INIT_LIST_HEAD(&ctx->timeout_list);
1192 spin_lock_init(&ctx->rsrc_ref_lock);
1193 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1194 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1195 init_llist_head(&ctx->rsrc_put_llist);
1196 INIT_LIST_HEAD(&ctx->tctx_list);
1197 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1198 INIT_LIST_HEAD(&ctx->locked_free_list);
1201 kfree(ctx->dummy_ubuf);
1202 kfree(ctx->cancel_hash);
1207 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1209 struct io_rings *r = ctx->rings;
1211 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1215 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1217 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1218 struct io_ring_ctx *ctx = req->ctx;
1220 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1226 static void io_req_track_inflight(struct io_kiocb *req)
1228 if (!(req->flags & REQ_F_INFLIGHT)) {
1229 req->flags |= REQ_F_INFLIGHT;
1230 atomic_inc(¤t->io_uring->inflight_tracked);
1234 static void io_prep_async_work(struct io_kiocb *req)
1236 const struct io_op_def *def = &io_op_defs[req->opcode];
1237 struct io_ring_ctx *ctx = req->ctx;
1239 if (!req->work.creds)
1240 req->work.creds = get_current_cred();
1242 req->work.list.next = NULL;
1243 req->work.flags = 0;
1244 if (req->flags & REQ_F_FORCE_ASYNC)
1245 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1247 if (req->flags & REQ_F_ISREG) {
1248 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1249 io_wq_hash_work(&req->work, file_inode(req->file));
1250 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1251 if (def->unbound_nonreg_file)
1252 req->work.flags |= IO_WQ_WORK_UNBOUND;
1255 switch (req->opcode) {
1256 case IORING_OP_SPLICE:
1258 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1259 req->work.flags |= IO_WQ_WORK_UNBOUND;
1264 static void io_prep_async_link(struct io_kiocb *req)
1266 struct io_kiocb *cur;
1268 io_for_each_link(cur, req)
1269 io_prep_async_work(cur);
1272 static void io_queue_async_work(struct io_kiocb *req)
1274 struct io_ring_ctx *ctx = req->ctx;
1275 struct io_kiocb *link = io_prep_linked_timeout(req);
1276 struct io_uring_task *tctx = req->task->io_uring;
1279 BUG_ON(!tctx->io_wq);
1281 /* init ->work of the whole link before punting */
1282 io_prep_async_link(req);
1283 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1284 &req->work, req->flags);
1285 io_wq_enqueue(tctx->io_wq, &req->work);
1287 io_queue_linked_timeout(link);
1290 static void io_kill_timeout(struct io_kiocb *req, int status)
1291 __must_hold(&req->ctx->completion_lock)
1293 struct io_timeout_data *io = req->async_data;
1295 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1296 atomic_set(&req->ctx->cq_timeouts,
1297 atomic_read(&req->ctx->cq_timeouts) + 1);
1298 list_del_init(&req->timeout.list);
1299 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1300 io_put_req_deferred(req, 1);
1304 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1307 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1308 struct io_defer_entry, list);
1310 if (req_need_defer(de->req, de->seq))
1312 list_del_init(&de->list);
1313 io_req_task_queue(de->req);
1315 } while (!list_empty(&ctx->defer_list));
1318 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1322 if (list_empty(&ctx->timeout_list))
1325 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1328 u32 events_needed, events_got;
1329 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1330 struct io_kiocb, timeout.list);
1332 if (io_is_timeout_noseq(req))
1336 * Since seq can easily wrap around over time, subtract
1337 * the last seq at which timeouts were flushed before comparing.
1338 * Assuming not more than 2^31-1 events have happened since,
1339 * these subtractions won't have wrapped, so we can check if
1340 * target is in [last_seq, current_seq] by comparing the two.
1342 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1343 events_got = seq - ctx->cq_last_tm_flush;
1344 if (events_got < events_needed)
1347 list_del_init(&req->timeout.list);
1348 io_kill_timeout(req, 0);
1349 } while (!list_empty(&ctx->timeout_list));
1351 ctx->cq_last_tm_flush = seq;
1354 static void io_commit_cqring(struct io_ring_ctx *ctx)
1356 io_flush_timeouts(ctx);
1358 /* order cqe stores with ring update */
1359 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1361 if (unlikely(!list_empty(&ctx->defer_list)))
1362 __io_queue_deferred(ctx);
1365 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1367 struct io_rings *r = ctx->rings;
1369 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1372 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1374 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1377 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1379 struct io_rings *rings = ctx->rings;
1380 unsigned tail, mask = ctx->cq_entries - 1;
1383 * writes to the cq entry need to come after reading head; the
1384 * control dependency is enough as we're using WRITE_ONCE to
1387 if (__io_cqring_events(ctx) == ctx->cq_entries)
1390 tail = ctx->cached_cq_tail++;
1391 return &rings->cqes[tail & mask];
1394 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1396 if (likely(!ctx->cq_ev_fd))
1398 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1400 return !ctx->eventfd_async || io_wq_current_is_worker();
1403 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1405 /* see waitqueue_active() comment */
1408 if (waitqueue_active(&ctx->wait))
1409 wake_up(&ctx->wait);
1410 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1411 wake_up(&ctx->sq_data->wait);
1412 if (io_should_trigger_evfd(ctx))
1413 eventfd_signal(ctx->cq_ev_fd, 1);
1414 if (waitqueue_active(&ctx->cq_wait)) {
1415 wake_up_interruptible(&ctx->cq_wait);
1416 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1420 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1422 /* see waitqueue_active() comment */
1425 if (ctx->flags & IORING_SETUP_SQPOLL) {
1426 if (waitqueue_active(&ctx->wait))
1427 wake_up(&ctx->wait);
1429 if (io_should_trigger_evfd(ctx))
1430 eventfd_signal(ctx->cq_ev_fd, 1);
1431 if (waitqueue_active(&ctx->cq_wait)) {
1432 wake_up_interruptible(&ctx->cq_wait);
1433 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1437 /* Returns true if there are no backlogged entries after the flush */
1438 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1440 unsigned long flags;
1441 bool all_flushed, posted;
1443 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1447 spin_lock_irqsave(&ctx->completion_lock, flags);
1448 while (!list_empty(&ctx->cq_overflow_list)) {
1449 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1450 struct io_overflow_cqe *ocqe;
1454 ocqe = list_first_entry(&ctx->cq_overflow_list,
1455 struct io_overflow_cqe, list);
1457 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1459 io_account_cq_overflow(ctx);
1462 list_del(&ocqe->list);
1466 all_flushed = list_empty(&ctx->cq_overflow_list);
1468 clear_bit(0, &ctx->sq_check_overflow);
1469 clear_bit(0, &ctx->cq_check_overflow);
1470 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1474 io_commit_cqring(ctx);
1475 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1477 io_cqring_ev_posted(ctx);
1481 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1485 if (test_bit(0, &ctx->cq_check_overflow)) {
1486 /* iopoll syncs against uring_lock, not completion_lock */
1487 if (ctx->flags & IORING_SETUP_IOPOLL)
1488 mutex_lock(&ctx->uring_lock);
1489 ret = __io_cqring_overflow_flush(ctx, force);
1490 if (ctx->flags & IORING_SETUP_IOPOLL)
1491 mutex_unlock(&ctx->uring_lock);
1498 * Shamelessly stolen from the mm implementation of page reference checking,
1499 * see commit f958d7b528b1 for details.
1501 #define req_ref_zero_or_close_to_overflow(req) \
1502 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1504 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1506 return atomic_inc_not_zero(&req->refs);
1509 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1511 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1512 return atomic_sub_and_test(refs, &req->refs);
1515 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1517 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1518 return atomic_dec_and_test(&req->refs);
1521 static inline void req_ref_put(struct io_kiocb *req)
1523 WARN_ON_ONCE(req_ref_put_and_test(req));
1526 static inline void req_ref_get(struct io_kiocb *req)
1528 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1529 atomic_inc(&req->refs);
1532 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1533 long res, unsigned int cflags)
1535 struct io_overflow_cqe *ocqe;
1537 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1540 * If we're in ring overflow flush mode, or in task cancel mode,
1541 * or cannot allocate an overflow entry, then we need to drop it
1544 io_account_cq_overflow(ctx);
1547 if (list_empty(&ctx->cq_overflow_list)) {
1548 set_bit(0, &ctx->sq_check_overflow);
1549 set_bit(0, &ctx->cq_check_overflow);
1550 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1552 ocqe->cqe.user_data = user_data;
1553 ocqe->cqe.res = res;
1554 ocqe->cqe.flags = cflags;
1555 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1559 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1560 long res, unsigned int cflags)
1562 struct io_uring_cqe *cqe;
1564 trace_io_uring_complete(ctx, user_data, res, cflags);
1567 * If we can't get a cq entry, userspace overflowed the
1568 * submission (by quite a lot). Increment the overflow count in
1571 cqe = io_get_cqe(ctx);
1573 WRITE_ONCE(cqe->user_data, user_data);
1574 WRITE_ONCE(cqe->res, res);
1575 WRITE_ONCE(cqe->flags, cflags);
1578 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1581 /* not as hot to bloat with inlining */
1582 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1583 long res, unsigned int cflags)
1585 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1588 static void io_req_complete_post(struct io_kiocb *req, long res,
1589 unsigned int cflags)
1591 struct io_ring_ctx *ctx = req->ctx;
1592 unsigned long flags;
1594 spin_lock_irqsave(&ctx->completion_lock, flags);
1595 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1597 * If we're the last reference to this request, add to our locked
1600 if (req_ref_put_and_test(req)) {
1601 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1602 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1603 io_disarm_next(req);
1605 io_req_task_queue(req->link);
1609 io_dismantle_req(req);
1610 io_put_task(req->task, 1);
1611 list_add(&req->compl.list, &ctx->locked_free_list);
1612 ctx->locked_free_nr++;
1614 if (!percpu_ref_tryget(&ctx->refs))
1617 io_commit_cqring(ctx);
1618 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1621 io_cqring_ev_posted(ctx);
1622 percpu_ref_put(&ctx->refs);
1626 static inline bool io_req_needs_clean(struct io_kiocb *req)
1628 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1629 REQ_F_POLLED | REQ_F_INFLIGHT);
1632 static void io_req_complete_state(struct io_kiocb *req, long res,
1633 unsigned int cflags)
1635 if (io_req_needs_clean(req))
1638 req->compl.cflags = cflags;
1639 req->flags |= REQ_F_COMPLETE_INLINE;
1642 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1643 long res, unsigned cflags)
1645 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1646 io_req_complete_state(req, res, cflags);
1648 io_req_complete_post(req, res, cflags);
1651 static inline void io_req_complete(struct io_kiocb *req, long res)
1653 __io_req_complete(req, 0, res, 0);
1656 static void io_req_complete_failed(struct io_kiocb *req, long res)
1660 io_req_complete_post(req, res, 0);
1663 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1664 struct io_comp_state *cs)
1666 spin_lock_irq(&ctx->completion_lock);
1667 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1668 ctx->locked_free_nr = 0;
1669 spin_unlock_irq(&ctx->completion_lock);
1672 /* Returns true IFF there are requests in the cache */
1673 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1675 struct io_submit_state *state = &ctx->submit_state;
1676 struct io_comp_state *cs = &state->comp;
1680 * If we have more than a batch's worth of requests in our IRQ side
1681 * locked cache, grab the lock and move them over to our submission
1684 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1685 io_flush_cached_locked_reqs(ctx, cs);
1687 nr = state->free_reqs;
1688 while (!list_empty(&cs->free_list)) {
1689 struct io_kiocb *req = list_first_entry(&cs->free_list,
1690 struct io_kiocb, compl.list);
1692 list_del(&req->compl.list);
1693 state->reqs[nr++] = req;
1694 if (nr == ARRAY_SIZE(state->reqs))
1698 state->free_reqs = nr;
1702 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1704 struct io_submit_state *state = &ctx->submit_state;
1706 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1708 if (!state->free_reqs) {
1709 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1712 if (io_flush_cached_reqs(ctx))
1715 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1719 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1720 * retry single alloc to be on the safe side.
1722 if (unlikely(ret <= 0)) {
1723 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1724 if (!state->reqs[0])
1728 state->free_reqs = ret;
1732 return state->reqs[state->free_reqs];
1735 static inline void io_put_file(struct file *file)
1741 static void io_dismantle_req(struct io_kiocb *req)
1743 unsigned int flags = req->flags;
1745 if (io_req_needs_clean(req))
1747 if (!(flags & REQ_F_FIXED_FILE))
1748 io_put_file(req->file);
1749 if (req->fixed_rsrc_refs)
1750 percpu_ref_put(req->fixed_rsrc_refs);
1751 if (req->async_data)
1752 kfree(req->async_data);
1753 if (req->work.creds) {
1754 put_cred(req->work.creds);
1755 req->work.creds = NULL;
1759 /* must to be called somewhat shortly after putting a request */
1760 static inline void io_put_task(struct task_struct *task, int nr)
1762 struct io_uring_task *tctx = task->io_uring;
1764 percpu_counter_sub(&tctx->inflight, nr);
1765 if (unlikely(atomic_read(&tctx->in_idle)))
1766 wake_up(&tctx->wait);
1767 put_task_struct_many(task, nr);
1770 static void __io_free_req(struct io_kiocb *req)
1772 struct io_ring_ctx *ctx = req->ctx;
1774 io_dismantle_req(req);
1775 io_put_task(req->task, 1);
1777 kmem_cache_free(req_cachep, req);
1778 percpu_ref_put(&ctx->refs);
1781 static inline void io_remove_next_linked(struct io_kiocb *req)
1783 struct io_kiocb *nxt = req->link;
1785 req->link = nxt->link;
1789 static bool io_kill_linked_timeout(struct io_kiocb *req)
1790 __must_hold(&req->ctx->completion_lock)
1792 struct io_kiocb *link = req->link;
1795 * Can happen if a linked timeout fired and link had been like
1796 * req -> link t-out -> link t-out [-> ...]
1798 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1799 struct io_timeout_data *io = link->async_data;
1801 io_remove_next_linked(req);
1802 link->timeout.head = NULL;
1803 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1804 io_cqring_fill_event(link->ctx, link->user_data,
1806 io_put_req_deferred(link, 1);
1813 static void io_fail_links(struct io_kiocb *req)
1814 __must_hold(&req->ctx->completion_lock)
1816 struct io_kiocb *nxt, *link = req->link;
1823 trace_io_uring_fail_link(req, link);
1824 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1825 io_put_req_deferred(link, 2);
1830 static bool io_disarm_next(struct io_kiocb *req)
1831 __must_hold(&req->ctx->completion_lock)
1833 bool posted = false;
1835 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1836 posted = io_kill_linked_timeout(req);
1837 if (unlikely((req->flags & REQ_F_FAIL) &&
1838 !(req->flags & REQ_F_HARDLINK))) {
1839 posted |= (req->link != NULL);
1845 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1847 struct io_kiocb *nxt;
1850 * If LINK is set, we have dependent requests in this chain. If we
1851 * didn't fail this request, queue the first one up, moving any other
1852 * dependencies to the next request. In case of failure, fail the rest
1855 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1856 struct io_ring_ctx *ctx = req->ctx;
1857 unsigned long flags;
1860 spin_lock_irqsave(&ctx->completion_lock, flags);
1861 posted = io_disarm_next(req);
1863 io_commit_cqring(req->ctx);
1864 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1866 io_cqring_ev_posted(ctx);
1873 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1875 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1877 return __io_req_find_next(req);
1880 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1884 if (ctx->submit_state.comp.nr) {
1885 mutex_lock(&ctx->uring_lock);
1886 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1887 mutex_unlock(&ctx->uring_lock);
1889 percpu_ref_put(&ctx->refs);
1892 static bool __tctx_task_work(struct io_uring_task *tctx)
1894 struct io_ring_ctx *ctx = NULL;
1895 struct io_wq_work_list list;
1896 struct io_wq_work_node *node;
1898 if (wq_list_empty(&tctx->task_list))
1901 spin_lock_irq(&tctx->task_lock);
1902 list = tctx->task_list;
1903 INIT_WQ_LIST(&tctx->task_list);
1904 spin_unlock_irq(&tctx->task_lock);
1908 struct io_wq_work_node *next = node->next;
1909 struct io_kiocb *req;
1911 req = container_of(node, struct io_kiocb, io_task_work.node);
1912 if (req->ctx != ctx) {
1913 ctx_flush_and_put(ctx);
1915 percpu_ref_get(&ctx->refs);
1918 req->task_work.func(&req->task_work);
1922 ctx_flush_and_put(ctx);
1923 return list.first != NULL;
1926 static void tctx_task_work(struct callback_head *cb)
1928 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1930 clear_bit(0, &tctx->task_state);
1932 while (__tctx_task_work(tctx))
1936 static int io_req_task_work_add(struct io_kiocb *req)
1938 struct task_struct *tsk = req->task;
1939 struct io_uring_task *tctx = tsk->io_uring;
1940 enum task_work_notify_mode notify;
1941 struct io_wq_work_node *node, *prev;
1942 unsigned long flags;
1945 if (unlikely(tsk->flags & PF_EXITING))
1948 WARN_ON_ONCE(!tctx);
1950 spin_lock_irqsave(&tctx->task_lock, flags);
1951 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1952 spin_unlock_irqrestore(&tctx->task_lock, flags);
1954 /* task_work already pending, we're done */
1955 if (test_bit(0, &tctx->task_state) ||
1956 test_and_set_bit(0, &tctx->task_state))
1960 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1961 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1962 * processing task_work. There's no reliable way to tell if TWA_RESUME
1965 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1967 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1968 wake_up_process(tsk);
1973 * Slow path - we failed, find and delete work. if the work is not
1974 * in the list, it got run and we're fine.
1976 spin_lock_irqsave(&tctx->task_lock, flags);
1977 wq_list_for_each(node, prev, &tctx->task_list) {
1978 if (&req->io_task_work.node == node) {
1979 wq_list_del(&tctx->task_list, node, prev);
1984 spin_unlock_irqrestore(&tctx->task_lock, flags);
1985 clear_bit(0, &tctx->task_state);
1989 static bool io_run_task_work_head(struct callback_head **work_head)
1991 struct callback_head *work, *next;
1992 bool executed = false;
1995 work = xchg(work_head, NULL);
2011 static void io_task_work_add_head(struct callback_head **work_head,
2012 struct callback_head *task_work)
2014 struct callback_head *head;
2017 head = READ_ONCE(*work_head);
2018 task_work->next = head;
2019 } while (cmpxchg(work_head, head, task_work) != head);
2022 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2023 task_work_func_t cb)
2025 init_task_work(&req->task_work, cb);
2026 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2029 static void io_req_task_cancel(struct callback_head *cb)
2031 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2032 struct io_ring_ctx *ctx = req->ctx;
2034 /* ctx is guaranteed to stay alive while we hold uring_lock */
2035 mutex_lock(&ctx->uring_lock);
2036 io_req_complete_failed(req, req->result);
2037 mutex_unlock(&ctx->uring_lock);
2040 static void __io_req_task_submit(struct io_kiocb *req)
2042 struct io_ring_ctx *ctx = req->ctx;
2044 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2045 mutex_lock(&ctx->uring_lock);
2046 if (!(current->flags & PF_EXITING) && !current->in_execve)
2047 __io_queue_sqe(req);
2049 io_req_complete_failed(req, -EFAULT);
2050 mutex_unlock(&ctx->uring_lock);
2053 static void io_req_task_submit(struct callback_head *cb)
2055 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2057 __io_req_task_submit(req);
2060 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2063 req->task_work.func = io_req_task_cancel;
2065 if (unlikely(io_req_task_work_add(req)))
2066 io_req_task_work_add_fallback(req, io_req_task_cancel);
2069 static void io_req_task_queue(struct io_kiocb *req)
2071 req->task_work.func = io_req_task_submit;
2073 if (unlikely(io_req_task_work_add(req)))
2074 io_req_task_queue_fail(req, -ECANCELED);
2077 static inline void io_queue_next(struct io_kiocb *req)
2079 struct io_kiocb *nxt = io_req_find_next(req);
2082 io_req_task_queue(nxt);
2085 static void io_free_req(struct io_kiocb *req)
2092 struct task_struct *task;
2097 static inline void io_init_req_batch(struct req_batch *rb)
2104 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2105 struct req_batch *rb)
2108 io_put_task(rb->task, rb->task_refs);
2110 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2113 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2114 struct io_submit_state *state)
2117 io_dismantle_req(req);
2119 if (req->task != rb->task) {
2121 io_put_task(rb->task, rb->task_refs);
2122 rb->task = req->task;
2128 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2129 state->reqs[state->free_reqs++] = req;
2131 list_add(&req->compl.list, &state->comp.free_list);
2134 static void io_submit_flush_completions(struct io_comp_state *cs,
2135 struct io_ring_ctx *ctx)
2138 struct io_kiocb *req;
2139 struct req_batch rb;
2141 io_init_req_batch(&rb);
2142 spin_lock_irq(&ctx->completion_lock);
2143 for (i = 0; i < nr; i++) {
2145 __io_cqring_fill_event(ctx, req->user_data, req->result,
2148 io_commit_cqring(ctx);
2149 spin_unlock_irq(&ctx->completion_lock);
2151 io_cqring_ev_posted(ctx);
2152 for (i = 0; i < nr; i++) {
2155 /* submission and completion refs */
2156 if (req_ref_sub_and_test(req, 2))
2157 io_req_free_batch(&rb, req, &ctx->submit_state);
2160 io_req_free_batch_finish(ctx, &rb);
2165 * Drop reference to request, return next in chain (if there is one) if this
2166 * was the last reference to this request.
2168 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2170 struct io_kiocb *nxt = NULL;
2172 if (req_ref_put_and_test(req)) {
2173 nxt = io_req_find_next(req);
2179 static inline void io_put_req(struct io_kiocb *req)
2181 if (req_ref_put_and_test(req))
2185 static void io_put_req_deferred_cb(struct callback_head *cb)
2187 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2192 static void io_free_req_deferred(struct io_kiocb *req)
2194 req->task_work.func = io_put_req_deferred_cb;
2195 if (unlikely(io_req_task_work_add(req)))
2196 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2199 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2201 if (req_ref_sub_and_test(req, refs))
2202 io_free_req_deferred(req);
2205 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2207 /* See comment at the top of this file */
2209 return __io_cqring_events(ctx);
2212 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2214 struct io_rings *rings = ctx->rings;
2216 /* make sure SQ entry isn't read before tail */
2217 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2220 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2222 unsigned int cflags;
2224 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2225 cflags |= IORING_CQE_F_BUFFER;
2226 req->flags &= ~REQ_F_BUFFER_SELECTED;
2231 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2233 struct io_buffer *kbuf;
2235 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2236 return io_put_kbuf(req, kbuf);
2239 static inline bool io_run_task_work(void)
2242 * Not safe to run on exiting task, and the task_work handling will
2243 * not add work to such a task.
2245 if (unlikely(current->flags & PF_EXITING))
2247 if (current->task_works) {
2248 __set_current_state(TASK_RUNNING);
2257 * Find and free completed poll iocbs
2259 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2260 struct list_head *done)
2262 struct req_batch rb;
2263 struct io_kiocb *req;
2265 /* order with ->result store in io_complete_rw_iopoll() */
2268 io_init_req_batch(&rb);
2269 while (!list_empty(done)) {
2272 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2273 list_del(&req->inflight_entry);
2275 if (READ_ONCE(req->result) == -EAGAIN &&
2276 !(req->flags & REQ_F_DONT_REISSUE)) {
2277 req->iopoll_completed = 0;
2279 io_queue_async_work(req);
2283 if (req->flags & REQ_F_BUFFER_SELECTED)
2284 cflags = io_put_rw_kbuf(req);
2286 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2289 if (req_ref_put_and_test(req))
2290 io_req_free_batch(&rb, req, &ctx->submit_state);
2293 io_commit_cqring(ctx);
2294 io_cqring_ev_posted_iopoll(ctx);
2295 io_req_free_batch_finish(ctx, &rb);
2298 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2301 struct io_kiocb *req, *tmp;
2307 * Only spin for completions if we don't have multiple devices hanging
2308 * off our complete list, and we're under the requested amount.
2310 spin = !ctx->poll_multi_file && *nr_events < min;
2313 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2314 struct kiocb *kiocb = &req->rw.kiocb;
2317 * Move completed and retryable entries to our local lists.
2318 * If we find a request that requires polling, break out
2319 * and complete those lists first, if we have entries there.
2321 if (READ_ONCE(req->iopoll_completed)) {
2322 list_move_tail(&req->inflight_entry, &done);
2325 if (!list_empty(&done))
2328 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2332 /* iopoll may have completed current req */
2333 if (READ_ONCE(req->iopoll_completed))
2334 list_move_tail(&req->inflight_entry, &done);
2341 if (!list_empty(&done))
2342 io_iopoll_complete(ctx, nr_events, &done);
2348 * We can't just wait for polled events to come to us, we have to actively
2349 * find and complete them.
2351 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2353 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2356 mutex_lock(&ctx->uring_lock);
2357 while (!list_empty(&ctx->iopoll_list)) {
2358 unsigned int nr_events = 0;
2360 io_do_iopoll(ctx, &nr_events, 0);
2362 /* let it sleep and repeat later if can't complete a request */
2366 * Ensure we allow local-to-the-cpu processing to take place,
2367 * in this case we need to ensure that we reap all events.
2368 * Also let task_work, etc. to progress by releasing the mutex
2370 if (need_resched()) {
2371 mutex_unlock(&ctx->uring_lock);
2373 mutex_lock(&ctx->uring_lock);
2376 mutex_unlock(&ctx->uring_lock);
2379 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2381 unsigned int nr_events = 0;
2385 * We disallow the app entering submit/complete with polling, but we
2386 * still need to lock the ring to prevent racing with polled issue
2387 * that got punted to a workqueue.
2389 mutex_lock(&ctx->uring_lock);
2391 * Don't enter poll loop if we already have events pending.
2392 * If we do, we can potentially be spinning for commands that
2393 * already triggered a CQE (eg in error).
2395 if (test_bit(0, &ctx->cq_check_overflow))
2396 __io_cqring_overflow_flush(ctx, false);
2397 if (io_cqring_events(ctx))
2401 * If a submit got punted to a workqueue, we can have the
2402 * application entering polling for a command before it gets
2403 * issued. That app will hold the uring_lock for the duration
2404 * of the poll right here, so we need to take a breather every
2405 * now and then to ensure that the issue has a chance to add
2406 * the poll to the issued list. Otherwise we can spin here
2407 * forever, while the workqueue is stuck trying to acquire the
2410 if (list_empty(&ctx->iopoll_list)) {
2411 mutex_unlock(&ctx->uring_lock);
2413 mutex_lock(&ctx->uring_lock);
2415 if (list_empty(&ctx->iopoll_list))
2418 ret = io_do_iopoll(ctx, &nr_events, min);
2419 } while (!ret && nr_events < min && !need_resched());
2421 mutex_unlock(&ctx->uring_lock);
2425 static void kiocb_end_write(struct io_kiocb *req)
2428 * Tell lockdep we inherited freeze protection from submission
2431 if (req->flags & REQ_F_ISREG) {
2432 struct super_block *sb = file_inode(req->file)->i_sb;
2434 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2440 static bool io_resubmit_prep(struct io_kiocb *req)
2442 struct io_async_rw *rw = req->async_data;
2445 return !io_req_prep_async(req);
2446 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2447 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2451 static bool io_rw_should_reissue(struct io_kiocb *req)
2453 umode_t mode = file_inode(req->file)->i_mode;
2454 struct io_ring_ctx *ctx = req->ctx;
2456 if (!S_ISBLK(mode) && !S_ISREG(mode))
2458 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2459 !(ctx->flags & IORING_SETUP_IOPOLL)))
2462 * If ref is dying, we might be running poll reap from the exit work.
2463 * Don't attempt to reissue from that path, just let it fail with
2466 if (percpu_ref_is_dying(&ctx->refs))
2471 static bool io_resubmit_prep(struct io_kiocb *req)
2475 static bool io_rw_should_reissue(struct io_kiocb *req)
2481 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2482 unsigned int issue_flags)
2486 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2487 kiocb_end_write(req);
2488 if (res != req->result) {
2489 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2490 io_rw_should_reissue(req)) {
2491 req->flags |= REQ_F_REISSUE;
2496 if (req->flags & REQ_F_BUFFER_SELECTED)
2497 cflags = io_put_rw_kbuf(req);
2498 __io_req_complete(req, issue_flags, res, cflags);
2501 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2503 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2505 __io_complete_rw(req, res, res2, 0);
2508 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2510 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2512 if (kiocb->ki_flags & IOCB_WRITE)
2513 kiocb_end_write(req);
2514 if (unlikely(res != req->result)) {
2515 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2516 io_resubmit_prep(req))) {
2518 req->flags |= REQ_F_DONT_REISSUE;
2522 WRITE_ONCE(req->result, res);
2523 /* order with io_iopoll_complete() checking ->result */
2525 WRITE_ONCE(req->iopoll_completed, 1);
2529 * After the iocb has been issued, it's safe to be found on the poll list.
2530 * Adding the kiocb to the list AFTER submission ensures that we don't
2531 * find it from a io_do_iopoll() thread before the issuer is done
2532 * accessing the kiocb cookie.
2534 static void io_iopoll_req_issued(struct io_kiocb *req)
2536 struct io_ring_ctx *ctx = req->ctx;
2537 const bool in_async = io_wq_current_is_worker();
2539 /* workqueue context doesn't hold uring_lock, grab it now */
2540 if (unlikely(in_async))
2541 mutex_lock(&ctx->uring_lock);
2544 * Track whether we have multiple files in our lists. This will impact
2545 * how we do polling eventually, not spinning if we're on potentially
2546 * different devices.
2548 if (list_empty(&ctx->iopoll_list)) {
2549 ctx->poll_multi_file = false;
2550 } else if (!ctx->poll_multi_file) {
2551 struct io_kiocb *list_req;
2553 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2555 if (list_req->file != req->file)
2556 ctx->poll_multi_file = true;
2560 * For fast devices, IO may have already completed. If it has, add
2561 * it to the front so we find it first.
2563 if (READ_ONCE(req->iopoll_completed))
2564 list_add(&req->inflight_entry, &ctx->iopoll_list);
2566 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2568 if (unlikely(in_async)) {
2570 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2571 * in sq thread task context or in io worker task context. If
2572 * current task context is sq thread, we don't need to check
2573 * whether should wake up sq thread.
2575 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2576 wq_has_sleeper(&ctx->sq_data->wait))
2577 wake_up(&ctx->sq_data->wait);
2579 mutex_unlock(&ctx->uring_lock);
2583 static inline void io_state_file_put(struct io_submit_state *state)
2585 if (state->file_refs) {
2586 fput_many(state->file, state->file_refs);
2587 state->file_refs = 0;
2592 * Get as many references to a file as we have IOs left in this submission,
2593 * assuming most submissions are for one file, or at least that each file
2594 * has more than one submission.
2596 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2601 if (state->file_refs) {
2602 if (state->fd == fd) {
2606 io_state_file_put(state);
2608 state->file = fget_many(fd, state->ios_left);
2609 if (unlikely(!state->file))
2613 state->file_refs = state->ios_left - 1;
2617 static bool io_bdev_nowait(struct block_device *bdev)
2619 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2623 * If we tracked the file through the SCM inflight mechanism, we could support
2624 * any file. For now, just ensure that anything potentially problematic is done
2627 static bool __io_file_supports_async(struct file *file, int rw)
2629 umode_t mode = file_inode(file)->i_mode;
2631 if (S_ISBLK(mode)) {
2632 if (IS_ENABLED(CONFIG_BLOCK) &&
2633 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2639 if (S_ISREG(mode)) {
2640 if (IS_ENABLED(CONFIG_BLOCK) &&
2641 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2642 file->f_op != &io_uring_fops)
2647 /* any ->read/write should understand O_NONBLOCK */
2648 if (file->f_flags & O_NONBLOCK)
2651 if (!(file->f_mode & FMODE_NOWAIT))
2655 return file->f_op->read_iter != NULL;
2657 return file->f_op->write_iter != NULL;
2660 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2662 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2664 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2667 return __io_file_supports_async(req->file, rw);
2670 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2672 struct io_ring_ctx *ctx = req->ctx;
2673 struct kiocb *kiocb = &req->rw.kiocb;
2674 struct file *file = req->file;
2678 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2679 req->flags |= REQ_F_ISREG;
2681 kiocb->ki_pos = READ_ONCE(sqe->off);
2682 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2683 req->flags |= REQ_F_CUR_POS;
2684 kiocb->ki_pos = file->f_pos;
2686 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2687 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2688 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2692 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2693 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2694 req->flags |= REQ_F_NOWAIT;
2696 ioprio = READ_ONCE(sqe->ioprio);
2698 ret = ioprio_check_cap(ioprio);
2702 kiocb->ki_ioprio = ioprio;
2704 kiocb->ki_ioprio = get_current_ioprio();
2706 if (ctx->flags & IORING_SETUP_IOPOLL) {
2707 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2708 !kiocb->ki_filp->f_op->iopoll)
2711 kiocb->ki_flags |= IOCB_HIPRI;
2712 kiocb->ki_complete = io_complete_rw_iopoll;
2713 req->iopoll_completed = 0;
2715 if (kiocb->ki_flags & IOCB_HIPRI)
2717 kiocb->ki_complete = io_complete_rw;
2720 if (req->opcode == IORING_OP_READ_FIXED ||
2721 req->opcode == IORING_OP_WRITE_FIXED) {
2723 io_req_set_rsrc_node(req);
2726 req->rw.addr = READ_ONCE(sqe->addr);
2727 req->rw.len = READ_ONCE(sqe->len);
2728 req->buf_index = READ_ONCE(sqe->buf_index);
2732 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2738 case -ERESTARTNOINTR:
2739 case -ERESTARTNOHAND:
2740 case -ERESTART_RESTARTBLOCK:
2742 * We can't just restart the syscall, since previously
2743 * submitted sqes may already be in progress. Just fail this
2749 kiocb->ki_complete(kiocb, ret, 0);
2753 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2754 unsigned int issue_flags)
2756 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2757 struct io_async_rw *io = req->async_data;
2758 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2760 /* add previously done IO, if any */
2761 if (io && io->bytes_done > 0) {
2763 ret = io->bytes_done;
2765 ret += io->bytes_done;
2768 if (req->flags & REQ_F_CUR_POS)
2769 req->file->f_pos = kiocb->ki_pos;
2770 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2771 __io_complete_rw(req, ret, 0, issue_flags);
2773 io_rw_done(kiocb, ret);
2775 if (check_reissue && req->flags & REQ_F_REISSUE) {
2776 req->flags &= ~REQ_F_REISSUE;
2777 if (io_resubmit_prep(req)) {
2779 io_queue_async_work(req);
2784 if (req->flags & REQ_F_BUFFER_SELECTED)
2785 cflags = io_put_rw_kbuf(req);
2786 __io_req_complete(req, issue_flags, ret, cflags);
2791 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2792 struct io_mapped_ubuf *imu)
2794 size_t len = req->rw.len;
2795 u64 buf_end, buf_addr = req->rw.addr;
2798 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2800 /* not inside the mapped region */
2801 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2805 * May not be a start of buffer, set size appropriately
2806 * and advance us to the beginning.
2808 offset = buf_addr - imu->ubuf;
2809 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2813 * Don't use iov_iter_advance() here, as it's really slow for
2814 * using the latter parts of a big fixed buffer - it iterates
2815 * over each segment manually. We can cheat a bit here, because
2818 * 1) it's a BVEC iter, we set it up
2819 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2820 * first and last bvec
2822 * So just find our index, and adjust the iterator afterwards.
2823 * If the offset is within the first bvec (or the whole first
2824 * bvec, just use iov_iter_advance(). This makes it easier
2825 * since we can just skip the first segment, which may not
2826 * be PAGE_SIZE aligned.
2828 const struct bio_vec *bvec = imu->bvec;
2830 if (offset <= bvec->bv_len) {
2831 iov_iter_advance(iter, offset);
2833 unsigned long seg_skip;
2835 /* skip first vec */
2836 offset -= bvec->bv_len;
2837 seg_skip = 1 + (offset >> PAGE_SHIFT);
2839 iter->bvec = bvec + seg_skip;
2840 iter->nr_segs -= seg_skip;
2841 iter->count -= bvec->bv_len + offset;
2842 iter->iov_offset = offset & ~PAGE_MASK;
2849 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2851 struct io_ring_ctx *ctx = req->ctx;
2852 struct io_mapped_ubuf *imu = req->imu;
2853 u16 index, buf_index = req->buf_index;
2856 if (unlikely(buf_index >= ctx->nr_user_bufs))
2858 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2859 imu = READ_ONCE(ctx->user_bufs[index]);
2862 return __io_import_fixed(req, rw, iter, imu);
2865 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2868 mutex_unlock(&ctx->uring_lock);
2871 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2874 * "Normal" inline submissions always hold the uring_lock, since we
2875 * grab it from the system call. Same is true for the SQPOLL offload.
2876 * The only exception is when we've detached the request and issue it
2877 * from an async worker thread, grab the lock for that case.
2880 mutex_lock(&ctx->uring_lock);
2883 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2884 int bgid, struct io_buffer *kbuf,
2887 struct io_buffer *head;
2889 if (req->flags & REQ_F_BUFFER_SELECTED)
2892 io_ring_submit_lock(req->ctx, needs_lock);
2894 lockdep_assert_held(&req->ctx->uring_lock);
2896 head = xa_load(&req->ctx->io_buffers, bgid);
2898 if (!list_empty(&head->list)) {
2899 kbuf = list_last_entry(&head->list, struct io_buffer,
2901 list_del(&kbuf->list);
2904 xa_erase(&req->ctx->io_buffers, bgid);
2906 if (*len > kbuf->len)
2909 kbuf = ERR_PTR(-ENOBUFS);
2912 io_ring_submit_unlock(req->ctx, needs_lock);
2917 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2920 struct io_buffer *kbuf;
2923 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2924 bgid = req->buf_index;
2925 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2928 req->rw.addr = (u64) (unsigned long) kbuf;
2929 req->flags |= REQ_F_BUFFER_SELECTED;
2930 return u64_to_user_ptr(kbuf->addr);
2933 #ifdef CONFIG_COMPAT
2934 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2937 struct compat_iovec __user *uiov;
2938 compat_ssize_t clen;
2942 uiov = u64_to_user_ptr(req->rw.addr);
2943 if (!access_ok(uiov, sizeof(*uiov)))
2945 if (__get_user(clen, &uiov->iov_len))
2951 buf = io_rw_buffer_select(req, &len, needs_lock);
2953 return PTR_ERR(buf);
2954 iov[0].iov_base = buf;
2955 iov[0].iov_len = (compat_size_t) len;
2960 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2963 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2967 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2970 len = iov[0].iov_len;
2973 buf = io_rw_buffer_select(req, &len, needs_lock);
2975 return PTR_ERR(buf);
2976 iov[0].iov_base = buf;
2977 iov[0].iov_len = len;
2981 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2984 if (req->flags & REQ_F_BUFFER_SELECTED) {
2985 struct io_buffer *kbuf;
2987 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2988 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2989 iov[0].iov_len = kbuf->len;
2992 if (req->rw.len != 1)
2995 #ifdef CONFIG_COMPAT
2996 if (req->ctx->compat)
2997 return io_compat_import(req, iov, needs_lock);
3000 return __io_iov_buffer_select(req, iov, needs_lock);
3003 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3004 struct iov_iter *iter, bool needs_lock)
3006 void __user *buf = u64_to_user_ptr(req->rw.addr);
3007 size_t sqe_len = req->rw.len;
3008 u8 opcode = req->opcode;
3011 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3013 return io_import_fixed(req, rw, iter);
3016 /* buffer index only valid with fixed read/write, or buffer select */
3017 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3020 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3021 if (req->flags & REQ_F_BUFFER_SELECT) {
3022 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3024 return PTR_ERR(buf);
3025 req->rw.len = sqe_len;
3028 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3033 if (req->flags & REQ_F_BUFFER_SELECT) {
3034 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3036 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3041 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3045 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3047 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3051 * For files that don't have ->read_iter() and ->write_iter(), handle them
3052 * by looping over ->read() or ->write() manually.
3054 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3056 struct kiocb *kiocb = &req->rw.kiocb;
3057 struct file *file = req->file;
3061 * Don't support polled IO through this interface, and we can't
3062 * support non-blocking either. For the latter, this just causes
3063 * the kiocb to be handled from an async context.
3065 if (kiocb->ki_flags & IOCB_HIPRI)
3067 if (kiocb->ki_flags & IOCB_NOWAIT)
3070 while (iov_iter_count(iter)) {
3074 if (!iov_iter_is_bvec(iter)) {
3075 iovec = iov_iter_iovec(iter);
3077 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3078 iovec.iov_len = req->rw.len;
3082 nr = file->f_op->read(file, iovec.iov_base,
3083 iovec.iov_len, io_kiocb_ppos(kiocb));
3085 nr = file->f_op->write(file, iovec.iov_base,
3086 iovec.iov_len, io_kiocb_ppos(kiocb));
3095 if (nr != iovec.iov_len)
3099 iov_iter_advance(iter, nr);
3105 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3106 const struct iovec *fast_iov, struct iov_iter *iter)
3108 struct io_async_rw *rw = req->async_data;
3110 memcpy(&rw->iter, iter, sizeof(*iter));
3111 rw->free_iovec = iovec;
3113 /* can only be fixed buffers, no need to do anything */
3114 if (iov_iter_is_bvec(iter))
3117 unsigned iov_off = 0;
3119 rw->iter.iov = rw->fast_iov;
3120 if (iter->iov != fast_iov) {
3121 iov_off = iter->iov - fast_iov;
3122 rw->iter.iov += iov_off;
3124 if (rw->fast_iov != fast_iov)
3125 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3126 sizeof(struct iovec) * iter->nr_segs);
3128 req->flags |= REQ_F_NEED_CLEANUP;
3132 static inline int io_alloc_async_data(struct io_kiocb *req)
3134 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3135 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3136 return req->async_data == NULL;
3139 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3140 const struct iovec *fast_iov,
3141 struct iov_iter *iter, bool force)
3143 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3145 if (!req->async_data) {
3146 if (io_alloc_async_data(req)) {
3151 io_req_map_rw(req, iovec, fast_iov, iter);
3156 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3158 struct io_async_rw *iorw = req->async_data;
3159 struct iovec *iov = iorw->fast_iov;
3162 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3163 if (unlikely(ret < 0))
3166 iorw->bytes_done = 0;
3167 iorw->free_iovec = iov;
3169 req->flags |= REQ_F_NEED_CLEANUP;
3173 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3175 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3177 return io_prep_rw(req, sqe);
3181 * This is our waitqueue callback handler, registered through lock_page_async()
3182 * when we initially tried to do the IO with the iocb armed our waitqueue.
3183 * This gets called when the page is unlocked, and we generally expect that to
3184 * happen when the page IO is completed and the page is now uptodate. This will
3185 * queue a task_work based retry of the operation, attempting to copy the data
3186 * again. If the latter fails because the page was NOT uptodate, then we will
3187 * do a thread based blocking retry of the operation. That's the unexpected
3190 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3191 int sync, void *arg)
3193 struct wait_page_queue *wpq;
3194 struct io_kiocb *req = wait->private;
3195 struct wait_page_key *key = arg;
3197 wpq = container_of(wait, struct wait_page_queue, wait);
3199 if (!wake_page_match(wpq, key))
3202 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3203 list_del_init(&wait->entry);
3205 /* submit ref gets dropped, acquire a new one */
3207 io_req_task_queue(req);
3212 * This controls whether a given IO request should be armed for async page
3213 * based retry. If we return false here, the request is handed to the async
3214 * worker threads for retry. If we're doing buffered reads on a regular file,
3215 * we prepare a private wait_page_queue entry and retry the operation. This
3216 * will either succeed because the page is now uptodate and unlocked, or it
3217 * will register a callback when the page is unlocked at IO completion. Through
3218 * that callback, io_uring uses task_work to setup a retry of the operation.
3219 * That retry will attempt the buffered read again. The retry will generally
3220 * succeed, or in rare cases where it fails, we then fall back to using the
3221 * async worker threads for a blocking retry.
3223 static bool io_rw_should_retry(struct io_kiocb *req)
3225 struct io_async_rw *rw = req->async_data;
3226 struct wait_page_queue *wait = &rw->wpq;
3227 struct kiocb *kiocb = &req->rw.kiocb;
3229 /* never retry for NOWAIT, we just complete with -EAGAIN */
3230 if (req->flags & REQ_F_NOWAIT)
3233 /* Only for buffered IO */
3234 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3238 * just use poll if we can, and don't attempt if the fs doesn't
3239 * support callback based unlocks
3241 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3244 wait->wait.func = io_async_buf_func;
3245 wait->wait.private = req;
3246 wait->wait.flags = 0;
3247 INIT_LIST_HEAD(&wait->wait.entry);
3248 kiocb->ki_flags |= IOCB_WAITQ;
3249 kiocb->ki_flags &= ~IOCB_NOWAIT;
3250 kiocb->ki_waitq = wait;
3254 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3256 if (req->file->f_op->read_iter)
3257 return call_read_iter(req->file, &req->rw.kiocb, iter);
3258 else if (req->file->f_op->read)
3259 return loop_rw_iter(READ, req, iter);
3264 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3266 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3267 struct kiocb *kiocb = &req->rw.kiocb;
3268 struct iov_iter __iter, *iter = &__iter;
3269 struct io_async_rw *rw = req->async_data;
3270 ssize_t io_size, ret, ret2;
3271 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3277 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3281 io_size = iov_iter_count(iter);
3282 req->result = io_size;
3284 /* Ensure we clear previously set non-block flag */
3285 if (!force_nonblock)
3286 kiocb->ki_flags &= ~IOCB_NOWAIT;
3288 kiocb->ki_flags |= IOCB_NOWAIT;
3290 /* If the file doesn't support async, just async punt */
3291 if (force_nonblock && !io_file_supports_async(req, READ)) {
3292 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3293 return ret ?: -EAGAIN;
3296 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3297 if (unlikely(ret)) {
3302 ret = io_iter_do_read(req, iter);
3304 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3305 req->flags &= ~REQ_F_REISSUE;
3306 /* IOPOLL retry should happen for io-wq threads */
3307 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3309 /* no retry on NONBLOCK nor RWF_NOWAIT */
3310 if (req->flags & REQ_F_NOWAIT)
3312 /* some cases will consume bytes even on error returns */
3313 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3315 } else if (ret == -EIOCBQUEUED) {
3317 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3318 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3319 /* read all, failed, already did sync or don't want to retry */
3323 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3328 rw = req->async_data;
3329 /* now use our persistent iterator, if we aren't already */
3334 rw->bytes_done += ret;
3335 /* if we can retry, do so with the callbacks armed */
3336 if (!io_rw_should_retry(req)) {
3337 kiocb->ki_flags &= ~IOCB_WAITQ;
3342 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3343 * we get -EIOCBQUEUED, then we'll get a notification when the
3344 * desired page gets unlocked. We can also get a partial read
3345 * here, and if we do, then just retry at the new offset.
3347 ret = io_iter_do_read(req, iter);
3348 if (ret == -EIOCBQUEUED)
3350 /* we got some bytes, but not all. retry. */
3351 kiocb->ki_flags &= ~IOCB_WAITQ;
3352 } while (ret > 0 && ret < io_size);
3354 kiocb_done(kiocb, ret, issue_flags);
3356 /* it's faster to check here then delegate to kfree */
3362 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3364 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3366 return io_prep_rw(req, sqe);
3369 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3371 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3372 struct kiocb *kiocb = &req->rw.kiocb;
3373 struct iov_iter __iter, *iter = &__iter;
3374 struct io_async_rw *rw = req->async_data;
3375 ssize_t ret, ret2, io_size;
3376 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3382 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3386 io_size = iov_iter_count(iter);
3387 req->result = io_size;
3389 /* Ensure we clear previously set non-block flag */
3390 if (!force_nonblock)
3391 kiocb->ki_flags &= ~IOCB_NOWAIT;
3393 kiocb->ki_flags |= IOCB_NOWAIT;
3395 /* If the file doesn't support async, just async punt */
3396 if (force_nonblock && !io_file_supports_async(req, WRITE))
3399 /* file path doesn't support NOWAIT for non-direct_IO */
3400 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3401 (req->flags & REQ_F_ISREG))
3404 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3409 * Open-code file_start_write here to grab freeze protection,
3410 * which will be released by another thread in
3411 * io_complete_rw(). Fool lockdep by telling it the lock got
3412 * released so that it doesn't complain about the held lock when
3413 * we return to userspace.
3415 if (req->flags & REQ_F_ISREG) {
3416 sb_start_write(file_inode(req->file)->i_sb);
3417 __sb_writers_release(file_inode(req->file)->i_sb,
3420 kiocb->ki_flags |= IOCB_WRITE;
3422 if (req->file->f_op->write_iter)
3423 ret2 = call_write_iter(req->file, kiocb, iter);
3424 else if (req->file->f_op->write)
3425 ret2 = loop_rw_iter(WRITE, req, iter);
3429 if (req->flags & REQ_F_REISSUE) {
3430 req->flags &= ~REQ_F_REISSUE;
3435 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3436 * retry them without IOCB_NOWAIT.
3438 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3440 /* no retry on NONBLOCK nor RWF_NOWAIT */
3441 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3443 if (!force_nonblock || ret2 != -EAGAIN) {
3444 /* IOPOLL retry should happen for io-wq threads */
3445 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3448 kiocb_done(kiocb, ret2, issue_flags);
3451 /* some cases will consume bytes even on error returns */
3452 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3453 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3454 return ret ?: -EAGAIN;
3457 /* it's reportedly faster than delegating the null check to kfree() */
3463 static int io_renameat_prep(struct io_kiocb *req,
3464 const struct io_uring_sqe *sqe)
3466 struct io_rename *ren = &req->rename;
3467 const char __user *oldf, *newf;
3469 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3472 ren->old_dfd = READ_ONCE(sqe->fd);
3473 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3474 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3475 ren->new_dfd = READ_ONCE(sqe->len);
3476 ren->flags = READ_ONCE(sqe->rename_flags);
3478 ren->oldpath = getname(oldf);
3479 if (IS_ERR(ren->oldpath))
3480 return PTR_ERR(ren->oldpath);
3482 ren->newpath = getname(newf);
3483 if (IS_ERR(ren->newpath)) {
3484 putname(ren->oldpath);
3485 return PTR_ERR(ren->newpath);
3488 req->flags |= REQ_F_NEED_CLEANUP;
3492 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3494 struct io_rename *ren = &req->rename;
3497 if (issue_flags & IO_URING_F_NONBLOCK)
3500 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3501 ren->newpath, ren->flags);
3503 req->flags &= ~REQ_F_NEED_CLEANUP;
3506 io_req_complete(req, ret);
3510 static int io_unlinkat_prep(struct io_kiocb *req,
3511 const struct io_uring_sqe *sqe)
3513 struct io_unlink *un = &req->unlink;
3514 const char __user *fname;
3516 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3519 un->dfd = READ_ONCE(sqe->fd);
3521 un->flags = READ_ONCE(sqe->unlink_flags);
3522 if (un->flags & ~AT_REMOVEDIR)
3525 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3526 un->filename = getname(fname);
3527 if (IS_ERR(un->filename))
3528 return PTR_ERR(un->filename);
3530 req->flags |= REQ_F_NEED_CLEANUP;
3534 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3536 struct io_unlink *un = &req->unlink;
3539 if (issue_flags & IO_URING_F_NONBLOCK)
3542 if (un->flags & AT_REMOVEDIR)
3543 ret = do_rmdir(un->dfd, un->filename);
3545 ret = do_unlinkat(un->dfd, un->filename);
3547 req->flags &= ~REQ_F_NEED_CLEANUP;
3550 io_req_complete(req, ret);
3554 static int io_shutdown_prep(struct io_kiocb *req,
3555 const struct io_uring_sqe *sqe)
3557 #if defined(CONFIG_NET)
3558 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3560 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3564 req->shutdown.how = READ_ONCE(sqe->len);
3571 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3573 #if defined(CONFIG_NET)
3574 struct socket *sock;
3577 if (issue_flags & IO_URING_F_NONBLOCK)
3580 sock = sock_from_file(req->file);
3581 if (unlikely(!sock))
3584 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3587 io_req_complete(req, ret);
3594 static int __io_splice_prep(struct io_kiocb *req,
3595 const struct io_uring_sqe *sqe)
3597 struct io_splice* sp = &req->splice;
3598 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3600 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3604 sp->len = READ_ONCE(sqe->len);
3605 sp->flags = READ_ONCE(sqe->splice_flags);
3607 if (unlikely(sp->flags & ~valid_flags))
3610 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3611 (sp->flags & SPLICE_F_FD_IN_FIXED));
3614 req->flags |= REQ_F_NEED_CLEANUP;
3618 static int io_tee_prep(struct io_kiocb *req,
3619 const struct io_uring_sqe *sqe)
3621 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3623 return __io_splice_prep(req, sqe);
3626 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3628 struct io_splice *sp = &req->splice;
3629 struct file *in = sp->file_in;
3630 struct file *out = sp->file_out;
3631 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3634 if (issue_flags & IO_URING_F_NONBLOCK)
3637 ret = do_tee(in, out, sp->len, flags);
3639 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3641 req->flags &= ~REQ_F_NEED_CLEANUP;
3645 io_req_complete(req, ret);
3649 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3651 struct io_splice* sp = &req->splice;
3653 sp->off_in = READ_ONCE(sqe->splice_off_in);
3654 sp->off_out = READ_ONCE(sqe->off);
3655 return __io_splice_prep(req, sqe);
3658 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3660 struct io_splice *sp = &req->splice;
3661 struct file *in = sp->file_in;
3662 struct file *out = sp->file_out;
3663 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3664 loff_t *poff_in, *poff_out;
3667 if (issue_flags & IO_URING_F_NONBLOCK)
3670 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3671 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3674 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3676 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3678 req->flags &= ~REQ_F_NEED_CLEANUP;
3682 io_req_complete(req, ret);
3687 * IORING_OP_NOP just posts a completion event, nothing else.
3689 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3691 struct io_ring_ctx *ctx = req->ctx;
3693 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3696 __io_req_complete(req, issue_flags, 0, 0);
3700 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3702 struct io_ring_ctx *ctx = req->ctx;
3707 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3709 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3712 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3713 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3716 req->sync.off = READ_ONCE(sqe->off);
3717 req->sync.len = READ_ONCE(sqe->len);
3721 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3723 loff_t end = req->sync.off + req->sync.len;
3726 /* fsync always requires a blocking context */
3727 if (issue_flags & IO_URING_F_NONBLOCK)
3730 ret = vfs_fsync_range(req->file, req->sync.off,
3731 end > 0 ? end : LLONG_MAX,
3732 req->sync.flags & IORING_FSYNC_DATASYNC);
3735 io_req_complete(req, ret);
3739 static int io_fallocate_prep(struct io_kiocb *req,
3740 const struct io_uring_sqe *sqe)
3742 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3744 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3747 req->sync.off = READ_ONCE(sqe->off);
3748 req->sync.len = READ_ONCE(sqe->addr);
3749 req->sync.mode = READ_ONCE(sqe->len);
3753 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3757 /* fallocate always requiring blocking context */
3758 if (issue_flags & IO_URING_F_NONBLOCK)
3760 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3764 io_req_complete(req, ret);
3768 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3770 const char __user *fname;
3773 if (unlikely(sqe->ioprio || sqe->buf_index))
3775 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3778 /* open.how should be already initialised */
3779 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3780 req->open.how.flags |= O_LARGEFILE;
3782 req->open.dfd = READ_ONCE(sqe->fd);
3783 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3784 req->open.filename = getname(fname);
3785 if (IS_ERR(req->open.filename)) {
3786 ret = PTR_ERR(req->open.filename);
3787 req->open.filename = NULL;
3790 req->open.nofile = rlimit(RLIMIT_NOFILE);
3791 req->flags |= REQ_F_NEED_CLEANUP;
3795 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3799 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3801 mode = READ_ONCE(sqe->len);
3802 flags = READ_ONCE(sqe->open_flags);
3803 req->open.how = build_open_how(flags, mode);
3804 return __io_openat_prep(req, sqe);
3807 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3809 struct open_how __user *how;
3813 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3815 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3816 len = READ_ONCE(sqe->len);
3817 if (len < OPEN_HOW_SIZE_VER0)
3820 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3825 return __io_openat_prep(req, sqe);
3828 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3830 struct open_flags op;
3833 bool resolve_nonblock;
3836 ret = build_open_flags(&req->open.how, &op);
3839 nonblock_set = op.open_flag & O_NONBLOCK;
3840 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3841 if (issue_flags & IO_URING_F_NONBLOCK) {
3843 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3844 * it'll always -EAGAIN
3846 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3848 op.lookup_flags |= LOOKUP_CACHED;
3849 op.open_flag |= O_NONBLOCK;
3852 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3856 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3857 /* only retry if RESOLVE_CACHED wasn't already set by application */
3858 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3859 file == ERR_PTR(-EAGAIN)) {
3861 * We could hang on to this 'fd', but seems like marginal
3862 * gain for something that is now known to be a slower path.
3863 * So just put it, and we'll get a new one when we retry.
3871 ret = PTR_ERR(file);
3873 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3874 file->f_flags &= ~O_NONBLOCK;
3875 fsnotify_open(file);
3876 fd_install(ret, file);
3879 putname(req->open.filename);
3880 req->flags &= ~REQ_F_NEED_CLEANUP;
3883 __io_req_complete(req, issue_flags, ret, 0);
3887 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3889 return io_openat2(req, issue_flags);
3892 static int io_remove_buffers_prep(struct io_kiocb *req,
3893 const struct io_uring_sqe *sqe)
3895 struct io_provide_buf *p = &req->pbuf;
3898 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3901 tmp = READ_ONCE(sqe->fd);
3902 if (!tmp || tmp > USHRT_MAX)
3905 memset(p, 0, sizeof(*p));
3907 p->bgid = READ_ONCE(sqe->buf_group);
3911 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3912 int bgid, unsigned nbufs)
3916 /* shouldn't happen */
3920 /* the head kbuf is the list itself */
3921 while (!list_empty(&buf->list)) {
3922 struct io_buffer *nxt;
3924 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3925 list_del(&nxt->list);
3932 xa_erase(&ctx->io_buffers, bgid);
3937 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3939 struct io_provide_buf *p = &req->pbuf;
3940 struct io_ring_ctx *ctx = req->ctx;
3941 struct io_buffer *head;
3943 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3945 io_ring_submit_lock(ctx, !force_nonblock);
3947 lockdep_assert_held(&ctx->uring_lock);
3950 head = xa_load(&ctx->io_buffers, p->bgid);
3952 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3956 /* complete before unlock, IOPOLL may need the lock */
3957 __io_req_complete(req, issue_flags, ret, 0);
3958 io_ring_submit_unlock(ctx, !force_nonblock);
3962 static int io_provide_buffers_prep(struct io_kiocb *req,
3963 const struct io_uring_sqe *sqe)
3965 unsigned long size, tmp_check;
3966 struct io_provide_buf *p = &req->pbuf;
3969 if (sqe->ioprio || sqe->rw_flags)
3972 tmp = READ_ONCE(sqe->fd);
3973 if (!tmp || tmp > USHRT_MAX)
3976 p->addr = READ_ONCE(sqe->addr);
3977 p->len = READ_ONCE(sqe->len);
3979 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3982 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3985 size = (unsigned long)p->len * p->nbufs;
3986 if (!access_ok(u64_to_user_ptr(p->addr), size))
3989 p->bgid = READ_ONCE(sqe->buf_group);
3990 tmp = READ_ONCE(sqe->off);
3991 if (tmp > USHRT_MAX)
3997 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3999 struct io_buffer *buf;
4000 u64 addr = pbuf->addr;
4001 int i, bid = pbuf->bid;
4003 for (i = 0; i < pbuf->nbufs; i++) {
4004 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4009 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4014 INIT_LIST_HEAD(&buf->list);
4017 list_add_tail(&buf->list, &(*head)->list);
4021 return i ? i : -ENOMEM;
4024 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4026 struct io_provide_buf *p = &req->pbuf;
4027 struct io_ring_ctx *ctx = req->ctx;
4028 struct io_buffer *head, *list;
4030 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4032 io_ring_submit_lock(ctx, !force_nonblock);
4034 lockdep_assert_held(&ctx->uring_lock);
4036 list = head = xa_load(&ctx->io_buffers, p->bgid);
4038 ret = io_add_buffers(p, &head);
4039 if (ret >= 0 && !list) {
4040 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4042 __io_remove_buffers(ctx, head, p->bgid, -1U);
4046 /* complete before unlock, IOPOLL may need the lock */
4047 __io_req_complete(req, issue_flags, ret, 0);
4048 io_ring_submit_unlock(ctx, !force_nonblock);
4052 static int io_epoll_ctl_prep(struct io_kiocb *req,
4053 const struct io_uring_sqe *sqe)
4055 #if defined(CONFIG_EPOLL)
4056 if (sqe->ioprio || sqe->buf_index)
4058 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4061 req->epoll.epfd = READ_ONCE(sqe->fd);
4062 req->epoll.op = READ_ONCE(sqe->len);
4063 req->epoll.fd = READ_ONCE(sqe->off);
4065 if (ep_op_has_event(req->epoll.op)) {
4066 struct epoll_event __user *ev;
4068 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4069 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4079 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4081 #if defined(CONFIG_EPOLL)
4082 struct io_epoll *ie = &req->epoll;
4084 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4086 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4087 if (force_nonblock && ret == -EAGAIN)
4092 __io_req_complete(req, issue_flags, ret, 0);
4099 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4101 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4102 if (sqe->ioprio || sqe->buf_index || sqe->off)
4104 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4107 req->madvise.addr = READ_ONCE(sqe->addr);
4108 req->madvise.len = READ_ONCE(sqe->len);
4109 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4116 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4118 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4119 struct io_madvise *ma = &req->madvise;
4122 if (issue_flags & IO_URING_F_NONBLOCK)
4125 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4128 io_req_complete(req, ret);
4135 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4137 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4139 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4142 req->fadvise.offset = READ_ONCE(sqe->off);
4143 req->fadvise.len = READ_ONCE(sqe->len);
4144 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4148 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4150 struct io_fadvise *fa = &req->fadvise;
4153 if (issue_flags & IO_URING_F_NONBLOCK) {
4154 switch (fa->advice) {
4155 case POSIX_FADV_NORMAL:
4156 case POSIX_FADV_RANDOM:
4157 case POSIX_FADV_SEQUENTIAL:
4164 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4167 __io_req_complete(req, issue_flags, ret, 0);
4171 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4173 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4175 if (sqe->ioprio || sqe->buf_index)
4177 if (req->flags & REQ_F_FIXED_FILE)
4180 req->statx.dfd = READ_ONCE(sqe->fd);
4181 req->statx.mask = READ_ONCE(sqe->len);
4182 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4183 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4184 req->statx.flags = READ_ONCE(sqe->statx_flags);
4189 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4191 struct io_statx *ctx = &req->statx;
4194 if (issue_flags & IO_URING_F_NONBLOCK)
4197 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4202 io_req_complete(req, ret);
4206 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4208 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4210 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4211 sqe->rw_flags || sqe->buf_index)
4213 if (req->flags & REQ_F_FIXED_FILE)
4216 req->close.fd = READ_ONCE(sqe->fd);
4220 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4222 struct files_struct *files = current->files;
4223 struct io_close *close = &req->close;
4224 struct fdtable *fdt;
4225 struct file *file = NULL;
4228 spin_lock(&files->file_lock);
4229 fdt = files_fdtable(files);
4230 if (close->fd >= fdt->max_fds) {
4231 spin_unlock(&files->file_lock);
4234 file = fdt->fd[close->fd];
4235 if (!file || file->f_op == &io_uring_fops) {
4236 spin_unlock(&files->file_lock);
4241 /* if the file has a flush method, be safe and punt to async */
4242 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4243 spin_unlock(&files->file_lock);
4247 ret = __close_fd_get_file(close->fd, &file);
4248 spin_unlock(&files->file_lock);
4255 /* No ->flush() or already async, safely close from here */
4256 ret = filp_close(file, current->files);
4262 __io_req_complete(req, issue_flags, ret, 0);
4266 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4268 struct io_ring_ctx *ctx = req->ctx;
4270 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4272 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4275 req->sync.off = READ_ONCE(sqe->off);
4276 req->sync.len = READ_ONCE(sqe->len);
4277 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4281 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4285 /* sync_file_range always requires a blocking context */
4286 if (issue_flags & IO_URING_F_NONBLOCK)
4289 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4293 io_req_complete(req, ret);
4297 #if defined(CONFIG_NET)
4298 static int io_setup_async_msg(struct io_kiocb *req,
4299 struct io_async_msghdr *kmsg)
4301 struct io_async_msghdr *async_msg = req->async_data;
4305 if (io_alloc_async_data(req)) {
4306 kfree(kmsg->free_iov);
4309 async_msg = req->async_data;
4310 req->flags |= REQ_F_NEED_CLEANUP;
4311 memcpy(async_msg, kmsg, sizeof(*kmsg));
4312 async_msg->msg.msg_name = &async_msg->addr;
4313 /* if were using fast_iov, set it to the new one */
4314 if (!async_msg->free_iov)
4315 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4320 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4321 struct io_async_msghdr *iomsg)
4323 iomsg->msg.msg_name = &iomsg->addr;
4324 iomsg->free_iov = iomsg->fast_iov;
4325 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4326 req->sr_msg.msg_flags, &iomsg->free_iov);
4329 static int io_sendmsg_prep_async(struct io_kiocb *req)
4333 ret = io_sendmsg_copy_hdr(req, req->async_data);
4335 req->flags |= REQ_F_NEED_CLEANUP;
4339 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4341 struct io_sr_msg *sr = &req->sr_msg;
4343 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4346 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4347 sr->len = READ_ONCE(sqe->len);
4348 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4349 if (sr->msg_flags & MSG_DONTWAIT)
4350 req->flags |= REQ_F_NOWAIT;
4352 #ifdef CONFIG_COMPAT
4353 if (req->ctx->compat)
4354 sr->msg_flags |= MSG_CMSG_COMPAT;
4359 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4361 struct io_async_msghdr iomsg, *kmsg;
4362 struct socket *sock;
4367 sock = sock_from_file(req->file);
4368 if (unlikely(!sock))
4371 kmsg = req->async_data;
4373 ret = io_sendmsg_copy_hdr(req, &iomsg);
4379 flags = req->sr_msg.msg_flags;
4380 if (issue_flags & IO_URING_F_NONBLOCK)
4381 flags |= MSG_DONTWAIT;
4382 if (flags & MSG_WAITALL)
4383 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4385 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4386 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4387 return io_setup_async_msg(req, kmsg);
4388 if (ret == -ERESTARTSYS)
4391 /* fast path, check for non-NULL to avoid function call */
4393 kfree(kmsg->free_iov);
4394 req->flags &= ~REQ_F_NEED_CLEANUP;
4397 __io_req_complete(req, issue_flags, ret, 0);
4401 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4403 struct io_sr_msg *sr = &req->sr_msg;
4406 struct socket *sock;
4411 sock = sock_from_file(req->file);
4412 if (unlikely(!sock))
4415 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4419 msg.msg_name = NULL;
4420 msg.msg_control = NULL;
4421 msg.msg_controllen = 0;
4422 msg.msg_namelen = 0;
4424 flags = req->sr_msg.msg_flags;
4425 if (issue_flags & IO_URING_F_NONBLOCK)
4426 flags |= MSG_DONTWAIT;
4427 if (flags & MSG_WAITALL)
4428 min_ret = iov_iter_count(&msg.msg_iter);
4430 msg.msg_flags = flags;
4431 ret = sock_sendmsg(sock, &msg);
4432 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4434 if (ret == -ERESTARTSYS)
4439 __io_req_complete(req, issue_flags, ret, 0);
4443 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4444 struct io_async_msghdr *iomsg)
4446 struct io_sr_msg *sr = &req->sr_msg;
4447 struct iovec __user *uiov;
4451 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4452 &iomsg->uaddr, &uiov, &iov_len);
4456 if (req->flags & REQ_F_BUFFER_SELECT) {
4459 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4461 sr->len = iomsg->fast_iov[0].iov_len;
4462 iomsg->free_iov = NULL;
4464 iomsg->free_iov = iomsg->fast_iov;
4465 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4466 &iomsg->free_iov, &iomsg->msg.msg_iter,
4475 #ifdef CONFIG_COMPAT
4476 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4477 struct io_async_msghdr *iomsg)
4479 struct io_sr_msg *sr = &req->sr_msg;
4480 struct compat_iovec __user *uiov;
4485 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4490 uiov = compat_ptr(ptr);
4491 if (req->flags & REQ_F_BUFFER_SELECT) {
4492 compat_ssize_t clen;
4496 if (!access_ok(uiov, sizeof(*uiov)))
4498 if (__get_user(clen, &uiov->iov_len))
4503 iomsg->free_iov = NULL;
4505 iomsg->free_iov = iomsg->fast_iov;
4506 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4507 UIO_FASTIOV, &iomsg->free_iov,
4508 &iomsg->msg.msg_iter, true);
4517 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4518 struct io_async_msghdr *iomsg)
4520 iomsg->msg.msg_name = &iomsg->addr;
4522 #ifdef CONFIG_COMPAT
4523 if (req->ctx->compat)
4524 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4527 return __io_recvmsg_copy_hdr(req, iomsg);
4530 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4533 struct io_sr_msg *sr = &req->sr_msg;
4534 struct io_buffer *kbuf;
4536 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4541 req->flags |= REQ_F_BUFFER_SELECTED;
4545 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4547 return io_put_kbuf(req, req->sr_msg.kbuf);
4550 static int io_recvmsg_prep_async(struct io_kiocb *req)
4554 ret = io_recvmsg_copy_hdr(req, req->async_data);
4556 req->flags |= REQ_F_NEED_CLEANUP;
4560 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4562 struct io_sr_msg *sr = &req->sr_msg;
4564 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4567 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4568 sr->len = READ_ONCE(sqe->len);
4569 sr->bgid = READ_ONCE(sqe->buf_group);
4570 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4571 if (sr->msg_flags & MSG_DONTWAIT)
4572 req->flags |= REQ_F_NOWAIT;
4574 #ifdef CONFIG_COMPAT
4575 if (req->ctx->compat)
4576 sr->msg_flags |= MSG_CMSG_COMPAT;
4581 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4583 struct io_async_msghdr iomsg, *kmsg;
4584 struct socket *sock;
4585 struct io_buffer *kbuf;
4588 int ret, cflags = 0;
4589 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4591 sock = sock_from_file(req->file);
4592 if (unlikely(!sock))
4595 kmsg = req->async_data;
4597 ret = io_recvmsg_copy_hdr(req, &iomsg);
4603 if (req->flags & REQ_F_BUFFER_SELECT) {
4604 kbuf = io_recv_buffer_select(req, !force_nonblock);
4606 return PTR_ERR(kbuf);
4607 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4608 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4609 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4610 1, req->sr_msg.len);
4613 flags = req->sr_msg.msg_flags;
4615 flags |= MSG_DONTWAIT;
4616 if (flags & MSG_WAITALL)
4617 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4619 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4620 kmsg->uaddr, flags);
4621 if (force_nonblock && ret == -EAGAIN)
4622 return io_setup_async_msg(req, kmsg);
4623 if (ret == -ERESTARTSYS)
4626 if (req->flags & REQ_F_BUFFER_SELECTED)
4627 cflags = io_put_recv_kbuf(req);
4628 /* fast path, check for non-NULL to avoid function call */
4630 kfree(kmsg->free_iov);
4631 req->flags &= ~REQ_F_NEED_CLEANUP;
4632 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4634 __io_req_complete(req, issue_flags, ret, cflags);
4638 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4640 struct io_buffer *kbuf;
4641 struct io_sr_msg *sr = &req->sr_msg;
4643 void __user *buf = sr->buf;
4644 struct socket *sock;
4648 int ret, cflags = 0;
4649 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4651 sock = sock_from_file(req->file);
4652 if (unlikely(!sock))
4655 if (req->flags & REQ_F_BUFFER_SELECT) {
4656 kbuf = io_recv_buffer_select(req, !force_nonblock);
4658 return PTR_ERR(kbuf);
4659 buf = u64_to_user_ptr(kbuf->addr);
4662 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4666 msg.msg_name = NULL;
4667 msg.msg_control = NULL;
4668 msg.msg_controllen = 0;
4669 msg.msg_namelen = 0;
4670 msg.msg_iocb = NULL;
4673 flags = req->sr_msg.msg_flags;
4675 flags |= MSG_DONTWAIT;
4676 if (flags & MSG_WAITALL)
4677 min_ret = iov_iter_count(&msg.msg_iter);
4679 ret = sock_recvmsg(sock, &msg, flags);
4680 if (force_nonblock && ret == -EAGAIN)
4682 if (ret == -ERESTARTSYS)
4685 if (req->flags & REQ_F_BUFFER_SELECTED)
4686 cflags = io_put_recv_kbuf(req);
4687 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4689 __io_req_complete(req, issue_flags, ret, cflags);
4693 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4695 struct io_accept *accept = &req->accept;
4697 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4699 if (sqe->ioprio || sqe->len || sqe->buf_index)
4702 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4703 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4704 accept->flags = READ_ONCE(sqe->accept_flags);
4705 accept->nofile = rlimit(RLIMIT_NOFILE);
4709 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4711 struct io_accept *accept = &req->accept;
4712 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4713 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4716 if (req->file->f_flags & O_NONBLOCK)
4717 req->flags |= REQ_F_NOWAIT;
4719 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4720 accept->addr_len, accept->flags,
4722 if (ret == -EAGAIN && force_nonblock)
4725 if (ret == -ERESTARTSYS)
4729 __io_req_complete(req, issue_flags, ret, 0);
4733 static int io_connect_prep_async(struct io_kiocb *req)
4735 struct io_async_connect *io = req->async_data;
4736 struct io_connect *conn = &req->connect;
4738 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4741 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4743 struct io_connect *conn = &req->connect;
4745 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4747 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4750 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4751 conn->addr_len = READ_ONCE(sqe->addr2);
4755 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4757 struct io_async_connect __io, *io;
4758 unsigned file_flags;
4760 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4762 if (req->async_data) {
4763 io = req->async_data;
4765 ret = move_addr_to_kernel(req->connect.addr,
4766 req->connect.addr_len,
4773 file_flags = force_nonblock ? O_NONBLOCK : 0;
4775 ret = __sys_connect_file(req->file, &io->address,
4776 req->connect.addr_len, file_flags);
4777 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4778 if (req->async_data)
4780 if (io_alloc_async_data(req)) {
4784 memcpy(req->async_data, &__io, sizeof(__io));
4787 if (ret == -ERESTARTSYS)
4792 __io_req_complete(req, issue_flags, ret, 0);
4795 #else /* !CONFIG_NET */
4796 #define IO_NETOP_FN(op) \
4797 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4799 return -EOPNOTSUPP; \
4802 #define IO_NETOP_PREP(op) \
4804 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4806 return -EOPNOTSUPP; \
4809 #define IO_NETOP_PREP_ASYNC(op) \
4811 static int io_##op##_prep_async(struct io_kiocb *req) \
4813 return -EOPNOTSUPP; \
4816 IO_NETOP_PREP_ASYNC(sendmsg);
4817 IO_NETOP_PREP_ASYNC(recvmsg);
4818 IO_NETOP_PREP_ASYNC(connect);
4819 IO_NETOP_PREP(accept);
4822 #endif /* CONFIG_NET */
4824 struct io_poll_table {
4825 struct poll_table_struct pt;
4826 struct io_kiocb *req;
4830 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4831 __poll_t mask, task_work_func_t func)
4835 /* for instances that support it check for an event match first: */
4836 if (mask && !(mask & poll->events))
4839 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4841 list_del_init(&poll->wait.entry);
4844 req->task_work.func = func;
4847 * If this fails, then the task is exiting. When a task exits, the
4848 * work gets canceled, so just cancel this request as well instead
4849 * of executing it. We can't safely execute it anyway, as we may not
4850 * have the needed state needed for it anyway.
4852 ret = io_req_task_work_add(req);
4853 if (unlikely(ret)) {
4854 WRITE_ONCE(poll->canceled, true);
4855 io_req_task_work_add_fallback(req, func);
4860 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4861 __acquires(&req->ctx->completion_lock)
4863 struct io_ring_ctx *ctx = req->ctx;
4865 if (!req->result && !READ_ONCE(poll->canceled)) {
4866 struct poll_table_struct pt = { ._key = poll->events };
4868 req->result = vfs_poll(req->file, &pt) & poll->events;
4871 spin_lock_irq(&ctx->completion_lock);
4872 if (!req->result && !READ_ONCE(poll->canceled)) {
4873 add_wait_queue(poll->head, &poll->wait);
4880 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4882 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4883 if (req->opcode == IORING_OP_POLL_ADD)
4884 return req->async_data;
4885 return req->apoll->double_poll;
4888 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4890 if (req->opcode == IORING_OP_POLL_ADD)
4892 return &req->apoll->poll;
4895 static void io_poll_remove_double(struct io_kiocb *req)
4896 __must_hold(&req->ctx->completion_lock)
4898 struct io_poll_iocb *poll = io_poll_get_double(req);
4900 lockdep_assert_held(&req->ctx->completion_lock);
4902 if (poll && poll->head) {
4903 struct wait_queue_head *head = poll->head;
4905 spin_lock(&head->lock);
4906 list_del_init(&poll->wait.entry);
4907 if (poll->wait.private)
4910 spin_unlock(&head->lock);
4914 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4915 __must_hold(&req->ctx->completion_lock)
4917 struct io_ring_ctx *ctx = req->ctx;
4918 unsigned flags = IORING_CQE_F_MORE;
4921 if (READ_ONCE(req->poll.canceled)) {
4923 req->poll.events |= EPOLLONESHOT;
4925 error = mangle_poll(mask);
4927 if (req->poll.events & EPOLLONESHOT)
4929 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4930 io_poll_remove_waitqs(req);
4931 req->poll.done = true;
4934 if (flags & IORING_CQE_F_MORE)
4937 io_commit_cqring(ctx);
4938 return !(flags & IORING_CQE_F_MORE);
4941 static void io_poll_task_func(struct callback_head *cb)
4943 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4944 struct io_ring_ctx *ctx = req->ctx;
4945 struct io_kiocb *nxt;
4947 if (io_poll_rewait(req, &req->poll)) {
4948 spin_unlock_irq(&ctx->completion_lock);
4952 done = io_poll_complete(req, req->result);
4954 hash_del(&req->hash_node);
4957 add_wait_queue(req->poll.head, &req->poll.wait);
4959 spin_unlock_irq(&ctx->completion_lock);
4960 io_cqring_ev_posted(ctx);
4963 nxt = io_put_req_find_next(req);
4965 __io_req_task_submit(nxt);
4970 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4971 int sync, void *key)
4973 struct io_kiocb *req = wait->private;
4974 struct io_poll_iocb *poll = io_poll_get_single(req);
4975 __poll_t mask = key_to_poll(key);
4977 /* for instances that support it check for an event match first: */
4978 if (mask && !(mask & poll->events))
4980 if (!(poll->events & EPOLLONESHOT))
4981 return poll->wait.func(&poll->wait, mode, sync, key);
4983 list_del_init(&wait->entry);
4985 if (poll && poll->head) {
4988 spin_lock(&poll->head->lock);
4989 done = list_empty(&poll->wait.entry);
4991 list_del_init(&poll->wait.entry);
4992 /* make sure double remove sees this as being gone */
4993 wait->private = NULL;
4994 spin_unlock(&poll->head->lock);
4996 /* use wait func handler, so it matches the rq type */
4997 poll->wait.func(&poll->wait, mode, sync, key);
5004 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5005 wait_queue_func_t wake_func)
5009 poll->canceled = false;
5010 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5011 /* mask in events that we always want/need */
5012 poll->events = events | IO_POLL_UNMASK;
5013 INIT_LIST_HEAD(&poll->wait.entry);
5014 init_waitqueue_func_entry(&poll->wait, wake_func);
5017 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5018 struct wait_queue_head *head,
5019 struct io_poll_iocb **poll_ptr)
5021 struct io_kiocb *req = pt->req;
5024 * If poll->head is already set, it's because the file being polled
5025 * uses multiple waitqueues for poll handling (eg one for read, one
5026 * for write). Setup a separate io_poll_iocb if this happens.
5028 if (unlikely(poll->head)) {
5029 struct io_poll_iocb *poll_one = poll;
5031 /* already have a 2nd entry, fail a third attempt */
5033 pt->error = -EINVAL;
5037 * Can't handle multishot for double wait for now, turn it
5038 * into one-shot mode.
5040 if (!(poll_one->events & EPOLLONESHOT))
5041 poll_one->events |= EPOLLONESHOT;
5042 /* double add on the same waitqueue head, ignore */
5043 if (poll_one->head == head)
5045 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5047 pt->error = -ENOMEM;
5050 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5052 poll->wait.private = req;
5059 if (poll->events & EPOLLEXCLUSIVE)
5060 add_wait_queue_exclusive(head, &poll->wait);
5062 add_wait_queue(head, &poll->wait);
5065 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5066 struct poll_table_struct *p)
5068 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5069 struct async_poll *apoll = pt->req->apoll;
5071 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5074 static void io_async_task_func(struct callback_head *cb)
5076 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5077 struct async_poll *apoll = req->apoll;
5078 struct io_ring_ctx *ctx = req->ctx;
5080 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5082 if (io_poll_rewait(req, &apoll->poll)) {
5083 spin_unlock_irq(&ctx->completion_lock);
5087 hash_del(&req->hash_node);
5088 io_poll_remove_double(req);
5089 spin_unlock_irq(&ctx->completion_lock);
5091 if (!READ_ONCE(apoll->poll.canceled))
5092 __io_req_task_submit(req);
5094 io_req_complete_failed(req, -ECANCELED);
5097 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5100 struct io_kiocb *req = wait->private;
5101 struct io_poll_iocb *poll = &req->apoll->poll;
5103 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5106 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5109 static void io_poll_req_insert(struct io_kiocb *req)
5111 struct io_ring_ctx *ctx = req->ctx;
5112 struct hlist_head *list;
5114 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5115 hlist_add_head(&req->hash_node, list);
5118 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5119 struct io_poll_iocb *poll,
5120 struct io_poll_table *ipt, __poll_t mask,
5121 wait_queue_func_t wake_func)
5122 __acquires(&ctx->completion_lock)
5124 struct io_ring_ctx *ctx = req->ctx;
5125 bool cancel = false;
5127 INIT_HLIST_NODE(&req->hash_node);
5128 io_init_poll_iocb(poll, mask, wake_func);
5129 poll->file = req->file;
5130 poll->wait.private = req;
5132 ipt->pt._key = mask;
5134 ipt->error = -EINVAL;
5136 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5138 spin_lock_irq(&ctx->completion_lock);
5139 if (likely(poll->head)) {
5140 spin_lock(&poll->head->lock);
5141 if (unlikely(list_empty(&poll->wait.entry))) {
5147 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5148 list_del_init(&poll->wait.entry);
5150 WRITE_ONCE(poll->canceled, true);
5151 else if (!poll->done) /* actually waiting for an event */
5152 io_poll_req_insert(req);
5153 spin_unlock(&poll->head->lock);
5159 static bool io_arm_poll_handler(struct io_kiocb *req)
5161 const struct io_op_def *def = &io_op_defs[req->opcode];
5162 struct io_ring_ctx *ctx = req->ctx;
5163 struct async_poll *apoll;
5164 struct io_poll_table ipt;
5168 if (!req->file || !file_can_poll(req->file))
5170 if (req->flags & REQ_F_POLLED)
5174 else if (def->pollout)
5178 /* if we can't nonblock try, then no point in arming a poll handler */
5179 if (!io_file_supports_async(req, rw))
5182 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5183 if (unlikely(!apoll))
5185 apoll->double_poll = NULL;
5187 req->flags |= REQ_F_POLLED;
5190 mask = EPOLLONESHOT;
5192 mask |= POLLIN | POLLRDNORM;
5194 mask |= POLLOUT | POLLWRNORM;
5196 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5197 if ((req->opcode == IORING_OP_RECVMSG) &&
5198 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5201 mask |= POLLERR | POLLPRI;
5203 ipt.pt._qproc = io_async_queue_proc;
5205 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5207 if (ret || ipt.error) {
5208 io_poll_remove_double(req);
5209 spin_unlock_irq(&ctx->completion_lock);
5212 spin_unlock_irq(&ctx->completion_lock);
5213 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5214 apoll->poll.events);
5218 static bool __io_poll_remove_one(struct io_kiocb *req,
5219 struct io_poll_iocb *poll, bool do_cancel)
5220 __must_hold(&req->ctx->completion_lock)
5222 bool do_complete = false;
5226 spin_lock(&poll->head->lock);
5228 WRITE_ONCE(poll->canceled, true);
5229 if (!list_empty(&poll->wait.entry)) {
5230 list_del_init(&poll->wait.entry);
5233 spin_unlock(&poll->head->lock);
5234 hash_del(&req->hash_node);
5238 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5239 __must_hold(&req->ctx->completion_lock)
5243 io_poll_remove_double(req);
5244 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5246 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5247 /* non-poll requests have submit ref still */
5253 static bool io_poll_remove_one(struct io_kiocb *req)
5254 __must_hold(&req->ctx->completion_lock)
5258 do_complete = io_poll_remove_waitqs(req);
5260 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5261 io_commit_cqring(req->ctx);
5263 io_put_req_deferred(req, 1);
5270 * Returns true if we found and killed one or more poll requests
5272 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5275 struct hlist_node *tmp;
5276 struct io_kiocb *req;
5279 spin_lock_irq(&ctx->completion_lock);
5280 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5281 struct hlist_head *list;
5283 list = &ctx->cancel_hash[i];
5284 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5285 if (io_match_task(req, tsk, cancel_all))
5286 posted += io_poll_remove_one(req);
5289 spin_unlock_irq(&ctx->completion_lock);
5292 io_cqring_ev_posted(ctx);
5297 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5299 __must_hold(&ctx->completion_lock)
5301 struct hlist_head *list;
5302 struct io_kiocb *req;
5304 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5305 hlist_for_each_entry(req, list, hash_node) {
5306 if (sqe_addr != req->user_data)
5308 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5315 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5317 __must_hold(&ctx->completion_lock)
5319 struct io_kiocb *req;
5321 req = io_poll_find(ctx, sqe_addr, poll_only);
5324 if (io_poll_remove_one(req))
5330 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5335 events = READ_ONCE(sqe->poll32_events);
5337 events = swahw32(events);
5339 if (!(flags & IORING_POLL_ADD_MULTI))
5340 events |= EPOLLONESHOT;
5341 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5344 static int io_poll_update_prep(struct io_kiocb *req,
5345 const struct io_uring_sqe *sqe)
5347 struct io_poll_update *upd = &req->poll_update;
5350 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5352 if (sqe->ioprio || sqe->buf_index)
5354 flags = READ_ONCE(sqe->len);
5355 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5356 IORING_POLL_ADD_MULTI))
5358 /* meaningless without update */
5359 if (flags == IORING_POLL_ADD_MULTI)
5362 upd->old_user_data = READ_ONCE(sqe->addr);
5363 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5364 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5366 upd->new_user_data = READ_ONCE(sqe->off);
5367 if (!upd->update_user_data && upd->new_user_data)
5369 if (upd->update_events)
5370 upd->events = io_poll_parse_events(sqe, flags);
5371 else if (sqe->poll32_events)
5377 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5380 struct io_kiocb *req = wait->private;
5381 struct io_poll_iocb *poll = &req->poll;
5383 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5386 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5387 struct poll_table_struct *p)
5389 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5391 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5394 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5396 struct io_poll_iocb *poll = &req->poll;
5399 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5401 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5403 flags = READ_ONCE(sqe->len);
5404 if (flags & ~IORING_POLL_ADD_MULTI)
5407 poll->events = io_poll_parse_events(sqe, flags);
5411 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5413 struct io_poll_iocb *poll = &req->poll;
5414 struct io_ring_ctx *ctx = req->ctx;
5415 struct io_poll_table ipt;
5418 ipt.pt._qproc = io_poll_queue_proc;
5420 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5423 if (mask) { /* no async, we'd stolen it */
5425 io_poll_complete(req, mask);
5427 spin_unlock_irq(&ctx->completion_lock);
5430 io_cqring_ev_posted(ctx);
5431 if (poll->events & EPOLLONESHOT)
5437 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5439 struct io_ring_ctx *ctx = req->ctx;
5440 struct io_kiocb *preq;
5444 spin_lock_irq(&ctx->completion_lock);
5445 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5451 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5453 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5458 * Don't allow racy completion with singleshot, as we cannot safely
5459 * update those. For multishot, if we're racing with completion, just
5460 * let completion re-add it.
5462 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5463 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5467 /* we now have a detached poll request. reissue. */
5471 spin_unlock_irq(&ctx->completion_lock);
5473 io_req_complete(req, ret);
5476 /* only mask one event flags, keep behavior flags */
5477 if (req->poll_update.update_events) {
5478 preq->poll.events &= ~0xffff;
5479 preq->poll.events |= req->poll_update.events & 0xffff;
5480 preq->poll.events |= IO_POLL_UNMASK;
5482 if (req->poll_update.update_user_data)
5483 preq->user_data = req->poll_update.new_user_data;
5484 spin_unlock_irq(&ctx->completion_lock);
5486 /* complete update request, we're done with it */
5487 io_req_complete(req, ret);
5490 ret = io_poll_add(preq, issue_flags);
5493 io_req_complete(preq, ret);
5499 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5501 struct io_timeout_data *data = container_of(timer,
5502 struct io_timeout_data, timer);
5503 struct io_kiocb *req = data->req;
5504 struct io_ring_ctx *ctx = req->ctx;
5505 unsigned long flags;
5507 spin_lock_irqsave(&ctx->completion_lock, flags);
5508 list_del_init(&req->timeout.list);
5509 atomic_set(&req->ctx->cq_timeouts,
5510 atomic_read(&req->ctx->cq_timeouts) + 1);
5512 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5513 io_commit_cqring(ctx);
5514 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5516 io_cqring_ev_posted(ctx);
5519 return HRTIMER_NORESTART;
5522 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5524 __must_hold(&ctx->completion_lock)
5526 struct io_timeout_data *io;
5527 struct io_kiocb *req;
5530 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5531 found = user_data == req->user_data;
5536 return ERR_PTR(-ENOENT);
5538 io = req->async_data;
5539 if (hrtimer_try_to_cancel(&io->timer) == -1)
5540 return ERR_PTR(-EALREADY);
5541 list_del_init(&req->timeout.list);
5545 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5546 __must_hold(&ctx->completion_lock)
5548 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5551 return PTR_ERR(req);
5554 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5555 io_put_req_deferred(req, 1);
5559 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5560 struct timespec64 *ts, enum hrtimer_mode mode)
5561 __must_hold(&ctx->completion_lock)
5563 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5564 struct io_timeout_data *data;
5567 return PTR_ERR(req);
5569 req->timeout.off = 0; /* noseq */
5570 data = req->async_data;
5571 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5572 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5573 data->timer.function = io_timeout_fn;
5574 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5578 static int io_timeout_remove_prep(struct io_kiocb *req,
5579 const struct io_uring_sqe *sqe)
5581 struct io_timeout_rem *tr = &req->timeout_rem;
5583 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5585 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5587 if (sqe->ioprio || sqe->buf_index || sqe->len)
5590 tr->addr = READ_ONCE(sqe->addr);
5591 tr->flags = READ_ONCE(sqe->timeout_flags);
5592 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5593 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5595 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5597 } else if (tr->flags) {
5598 /* timeout removal doesn't support flags */
5605 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5607 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5612 * Remove or update an existing timeout command
5614 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5616 struct io_timeout_rem *tr = &req->timeout_rem;
5617 struct io_ring_ctx *ctx = req->ctx;
5620 spin_lock_irq(&ctx->completion_lock);
5621 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5622 ret = io_timeout_cancel(ctx, tr->addr);
5624 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5625 io_translate_timeout_mode(tr->flags));
5627 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5628 io_commit_cqring(ctx);
5629 spin_unlock_irq(&ctx->completion_lock);
5630 io_cqring_ev_posted(ctx);
5637 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5638 bool is_timeout_link)
5640 struct io_timeout_data *data;
5642 u32 off = READ_ONCE(sqe->off);
5644 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5646 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5648 if (off && is_timeout_link)
5650 flags = READ_ONCE(sqe->timeout_flags);
5651 if (flags & ~IORING_TIMEOUT_ABS)
5654 req->timeout.off = off;
5656 if (!req->async_data && io_alloc_async_data(req))
5659 data = req->async_data;
5662 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5665 data->mode = io_translate_timeout_mode(flags);
5666 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5667 if (is_timeout_link)
5668 io_req_track_inflight(req);
5672 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5674 struct io_ring_ctx *ctx = req->ctx;
5675 struct io_timeout_data *data = req->async_data;
5676 struct list_head *entry;
5677 u32 tail, off = req->timeout.off;
5679 spin_lock_irq(&ctx->completion_lock);
5682 * sqe->off holds how many events that need to occur for this
5683 * timeout event to be satisfied. If it isn't set, then this is
5684 * a pure timeout request, sequence isn't used.
5686 if (io_is_timeout_noseq(req)) {
5687 entry = ctx->timeout_list.prev;
5691 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5692 req->timeout.target_seq = tail + off;
5694 /* Update the last seq here in case io_flush_timeouts() hasn't.
5695 * This is safe because ->completion_lock is held, and submissions
5696 * and completions are never mixed in the same ->completion_lock section.
5698 ctx->cq_last_tm_flush = tail;
5701 * Insertion sort, ensuring the first entry in the list is always
5702 * the one we need first.
5704 list_for_each_prev(entry, &ctx->timeout_list) {
5705 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5708 if (io_is_timeout_noseq(nxt))
5710 /* nxt.seq is behind @tail, otherwise would've been completed */
5711 if (off >= nxt->timeout.target_seq - tail)
5715 list_add(&req->timeout.list, entry);
5716 data->timer.function = io_timeout_fn;
5717 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5718 spin_unlock_irq(&ctx->completion_lock);
5722 struct io_cancel_data {
5723 struct io_ring_ctx *ctx;
5727 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5729 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5730 struct io_cancel_data *cd = data;
5732 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5735 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5736 struct io_ring_ctx *ctx)
5738 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5739 enum io_wq_cancel cancel_ret;
5742 if (!tctx || !tctx->io_wq)
5745 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5746 switch (cancel_ret) {
5747 case IO_WQ_CANCEL_OK:
5750 case IO_WQ_CANCEL_RUNNING:
5753 case IO_WQ_CANCEL_NOTFOUND:
5761 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5762 struct io_kiocb *req, __u64 sqe_addr,
5765 unsigned long flags;
5768 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5769 spin_lock_irqsave(&ctx->completion_lock, flags);
5772 ret = io_timeout_cancel(ctx, sqe_addr);
5775 ret = io_poll_cancel(ctx, sqe_addr, false);
5779 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5780 io_commit_cqring(ctx);
5781 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5782 io_cqring_ev_posted(ctx);
5788 static int io_async_cancel_prep(struct io_kiocb *req,
5789 const struct io_uring_sqe *sqe)
5791 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5793 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5795 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5798 req->cancel.addr = READ_ONCE(sqe->addr);
5802 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5804 struct io_ring_ctx *ctx = req->ctx;
5805 u64 sqe_addr = req->cancel.addr;
5806 struct io_tctx_node *node;
5809 /* tasks should wait for their io-wq threads, so safe w/o sync */
5810 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5811 spin_lock_irq(&ctx->completion_lock);
5814 ret = io_timeout_cancel(ctx, sqe_addr);
5817 ret = io_poll_cancel(ctx, sqe_addr, false);
5820 spin_unlock_irq(&ctx->completion_lock);
5822 /* slow path, try all io-wq's */
5823 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5825 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5826 struct io_uring_task *tctx = node->task->io_uring;
5828 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5832 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5834 spin_lock_irq(&ctx->completion_lock);
5836 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5837 io_commit_cqring(ctx);
5838 spin_unlock_irq(&ctx->completion_lock);
5839 io_cqring_ev_posted(ctx);
5847 static int io_rsrc_update_prep(struct io_kiocb *req,
5848 const struct io_uring_sqe *sqe)
5850 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5852 if (sqe->ioprio || sqe->rw_flags)
5855 req->rsrc_update.offset = READ_ONCE(sqe->off);
5856 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5857 if (!req->rsrc_update.nr_args)
5859 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5863 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5865 struct io_ring_ctx *ctx = req->ctx;
5866 struct io_uring_rsrc_update2 up;
5869 if (issue_flags & IO_URING_F_NONBLOCK)
5872 up.offset = req->rsrc_update.offset;
5873 up.data = req->rsrc_update.arg;
5878 mutex_lock(&ctx->uring_lock);
5879 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5880 &up, req->rsrc_update.nr_args);
5881 mutex_unlock(&ctx->uring_lock);
5885 __io_req_complete(req, issue_flags, ret, 0);
5889 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5891 switch (req->opcode) {
5894 case IORING_OP_READV:
5895 case IORING_OP_READ_FIXED:
5896 case IORING_OP_READ:
5897 return io_read_prep(req, sqe);
5898 case IORING_OP_WRITEV:
5899 case IORING_OP_WRITE_FIXED:
5900 case IORING_OP_WRITE:
5901 return io_write_prep(req, sqe);
5902 case IORING_OP_POLL_ADD:
5903 return io_poll_add_prep(req, sqe);
5904 case IORING_OP_POLL_REMOVE:
5905 return io_poll_update_prep(req, sqe);
5906 case IORING_OP_FSYNC:
5907 return io_fsync_prep(req, sqe);
5908 case IORING_OP_SYNC_FILE_RANGE:
5909 return io_sfr_prep(req, sqe);
5910 case IORING_OP_SENDMSG:
5911 case IORING_OP_SEND:
5912 return io_sendmsg_prep(req, sqe);
5913 case IORING_OP_RECVMSG:
5914 case IORING_OP_RECV:
5915 return io_recvmsg_prep(req, sqe);
5916 case IORING_OP_CONNECT:
5917 return io_connect_prep(req, sqe);
5918 case IORING_OP_TIMEOUT:
5919 return io_timeout_prep(req, sqe, false);
5920 case IORING_OP_TIMEOUT_REMOVE:
5921 return io_timeout_remove_prep(req, sqe);
5922 case IORING_OP_ASYNC_CANCEL:
5923 return io_async_cancel_prep(req, sqe);
5924 case IORING_OP_LINK_TIMEOUT:
5925 return io_timeout_prep(req, sqe, true);
5926 case IORING_OP_ACCEPT:
5927 return io_accept_prep(req, sqe);
5928 case IORING_OP_FALLOCATE:
5929 return io_fallocate_prep(req, sqe);
5930 case IORING_OP_OPENAT:
5931 return io_openat_prep(req, sqe);
5932 case IORING_OP_CLOSE:
5933 return io_close_prep(req, sqe);
5934 case IORING_OP_FILES_UPDATE:
5935 return io_rsrc_update_prep(req, sqe);
5936 case IORING_OP_STATX:
5937 return io_statx_prep(req, sqe);
5938 case IORING_OP_FADVISE:
5939 return io_fadvise_prep(req, sqe);
5940 case IORING_OP_MADVISE:
5941 return io_madvise_prep(req, sqe);
5942 case IORING_OP_OPENAT2:
5943 return io_openat2_prep(req, sqe);
5944 case IORING_OP_EPOLL_CTL:
5945 return io_epoll_ctl_prep(req, sqe);
5946 case IORING_OP_SPLICE:
5947 return io_splice_prep(req, sqe);
5948 case IORING_OP_PROVIDE_BUFFERS:
5949 return io_provide_buffers_prep(req, sqe);
5950 case IORING_OP_REMOVE_BUFFERS:
5951 return io_remove_buffers_prep(req, sqe);
5953 return io_tee_prep(req, sqe);
5954 case IORING_OP_SHUTDOWN:
5955 return io_shutdown_prep(req, sqe);
5956 case IORING_OP_RENAMEAT:
5957 return io_renameat_prep(req, sqe);
5958 case IORING_OP_UNLINKAT:
5959 return io_unlinkat_prep(req, sqe);
5962 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5967 static int io_req_prep_async(struct io_kiocb *req)
5969 if (!io_op_defs[req->opcode].needs_async_setup)
5971 if (WARN_ON_ONCE(req->async_data))
5973 if (io_alloc_async_data(req))
5976 switch (req->opcode) {
5977 case IORING_OP_READV:
5978 return io_rw_prep_async(req, READ);
5979 case IORING_OP_WRITEV:
5980 return io_rw_prep_async(req, WRITE);
5981 case IORING_OP_SENDMSG:
5982 return io_sendmsg_prep_async(req);
5983 case IORING_OP_RECVMSG:
5984 return io_recvmsg_prep_async(req);
5985 case IORING_OP_CONNECT:
5986 return io_connect_prep_async(req);
5988 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5993 static u32 io_get_sequence(struct io_kiocb *req)
5995 struct io_kiocb *pos;
5996 struct io_ring_ctx *ctx = req->ctx;
5997 u32 total_submitted, nr_reqs = 0;
5999 io_for_each_link(pos, req)
6002 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6003 return total_submitted - nr_reqs;
6006 static int io_req_defer(struct io_kiocb *req)
6008 struct io_ring_ctx *ctx = req->ctx;
6009 struct io_defer_entry *de;
6013 /* Still need defer if there is pending req in defer list. */
6014 if (likely(list_empty_careful(&ctx->defer_list) &&
6015 !(req->flags & REQ_F_IO_DRAIN)))
6018 seq = io_get_sequence(req);
6019 /* Still a chance to pass the sequence check */
6020 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6023 ret = io_req_prep_async(req);
6026 io_prep_async_link(req);
6027 de = kmalloc(sizeof(*de), GFP_KERNEL);
6031 spin_lock_irq(&ctx->completion_lock);
6032 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6033 spin_unlock_irq(&ctx->completion_lock);
6035 io_queue_async_work(req);
6036 return -EIOCBQUEUED;
6039 trace_io_uring_defer(ctx, req, req->user_data);
6042 list_add_tail(&de->list, &ctx->defer_list);
6043 spin_unlock_irq(&ctx->completion_lock);
6044 return -EIOCBQUEUED;
6047 static void io_clean_op(struct io_kiocb *req)
6049 if (req->flags & REQ_F_BUFFER_SELECTED) {
6050 switch (req->opcode) {
6051 case IORING_OP_READV:
6052 case IORING_OP_READ_FIXED:
6053 case IORING_OP_READ:
6054 kfree((void *)(unsigned long)req->rw.addr);
6056 case IORING_OP_RECVMSG:
6057 case IORING_OP_RECV:
6058 kfree(req->sr_msg.kbuf);
6061 req->flags &= ~REQ_F_BUFFER_SELECTED;
6064 if (req->flags & REQ_F_NEED_CLEANUP) {
6065 switch (req->opcode) {
6066 case IORING_OP_READV:
6067 case IORING_OP_READ_FIXED:
6068 case IORING_OP_READ:
6069 case IORING_OP_WRITEV:
6070 case IORING_OP_WRITE_FIXED:
6071 case IORING_OP_WRITE: {
6072 struct io_async_rw *io = req->async_data;
6074 kfree(io->free_iovec);
6077 case IORING_OP_RECVMSG:
6078 case IORING_OP_SENDMSG: {
6079 struct io_async_msghdr *io = req->async_data;
6081 kfree(io->free_iov);
6084 case IORING_OP_SPLICE:
6086 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6087 io_put_file(req->splice.file_in);
6089 case IORING_OP_OPENAT:
6090 case IORING_OP_OPENAT2:
6091 if (req->open.filename)
6092 putname(req->open.filename);
6094 case IORING_OP_RENAMEAT:
6095 putname(req->rename.oldpath);
6096 putname(req->rename.newpath);
6098 case IORING_OP_UNLINKAT:
6099 putname(req->unlink.filename);
6102 req->flags &= ~REQ_F_NEED_CLEANUP;
6104 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6105 kfree(req->apoll->double_poll);
6109 if (req->flags & REQ_F_INFLIGHT) {
6110 struct io_uring_task *tctx = req->task->io_uring;
6112 atomic_dec(&tctx->inflight_tracked);
6113 req->flags &= ~REQ_F_INFLIGHT;
6117 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6119 struct io_ring_ctx *ctx = req->ctx;
6120 const struct cred *creds = NULL;
6123 if (req->work.creds && req->work.creds != current_cred())
6124 creds = override_creds(req->work.creds);
6126 switch (req->opcode) {
6128 ret = io_nop(req, issue_flags);
6130 case IORING_OP_READV:
6131 case IORING_OP_READ_FIXED:
6132 case IORING_OP_READ:
6133 ret = io_read(req, issue_flags);
6135 case IORING_OP_WRITEV:
6136 case IORING_OP_WRITE_FIXED:
6137 case IORING_OP_WRITE:
6138 ret = io_write(req, issue_flags);
6140 case IORING_OP_FSYNC:
6141 ret = io_fsync(req, issue_flags);
6143 case IORING_OP_POLL_ADD:
6144 ret = io_poll_add(req, issue_flags);
6146 case IORING_OP_POLL_REMOVE:
6147 ret = io_poll_update(req, issue_flags);
6149 case IORING_OP_SYNC_FILE_RANGE:
6150 ret = io_sync_file_range(req, issue_flags);
6152 case IORING_OP_SENDMSG:
6153 ret = io_sendmsg(req, issue_flags);
6155 case IORING_OP_SEND:
6156 ret = io_send(req, issue_flags);
6158 case IORING_OP_RECVMSG:
6159 ret = io_recvmsg(req, issue_flags);
6161 case IORING_OP_RECV:
6162 ret = io_recv(req, issue_flags);
6164 case IORING_OP_TIMEOUT:
6165 ret = io_timeout(req, issue_flags);
6167 case IORING_OP_TIMEOUT_REMOVE:
6168 ret = io_timeout_remove(req, issue_flags);
6170 case IORING_OP_ACCEPT:
6171 ret = io_accept(req, issue_flags);
6173 case IORING_OP_CONNECT:
6174 ret = io_connect(req, issue_flags);
6176 case IORING_OP_ASYNC_CANCEL:
6177 ret = io_async_cancel(req, issue_flags);
6179 case IORING_OP_FALLOCATE:
6180 ret = io_fallocate(req, issue_flags);
6182 case IORING_OP_OPENAT:
6183 ret = io_openat(req, issue_flags);
6185 case IORING_OP_CLOSE:
6186 ret = io_close(req, issue_flags);
6188 case IORING_OP_FILES_UPDATE:
6189 ret = io_files_update(req, issue_flags);
6191 case IORING_OP_STATX:
6192 ret = io_statx(req, issue_flags);
6194 case IORING_OP_FADVISE:
6195 ret = io_fadvise(req, issue_flags);
6197 case IORING_OP_MADVISE:
6198 ret = io_madvise(req, issue_flags);
6200 case IORING_OP_OPENAT2:
6201 ret = io_openat2(req, issue_flags);
6203 case IORING_OP_EPOLL_CTL:
6204 ret = io_epoll_ctl(req, issue_flags);
6206 case IORING_OP_SPLICE:
6207 ret = io_splice(req, issue_flags);
6209 case IORING_OP_PROVIDE_BUFFERS:
6210 ret = io_provide_buffers(req, issue_flags);
6212 case IORING_OP_REMOVE_BUFFERS:
6213 ret = io_remove_buffers(req, issue_flags);
6216 ret = io_tee(req, issue_flags);
6218 case IORING_OP_SHUTDOWN:
6219 ret = io_shutdown(req, issue_flags);
6221 case IORING_OP_RENAMEAT:
6222 ret = io_renameat(req, issue_flags);
6224 case IORING_OP_UNLINKAT:
6225 ret = io_unlinkat(req, issue_flags);
6233 revert_creds(creds);
6236 /* If the op doesn't have a file, we're not polling for it */
6237 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6238 io_iopoll_req_issued(req);
6243 static void io_wq_submit_work(struct io_wq_work *work)
6245 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6246 struct io_kiocb *timeout;
6249 timeout = io_prep_linked_timeout(req);
6251 io_queue_linked_timeout(timeout);
6253 if (work->flags & IO_WQ_WORK_CANCEL)
6258 ret = io_issue_sqe(req, 0);
6260 * We can get EAGAIN for polled IO even though we're
6261 * forcing a sync submission from here, since we can't
6262 * wait for request slots on the block side.
6270 /* avoid locking problems by failing it from a clean context */
6272 /* io-wq is going to take one down */
6274 io_req_task_queue_fail(req, ret);
6278 #define FFS_ASYNC_READ 0x1UL
6279 #define FFS_ASYNC_WRITE 0x2UL
6281 #define FFS_ISREG 0x4UL
6283 #define FFS_ISREG 0x0UL
6285 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6287 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6290 struct io_fixed_file *table_l2;
6292 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6293 return &table_l2[i & IORING_FILE_TABLE_MASK];
6296 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6299 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6301 return (struct file *) (slot->file_ptr & FFS_MASK);
6304 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6306 unsigned long file_ptr = (unsigned long) file;
6308 if (__io_file_supports_async(file, READ))
6309 file_ptr |= FFS_ASYNC_READ;
6310 if (__io_file_supports_async(file, WRITE))
6311 file_ptr |= FFS_ASYNC_WRITE;
6312 if (S_ISREG(file_inode(file)->i_mode))
6313 file_ptr |= FFS_ISREG;
6314 file_slot->file_ptr = file_ptr;
6317 static struct file *io_file_get(struct io_submit_state *state,
6318 struct io_kiocb *req, int fd, bool fixed)
6320 struct io_ring_ctx *ctx = req->ctx;
6324 unsigned long file_ptr;
6326 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6328 fd = array_index_nospec(fd, ctx->nr_user_files);
6329 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6330 file = (struct file *) (file_ptr & FFS_MASK);
6331 file_ptr &= ~FFS_MASK;
6332 /* mask in overlapping REQ_F and FFS bits */
6333 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6334 io_req_set_rsrc_node(req);
6336 trace_io_uring_file_get(ctx, fd);
6337 file = __io_file_get(state, fd);
6339 /* we don't allow fixed io_uring files */
6340 if (file && unlikely(file->f_op == &io_uring_fops))
6341 io_req_track_inflight(req);
6347 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6349 struct io_timeout_data *data = container_of(timer,
6350 struct io_timeout_data, timer);
6351 struct io_kiocb *prev, *req = data->req;
6352 struct io_ring_ctx *ctx = req->ctx;
6353 unsigned long flags;
6355 spin_lock_irqsave(&ctx->completion_lock, flags);
6356 prev = req->timeout.head;
6357 req->timeout.head = NULL;
6360 * We don't expect the list to be empty, that will only happen if we
6361 * race with the completion of the linked work.
6364 io_remove_next_linked(prev);
6365 if (!req_ref_inc_not_zero(prev))
6368 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6371 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6372 io_put_req_deferred(prev, 1);
6373 io_put_req_deferred(req, 1);
6375 io_req_complete_post(req, -ETIME, 0);
6377 return HRTIMER_NORESTART;
6380 static void io_queue_linked_timeout(struct io_kiocb *req)
6382 struct io_ring_ctx *ctx = req->ctx;
6384 spin_lock_irq(&ctx->completion_lock);
6386 * If the back reference is NULL, then our linked request finished
6387 * before we got a chance to setup the timer
6389 if (req->timeout.head) {
6390 struct io_timeout_data *data = req->async_data;
6392 data->timer.function = io_link_timeout_fn;
6393 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6396 spin_unlock_irq(&ctx->completion_lock);
6397 /* drop submission reference */
6401 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6403 struct io_kiocb *nxt = req->link;
6405 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6406 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6409 nxt->timeout.head = req;
6410 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6411 req->flags |= REQ_F_LINK_TIMEOUT;
6415 static void __io_queue_sqe(struct io_kiocb *req)
6417 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6420 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6423 * We async punt it if the file wasn't marked NOWAIT, or if the file
6424 * doesn't support non-blocking read/write attempts
6427 /* drop submission reference */
6428 if (req->flags & REQ_F_COMPLETE_INLINE) {
6429 struct io_ring_ctx *ctx = req->ctx;
6430 struct io_comp_state *cs = &ctx->submit_state.comp;
6432 cs->reqs[cs->nr++] = req;
6433 if (cs->nr == ARRAY_SIZE(cs->reqs))
6434 io_submit_flush_completions(cs, ctx);
6438 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6439 if (!io_arm_poll_handler(req)) {
6441 * Queued up for async execution, worker will release
6442 * submit reference when the iocb is actually submitted.
6444 io_queue_async_work(req);
6447 io_req_complete_failed(req, ret);
6450 io_queue_linked_timeout(linked_timeout);
6453 static void io_queue_sqe(struct io_kiocb *req)
6457 ret = io_req_defer(req);
6459 if (ret != -EIOCBQUEUED) {
6461 io_req_complete_failed(req, ret);
6463 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6464 ret = io_req_prep_async(req);
6467 io_queue_async_work(req);
6469 __io_queue_sqe(req);
6474 * Check SQE restrictions (opcode and flags).
6476 * Returns 'true' if SQE is allowed, 'false' otherwise.
6478 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6479 struct io_kiocb *req,
6480 unsigned int sqe_flags)
6482 if (!ctx->restricted)
6485 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6488 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6489 ctx->restrictions.sqe_flags_required)
6492 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6493 ctx->restrictions.sqe_flags_required))
6499 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6500 const struct io_uring_sqe *sqe)
6502 struct io_submit_state *state;
6503 unsigned int sqe_flags;
6504 int personality, ret = 0;
6506 req->opcode = READ_ONCE(sqe->opcode);
6507 /* same numerical values with corresponding REQ_F_*, safe to copy */
6508 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6509 req->user_data = READ_ONCE(sqe->user_data);
6510 req->async_data = NULL;
6514 req->fixed_rsrc_refs = NULL;
6515 /* one is dropped after submission, the other at completion */
6516 atomic_set(&req->refs, 2);
6517 req->task = current;
6519 req->work.creds = NULL;
6521 /* enforce forwards compatibility on users */
6522 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6524 if (unlikely(req->opcode >= IORING_OP_LAST))
6526 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6529 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6530 !io_op_defs[req->opcode].buffer_select)
6533 personality = READ_ONCE(sqe->personality);
6535 req->work.creds = xa_load(&ctx->personalities, personality);
6536 if (!req->work.creds)
6538 get_cred(req->work.creds);
6540 state = &ctx->submit_state;
6543 * Plug now if we have more than 1 IO left after this, and the target
6544 * is potentially a read/write to block based storage.
6546 if (!state->plug_started && state->ios_left > 1 &&
6547 io_op_defs[req->opcode].plug) {
6548 blk_start_plug(&state->plug);
6549 state->plug_started = true;
6552 if (io_op_defs[req->opcode].needs_file) {
6553 bool fixed = req->flags & REQ_F_FIXED_FILE;
6555 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6556 if (unlikely(!req->file))
6564 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6565 const struct io_uring_sqe *sqe)
6567 struct io_submit_link *link = &ctx->submit_state.link;
6570 ret = io_init_req(ctx, req, sqe);
6571 if (unlikely(ret)) {
6574 /* fail even hard links since we don't submit */
6575 req_set_fail(link->head);
6576 io_req_complete_failed(link->head, -ECANCELED);
6579 io_req_complete_failed(req, ret);
6582 ret = io_req_prep(req, sqe);
6586 /* don't need @sqe from now on */
6587 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6588 true, ctx->flags & IORING_SETUP_SQPOLL);
6591 * If we already have a head request, queue this one for async
6592 * submittal once the head completes. If we don't have a head but
6593 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6594 * submitted sync once the chain is complete. If none of those
6595 * conditions are true (normal request), then just queue it.
6598 struct io_kiocb *head = link->head;
6601 * Taking sequential execution of a link, draining both sides
6602 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6603 * requests in the link. So, it drains the head and the
6604 * next after the link request. The last one is done via
6605 * drain_next flag to persist the effect across calls.
6607 if (req->flags & REQ_F_IO_DRAIN) {
6608 head->flags |= REQ_F_IO_DRAIN;
6609 ctx->drain_next = 1;
6611 ret = io_req_prep_async(req);
6614 trace_io_uring_link(ctx, req, head);
6615 link->last->link = req;
6618 /* last request of a link, enqueue the link */
6619 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6624 if (unlikely(ctx->drain_next)) {
6625 req->flags |= REQ_F_IO_DRAIN;
6626 ctx->drain_next = 0;
6628 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6640 * Batched submission is done, ensure local IO is flushed out.
6642 static void io_submit_state_end(struct io_submit_state *state,
6643 struct io_ring_ctx *ctx)
6645 if (state->link.head)
6646 io_queue_sqe(state->link.head);
6648 io_submit_flush_completions(&state->comp, ctx);
6649 if (state->plug_started)
6650 blk_finish_plug(&state->plug);
6651 io_state_file_put(state);
6655 * Start submission side cache.
6657 static void io_submit_state_start(struct io_submit_state *state,
6658 unsigned int max_ios)
6660 state->plug_started = false;
6661 state->ios_left = max_ios;
6662 /* set only head, no need to init link_last in advance */
6663 state->link.head = NULL;
6666 static void io_commit_sqring(struct io_ring_ctx *ctx)
6668 struct io_rings *rings = ctx->rings;
6671 * Ensure any loads from the SQEs are done at this point,
6672 * since once we write the new head, the application could
6673 * write new data to them.
6675 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6679 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6680 * that is mapped by userspace. This means that care needs to be taken to
6681 * ensure that reads are stable, as we cannot rely on userspace always
6682 * being a good citizen. If members of the sqe are validated and then later
6683 * used, it's important that those reads are done through READ_ONCE() to
6684 * prevent a re-load down the line.
6686 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6688 u32 *sq_array = ctx->sq_array;
6689 unsigned head, mask = ctx->sq_entries - 1;
6692 * The cached sq head (or cq tail) serves two purposes:
6694 * 1) allows us to batch the cost of updating the user visible
6696 * 2) allows the kernel side to track the head on its own, even
6697 * though the application is the one updating it.
6699 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & mask]);
6700 if (likely(head < ctx->sq_entries))
6701 return &ctx->sq_sqes[head];
6703 /* drop invalid entries */
6704 ctx->cached_sq_dropped++;
6705 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6709 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6711 struct io_uring_task *tctx;
6714 /* make sure SQ entry isn't read before tail */
6715 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6716 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6719 tctx = current->io_uring;
6720 tctx->cached_refs -= nr;
6721 if (unlikely(tctx->cached_refs < 0)) {
6722 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6724 percpu_counter_add(&tctx->inflight, refill);
6725 refcount_add(refill, ¤t->usage);
6726 tctx->cached_refs += refill;
6728 io_submit_state_start(&ctx->submit_state, nr);
6730 while (submitted < nr) {
6731 const struct io_uring_sqe *sqe;
6732 struct io_kiocb *req;
6734 req = io_alloc_req(ctx);
6735 if (unlikely(!req)) {
6737 submitted = -EAGAIN;
6740 sqe = io_get_sqe(ctx);
6741 if (unlikely(!sqe)) {
6742 kmem_cache_free(req_cachep, req);
6745 /* will complete beyond this point, count as submitted */
6747 if (io_submit_sqe(ctx, req, sqe))
6751 if (unlikely(submitted != nr)) {
6752 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6753 int unused = nr - ref_used;
6755 current->io_uring->cached_refs += unused;
6756 percpu_ref_put_many(&ctx->refs, unused);
6759 io_submit_state_end(&ctx->submit_state, ctx);
6760 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6761 io_commit_sqring(ctx);
6766 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6768 return READ_ONCE(sqd->state);
6771 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6773 /* Tell userspace we may need a wakeup call */
6774 spin_lock_irq(&ctx->completion_lock);
6775 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6776 spin_unlock_irq(&ctx->completion_lock);
6779 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6781 spin_lock_irq(&ctx->completion_lock);
6782 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6783 spin_unlock_irq(&ctx->completion_lock);
6786 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6788 unsigned int to_submit;
6791 to_submit = io_sqring_entries(ctx);
6792 /* if we're handling multiple rings, cap submit size for fairness */
6793 if (cap_entries && to_submit > 8)
6796 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6797 unsigned nr_events = 0;
6799 mutex_lock(&ctx->uring_lock);
6800 if (!list_empty(&ctx->iopoll_list))
6801 io_do_iopoll(ctx, &nr_events, 0);
6804 * Don't submit if refs are dying, good for io_uring_register(),
6805 * but also it is relied upon by io_ring_exit_work()
6807 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6808 !(ctx->flags & IORING_SETUP_R_DISABLED))
6809 ret = io_submit_sqes(ctx, to_submit);
6810 mutex_unlock(&ctx->uring_lock);
6812 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6813 wake_up(&ctx->sqo_sq_wait);
6819 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6821 struct io_ring_ctx *ctx;
6822 unsigned sq_thread_idle = 0;
6824 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6825 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6826 sqd->sq_thread_idle = sq_thread_idle;
6829 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6831 bool did_sig = false;
6832 struct ksignal ksig;
6834 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6835 signal_pending(current)) {
6836 mutex_unlock(&sqd->lock);
6837 if (signal_pending(current))
6838 did_sig = get_signal(&ksig);
6840 mutex_lock(&sqd->lock);
6843 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6846 static int io_sq_thread(void *data)
6848 struct io_sq_data *sqd = data;
6849 struct io_ring_ctx *ctx;
6850 unsigned long timeout = 0;
6851 char buf[TASK_COMM_LEN];
6854 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6855 set_task_comm(current, buf);
6857 if (sqd->sq_cpu != -1)
6858 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6860 set_cpus_allowed_ptr(current, cpu_online_mask);
6861 current->flags |= PF_NO_SETAFFINITY;
6863 mutex_lock(&sqd->lock);
6866 bool cap_entries, sqt_spin, needs_sched;
6868 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6869 if (io_sqd_handle_event(sqd))
6871 timeout = jiffies + sqd->sq_thread_idle;
6876 cap_entries = !list_is_singular(&sqd->ctx_list);
6877 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6878 const struct cred *creds = NULL;
6880 if (ctx->sq_creds != current_cred())
6881 creds = override_creds(ctx->sq_creds);
6882 ret = __io_sq_thread(ctx, cap_entries);
6884 revert_creds(creds);
6885 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6889 if (sqt_spin || !time_after(jiffies, timeout)) {
6893 timeout = jiffies + sqd->sq_thread_idle;
6897 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6898 if (!io_sqd_events_pending(sqd)) {
6900 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6901 io_ring_set_wakeup_flag(ctx);
6903 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6904 !list_empty_careful(&ctx->iopoll_list)) {
6905 needs_sched = false;
6908 if (io_sqring_entries(ctx)) {
6909 needs_sched = false;
6915 mutex_unlock(&sqd->lock);
6917 mutex_lock(&sqd->lock);
6919 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6920 io_ring_clear_wakeup_flag(ctx);
6923 finish_wait(&sqd->wait, &wait);
6924 timeout = jiffies + sqd->sq_thread_idle;
6927 io_uring_cancel_generic(true, sqd);
6929 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6930 io_ring_set_wakeup_flag(ctx);
6932 mutex_unlock(&sqd->lock);
6934 complete(&sqd->exited);
6938 struct io_wait_queue {
6939 struct wait_queue_entry wq;
6940 struct io_ring_ctx *ctx;
6942 unsigned nr_timeouts;
6945 static inline bool io_should_wake(struct io_wait_queue *iowq)
6947 struct io_ring_ctx *ctx = iowq->ctx;
6950 * Wake up if we have enough events, or if a timeout occurred since we
6951 * started waiting. For timeouts, we always want to return to userspace,
6952 * regardless of event count.
6954 return io_cqring_events(ctx) >= iowq->to_wait ||
6955 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6958 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6959 int wake_flags, void *key)
6961 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6965 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6966 * the task, and the next invocation will do it.
6968 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6969 return autoremove_wake_function(curr, mode, wake_flags, key);
6973 static int io_run_task_work_sig(void)
6975 if (io_run_task_work())
6977 if (!signal_pending(current))
6979 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6980 return -ERESTARTSYS;
6984 /* when returns >0, the caller should retry */
6985 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6986 struct io_wait_queue *iowq,
6987 signed long *timeout)
6991 /* make sure we run task_work before checking for signals */
6992 ret = io_run_task_work_sig();
6993 if (ret || io_should_wake(iowq))
6995 /* let the caller flush overflows, retry */
6996 if (test_bit(0, &ctx->cq_check_overflow))
6999 *timeout = schedule_timeout(*timeout);
7000 return !*timeout ? -ETIME : 1;
7004 * Wait until events become available, if we don't already have some. The
7005 * application must reap them itself, as they reside on the shared cq ring.
7007 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7008 const sigset_t __user *sig, size_t sigsz,
7009 struct __kernel_timespec __user *uts)
7011 struct io_wait_queue iowq = {
7014 .func = io_wake_function,
7015 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7018 .to_wait = min_events,
7020 struct io_rings *rings = ctx->rings;
7021 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7025 io_cqring_overflow_flush(ctx, false);
7026 if (io_cqring_events(ctx) >= min_events)
7028 if (!io_run_task_work())
7033 #ifdef CONFIG_COMPAT
7034 if (in_compat_syscall())
7035 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7039 ret = set_user_sigmask(sig, sigsz);
7046 struct timespec64 ts;
7048 if (get_timespec64(&ts, uts))
7050 timeout = timespec64_to_jiffies(&ts);
7053 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7054 trace_io_uring_cqring_wait(ctx, min_events);
7056 /* if we can't even flush overflow, don't wait for more */
7057 if (!io_cqring_overflow_flush(ctx, false)) {
7061 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7062 TASK_INTERRUPTIBLE);
7063 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7064 finish_wait(&ctx->wait, &iowq.wq);
7068 restore_saved_sigmask_unless(ret == -EINTR);
7070 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7073 static void io_free_page_table(void **table, size_t size)
7075 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7077 for (i = 0; i < nr_tables; i++)
7082 static void **io_alloc_page_table(size_t size)
7084 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7085 size_t init_size = size;
7088 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7092 for (i = 0; i < nr_tables; i++) {
7093 unsigned int this_size = min(size, PAGE_SIZE);
7095 table[i] = kzalloc(this_size, GFP_KERNEL);
7097 io_free_page_table(table, init_size);
7105 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7107 spin_lock_bh(&ctx->rsrc_ref_lock);
7110 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7112 spin_unlock_bh(&ctx->rsrc_ref_lock);
7115 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7117 percpu_ref_exit(&ref_node->refs);
7121 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7122 struct io_rsrc_data *data_to_kill)
7124 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7125 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7128 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7130 rsrc_node->rsrc_data = data_to_kill;
7131 io_rsrc_ref_lock(ctx);
7132 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7133 io_rsrc_ref_unlock(ctx);
7135 atomic_inc(&data_to_kill->refs);
7136 percpu_ref_kill(&rsrc_node->refs);
7137 ctx->rsrc_node = NULL;
7140 if (!ctx->rsrc_node) {
7141 ctx->rsrc_node = ctx->rsrc_backup_node;
7142 ctx->rsrc_backup_node = NULL;
7146 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7148 if (ctx->rsrc_backup_node)
7150 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7151 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7154 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7158 /* As we may drop ->uring_lock, other task may have started quiesce */
7162 data->quiesce = true;
7164 ret = io_rsrc_node_switch_start(ctx);
7167 io_rsrc_node_switch(ctx, data);
7169 /* kill initial ref, already quiesced if zero */
7170 if (atomic_dec_and_test(&data->refs))
7172 flush_delayed_work(&ctx->rsrc_put_work);
7173 ret = wait_for_completion_interruptible(&data->done);
7177 atomic_inc(&data->refs);
7178 /* wait for all works potentially completing data->done */
7179 flush_delayed_work(&ctx->rsrc_put_work);
7180 reinit_completion(&data->done);
7182 mutex_unlock(&ctx->uring_lock);
7183 ret = io_run_task_work_sig();
7184 mutex_lock(&ctx->uring_lock);
7186 data->quiesce = false;
7191 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7193 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7194 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7196 return &data->tags[table_idx][off];
7199 static void io_rsrc_data_free(struct io_rsrc_data *data)
7201 size_t size = data->nr * sizeof(data->tags[0][0]);
7204 io_free_page_table((void **)data->tags, size);
7208 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7209 u64 __user *utags, unsigned nr,
7210 struct io_rsrc_data **pdata)
7212 struct io_rsrc_data *data;
7216 data = kzalloc(sizeof(*data), GFP_KERNEL);
7219 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7227 data->do_put = do_put;
7230 for (i = 0; i < nr; i++) {
7231 u64 *tag_slot = io_get_tag_slot(data, i);
7233 if (copy_from_user(tag_slot, &utags[i],
7239 atomic_set(&data->refs, 1);
7240 init_completion(&data->done);
7244 io_rsrc_data_free(data);
7248 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7250 size_t size = nr_files * sizeof(struct io_fixed_file);
7252 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7253 return !!table->files;
7256 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7258 size_t size = nr_files * sizeof(struct io_fixed_file);
7260 io_free_page_table((void **)table->files, size);
7261 table->files = NULL;
7264 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7266 #if defined(CONFIG_UNIX)
7267 if (ctx->ring_sock) {
7268 struct sock *sock = ctx->ring_sock->sk;
7269 struct sk_buff *skb;
7271 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7277 for (i = 0; i < ctx->nr_user_files; i++) {
7280 file = io_file_from_index(ctx, i);
7285 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7286 io_rsrc_data_free(ctx->file_data);
7287 ctx->file_data = NULL;
7288 ctx->nr_user_files = 0;
7291 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7295 if (!ctx->file_data)
7297 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7299 __io_sqe_files_unregister(ctx);
7303 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7304 __releases(&sqd->lock)
7306 WARN_ON_ONCE(sqd->thread == current);
7309 * Do the dance but not conditional clear_bit() because it'd race with
7310 * other threads incrementing park_pending and setting the bit.
7312 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7313 if (atomic_dec_return(&sqd->park_pending))
7314 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7315 mutex_unlock(&sqd->lock);
7318 static void io_sq_thread_park(struct io_sq_data *sqd)
7319 __acquires(&sqd->lock)
7321 WARN_ON_ONCE(sqd->thread == current);
7323 atomic_inc(&sqd->park_pending);
7324 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7325 mutex_lock(&sqd->lock);
7327 wake_up_process(sqd->thread);
7330 static void io_sq_thread_stop(struct io_sq_data *sqd)
7332 WARN_ON_ONCE(sqd->thread == current);
7333 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7335 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7336 mutex_lock(&sqd->lock);
7338 wake_up_process(sqd->thread);
7339 mutex_unlock(&sqd->lock);
7340 wait_for_completion(&sqd->exited);
7343 static void io_put_sq_data(struct io_sq_data *sqd)
7345 if (refcount_dec_and_test(&sqd->refs)) {
7346 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7348 io_sq_thread_stop(sqd);
7353 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7355 struct io_sq_data *sqd = ctx->sq_data;
7358 io_sq_thread_park(sqd);
7359 list_del_init(&ctx->sqd_list);
7360 io_sqd_update_thread_idle(sqd);
7361 io_sq_thread_unpark(sqd);
7363 io_put_sq_data(sqd);
7364 ctx->sq_data = NULL;
7368 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7370 struct io_ring_ctx *ctx_attach;
7371 struct io_sq_data *sqd;
7374 f = fdget(p->wq_fd);
7376 return ERR_PTR(-ENXIO);
7377 if (f.file->f_op != &io_uring_fops) {
7379 return ERR_PTR(-EINVAL);
7382 ctx_attach = f.file->private_data;
7383 sqd = ctx_attach->sq_data;
7386 return ERR_PTR(-EINVAL);
7388 if (sqd->task_tgid != current->tgid) {
7390 return ERR_PTR(-EPERM);
7393 refcount_inc(&sqd->refs);
7398 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7401 struct io_sq_data *sqd;
7404 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7405 sqd = io_attach_sq_data(p);
7410 /* fall through for EPERM case, setup new sqd/task */
7411 if (PTR_ERR(sqd) != -EPERM)
7415 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7417 return ERR_PTR(-ENOMEM);
7419 atomic_set(&sqd->park_pending, 0);
7420 refcount_set(&sqd->refs, 1);
7421 INIT_LIST_HEAD(&sqd->ctx_list);
7422 mutex_init(&sqd->lock);
7423 init_waitqueue_head(&sqd->wait);
7424 init_completion(&sqd->exited);
7428 #if defined(CONFIG_UNIX)
7430 * Ensure the UNIX gc is aware of our file set, so we are certain that
7431 * the io_uring can be safely unregistered on process exit, even if we have
7432 * loops in the file referencing.
7434 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7436 struct sock *sk = ctx->ring_sock->sk;
7437 struct scm_fp_list *fpl;
7438 struct sk_buff *skb;
7441 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7445 skb = alloc_skb(0, GFP_KERNEL);
7454 fpl->user = get_uid(current_user());
7455 for (i = 0; i < nr; i++) {
7456 struct file *file = io_file_from_index(ctx, i + offset);
7460 fpl->fp[nr_files] = get_file(file);
7461 unix_inflight(fpl->user, fpl->fp[nr_files]);
7466 fpl->max = SCM_MAX_FD;
7467 fpl->count = nr_files;
7468 UNIXCB(skb).fp = fpl;
7469 skb->destructor = unix_destruct_scm;
7470 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7471 skb_queue_head(&sk->sk_receive_queue, skb);
7473 for (i = 0; i < nr_files; i++)
7484 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7485 * causes regular reference counting to break down. We rely on the UNIX
7486 * garbage collection to take care of this problem for us.
7488 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7490 unsigned left, total;
7494 left = ctx->nr_user_files;
7496 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7498 ret = __io_sqe_files_scm(ctx, this_files, total);
7502 total += this_files;
7508 while (total < ctx->nr_user_files) {
7509 struct file *file = io_file_from_index(ctx, total);
7519 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7525 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7527 struct file *file = prsrc->file;
7528 #if defined(CONFIG_UNIX)
7529 struct sock *sock = ctx->ring_sock->sk;
7530 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7531 struct sk_buff *skb;
7534 __skb_queue_head_init(&list);
7537 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7538 * remove this entry and rearrange the file array.
7540 skb = skb_dequeue(head);
7542 struct scm_fp_list *fp;
7544 fp = UNIXCB(skb).fp;
7545 for (i = 0; i < fp->count; i++) {
7548 if (fp->fp[i] != file)
7551 unix_notinflight(fp->user, fp->fp[i]);
7552 left = fp->count - 1 - i;
7554 memmove(&fp->fp[i], &fp->fp[i + 1],
7555 left * sizeof(struct file *));
7562 __skb_queue_tail(&list, skb);
7572 __skb_queue_tail(&list, skb);
7574 skb = skb_dequeue(head);
7577 if (skb_peek(&list)) {
7578 spin_lock_irq(&head->lock);
7579 while ((skb = __skb_dequeue(&list)) != NULL)
7580 __skb_queue_tail(head, skb);
7581 spin_unlock_irq(&head->lock);
7588 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7590 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7591 struct io_ring_ctx *ctx = rsrc_data->ctx;
7592 struct io_rsrc_put *prsrc, *tmp;
7594 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7595 list_del(&prsrc->list);
7598 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7600 io_ring_submit_lock(ctx, lock_ring);
7601 spin_lock_irq(&ctx->completion_lock);
7602 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7604 io_commit_cqring(ctx);
7605 spin_unlock_irq(&ctx->completion_lock);
7606 io_cqring_ev_posted(ctx);
7607 io_ring_submit_unlock(ctx, lock_ring);
7610 rsrc_data->do_put(ctx, prsrc);
7614 io_rsrc_node_destroy(ref_node);
7615 if (atomic_dec_and_test(&rsrc_data->refs))
7616 complete(&rsrc_data->done);
7619 static void io_rsrc_put_work(struct work_struct *work)
7621 struct io_ring_ctx *ctx;
7622 struct llist_node *node;
7624 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7625 node = llist_del_all(&ctx->rsrc_put_llist);
7628 struct io_rsrc_node *ref_node;
7629 struct llist_node *next = node->next;
7631 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7632 __io_rsrc_put_work(ref_node);
7637 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7639 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7640 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7641 bool first_add = false;
7643 io_rsrc_ref_lock(ctx);
7646 while (!list_empty(&ctx->rsrc_ref_list)) {
7647 node = list_first_entry(&ctx->rsrc_ref_list,
7648 struct io_rsrc_node, node);
7649 /* recycle ref nodes in order */
7652 list_del(&node->node);
7653 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7655 io_rsrc_ref_unlock(ctx);
7658 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7661 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7663 struct io_rsrc_node *ref_node;
7665 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7669 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7674 INIT_LIST_HEAD(&ref_node->node);
7675 INIT_LIST_HEAD(&ref_node->rsrc_list);
7676 ref_node->done = false;
7680 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7681 unsigned nr_args, u64 __user *tags)
7683 __s32 __user *fds = (__s32 __user *) arg;
7692 if (nr_args > IORING_MAX_FIXED_FILES)
7694 ret = io_rsrc_node_switch_start(ctx);
7697 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7703 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7706 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7707 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7711 /* allow sparse sets */
7714 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7721 if (unlikely(!file))
7725 * Don't allow io_uring instances to be registered. If UNIX
7726 * isn't enabled, then this causes a reference cycle and this
7727 * instance can never get freed. If UNIX is enabled we'll
7728 * handle it just fine, but there's still no point in allowing
7729 * a ring fd as it doesn't support regular read/write anyway.
7731 if (file->f_op == &io_uring_fops) {
7735 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7738 ret = io_sqe_files_scm(ctx);
7740 __io_sqe_files_unregister(ctx);
7744 io_rsrc_node_switch(ctx, NULL);
7747 for (i = 0; i < ctx->nr_user_files; i++) {
7748 file = io_file_from_index(ctx, i);
7752 io_free_file_tables(&ctx->file_table, nr_args);
7753 ctx->nr_user_files = 0;
7755 io_rsrc_data_free(ctx->file_data);
7756 ctx->file_data = NULL;
7760 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7763 #if defined(CONFIG_UNIX)
7764 struct sock *sock = ctx->ring_sock->sk;
7765 struct sk_buff_head *head = &sock->sk_receive_queue;
7766 struct sk_buff *skb;
7769 * See if we can merge this file into an existing skb SCM_RIGHTS
7770 * file set. If there's no room, fall back to allocating a new skb
7771 * and filling it in.
7773 spin_lock_irq(&head->lock);
7774 skb = skb_peek(head);
7776 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7778 if (fpl->count < SCM_MAX_FD) {
7779 __skb_unlink(skb, head);
7780 spin_unlock_irq(&head->lock);
7781 fpl->fp[fpl->count] = get_file(file);
7782 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7784 spin_lock_irq(&head->lock);
7785 __skb_queue_head(head, skb);
7790 spin_unlock_irq(&head->lock);
7797 return __io_sqe_files_scm(ctx, 1, index);
7803 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7804 struct io_rsrc_node *node, void *rsrc)
7806 struct io_rsrc_put *prsrc;
7808 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7812 prsrc->tag = *io_get_tag_slot(data, idx);
7814 list_add(&prsrc->list, &node->rsrc_list);
7818 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7819 struct io_uring_rsrc_update2 *up,
7822 u64 __user *tags = u64_to_user_ptr(up->tags);
7823 __s32 __user *fds = u64_to_user_ptr(up->data);
7824 struct io_rsrc_data *data = ctx->file_data;
7825 struct io_fixed_file *file_slot;
7829 bool needs_switch = false;
7831 if (!ctx->file_data)
7833 if (up->offset + nr_args > ctx->nr_user_files)
7836 for (done = 0; done < nr_args; done++) {
7839 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7840 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7844 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7848 if (fd == IORING_REGISTER_FILES_SKIP)
7851 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7852 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7854 if (file_slot->file_ptr) {
7855 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7856 err = io_queue_rsrc_removal(data, up->offset + done,
7857 ctx->rsrc_node, file);
7860 file_slot->file_ptr = 0;
7861 needs_switch = true;
7870 * Don't allow io_uring instances to be registered. If
7871 * UNIX isn't enabled, then this causes a reference
7872 * cycle and this instance can never get freed. If UNIX
7873 * is enabled we'll handle it just fine, but there's
7874 * still no point in allowing a ring fd as it doesn't
7875 * support regular read/write anyway.
7877 if (file->f_op == &io_uring_fops) {
7882 *io_get_tag_slot(data, up->offset + done) = tag;
7883 io_fixed_file_set(file_slot, file);
7884 err = io_sqe_file_register(ctx, file, i);
7886 file_slot->file_ptr = 0;
7894 io_rsrc_node_switch(ctx, data);
7895 return done ? done : err;
7898 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7900 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7902 req = io_put_req_find_next(req);
7903 return req ? &req->work : NULL;
7906 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7907 struct task_struct *task)
7909 struct io_wq_hash *hash;
7910 struct io_wq_data data;
7911 unsigned int concurrency;
7913 hash = ctx->hash_map;
7915 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7917 return ERR_PTR(-ENOMEM);
7918 refcount_set(&hash->refs, 1);
7919 init_waitqueue_head(&hash->wait);
7920 ctx->hash_map = hash;
7925 data.free_work = io_free_work;
7926 data.do_work = io_wq_submit_work;
7928 /* Do QD, or 4 * CPUS, whatever is smallest */
7929 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7931 return io_wq_create(concurrency, &data);
7934 static int io_uring_alloc_task_context(struct task_struct *task,
7935 struct io_ring_ctx *ctx)
7937 struct io_uring_task *tctx;
7940 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7941 if (unlikely(!tctx))
7944 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7945 if (unlikely(ret)) {
7950 tctx->io_wq = io_init_wq_offload(ctx, task);
7951 if (IS_ERR(tctx->io_wq)) {
7952 ret = PTR_ERR(tctx->io_wq);
7953 percpu_counter_destroy(&tctx->inflight);
7959 init_waitqueue_head(&tctx->wait);
7960 atomic_set(&tctx->in_idle, 0);
7961 atomic_set(&tctx->inflight_tracked, 0);
7962 task->io_uring = tctx;
7963 spin_lock_init(&tctx->task_lock);
7964 INIT_WQ_LIST(&tctx->task_list);
7965 init_task_work(&tctx->task_work, tctx_task_work);
7969 void __io_uring_free(struct task_struct *tsk)
7971 struct io_uring_task *tctx = tsk->io_uring;
7973 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7974 WARN_ON_ONCE(tctx->io_wq);
7975 WARN_ON_ONCE(tctx->cached_refs);
7977 percpu_counter_destroy(&tctx->inflight);
7979 tsk->io_uring = NULL;
7982 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7983 struct io_uring_params *p)
7987 /* Retain compatibility with failing for an invalid attach attempt */
7988 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7989 IORING_SETUP_ATTACH_WQ) {
7992 f = fdget(p->wq_fd);
7996 if (f.file->f_op != &io_uring_fops)
7999 if (ctx->flags & IORING_SETUP_SQPOLL) {
8000 struct task_struct *tsk;
8001 struct io_sq_data *sqd;
8004 sqd = io_get_sq_data(p, &attached);
8010 ctx->sq_creds = get_current_cred();
8012 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8013 if (!ctx->sq_thread_idle)
8014 ctx->sq_thread_idle = HZ;
8016 io_sq_thread_park(sqd);
8017 list_add(&ctx->sqd_list, &sqd->ctx_list);
8018 io_sqd_update_thread_idle(sqd);
8019 /* don't attach to a dying SQPOLL thread, would be racy */
8020 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8021 io_sq_thread_unpark(sqd);
8028 if (p->flags & IORING_SETUP_SQ_AFF) {
8029 int cpu = p->sq_thread_cpu;
8032 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8039 sqd->task_pid = current->pid;
8040 sqd->task_tgid = current->tgid;
8041 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8048 ret = io_uring_alloc_task_context(tsk, ctx);
8049 wake_up_new_task(tsk);
8052 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8053 /* Can't have SQ_AFF without SQPOLL */
8060 complete(&ctx->sq_data->exited);
8062 io_sq_thread_finish(ctx);
8066 static inline void __io_unaccount_mem(struct user_struct *user,
8067 unsigned long nr_pages)
8069 atomic_long_sub(nr_pages, &user->locked_vm);
8072 static inline int __io_account_mem(struct user_struct *user,
8073 unsigned long nr_pages)
8075 unsigned long page_limit, cur_pages, new_pages;
8077 /* Don't allow more pages than we can safely lock */
8078 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8081 cur_pages = atomic_long_read(&user->locked_vm);
8082 new_pages = cur_pages + nr_pages;
8083 if (new_pages > page_limit)
8085 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8086 new_pages) != cur_pages);
8091 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8094 __io_unaccount_mem(ctx->user, nr_pages);
8096 if (ctx->mm_account)
8097 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8100 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8105 ret = __io_account_mem(ctx->user, nr_pages);
8110 if (ctx->mm_account)
8111 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8116 static void io_mem_free(void *ptr)
8123 page = virt_to_head_page(ptr);
8124 if (put_page_testzero(page))
8125 free_compound_page(page);
8128 static void *io_mem_alloc(size_t size)
8130 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8131 __GFP_NORETRY | __GFP_ACCOUNT;
8133 return (void *) __get_free_pages(gfp_flags, get_order(size));
8136 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8139 struct io_rings *rings;
8140 size_t off, sq_array_size;
8142 off = struct_size(rings, cqes, cq_entries);
8143 if (off == SIZE_MAX)
8147 off = ALIGN(off, SMP_CACHE_BYTES);
8155 sq_array_size = array_size(sizeof(u32), sq_entries);
8156 if (sq_array_size == SIZE_MAX)
8159 if (check_add_overflow(off, sq_array_size, &off))
8165 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8167 struct io_mapped_ubuf *imu = *slot;
8170 if (imu != ctx->dummy_ubuf) {
8171 for (i = 0; i < imu->nr_bvecs; i++)
8172 unpin_user_page(imu->bvec[i].bv_page);
8173 if (imu->acct_pages)
8174 io_unaccount_mem(ctx, imu->acct_pages);
8180 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8182 io_buffer_unmap(ctx, &prsrc->buf);
8186 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8190 for (i = 0; i < ctx->nr_user_bufs; i++)
8191 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8192 kfree(ctx->user_bufs);
8193 io_rsrc_data_free(ctx->buf_data);
8194 ctx->user_bufs = NULL;
8195 ctx->buf_data = NULL;
8196 ctx->nr_user_bufs = 0;
8199 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8206 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8208 __io_sqe_buffers_unregister(ctx);
8212 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8213 void __user *arg, unsigned index)
8215 struct iovec __user *src;
8217 #ifdef CONFIG_COMPAT
8219 struct compat_iovec __user *ciovs;
8220 struct compat_iovec ciov;
8222 ciovs = (struct compat_iovec __user *) arg;
8223 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8226 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8227 dst->iov_len = ciov.iov_len;
8231 src = (struct iovec __user *) arg;
8232 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8238 * Not super efficient, but this is just a registration time. And we do cache
8239 * the last compound head, so generally we'll only do a full search if we don't
8242 * We check if the given compound head page has already been accounted, to
8243 * avoid double accounting it. This allows us to account the full size of the
8244 * page, not just the constituent pages of a huge page.
8246 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8247 int nr_pages, struct page *hpage)
8251 /* check current page array */
8252 for (i = 0; i < nr_pages; i++) {
8253 if (!PageCompound(pages[i]))
8255 if (compound_head(pages[i]) == hpage)
8259 /* check previously registered pages */
8260 for (i = 0; i < ctx->nr_user_bufs; i++) {
8261 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8263 for (j = 0; j < imu->nr_bvecs; j++) {
8264 if (!PageCompound(imu->bvec[j].bv_page))
8266 if (compound_head(imu->bvec[j].bv_page) == hpage)
8274 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8275 int nr_pages, struct io_mapped_ubuf *imu,
8276 struct page **last_hpage)
8280 imu->acct_pages = 0;
8281 for (i = 0; i < nr_pages; i++) {
8282 if (!PageCompound(pages[i])) {
8287 hpage = compound_head(pages[i]);
8288 if (hpage == *last_hpage)
8290 *last_hpage = hpage;
8291 if (headpage_already_acct(ctx, pages, i, hpage))
8293 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8297 if (!imu->acct_pages)
8300 ret = io_account_mem(ctx, imu->acct_pages);
8302 imu->acct_pages = 0;
8306 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8307 struct io_mapped_ubuf **pimu,
8308 struct page **last_hpage)
8310 struct io_mapped_ubuf *imu = NULL;
8311 struct vm_area_struct **vmas = NULL;
8312 struct page **pages = NULL;
8313 unsigned long off, start, end, ubuf;
8315 int ret, pret, nr_pages, i;
8317 if (!iov->iov_base) {
8318 *pimu = ctx->dummy_ubuf;
8322 ubuf = (unsigned long) iov->iov_base;
8323 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8324 start = ubuf >> PAGE_SHIFT;
8325 nr_pages = end - start;
8330 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8334 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8339 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8344 mmap_read_lock(current->mm);
8345 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8347 if (pret == nr_pages) {
8348 /* don't support file backed memory */
8349 for (i = 0; i < nr_pages; i++) {
8350 struct vm_area_struct *vma = vmas[i];
8352 if (vma_is_shmem(vma))
8355 !is_file_hugepages(vma->vm_file)) {
8361 ret = pret < 0 ? pret : -EFAULT;
8363 mmap_read_unlock(current->mm);
8366 * if we did partial map, or found file backed vmas,
8367 * release any pages we did get
8370 unpin_user_pages(pages, pret);
8374 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8376 unpin_user_pages(pages, pret);
8380 off = ubuf & ~PAGE_MASK;
8381 size = iov->iov_len;
8382 for (i = 0; i < nr_pages; i++) {
8385 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8386 imu->bvec[i].bv_page = pages[i];
8387 imu->bvec[i].bv_len = vec_len;
8388 imu->bvec[i].bv_offset = off;
8392 /* store original address for later verification */
8394 imu->ubuf_end = ubuf + iov->iov_len;
8395 imu->nr_bvecs = nr_pages;
8406 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8408 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8409 return ctx->user_bufs ? 0 : -ENOMEM;
8412 static int io_buffer_validate(struct iovec *iov)
8414 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8417 * Don't impose further limits on the size and buffer
8418 * constraints here, we'll -EINVAL later when IO is
8419 * submitted if they are wrong.
8422 return iov->iov_len ? -EFAULT : 0;
8426 /* arbitrary limit, but we need something */
8427 if (iov->iov_len > SZ_1G)
8430 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8436 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8437 unsigned int nr_args, u64 __user *tags)
8439 struct page *last_hpage = NULL;
8440 struct io_rsrc_data *data;
8446 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8448 ret = io_rsrc_node_switch_start(ctx);
8451 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8454 ret = io_buffers_map_alloc(ctx, nr_args);
8456 io_rsrc_data_free(data);
8460 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8461 ret = io_copy_iov(ctx, &iov, arg, i);
8464 ret = io_buffer_validate(&iov);
8467 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8472 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8478 WARN_ON_ONCE(ctx->buf_data);
8480 ctx->buf_data = data;
8482 __io_sqe_buffers_unregister(ctx);
8484 io_rsrc_node_switch(ctx, NULL);
8488 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8489 struct io_uring_rsrc_update2 *up,
8490 unsigned int nr_args)
8492 u64 __user *tags = u64_to_user_ptr(up->tags);
8493 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8494 struct page *last_hpage = NULL;
8495 bool needs_switch = false;
8501 if (up->offset + nr_args > ctx->nr_user_bufs)
8504 for (done = 0; done < nr_args; done++) {
8505 struct io_mapped_ubuf *imu;
8506 int offset = up->offset + done;
8509 err = io_copy_iov(ctx, &iov, iovs, done);
8512 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8516 err = io_buffer_validate(&iov);
8519 if (!iov.iov_base && tag) {
8523 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8527 i = array_index_nospec(offset, ctx->nr_user_bufs);
8528 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8529 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8530 ctx->rsrc_node, ctx->user_bufs[i]);
8531 if (unlikely(err)) {
8532 io_buffer_unmap(ctx, &imu);
8535 ctx->user_bufs[i] = NULL;
8536 needs_switch = true;
8539 ctx->user_bufs[i] = imu;
8540 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8544 io_rsrc_node_switch(ctx, ctx->buf_data);
8545 return done ? done : err;
8548 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8550 __s32 __user *fds = arg;
8556 if (copy_from_user(&fd, fds, sizeof(*fds)))
8559 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8560 if (IS_ERR(ctx->cq_ev_fd)) {
8561 int ret = PTR_ERR(ctx->cq_ev_fd);
8562 ctx->cq_ev_fd = NULL;
8569 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8571 if (ctx->cq_ev_fd) {
8572 eventfd_ctx_put(ctx->cq_ev_fd);
8573 ctx->cq_ev_fd = NULL;
8580 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8582 struct io_buffer *buf;
8583 unsigned long index;
8585 xa_for_each(&ctx->io_buffers, index, buf)
8586 __io_remove_buffers(ctx, buf, index, -1U);
8589 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8591 struct io_kiocb *req, *nxt;
8593 list_for_each_entry_safe(req, nxt, list, compl.list) {
8594 if (tsk && req->task != tsk)
8596 list_del(&req->compl.list);
8597 kmem_cache_free(req_cachep, req);
8601 static void io_req_caches_free(struct io_ring_ctx *ctx)
8603 struct io_submit_state *submit_state = &ctx->submit_state;
8604 struct io_comp_state *cs = &ctx->submit_state.comp;
8606 mutex_lock(&ctx->uring_lock);
8608 if (submit_state->free_reqs) {
8609 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8610 submit_state->reqs);
8611 submit_state->free_reqs = 0;
8614 io_flush_cached_locked_reqs(ctx, cs);
8615 io_req_cache_free(&cs->free_list, NULL);
8616 mutex_unlock(&ctx->uring_lock);
8619 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8623 if (!atomic_dec_and_test(&data->refs))
8624 wait_for_completion(&data->done);
8628 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8630 io_sq_thread_finish(ctx);
8632 if (ctx->mm_account) {
8633 mmdrop(ctx->mm_account);
8634 ctx->mm_account = NULL;
8637 mutex_lock(&ctx->uring_lock);
8638 if (io_wait_rsrc_data(ctx->buf_data))
8639 __io_sqe_buffers_unregister(ctx);
8640 if (io_wait_rsrc_data(ctx->file_data))
8641 __io_sqe_files_unregister(ctx);
8643 __io_cqring_overflow_flush(ctx, true);
8644 mutex_unlock(&ctx->uring_lock);
8645 io_eventfd_unregister(ctx);
8646 io_destroy_buffers(ctx);
8648 put_cred(ctx->sq_creds);
8650 /* there are no registered resources left, nobody uses it */
8652 io_rsrc_node_destroy(ctx->rsrc_node);
8653 if (ctx->rsrc_backup_node)
8654 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8655 flush_delayed_work(&ctx->rsrc_put_work);
8657 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8658 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8660 #if defined(CONFIG_UNIX)
8661 if (ctx->ring_sock) {
8662 ctx->ring_sock->file = NULL; /* so that iput() is called */
8663 sock_release(ctx->ring_sock);
8667 io_mem_free(ctx->rings);
8668 io_mem_free(ctx->sq_sqes);
8670 percpu_ref_exit(&ctx->refs);
8671 free_uid(ctx->user);
8672 io_req_caches_free(ctx);
8674 io_wq_put_hash(ctx->hash_map);
8675 kfree(ctx->cancel_hash);
8676 kfree(ctx->dummy_ubuf);
8680 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8682 struct io_ring_ctx *ctx = file->private_data;
8685 poll_wait(file, &ctx->cq_wait, wait);
8687 * synchronizes with barrier from wq_has_sleeper call in
8691 if (!io_sqring_full(ctx))
8692 mask |= EPOLLOUT | EPOLLWRNORM;
8695 * Don't flush cqring overflow list here, just do a simple check.
8696 * Otherwise there could possible be ABBA deadlock:
8699 * lock(&ctx->uring_lock);
8701 * lock(&ctx->uring_lock);
8704 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8705 * pushs them to do the flush.
8707 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8708 mask |= EPOLLIN | EPOLLRDNORM;
8713 static int io_uring_fasync(int fd, struct file *file, int on)
8715 struct io_ring_ctx *ctx = file->private_data;
8717 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8720 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8722 const struct cred *creds;
8724 creds = xa_erase(&ctx->personalities, id);
8733 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8735 return io_run_task_work_head(&ctx->exit_task_work);
8738 struct io_tctx_exit {
8739 struct callback_head task_work;
8740 struct completion completion;
8741 struct io_ring_ctx *ctx;
8744 static void io_tctx_exit_cb(struct callback_head *cb)
8746 struct io_uring_task *tctx = current->io_uring;
8747 struct io_tctx_exit *work;
8749 work = container_of(cb, struct io_tctx_exit, task_work);
8751 * When @in_idle, we're in cancellation and it's racy to remove the
8752 * node. It'll be removed by the end of cancellation, just ignore it.
8754 if (!atomic_read(&tctx->in_idle))
8755 io_uring_del_tctx_node((unsigned long)work->ctx);
8756 complete(&work->completion);
8759 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8761 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8763 return req->ctx == data;
8766 static void io_ring_exit_work(struct work_struct *work)
8768 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8769 unsigned long timeout = jiffies + HZ * 60 * 5;
8770 struct io_tctx_exit exit;
8771 struct io_tctx_node *node;
8775 * If we're doing polled IO and end up having requests being
8776 * submitted async (out-of-line), then completions can come in while
8777 * we're waiting for refs to drop. We need to reap these manually,
8778 * as nobody else will be looking for them.
8781 io_uring_try_cancel_requests(ctx, NULL, true);
8783 struct io_sq_data *sqd = ctx->sq_data;
8784 struct task_struct *tsk;
8786 io_sq_thread_park(sqd);
8788 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8789 io_wq_cancel_cb(tsk->io_uring->io_wq,
8790 io_cancel_ctx_cb, ctx, true);
8791 io_sq_thread_unpark(sqd);
8794 WARN_ON_ONCE(time_after(jiffies, timeout));
8795 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8797 init_completion(&exit.completion);
8798 init_task_work(&exit.task_work, io_tctx_exit_cb);
8801 * Some may use context even when all refs and requests have been put,
8802 * and they are free to do so while still holding uring_lock or
8803 * completion_lock, see __io_req_task_submit(). Apart from other work,
8804 * this lock/unlock section also waits them to finish.
8806 mutex_lock(&ctx->uring_lock);
8807 while (!list_empty(&ctx->tctx_list)) {
8808 WARN_ON_ONCE(time_after(jiffies, timeout));
8810 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8812 /* don't spin on a single task if cancellation failed */
8813 list_rotate_left(&ctx->tctx_list);
8814 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8815 if (WARN_ON_ONCE(ret))
8817 wake_up_process(node->task);
8819 mutex_unlock(&ctx->uring_lock);
8820 wait_for_completion(&exit.completion);
8821 mutex_lock(&ctx->uring_lock);
8823 mutex_unlock(&ctx->uring_lock);
8824 spin_lock_irq(&ctx->completion_lock);
8825 spin_unlock_irq(&ctx->completion_lock);
8827 io_ring_ctx_free(ctx);
8830 /* Returns true if we found and killed one or more timeouts */
8831 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8834 struct io_kiocb *req, *tmp;
8837 spin_lock_irq(&ctx->completion_lock);
8838 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8839 if (io_match_task(req, tsk, cancel_all)) {
8840 io_kill_timeout(req, -ECANCELED);
8845 io_commit_cqring(ctx);
8846 spin_unlock_irq(&ctx->completion_lock);
8848 io_cqring_ev_posted(ctx);
8849 return canceled != 0;
8852 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8854 unsigned long index;
8855 struct creds *creds;
8857 mutex_lock(&ctx->uring_lock);
8858 percpu_ref_kill(&ctx->refs);
8860 __io_cqring_overflow_flush(ctx, true);
8861 xa_for_each(&ctx->personalities, index, creds)
8862 io_unregister_personality(ctx, index);
8863 mutex_unlock(&ctx->uring_lock);
8865 io_kill_timeouts(ctx, NULL, true);
8866 io_poll_remove_all(ctx, NULL, true);
8868 /* if we failed setting up the ctx, we might not have any rings */
8869 io_iopoll_try_reap_events(ctx);
8871 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8873 * Use system_unbound_wq to avoid spawning tons of event kworkers
8874 * if we're exiting a ton of rings at the same time. It just adds
8875 * noise and overhead, there's no discernable change in runtime
8876 * over using system_wq.
8878 queue_work(system_unbound_wq, &ctx->exit_work);
8881 static int io_uring_release(struct inode *inode, struct file *file)
8883 struct io_ring_ctx *ctx = file->private_data;
8885 file->private_data = NULL;
8886 io_ring_ctx_wait_and_kill(ctx);
8890 struct io_task_cancel {
8891 struct task_struct *task;
8895 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8897 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8898 struct io_task_cancel *cancel = data;
8901 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8902 unsigned long flags;
8903 struct io_ring_ctx *ctx = req->ctx;
8905 /* protect against races with linked timeouts */
8906 spin_lock_irqsave(&ctx->completion_lock, flags);
8907 ret = io_match_task(req, cancel->task, cancel->all);
8908 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8910 ret = io_match_task(req, cancel->task, cancel->all);
8915 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8916 struct task_struct *task, bool cancel_all)
8918 struct io_defer_entry *de;
8921 spin_lock_irq(&ctx->completion_lock);
8922 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8923 if (io_match_task(de->req, task, cancel_all)) {
8924 list_cut_position(&list, &ctx->defer_list, &de->list);
8928 spin_unlock_irq(&ctx->completion_lock);
8929 if (list_empty(&list))
8932 while (!list_empty(&list)) {
8933 de = list_first_entry(&list, struct io_defer_entry, list);
8934 list_del_init(&de->list);
8935 io_req_complete_failed(de->req, -ECANCELED);
8941 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8943 struct io_tctx_node *node;
8944 enum io_wq_cancel cret;
8947 mutex_lock(&ctx->uring_lock);
8948 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8949 struct io_uring_task *tctx = node->task->io_uring;
8952 * io_wq will stay alive while we hold uring_lock, because it's
8953 * killed after ctx nodes, which requires to take the lock.
8955 if (!tctx || !tctx->io_wq)
8957 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8958 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8960 mutex_unlock(&ctx->uring_lock);
8965 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8966 struct task_struct *task,
8969 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8970 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8973 enum io_wq_cancel cret;
8977 ret |= io_uring_try_cancel_iowq(ctx);
8978 } else if (tctx && tctx->io_wq) {
8980 * Cancels requests of all rings, not only @ctx, but
8981 * it's fine as the task is in exit/exec.
8983 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8985 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8988 /* SQPOLL thread does its own polling */
8989 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8990 (ctx->sq_data && ctx->sq_data->thread == current)) {
8991 while (!list_empty_careful(&ctx->iopoll_list)) {
8992 io_iopoll_try_reap_events(ctx);
8997 ret |= io_cancel_defer_files(ctx, task, cancel_all);
8998 ret |= io_poll_remove_all(ctx, task, cancel_all);
8999 ret |= io_kill_timeouts(ctx, task, cancel_all);
9000 ret |= io_run_task_work();
9001 ret |= io_run_ctx_fallback(ctx);
9008 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9010 struct io_uring_task *tctx = current->io_uring;
9011 struct io_tctx_node *node;
9014 if (unlikely(!tctx)) {
9015 ret = io_uring_alloc_task_context(current, ctx);
9018 tctx = current->io_uring;
9020 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9021 node = kmalloc(sizeof(*node), GFP_KERNEL);
9025 node->task = current;
9027 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9034 mutex_lock(&ctx->uring_lock);
9035 list_add(&node->ctx_node, &ctx->tctx_list);
9036 mutex_unlock(&ctx->uring_lock);
9043 * Note that this task has used io_uring. We use it for cancelation purposes.
9045 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9047 struct io_uring_task *tctx = current->io_uring;
9049 if (likely(tctx && tctx->last == ctx))
9051 return __io_uring_add_tctx_node(ctx);
9055 * Remove this io_uring_file -> task mapping.
9057 static void io_uring_del_tctx_node(unsigned long index)
9059 struct io_uring_task *tctx = current->io_uring;
9060 struct io_tctx_node *node;
9064 node = xa_erase(&tctx->xa, index);
9068 WARN_ON_ONCE(current != node->task);
9069 WARN_ON_ONCE(list_empty(&node->ctx_node));
9071 mutex_lock(&node->ctx->uring_lock);
9072 list_del(&node->ctx_node);
9073 mutex_unlock(&node->ctx->uring_lock);
9075 if (tctx->last == node->ctx)
9080 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9082 struct io_wq *wq = tctx->io_wq;
9083 struct io_tctx_node *node;
9084 unsigned long index;
9086 xa_for_each(&tctx->xa, index, node)
9087 io_uring_del_tctx_node(index);
9090 * Must be after io_uring_del_task_file() (removes nodes under
9091 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9094 io_wq_put_and_exit(wq);
9098 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9101 return atomic_read(&tctx->inflight_tracked);
9102 return percpu_counter_sum(&tctx->inflight);
9105 static void io_uring_drop_tctx_refs(struct task_struct *task)
9107 struct io_uring_task *tctx = task->io_uring;
9108 unsigned int refs = tctx->cached_refs;
9110 tctx->cached_refs = 0;
9111 percpu_counter_sub(&tctx->inflight, refs);
9112 put_task_struct_many(task, refs);
9116 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9117 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9119 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9121 struct io_uring_task *tctx = current->io_uring;
9122 struct io_ring_ctx *ctx;
9126 WARN_ON_ONCE(sqd && sqd->thread != current);
9128 if (!current->io_uring)
9131 io_wq_exit_start(tctx->io_wq);
9133 io_uring_drop_tctx_refs(current);
9134 atomic_inc(&tctx->in_idle);
9136 /* read completions before cancelations */
9137 inflight = tctx_inflight(tctx, !cancel_all);
9142 struct io_tctx_node *node;
9143 unsigned long index;
9145 xa_for_each(&tctx->xa, index, node) {
9146 /* sqpoll task will cancel all its requests */
9147 if (node->ctx->sq_data)
9149 io_uring_try_cancel_requests(node->ctx, current,
9153 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9154 io_uring_try_cancel_requests(ctx, current,
9158 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9160 * If we've seen completions, retry without waiting. This
9161 * avoids a race where a completion comes in before we did
9162 * prepare_to_wait().
9164 if (inflight == tctx_inflight(tctx, !cancel_all))
9166 finish_wait(&tctx->wait, &wait);
9168 atomic_dec(&tctx->in_idle);
9170 io_uring_clean_tctx(tctx);
9172 /* for exec all current's requests should be gone, kill tctx */
9173 __io_uring_free(current);
9177 void __io_uring_cancel(struct files_struct *files)
9179 io_uring_cancel_generic(!files, NULL);
9182 static void *io_uring_validate_mmap_request(struct file *file,
9183 loff_t pgoff, size_t sz)
9185 struct io_ring_ctx *ctx = file->private_data;
9186 loff_t offset = pgoff << PAGE_SHIFT;
9191 case IORING_OFF_SQ_RING:
9192 case IORING_OFF_CQ_RING:
9195 case IORING_OFF_SQES:
9199 return ERR_PTR(-EINVAL);
9202 page = virt_to_head_page(ptr);
9203 if (sz > page_size(page))
9204 return ERR_PTR(-EINVAL);
9211 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9213 size_t sz = vma->vm_end - vma->vm_start;
9217 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9219 return PTR_ERR(ptr);
9221 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9222 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9225 #else /* !CONFIG_MMU */
9227 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9229 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9232 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9234 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9237 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9238 unsigned long addr, unsigned long len,
9239 unsigned long pgoff, unsigned long flags)
9243 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9245 return PTR_ERR(ptr);
9247 return (unsigned long) ptr;
9250 #endif /* !CONFIG_MMU */
9252 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9257 if (!io_sqring_full(ctx))
9259 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9261 if (!io_sqring_full(ctx))
9264 } while (!signal_pending(current));
9266 finish_wait(&ctx->sqo_sq_wait, &wait);
9270 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9271 struct __kernel_timespec __user **ts,
9272 const sigset_t __user **sig)
9274 struct io_uring_getevents_arg arg;
9277 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9278 * is just a pointer to the sigset_t.
9280 if (!(flags & IORING_ENTER_EXT_ARG)) {
9281 *sig = (const sigset_t __user *) argp;
9287 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9288 * timespec and sigset_t pointers if good.
9290 if (*argsz != sizeof(arg))
9292 if (copy_from_user(&arg, argp, sizeof(arg)))
9294 *sig = u64_to_user_ptr(arg.sigmask);
9295 *argsz = arg.sigmask_sz;
9296 *ts = u64_to_user_ptr(arg.ts);
9300 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9301 u32, min_complete, u32, flags, const void __user *, argp,
9304 struct io_ring_ctx *ctx;
9311 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9312 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9316 if (unlikely(!f.file))
9320 if (unlikely(f.file->f_op != &io_uring_fops))
9324 ctx = f.file->private_data;
9325 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9329 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9333 * For SQ polling, the thread will do all submissions and completions.
9334 * Just return the requested submit count, and wake the thread if
9338 if (ctx->flags & IORING_SETUP_SQPOLL) {
9339 io_cqring_overflow_flush(ctx, false);
9342 if (unlikely(ctx->sq_data->thread == NULL)) {
9345 if (flags & IORING_ENTER_SQ_WAKEUP)
9346 wake_up(&ctx->sq_data->wait);
9347 if (flags & IORING_ENTER_SQ_WAIT) {
9348 ret = io_sqpoll_wait_sq(ctx);
9352 submitted = to_submit;
9353 } else if (to_submit) {
9354 ret = io_uring_add_tctx_node(ctx);
9357 mutex_lock(&ctx->uring_lock);
9358 submitted = io_submit_sqes(ctx, to_submit);
9359 mutex_unlock(&ctx->uring_lock);
9361 if (submitted != to_submit)
9364 if (flags & IORING_ENTER_GETEVENTS) {
9365 const sigset_t __user *sig;
9366 struct __kernel_timespec __user *ts;
9368 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9372 min_complete = min(min_complete, ctx->cq_entries);
9375 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9376 * space applications don't need to do io completion events
9377 * polling again, they can rely on io_sq_thread to do polling
9378 * work, which can reduce cpu usage and uring_lock contention.
9380 if (ctx->flags & IORING_SETUP_IOPOLL &&
9381 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9382 ret = io_iopoll_check(ctx, min_complete);
9384 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9389 percpu_ref_put(&ctx->refs);
9392 return submitted ? submitted : ret;
9395 #ifdef CONFIG_PROC_FS
9396 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9397 const struct cred *cred)
9399 struct user_namespace *uns = seq_user_ns(m);
9400 struct group_info *gi;
9405 seq_printf(m, "%5d\n", id);
9406 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9407 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9408 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9409 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9410 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9411 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9412 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9413 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9414 seq_puts(m, "\n\tGroups:\t");
9415 gi = cred->group_info;
9416 for (g = 0; g < gi->ngroups; g++) {
9417 seq_put_decimal_ull(m, g ? " " : "",
9418 from_kgid_munged(uns, gi->gid[g]));
9420 seq_puts(m, "\n\tCapEff:\t");
9421 cap = cred->cap_effective;
9422 CAP_FOR_EACH_U32(__capi)
9423 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9428 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9430 struct io_sq_data *sq = NULL;
9435 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9436 * since fdinfo case grabs it in the opposite direction of normal use
9437 * cases. If we fail to get the lock, we just don't iterate any
9438 * structures that could be going away outside the io_uring mutex.
9440 has_lock = mutex_trylock(&ctx->uring_lock);
9442 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9448 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9449 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9450 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9451 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9452 struct file *f = io_file_from_index(ctx, i);
9455 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9457 seq_printf(m, "%5u: <none>\n", i);
9459 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9460 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9461 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9462 unsigned int len = buf->ubuf_end - buf->ubuf;
9464 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9466 if (has_lock && !xa_empty(&ctx->personalities)) {
9467 unsigned long index;
9468 const struct cred *cred;
9470 seq_printf(m, "Personalities:\n");
9471 xa_for_each(&ctx->personalities, index, cred)
9472 io_uring_show_cred(m, index, cred);
9474 seq_printf(m, "PollList:\n");
9475 spin_lock_irq(&ctx->completion_lock);
9476 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9477 struct hlist_head *list = &ctx->cancel_hash[i];
9478 struct io_kiocb *req;
9480 hlist_for_each_entry(req, list, hash_node)
9481 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9482 req->task->task_works != NULL);
9484 spin_unlock_irq(&ctx->completion_lock);
9486 mutex_unlock(&ctx->uring_lock);
9489 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9491 struct io_ring_ctx *ctx = f->private_data;
9493 if (percpu_ref_tryget(&ctx->refs)) {
9494 __io_uring_show_fdinfo(ctx, m);
9495 percpu_ref_put(&ctx->refs);
9500 static const struct file_operations io_uring_fops = {
9501 .release = io_uring_release,
9502 .mmap = io_uring_mmap,
9504 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9505 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9507 .poll = io_uring_poll,
9508 .fasync = io_uring_fasync,
9509 #ifdef CONFIG_PROC_FS
9510 .show_fdinfo = io_uring_show_fdinfo,
9514 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9515 struct io_uring_params *p)
9517 struct io_rings *rings;
9518 size_t size, sq_array_offset;
9520 /* make sure these are sane, as we already accounted them */
9521 ctx->sq_entries = p->sq_entries;
9522 ctx->cq_entries = p->cq_entries;
9524 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9525 if (size == SIZE_MAX)
9528 rings = io_mem_alloc(size);
9533 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9534 rings->sq_ring_mask = p->sq_entries - 1;
9535 rings->cq_ring_mask = p->cq_entries - 1;
9536 rings->sq_ring_entries = p->sq_entries;
9537 rings->cq_ring_entries = p->cq_entries;
9539 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9540 if (size == SIZE_MAX) {
9541 io_mem_free(ctx->rings);
9546 ctx->sq_sqes = io_mem_alloc(size);
9547 if (!ctx->sq_sqes) {
9548 io_mem_free(ctx->rings);
9556 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9560 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9564 ret = io_uring_add_tctx_node(ctx);
9569 fd_install(fd, file);
9574 * Allocate an anonymous fd, this is what constitutes the application
9575 * visible backing of an io_uring instance. The application mmaps this
9576 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9577 * we have to tie this fd to a socket for file garbage collection purposes.
9579 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9582 #if defined(CONFIG_UNIX)
9585 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9588 return ERR_PTR(ret);
9591 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9592 O_RDWR | O_CLOEXEC);
9593 #if defined(CONFIG_UNIX)
9595 sock_release(ctx->ring_sock);
9596 ctx->ring_sock = NULL;
9598 ctx->ring_sock->file = file;
9604 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9605 struct io_uring_params __user *params)
9607 struct io_ring_ctx *ctx;
9613 if (entries > IORING_MAX_ENTRIES) {
9614 if (!(p->flags & IORING_SETUP_CLAMP))
9616 entries = IORING_MAX_ENTRIES;
9620 * Use twice as many entries for the CQ ring. It's possible for the
9621 * application to drive a higher depth than the size of the SQ ring,
9622 * since the sqes are only used at submission time. This allows for
9623 * some flexibility in overcommitting a bit. If the application has
9624 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9625 * of CQ ring entries manually.
9627 p->sq_entries = roundup_pow_of_two(entries);
9628 if (p->flags & IORING_SETUP_CQSIZE) {
9630 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9631 * to a power-of-two, if it isn't already. We do NOT impose
9632 * any cq vs sq ring sizing.
9636 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9637 if (!(p->flags & IORING_SETUP_CLAMP))
9639 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9641 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9642 if (p->cq_entries < p->sq_entries)
9645 p->cq_entries = 2 * p->sq_entries;
9648 ctx = io_ring_ctx_alloc(p);
9651 ctx->compat = in_compat_syscall();
9652 if (!capable(CAP_IPC_LOCK))
9653 ctx->user = get_uid(current_user());
9656 * This is just grabbed for accounting purposes. When a process exits,
9657 * the mm is exited and dropped before the files, hence we need to hang
9658 * on to this mm purely for the purposes of being able to unaccount
9659 * memory (locked/pinned vm). It's not used for anything else.
9661 mmgrab(current->mm);
9662 ctx->mm_account = current->mm;
9664 ret = io_allocate_scq_urings(ctx, p);
9668 ret = io_sq_offload_create(ctx, p);
9671 /* always set a rsrc node */
9672 ret = io_rsrc_node_switch_start(ctx);
9675 io_rsrc_node_switch(ctx, NULL);
9677 memset(&p->sq_off, 0, sizeof(p->sq_off));
9678 p->sq_off.head = offsetof(struct io_rings, sq.head);
9679 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9680 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9681 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9682 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9683 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9684 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9686 memset(&p->cq_off, 0, sizeof(p->cq_off));
9687 p->cq_off.head = offsetof(struct io_rings, cq.head);
9688 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9689 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9690 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9691 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9692 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9693 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9695 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9696 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9697 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9698 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9699 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9700 IORING_FEAT_RSRC_TAGS;
9702 if (copy_to_user(params, p, sizeof(*p))) {
9707 file = io_uring_get_file(ctx);
9709 ret = PTR_ERR(file);
9714 * Install ring fd as the very last thing, so we don't risk someone
9715 * having closed it before we finish setup
9717 ret = io_uring_install_fd(ctx, file);
9719 /* fput will clean it up */
9724 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9727 io_ring_ctx_wait_and_kill(ctx);
9732 * Sets up an aio uring context, and returns the fd. Applications asks for a
9733 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9734 * params structure passed in.
9736 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9738 struct io_uring_params p;
9741 if (copy_from_user(&p, params, sizeof(p)))
9743 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9748 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9749 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9750 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9751 IORING_SETUP_R_DISABLED))
9754 return io_uring_create(entries, &p, params);
9757 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9758 struct io_uring_params __user *, params)
9760 return io_uring_setup(entries, params);
9763 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9765 struct io_uring_probe *p;
9769 size = struct_size(p, ops, nr_args);
9770 if (size == SIZE_MAX)
9772 p = kzalloc(size, GFP_KERNEL);
9777 if (copy_from_user(p, arg, size))
9780 if (memchr_inv(p, 0, size))
9783 p->last_op = IORING_OP_LAST - 1;
9784 if (nr_args > IORING_OP_LAST)
9785 nr_args = IORING_OP_LAST;
9787 for (i = 0; i < nr_args; i++) {
9789 if (!io_op_defs[i].not_supported)
9790 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9795 if (copy_to_user(arg, p, size))
9802 static int io_register_personality(struct io_ring_ctx *ctx)
9804 const struct cred *creds;
9808 creds = get_current_cred();
9810 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9811 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9818 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9819 unsigned int nr_args)
9821 struct io_uring_restriction *res;
9825 /* Restrictions allowed only if rings started disabled */
9826 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9829 /* We allow only a single restrictions registration */
9830 if (ctx->restrictions.registered)
9833 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9836 size = array_size(nr_args, sizeof(*res));
9837 if (size == SIZE_MAX)
9840 res = memdup_user(arg, size);
9842 return PTR_ERR(res);
9846 for (i = 0; i < nr_args; i++) {
9847 switch (res[i].opcode) {
9848 case IORING_RESTRICTION_REGISTER_OP:
9849 if (res[i].register_op >= IORING_REGISTER_LAST) {
9854 __set_bit(res[i].register_op,
9855 ctx->restrictions.register_op);
9857 case IORING_RESTRICTION_SQE_OP:
9858 if (res[i].sqe_op >= IORING_OP_LAST) {
9863 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9865 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9866 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9868 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9869 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9878 /* Reset all restrictions if an error happened */
9880 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9882 ctx->restrictions.registered = true;
9888 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9890 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9893 if (ctx->restrictions.registered)
9894 ctx->restricted = 1;
9896 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9897 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9898 wake_up(&ctx->sq_data->wait);
9902 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9903 struct io_uring_rsrc_update2 *up,
9911 if (check_add_overflow(up->offset, nr_args, &tmp))
9913 err = io_rsrc_node_switch_start(ctx);
9918 case IORING_RSRC_FILE:
9919 return __io_sqe_files_update(ctx, up, nr_args);
9920 case IORING_RSRC_BUFFER:
9921 return __io_sqe_buffers_update(ctx, up, nr_args);
9926 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9929 struct io_uring_rsrc_update2 up;
9933 memset(&up, 0, sizeof(up));
9934 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9936 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9939 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9940 unsigned size, unsigned type)
9942 struct io_uring_rsrc_update2 up;
9944 if (size != sizeof(up))
9946 if (copy_from_user(&up, arg, sizeof(up)))
9948 if (!up.nr || up.resv)
9950 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9953 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9954 unsigned int size, unsigned int type)
9956 struct io_uring_rsrc_register rr;
9958 /* keep it extendible */
9959 if (size != sizeof(rr))
9962 memset(&rr, 0, sizeof(rr));
9963 if (copy_from_user(&rr, arg, size))
9965 if (!rr.nr || rr.resv || rr.resv2)
9969 case IORING_RSRC_FILE:
9970 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9971 rr.nr, u64_to_user_ptr(rr.tags));
9972 case IORING_RSRC_BUFFER:
9973 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9974 rr.nr, u64_to_user_ptr(rr.tags));
9979 static bool io_register_op_must_quiesce(int op)
9982 case IORING_REGISTER_BUFFERS:
9983 case IORING_UNREGISTER_BUFFERS:
9984 case IORING_REGISTER_FILES:
9985 case IORING_UNREGISTER_FILES:
9986 case IORING_REGISTER_FILES_UPDATE:
9987 case IORING_REGISTER_PROBE:
9988 case IORING_REGISTER_PERSONALITY:
9989 case IORING_UNREGISTER_PERSONALITY:
9990 case IORING_REGISTER_FILES2:
9991 case IORING_REGISTER_FILES_UPDATE2:
9992 case IORING_REGISTER_BUFFERS2:
9993 case IORING_REGISTER_BUFFERS_UPDATE:
10000 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10001 void __user *arg, unsigned nr_args)
10002 __releases(ctx->uring_lock)
10003 __acquires(ctx->uring_lock)
10008 * We're inside the ring mutex, if the ref is already dying, then
10009 * someone else killed the ctx or is already going through
10010 * io_uring_register().
10012 if (percpu_ref_is_dying(&ctx->refs))
10015 if (ctx->restricted) {
10016 if (opcode >= IORING_REGISTER_LAST)
10018 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10019 if (!test_bit(opcode, ctx->restrictions.register_op))
10023 if (io_register_op_must_quiesce(opcode)) {
10024 percpu_ref_kill(&ctx->refs);
10027 * Drop uring mutex before waiting for references to exit. If
10028 * another thread is currently inside io_uring_enter() it might
10029 * need to grab the uring_lock to make progress. If we hold it
10030 * here across the drain wait, then we can deadlock. It's safe
10031 * to drop the mutex here, since no new references will come in
10032 * after we've killed the percpu ref.
10034 mutex_unlock(&ctx->uring_lock);
10036 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10039 ret = io_run_task_work_sig();
10043 mutex_lock(&ctx->uring_lock);
10046 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10052 case IORING_REGISTER_BUFFERS:
10053 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10055 case IORING_UNREGISTER_BUFFERS:
10057 if (arg || nr_args)
10059 ret = io_sqe_buffers_unregister(ctx);
10061 case IORING_REGISTER_FILES:
10062 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10064 case IORING_UNREGISTER_FILES:
10066 if (arg || nr_args)
10068 ret = io_sqe_files_unregister(ctx);
10070 case IORING_REGISTER_FILES_UPDATE:
10071 ret = io_register_files_update(ctx, arg, nr_args);
10073 case IORING_REGISTER_EVENTFD:
10074 case IORING_REGISTER_EVENTFD_ASYNC:
10078 ret = io_eventfd_register(ctx, arg);
10081 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10082 ctx->eventfd_async = 1;
10084 ctx->eventfd_async = 0;
10086 case IORING_UNREGISTER_EVENTFD:
10088 if (arg || nr_args)
10090 ret = io_eventfd_unregister(ctx);
10092 case IORING_REGISTER_PROBE:
10094 if (!arg || nr_args > 256)
10096 ret = io_probe(ctx, arg, nr_args);
10098 case IORING_REGISTER_PERSONALITY:
10100 if (arg || nr_args)
10102 ret = io_register_personality(ctx);
10104 case IORING_UNREGISTER_PERSONALITY:
10108 ret = io_unregister_personality(ctx, nr_args);
10110 case IORING_REGISTER_ENABLE_RINGS:
10112 if (arg || nr_args)
10114 ret = io_register_enable_rings(ctx);
10116 case IORING_REGISTER_RESTRICTIONS:
10117 ret = io_register_restrictions(ctx, arg, nr_args);
10119 case IORING_REGISTER_FILES2:
10120 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10122 case IORING_REGISTER_FILES_UPDATE2:
10123 ret = io_register_rsrc_update(ctx, arg, nr_args,
10126 case IORING_REGISTER_BUFFERS2:
10127 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10129 case IORING_REGISTER_BUFFERS_UPDATE:
10130 ret = io_register_rsrc_update(ctx, arg, nr_args,
10131 IORING_RSRC_BUFFER);
10138 if (io_register_op_must_quiesce(opcode)) {
10139 /* bring the ctx back to life */
10140 percpu_ref_reinit(&ctx->refs);
10141 reinit_completion(&ctx->ref_comp);
10146 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10147 void __user *, arg, unsigned int, nr_args)
10149 struct io_ring_ctx *ctx;
10158 if (f.file->f_op != &io_uring_fops)
10161 ctx = f.file->private_data;
10163 io_run_task_work();
10165 mutex_lock(&ctx->uring_lock);
10166 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10167 mutex_unlock(&ctx->uring_lock);
10168 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10169 ctx->cq_ev_fd != NULL, ret);
10175 static int __init io_uring_init(void)
10177 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10178 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10179 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10182 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10183 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10184 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10185 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10186 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10187 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10188 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10189 BUILD_BUG_SQE_ELEM(8, __u64, off);
10190 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10191 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10192 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10193 BUILD_BUG_SQE_ELEM(24, __u32, len);
10194 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10195 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10196 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10197 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10198 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10199 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10200 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10201 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10202 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10203 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10204 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10205 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10206 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10207 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10208 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10209 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10210 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10211 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10212 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10214 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10215 sizeof(struct io_uring_rsrc_update));
10216 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10217 sizeof(struct io_uring_rsrc_update2));
10218 /* should fit into one byte */
10219 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10221 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10222 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10223 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10227 __initcall(io_uring_init);