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>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head ltimeout_list;
379 struct list_head cq_overflow_list;
380 struct xarray io_buffers;
381 struct xarray personalities;
383 unsigned sq_thread_idle;
384 } ____cacheline_aligned_in_smp;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list;
388 unsigned int locked_free_nr;
390 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
396 unsigned long check_cq_overflow;
399 unsigned cached_cq_tail;
401 struct eventfd_ctx *cq_ev_fd;
402 struct wait_queue_head poll_wait;
403 struct wait_queue_head cq_wait;
405 atomic_t cq_timeouts;
406 struct fasync_struct *cq_fasync;
407 unsigned cq_last_tm_flush;
408 } ____cacheline_aligned_in_smp;
411 spinlock_t completion_lock;
413 spinlock_t timeout_lock;
416 * ->iopoll_list is protected by the ctx->uring_lock for
417 * io_uring instances that don't use IORING_SETUP_SQPOLL.
418 * For SQPOLL, only the single threaded io_sq_thread() will
419 * manipulate the list, hence no extra locking is needed there.
421 struct list_head iopoll_list;
422 struct hlist_head *cancel_hash;
423 unsigned cancel_hash_bits;
424 bool poll_multi_queue;
425 } ____cacheline_aligned_in_smp;
427 struct io_restriction restrictions;
429 /* slow path rsrc auxilary data, used by update/register */
431 struct io_rsrc_node *rsrc_backup_node;
432 struct io_mapped_ubuf *dummy_ubuf;
433 struct io_rsrc_data *file_data;
434 struct io_rsrc_data *buf_data;
436 struct delayed_work rsrc_put_work;
437 struct llist_head rsrc_put_llist;
438 struct list_head rsrc_ref_list;
439 spinlock_t rsrc_ref_lock;
442 /* Keep this last, we don't need it for the fast path */
444 #if defined(CONFIG_UNIX)
445 struct socket *ring_sock;
447 /* hashed buffered write serialization */
448 struct io_wq_hash *hash_map;
450 /* Only used for accounting purposes */
451 struct user_struct *user;
452 struct mm_struct *mm_account;
454 /* ctx exit and cancelation */
455 struct llist_head fallback_llist;
456 struct delayed_work fallback_work;
457 struct work_struct exit_work;
458 struct list_head tctx_list;
459 struct completion ref_comp;
463 struct io_uring_task {
464 /* submission side */
467 struct wait_queue_head wait;
468 const struct io_ring_ctx *last;
470 struct percpu_counter inflight;
471 atomic_t inflight_tracked;
474 spinlock_t task_lock;
475 struct io_wq_work_list task_list;
476 struct callback_head task_work;
481 * First field must be the file pointer in all the
482 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
484 struct io_poll_iocb {
486 struct wait_queue_head *head;
490 struct wait_queue_entry wait;
493 struct io_poll_update {
499 bool update_user_data;
507 struct io_timeout_data {
508 struct io_kiocb *req;
509 struct hrtimer timer;
510 struct timespec64 ts;
511 enum hrtimer_mode mode;
517 struct sockaddr __user *addr;
518 int __user *addr_len;
521 unsigned long nofile;
541 struct list_head list;
542 /* head of the link, used by linked timeouts only */
543 struct io_kiocb *head;
544 /* for linked completions */
545 struct io_kiocb *prev;
548 struct io_timeout_rem {
553 struct timespec64 ts;
559 /* NOTE: kiocb has the file as the first member, so don't do it here */
567 struct sockaddr __user *addr;
574 struct compat_msghdr __user *umsg_compat;
575 struct user_msghdr __user *umsg;
581 struct io_buffer *kbuf;
588 struct filename *filename;
590 unsigned long nofile;
593 struct io_rsrc_update {
619 struct epoll_event event;
623 struct file *file_out;
624 struct file *file_in;
631 struct io_provide_buf {
645 const char __user *filename;
646 struct statx __user *buffer;
658 struct filename *oldpath;
659 struct filename *newpath;
667 struct filename *filename;
674 struct filename *filename;
680 struct filename *oldpath;
681 struct filename *newpath;
688 struct filename *oldpath;
689 struct filename *newpath;
693 struct io_completion {
698 struct io_async_connect {
699 struct sockaddr_storage address;
702 struct io_async_msghdr {
703 struct iovec fast_iov[UIO_FASTIOV];
704 /* points to an allocated iov, if NULL we use fast_iov instead */
705 struct iovec *free_iov;
706 struct sockaddr __user *uaddr;
708 struct sockaddr_storage addr;
712 struct iovec fast_iov[UIO_FASTIOV];
713 const struct iovec *free_iovec;
714 struct iov_iter iter;
716 struct wait_page_queue wpq;
720 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
721 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
722 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
723 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
724 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
725 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
727 /* first byte is taken by user flags, shift it to not overlap */
732 REQ_F_LINK_TIMEOUT_BIT,
733 REQ_F_NEED_CLEANUP_BIT,
735 REQ_F_BUFFER_SELECTED_BIT,
736 REQ_F_COMPLETE_INLINE_BIT,
738 REQ_F_DONT_REISSUE_BIT,
741 REQ_F_ARM_LTIMEOUT_BIT,
742 /* keep async read/write and isreg together and in order */
743 REQ_F_NOWAIT_READ_BIT,
744 REQ_F_NOWAIT_WRITE_BIT,
747 /* not a real bit, just to check we're not overflowing the space */
753 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
754 /* drain existing IO first */
755 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
757 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
758 /* doesn't sever on completion < 0 */
759 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
761 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
762 /* IOSQE_BUFFER_SELECT */
763 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
765 /* fail rest of links */
766 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
767 /* on inflight list, should be cancelled and waited on exit reliably */
768 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
769 /* read/write uses file position */
770 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
771 /* must not punt to workers */
772 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
773 /* has or had linked timeout */
774 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
776 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
777 /* already went through poll handler */
778 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
779 /* buffer already selected */
780 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
781 /* completion is deferred through io_comp_state */
782 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
783 /* caller should reissue async */
784 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
785 /* don't attempt request reissue, see io_rw_reissue() */
786 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
787 /* supports async reads */
788 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
789 /* supports async writes */
790 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
792 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
793 /* has creds assigned */
794 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
795 /* skip refcounting if not set */
796 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
797 /* there is a linked timeout that has to be armed */
798 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
802 struct io_poll_iocb poll;
803 struct io_poll_iocb *double_poll;
806 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
808 struct io_task_work {
810 struct io_wq_work_node node;
811 struct llist_node fallback_node;
813 io_req_tw_func_t func;
817 IORING_RSRC_FILE = 0,
818 IORING_RSRC_BUFFER = 1,
822 * NOTE! Each of the iocb union members has the file pointer
823 * as the first entry in their struct definition. So you can
824 * access the file pointer through any of the sub-structs,
825 * or directly as just 'ki_filp' in this struct.
831 struct io_poll_iocb poll;
832 struct io_poll_update poll_update;
833 struct io_accept accept;
835 struct io_cancel cancel;
836 struct io_timeout timeout;
837 struct io_timeout_rem timeout_rem;
838 struct io_connect connect;
839 struct io_sr_msg sr_msg;
841 struct io_close close;
842 struct io_rsrc_update rsrc_update;
843 struct io_fadvise fadvise;
844 struct io_madvise madvise;
845 struct io_epoll epoll;
846 struct io_splice splice;
847 struct io_provide_buf pbuf;
848 struct io_statx statx;
849 struct io_shutdown shutdown;
850 struct io_rename rename;
851 struct io_unlink unlink;
852 struct io_mkdir mkdir;
853 struct io_symlink symlink;
854 struct io_hardlink hardlink;
855 /* use only after cleaning per-op data, see io_clean_op() */
856 struct io_completion compl;
859 /* opcode allocated if it needs to store data for async defer */
862 /* polled IO has completed */
868 struct io_ring_ctx *ctx;
871 struct task_struct *task;
874 struct io_kiocb *link;
875 struct percpu_ref *fixed_rsrc_refs;
877 /* used with ctx->iopoll_list with reads/writes */
878 struct list_head inflight_entry;
879 struct io_task_work io_task_work;
880 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
881 struct hlist_node hash_node;
882 struct async_poll *apoll;
883 struct io_wq_work work;
884 const struct cred *creds;
886 /* store used ubuf, so we can prevent reloading */
887 struct io_mapped_ubuf *imu;
890 struct io_tctx_node {
891 struct list_head ctx_node;
892 struct task_struct *task;
893 struct io_ring_ctx *ctx;
896 struct io_defer_entry {
897 struct list_head list;
898 struct io_kiocb *req;
903 /* needs req->file assigned */
904 unsigned needs_file : 1;
905 /* hash wq insertion if file is a regular file */
906 unsigned hash_reg_file : 1;
907 /* unbound wq insertion if file is a non-regular file */
908 unsigned unbound_nonreg_file : 1;
909 /* opcode is not supported by this kernel */
910 unsigned not_supported : 1;
911 /* set if opcode supports polled "wait" */
913 unsigned pollout : 1;
914 /* op supports buffer selection */
915 unsigned buffer_select : 1;
916 /* do prep async if is going to be punted */
917 unsigned needs_async_setup : 1;
918 /* should block plug */
920 /* size of async data needed, if any */
921 unsigned short async_size;
924 static const struct io_op_def io_op_defs[] = {
925 [IORING_OP_NOP] = {},
926 [IORING_OP_READV] = {
928 .unbound_nonreg_file = 1,
931 .needs_async_setup = 1,
933 .async_size = sizeof(struct io_async_rw),
935 [IORING_OP_WRITEV] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_setup = 1,
942 .async_size = sizeof(struct io_async_rw),
944 [IORING_OP_FSYNC] = {
947 [IORING_OP_READ_FIXED] = {
949 .unbound_nonreg_file = 1,
952 .async_size = sizeof(struct io_async_rw),
954 [IORING_OP_WRITE_FIXED] = {
957 .unbound_nonreg_file = 1,
960 .async_size = sizeof(struct io_async_rw),
962 [IORING_OP_POLL_ADD] = {
964 .unbound_nonreg_file = 1,
966 [IORING_OP_POLL_REMOVE] = {},
967 [IORING_OP_SYNC_FILE_RANGE] = {
970 [IORING_OP_SENDMSG] = {
972 .unbound_nonreg_file = 1,
974 .needs_async_setup = 1,
975 .async_size = sizeof(struct io_async_msghdr),
977 [IORING_OP_RECVMSG] = {
979 .unbound_nonreg_file = 1,
982 .needs_async_setup = 1,
983 .async_size = sizeof(struct io_async_msghdr),
985 [IORING_OP_TIMEOUT] = {
986 .async_size = sizeof(struct io_timeout_data),
988 [IORING_OP_TIMEOUT_REMOVE] = {
989 /* used by timeout updates' prep() */
991 [IORING_OP_ACCEPT] = {
993 .unbound_nonreg_file = 1,
996 [IORING_OP_ASYNC_CANCEL] = {},
997 [IORING_OP_LINK_TIMEOUT] = {
998 .async_size = sizeof(struct io_timeout_data),
1000 [IORING_OP_CONNECT] = {
1002 .unbound_nonreg_file = 1,
1004 .needs_async_setup = 1,
1005 .async_size = sizeof(struct io_async_connect),
1007 [IORING_OP_FALLOCATE] = {
1010 [IORING_OP_OPENAT] = {},
1011 [IORING_OP_CLOSE] = {},
1012 [IORING_OP_FILES_UPDATE] = {},
1013 [IORING_OP_STATX] = {},
1014 [IORING_OP_READ] = {
1016 .unbound_nonreg_file = 1,
1020 .async_size = sizeof(struct io_async_rw),
1022 [IORING_OP_WRITE] = {
1025 .unbound_nonreg_file = 1,
1028 .async_size = sizeof(struct io_async_rw),
1030 [IORING_OP_FADVISE] = {
1033 [IORING_OP_MADVISE] = {},
1034 [IORING_OP_SEND] = {
1036 .unbound_nonreg_file = 1,
1039 [IORING_OP_RECV] = {
1041 .unbound_nonreg_file = 1,
1045 [IORING_OP_OPENAT2] = {
1047 [IORING_OP_EPOLL_CTL] = {
1048 .unbound_nonreg_file = 1,
1050 [IORING_OP_SPLICE] = {
1053 .unbound_nonreg_file = 1,
1055 [IORING_OP_PROVIDE_BUFFERS] = {},
1056 [IORING_OP_REMOVE_BUFFERS] = {},
1060 .unbound_nonreg_file = 1,
1062 [IORING_OP_SHUTDOWN] = {
1065 [IORING_OP_RENAMEAT] = {},
1066 [IORING_OP_UNLINKAT] = {},
1067 [IORING_OP_MKDIRAT] = {},
1068 [IORING_OP_SYMLINKAT] = {},
1069 [IORING_OP_LINKAT] = {},
1072 /* requests with any of those set should undergo io_disarm_next() */
1073 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1075 static bool io_disarm_next(struct io_kiocb *req);
1076 static void io_uring_del_tctx_node(unsigned long index);
1077 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1078 struct task_struct *task,
1080 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1082 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1083 long res, unsigned int cflags);
1084 static void io_put_req(struct io_kiocb *req);
1085 static void io_put_req_deferred(struct io_kiocb *req);
1086 static void io_dismantle_req(struct io_kiocb *req);
1087 static void io_queue_linked_timeout(struct io_kiocb *req);
1088 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1089 struct io_uring_rsrc_update2 *up,
1091 static void io_clean_op(struct io_kiocb *req);
1092 static struct file *io_file_get(struct io_ring_ctx *ctx,
1093 struct io_kiocb *req, int fd, bool fixed);
1094 static void __io_queue_sqe(struct io_kiocb *req);
1095 static void io_rsrc_put_work(struct work_struct *work);
1097 static void io_req_task_queue(struct io_kiocb *req);
1098 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1099 static int io_req_prep_async(struct io_kiocb *req);
1101 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1102 unsigned int issue_flags, u32 slot_index);
1103 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1105 static struct kmem_cache *req_cachep;
1107 static const struct file_operations io_uring_fops;
1109 struct sock *io_uring_get_socket(struct file *file)
1111 #if defined(CONFIG_UNIX)
1112 if (file->f_op == &io_uring_fops) {
1113 struct io_ring_ctx *ctx = file->private_data;
1115 return ctx->ring_sock->sk;
1120 EXPORT_SYMBOL(io_uring_get_socket);
1122 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1125 mutex_lock(&ctx->uring_lock);
1130 #define io_for_each_link(pos, head) \
1131 for (pos = (head); pos; pos = pos->link)
1134 * Shamelessly stolen from the mm implementation of page reference checking,
1135 * see commit f958d7b528b1 for details.
1137 #define req_ref_zero_or_close_to_overflow(req) \
1138 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1140 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1142 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1143 return atomic_inc_not_zero(&req->refs);
1146 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1148 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1151 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1152 return atomic_dec_and_test(&req->refs);
1155 static inline void req_ref_put(struct io_kiocb *req)
1157 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1158 WARN_ON_ONCE(req_ref_put_and_test(req));
1161 static inline void req_ref_get(struct io_kiocb *req)
1163 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1164 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1165 atomic_inc(&req->refs);
1168 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1170 if (!(req->flags & REQ_F_REFCOUNT)) {
1171 req->flags |= REQ_F_REFCOUNT;
1172 atomic_set(&req->refs, nr);
1176 static inline void io_req_set_refcount(struct io_kiocb *req)
1178 __io_req_set_refcount(req, 1);
1181 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1183 struct io_ring_ctx *ctx = req->ctx;
1185 if (!req->fixed_rsrc_refs) {
1186 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1187 percpu_ref_get(req->fixed_rsrc_refs);
1191 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1193 bool got = percpu_ref_tryget(ref);
1195 /* already at zero, wait for ->release() */
1197 wait_for_completion(compl);
1198 percpu_ref_resurrect(ref);
1200 percpu_ref_put(ref);
1203 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1206 struct io_kiocb *req;
1208 if (task && head->task != task)
1213 io_for_each_link(req, head) {
1214 if (req->flags & REQ_F_INFLIGHT)
1220 static inline void req_set_fail(struct io_kiocb *req)
1222 req->flags |= REQ_F_FAIL;
1225 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1231 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1233 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1235 complete(&ctx->ref_comp);
1238 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1240 return !req->timeout.off;
1243 static void io_fallback_req_func(struct work_struct *work)
1245 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1246 fallback_work.work);
1247 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1248 struct io_kiocb *req, *tmp;
1249 bool locked = false;
1251 percpu_ref_get(&ctx->refs);
1252 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1253 req->io_task_work.func(req, &locked);
1256 if (ctx->submit_state.compl_nr)
1257 io_submit_flush_completions(ctx);
1258 mutex_unlock(&ctx->uring_lock);
1260 percpu_ref_put(&ctx->refs);
1264 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1266 struct io_ring_ctx *ctx;
1269 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1274 * Use 5 bits less than the max cq entries, that should give us around
1275 * 32 entries per hash list if totally full and uniformly spread.
1277 hash_bits = ilog2(p->cq_entries);
1281 ctx->cancel_hash_bits = hash_bits;
1282 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1284 if (!ctx->cancel_hash)
1286 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1288 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1289 if (!ctx->dummy_ubuf)
1291 /* set invalid range, so io_import_fixed() fails meeting it */
1292 ctx->dummy_ubuf->ubuf = -1UL;
1294 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1295 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1298 ctx->flags = p->flags;
1299 init_waitqueue_head(&ctx->sqo_sq_wait);
1300 INIT_LIST_HEAD(&ctx->sqd_list);
1301 init_waitqueue_head(&ctx->poll_wait);
1302 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1303 init_completion(&ctx->ref_comp);
1304 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1305 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1306 mutex_init(&ctx->uring_lock);
1307 init_waitqueue_head(&ctx->cq_wait);
1308 spin_lock_init(&ctx->completion_lock);
1309 spin_lock_init(&ctx->timeout_lock);
1310 INIT_LIST_HEAD(&ctx->iopoll_list);
1311 INIT_LIST_HEAD(&ctx->defer_list);
1312 INIT_LIST_HEAD(&ctx->timeout_list);
1313 INIT_LIST_HEAD(&ctx->ltimeout_list);
1314 spin_lock_init(&ctx->rsrc_ref_lock);
1315 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1316 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1317 init_llist_head(&ctx->rsrc_put_llist);
1318 INIT_LIST_HEAD(&ctx->tctx_list);
1319 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1320 INIT_LIST_HEAD(&ctx->locked_free_list);
1321 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1324 kfree(ctx->dummy_ubuf);
1325 kfree(ctx->cancel_hash);
1330 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1332 struct io_rings *r = ctx->rings;
1334 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1338 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1340 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1341 struct io_ring_ctx *ctx = req->ctx;
1343 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1349 #define FFS_ASYNC_READ 0x1UL
1350 #define FFS_ASYNC_WRITE 0x2UL
1352 #define FFS_ISREG 0x4UL
1354 #define FFS_ISREG 0x0UL
1356 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1358 static inline bool io_req_ffs_set(struct io_kiocb *req)
1360 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1363 static void io_req_track_inflight(struct io_kiocb *req)
1365 if (!(req->flags & REQ_F_INFLIGHT)) {
1366 req->flags |= REQ_F_INFLIGHT;
1367 atomic_inc(¤t->io_uring->inflight_tracked);
1371 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1373 req->flags &= ~REQ_F_LINK_TIMEOUT;
1376 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1378 if (WARN_ON_ONCE(!req->link))
1381 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1382 req->flags |= REQ_F_LINK_TIMEOUT;
1384 /* linked timeouts should have two refs once prep'ed */
1385 io_req_set_refcount(req);
1386 __io_req_set_refcount(req->link, 2);
1390 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1392 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1394 return __io_prep_linked_timeout(req);
1397 static void io_prep_async_work(struct io_kiocb *req)
1399 const struct io_op_def *def = &io_op_defs[req->opcode];
1400 struct io_ring_ctx *ctx = req->ctx;
1402 if (!(req->flags & REQ_F_CREDS)) {
1403 req->flags |= REQ_F_CREDS;
1404 req->creds = get_current_cred();
1407 req->work.list.next = NULL;
1408 req->work.flags = 0;
1409 if (req->flags & REQ_F_FORCE_ASYNC)
1410 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1412 if (req->flags & REQ_F_ISREG) {
1413 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1414 io_wq_hash_work(&req->work, file_inode(req->file));
1415 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1416 if (def->unbound_nonreg_file)
1417 req->work.flags |= IO_WQ_WORK_UNBOUND;
1420 switch (req->opcode) {
1421 case IORING_OP_SPLICE:
1423 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1424 req->work.flags |= IO_WQ_WORK_UNBOUND;
1429 static void io_prep_async_link(struct io_kiocb *req)
1431 struct io_kiocb *cur;
1433 if (req->flags & REQ_F_LINK_TIMEOUT) {
1434 struct io_ring_ctx *ctx = req->ctx;
1436 spin_lock(&ctx->completion_lock);
1437 io_for_each_link(cur, req)
1438 io_prep_async_work(cur);
1439 spin_unlock(&ctx->completion_lock);
1441 io_for_each_link(cur, req)
1442 io_prep_async_work(cur);
1446 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1448 struct io_ring_ctx *ctx = req->ctx;
1449 struct io_kiocb *link = io_prep_linked_timeout(req);
1450 struct io_uring_task *tctx = req->task->io_uring;
1452 /* must not take the lock, NULL it as a precaution */
1456 BUG_ON(!tctx->io_wq);
1458 /* init ->work of the whole link before punting */
1459 io_prep_async_link(req);
1462 * Not expected to happen, but if we do have a bug where this _can_
1463 * happen, catch it here and ensure the request is marked as
1464 * canceled. That will make io-wq go through the usual work cancel
1465 * procedure rather than attempt to run this request (or create a new
1468 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1469 req->work.flags |= IO_WQ_WORK_CANCEL;
1471 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1472 &req->work, req->flags);
1473 io_wq_enqueue(tctx->io_wq, &req->work);
1475 io_queue_linked_timeout(link);
1478 static void io_kill_timeout(struct io_kiocb *req, int status)
1479 __must_hold(&req->ctx->completion_lock)
1480 __must_hold(&req->ctx->timeout_lock)
1482 struct io_timeout_data *io = req->async_data;
1484 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1485 atomic_set(&req->ctx->cq_timeouts,
1486 atomic_read(&req->ctx->cq_timeouts) + 1);
1487 list_del_init(&req->timeout.list);
1488 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1489 io_put_req_deferred(req);
1493 static void io_queue_deferred(struct io_ring_ctx *ctx)
1495 while (!list_empty(&ctx->defer_list)) {
1496 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1497 struct io_defer_entry, list);
1499 if (req_need_defer(de->req, de->seq))
1501 list_del_init(&de->list);
1502 io_req_task_queue(de->req);
1507 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1508 __must_hold(&ctx->completion_lock)
1510 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1512 spin_lock_irq(&ctx->timeout_lock);
1513 while (!list_empty(&ctx->timeout_list)) {
1514 u32 events_needed, events_got;
1515 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1516 struct io_kiocb, timeout.list);
1518 if (io_is_timeout_noseq(req))
1522 * Since seq can easily wrap around over time, subtract
1523 * the last seq at which timeouts were flushed before comparing.
1524 * Assuming not more than 2^31-1 events have happened since,
1525 * these subtractions won't have wrapped, so we can check if
1526 * target is in [last_seq, current_seq] by comparing the two.
1528 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1529 events_got = seq - ctx->cq_last_tm_flush;
1530 if (events_got < events_needed)
1533 list_del_init(&req->timeout.list);
1534 io_kill_timeout(req, 0);
1536 ctx->cq_last_tm_flush = seq;
1537 spin_unlock_irq(&ctx->timeout_lock);
1540 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1542 if (ctx->off_timeout_used)
1543 io_flush_timeouts(ctx);
1544 if (ctx->drain_active)
1545 io_queue_deferred(ctx);
1548 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1550 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1551 __io_commit_cqring_flush(ctx);
1552 /* order cqe stores with ring update */
1553 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1556 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1558 struct io_rings *r = ctx->rings;
1560 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1563 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1565 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1568 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1570 struct io_rings *rings = ctx->rings;
1571 unsigned tail, mask = ctx->cq_entries - 1;
1574 * writes to the cq entry need to come after reading head; the
1575 * control dependency is enough as we're using WRITE_ONCE to
1578 if (__io_cqring_events(ctx) == ctx->cq_entries)
1581 tail = ctx->cached_cq_tail++;
1582 return &rings->cqes[tail & mask];
1585 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1587 if (likely(!ctx->cq_ev_fd))
1589 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1591 return !ctx->eventfd_async || io_wq_current_is_worker();
1595 * This should only get called when at least one event has been posted.
1596 * Some applications rely on the eventfd notification count only changing
1597 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1598 * 1:1 relationship between how many times this function is called (and
1599 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1601 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1604 * wake_up_all() may seem excessive, but io_wake_function() and
1605 * io_should_wake() handle the termination of the loop and only
1606 * wake as many waiters as we need to.
1608 if (wq_has_sleeper(&ctx->cq_wait))
1609 wake_up_all(&ctx->cq_wait);
1610 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1611 wake_up(&ctx->sq_data->wait);
1612 if (io_should_trigger_evfd(ctx))
1613 eventfd_signal(ctx->cq_ev_fd, 1);
1614 if (waitqueue_active(&ctx->poll_wait)) {
1615 wake_up_interruptible(&ctx->poll_wait);
1616 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1620 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1622 if (ctx->flags & IORING_SETUP_SQPOLL) {
1623 if (wq_has_sleeper(&ctx->cq_wait))
1624 wake_up_all(&ctx->cq_wait);
1626 if (io_should_trigger_evfd(ctx))
1627 eventfd_signal(ctx->cq_ev_fd, 1);
1628 if (waitqueue_active(&ctx->poll_wait)) {
1629 wake_up_interruptible(&ctx->poll_wait);
1630 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1634 /* Returns true if there are no backlogged entries after the flush */
1635 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1637 bool all_flushed, posted;
1639 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1643 spin_lock(&ctx->completion_lock);
1644 while (!list_empty(&ctx->cq_overflow_list)) {
1645 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1646 struct io_overflow_cqe *ocqe;
1650 ocqe = list_first_entry(&ctx->cq_overflow_list,
1651 struct io_overflow_cqe, list);
1653 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1655 io_account_cq_overflow(ctx);
1658 list_del(&ocqe->list);
1662 all_flushed = list_empty(&ctx->cq_overflow_list);
1664 clear_bit(0, &ctx->check_cq_overflow);
1665 WRITE_ONCE(ctx->rings->sq_flags,
1666 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1670 io_commit_cqring(ctx);
1671 spin_unlock(&ctx->completion_lock);
1673 io_cqring_ev_posted(ctx);
1677 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1681 if (test_bit(0, &ctx->check_cq_overflow)) {
1682 /* iopoll syncs against uring_lock, not completion_lock */
1683 if (ctx->flags & IORING_SETUP_IOPOLL)
1684 mutex_lock(&ctx->uring_lock);
1685 ret = __io_cqring_overflow_flush(ctx, false);
1686 if (ctx->flags & IORING_SETUP_IOPOLL)
1687 mutex_unlock(&ctx->uring_lock);
1693 /* must to be called somewhat shortly after putting a request */
1694 static inline void io_put_task(struct task_struct *task, int nr)
1696 struct io_uring_task *tctx = task->io_uring;
1698 if (likely(task == current)) {
1699 tctx->cached_refs += nr;
1701 percpu_counter_sub(&tctx->inflight, nr);
1702 if (unlikely(atomic_read(&tctx->in_idle)))
1703 wake_up(&tctx->wait);
1704 put_task_struct_many(task, nr);
1708 static void io_task_refs_refill(struct io_uring_task *tctx)
1710 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1712 percpu_counter_add(&tctx->inflight, refill);
1713 refcount_add(refill, ¤t->usage);
1714 tctx->cached_refs += refill;
1717 static inline void io_get_task_refs(int nr)
1719 struct io_uring_task *tctx = current->io_uring;
1721 tctx->cached_refs -= nr;
1722 if (unlikely(tctx->cached_refs < 0))
1723 io_task_refs_refill(tctx);
1726 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1727 long res, unsigned int cflags)
1729 struct io_overflow_cqe *ocqe;
1731 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1734 * If we're in ring overflow flush mode, or in task cancel mode,
1735 * or cannot allocate an overflow entry, then we need to drop it
1738 io_account_cq_overflow(ctx);
1741 if (list_empty(&ctx->cq_overflow_list)) {
1742 set_bit(0, &ctx->check_cq_overflow);
1743 WRITE_ONCE(ctx->rings->sq_flags,
1744 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1747 ocqe->cqe.user_data = user_data;
1748 ocqe->cqe.res = res;
1749 ocqe->cqe.flags = cflags;
1750 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1754 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1755 long res, unsigned int cflags)
1757 struct io_uring_cqe *cqe;
1759 trace_io_uring_complete(ctx, user_data, res, cflags);
1762 * If we can't get a cq entry, userspace overflowed the
1763 * submission (by quite a lot). Increment the overflow count in
1766 cqe = io_get_cqe(ctx);
1768 WRITE_ONCE(cqe->user_data, user_data);
1769 WRITE_ONCE(cqe->res, res);
1770 WRITE_ONCE(cqe->flags, cflags);
1773 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1776 /* not as hot to bloat with inlining */
1777 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1778 long res, unsigned int cflags)
1780 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1783 static void io_req_complete_post(struct io_kiocb *req, long res,
1784 unsigned int cflags)
1786 struct io_ring_ctx *ctx = req->ctx;
1788 spin_lock(&ctx->completion_lock);
1789 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1791 * If we're the last reference to this request, add to our locked
1794 if (req_ref_put_and_test(req)) {
1795 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1796 if (req->flags & IO_DISARM_MASK)
1797 io_disarm_next(req);
1799 io_req_task_queue(req->link);
1803 io_dismantle_req(req);
1804 io_put_task(req->task, 1);
1805 list_add(&req->inflight_entry, &ctx->locked_free_list);
1806 ctx->locked_free_nr++;
1808 if (!percpu_ref_tryget(&ctx->refs))
1811 io_commit_cqring(ctx);
1812 spin_unlock(&ctx->completion_lock);
1815 io_cqring_ev_posted(ctx);
1816 percpu_ref_put(&ctx->refs);
1820 static inline bool io_req_needs_clean(struct io_kiocb *req)
1822 return req->flags & IO_REQ_CLEAN_FLAGS;
1825 static void io_req_complete_state(struct io_kiocb *req, long res,
1826 unsigned int cflags)
1828 if (io_req_needs_clean(req))
1831 req->compl.cflags = cflags;
1832 req->flags |= REQ_F_COMPLETE_INLINE;
1835 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1836 long res, unsigned cflags)
1838 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1839 io_req_complete_state(req, res, cflags);
1841 io_req_complete_post(req, res, cflags);
1844 static inline void io_req_complete(struct io_kiocb *req, long res)
1846 __io_req_complete(req, 0, res, 0);
1849 static void io_req_complete_failed(struct io_kiocb *req, long res)
1852 io_req_complete_post(req, res, 0);
1855 static void io_req_complete_fail_submit(struct io_kiocb *req)
1858 * We don't submit, fail them all, for that replace hardlinks with
1859 * normal links. Extra REQ_F_LINK is tolerated.
1861 req->flags &= ~REQ_F_HARDLINK;
1862 req->flags |= REQ_F_LINK;
1863 io_req_complete_failed(req, req->result);
1867 * Don't initialise the fields below on every allocation, but do that in
1868 * advance and keep them valid across allocations.
1870 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1874 req->async_data = NULL;
1875 /* not necessary, but safer to zero */
1879 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1880 struct io_submit_state *state)
1882 spin_lock(&ctx->completion_lock);
1883 list_splice_init(&ctx->locked_free_list, &state->free_list);
1884 ctx->locked_free_nr = 0;
1885 spin_unlock(&ctx->completion_lock);
1888 /* Returns true IFF there are requests in the cache */
1889 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1891 struct io_submit_state *state = &ctx->submit_state;
1895 * If we have more than a batch's worth of requests in our IRQ side
1896 * locked cache, grab the lock and move them over to our submission
1899 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1900 io_flush_cached_locked_reqs(ctx, state);
1902 nr = state->free_reqs;
1903 while (!list_empty(&state->free_list)) {
1904 struct io_kiocb *req = list_first_entry(&state->free_list,
1905 struct io_kiocb, inflight_entry);
1907 list_del(&req->inflight_entry);
1908 state->reqs[nr++] = req;
1909 if (nr == ARRAY_SIZE(state->reqs))
1913 state->free_reqs = nr;
1918 * A request might get retired back into the request caches even before opcode
1919 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1920 * Because of that, io_alloc_req() should be called only under ->uring_lock
1921 * and with extra caution to not get a request that is still worked on.
1923 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1924 __must_hold(&ctx->uring_lock)
1926 struct io_submit_state *state = &ctx->submit_state;
1927 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1930 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1932 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1935 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1939 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1940 * retry single alloc to be on the safe side.
1942 if (unlikely(ret <= 0)) {
1943 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1944 if (!state->reqs[0])
1949 for (i = 0; i < ret; i++)
1950 io_preinit_req(state->reqs[i], ctx);
1951 state->free_reqs = ret;
1954 return state->reqs[state->free_reqs];
1957 static inline void io_put_file(struct file *file)
1963 static void io_dismantle_req(struct io_kiocb *req)
1965 unsigned int flags = req->flags;
1967 if (io_req_needs_clean(req))
1969 if (!(flags & REQ_F_FIXED_FILE))
1970 io_put_file(req->file);
1971 if (req->fixed_rsrc_refs)
1972 percpu_ref_put(req->fixed_rsrc_refs);
1973 if (req->async_data) {
1974 kfree(req->async_data);
1975 req->async_data = NULL;
1979 static void __io_free_req(struct io_kiocb *req)
1981 struct io_ring_ctx *ctx = req->ctx;
1983 io_dismantle_req(req);
1984 io_put_task(req->task, 1);
1986 spin_lock(&ctx->completion_lock);
1987 list_add(&req->inflight_entry, &ctx->locked_free_list);
1988 ctx->locked_free_nr++;
1989 spin_unlock(&ctx->completion_lock);
1991 percpu_ref_put(&ctx->refs);
1994 static inline void io_remove_next_linked(struct io_kiocb *req)
1996 struct io_kiocb *nxt = req->link;
1998 req->link = nxt->link;
2002 static bool io_kill_linked_timeout(struct io_kiocb *req)
2003 __must_hold(&req->ctx->completion_lock)
2004 __must_hold(&req->ctx->timeout_lock)
2006 struct io_kiocb *link = req->link;
2008 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2009 struct io_timeout_data *io = link->async_data;
2011 io_remove_next_linked(req);
2012 link->timeout.head = NULL;
2013 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2014 list_del(&link->timeout.list);
2015 io_cqring_fill_event(link->ctx, link->user_data,
2017 io_put_req_deferred(link);
2024 static void io_fail_links(struct io_kiocb *req)
2025 __must_hold(&req->ctx->completion_lock)
2027 struct io_kiocb *nxt, *link = req->link;
2031 long res = -ECANCELED;
2033 if (link->flags & REQ_F_FAIL)
2039 trace_io_uring_fail_link(req, link);
2040 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2041 io_put_req_deferred(link);
2046 static bool io_disarm_next(struct io_kiocb *req)
2047 __must_hold(&req->ctx->completion_lock)
2049 bool posted = false;
2051 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2052 struct io_kiocb *link = req->link;
2054 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2055 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2056 io_remove_next_linked(req);
2057 io_cqring_fill_event(link->ctx, link->user_data,
2059 io_put_req_deferred(link);
2062 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2063 struct io_ring_ctx *ctx = req->ctx;
2065 spin_lock_irq(&ctx->timeout_lock);
2066 posted = io_kill_linked_timeout(req);
2067 spin_unlock_irq(&ctx->timeout_lock);
2069 if (unlikely((req->flags & REQ_F_FAIL) &&
2070 !(req->flags & REQ_F_HARDLINK))) {
2071 posted |= (req->link != NULL);
2077 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2079 struct io_kiocb *nxt;
2082 * If LINK is set, we have dependent requests in this chain. If we
2083 * didn't fail this request, queue the first one up, moving any other
2084 * dependencies to the next request. In case of failure, fail the rest
2087 if (req->flags & IO_DISARM_MASK) {
2088 struct io_ring_ctx *ctx = req->ctx;
2091 spin_lock(&ctx->completion_lock);
2092 posted = io_disarm_next(req);
2094 io_commit_cqring(req->ctx);
2095 spin_unlock(&ctx->completion_lock);
2097 io_cqring_ev_posted(ctx);
2104 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2106 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2108 return __io_req_find_next(req);
2111 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2116 if (ctx->submit_state.compl_nr)
2117 io_submit_flush_completions(ctx);
2118 mutex_unlock(&ctx->uring_lock);
2121 percpu_ref_put(&ctx->refs);
2124 static void tctx_task_work(struct callback_head *cb)
2126 bool locked = false;
2127 struct io_ring_ctx *ctx = NULL;
2128 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2132 struct io_wq_work_node *node;
2134 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2135 io_submit_flush_completions(ctx);
2137 spin_lock_irq(&tctx->task_lock);
2138 node = tctx->task_list.first;
2139 INIT_WQ_LIST(&tctx->task_list);
2141 tctx->task_running = false;
2142 spin_unlock_irq(&tctx->task_lock);
2147 struct io_wq_work_node *next = node->next;
2148 struct io_kiocb *req = container_of(node, struct io_kiocb,
2151 if (req->ctx != ctx) {
2152 ctx_flush_and_put(ctx, &locked);
2154 /* if not contended, grab and improve batching */
2155 locked = mutex_trylock(&ctx->uring_lock);
2156 percpu_ref_get(&ctx->refs);
2158 req->io_task_work.func(req, &locked);
2165 ctx_flush_and_put(ctx, &locked);
2168 static void io_req_task_work_add(struct io_kiocb *req)
2170 struct task_struct *tsk = req->task;
2171 struct io_uring_task *tctx = tsk->io_uring;
2172 enum task_work_notify_mode notify;
2173 struct io_wq_work_node *node;
2174 unsigned long flags;
2177 WARN_ON_ONCE(!tctx);
2179 spin_lock_irqsave(&tctx->task_lock, flags);
2180 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2181 running = tctx->task_running;
2183 tctx->task_running = true;
2184 spin_unlock_irqrestore(&tctx->task_lock, flags);
2186 /* task_work already pending, we're done */
2191 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2192 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2193 * processing task_work. There's no reliable way to tell if TWA_RESUME
2196 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2197 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2198 wake_up_process(tsk);
2202 spin_lock_irqsave(&tctx->task_lock, flags);
2203 tctx->task_running = false;
2204 node = tctx->task_list.first;
2205 INIT_WQ_LIST(&tctx->task_list);
2206 spin_unlock_irqrestore(&tctx->task_lock, flags);
2209 req = container_of(node, struct io_kiocb, io_task_work.node);
2211 if (llist_add(&req->io_task_work.fallback_node,
2212 &req->ctx->fallback_llist))
2213 schedule_delayed_work(&req->ctx->fallback_work, 1);
2217 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2219 struct io_ring_ctx *ctx = req->ctx;
2221 /* not needed for normal modes, but SQPOLL depends on it */
2222 io_tw_lock(ctx, locked);
2223 io_req_complete_failed(req, req->result);
2226 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2228 struct io_ring_ctx *ctx = req->ctx;
2230 io_tw_lock(ctx, locked);
2231 /* req->task == current here, checking PF_EXITING is safe */
2232 if (likely(!(req->task->flags & PF_EXITING)))
2233 __io_queue_sqe(req);
2235 io_req_complete_failed(req, -EFAULT);
2238 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2241 req->io_task_work.func = io_req_task_cancel;
2242 io_req_task_work_add(req);
2245 static void io_req_task_queue(struct io_kiocb *req)
2247 req->io_task_work.func = io_req_task_submit;
2248 io_req_task_work_add(req);
2251 static void io_req_task_queue_reissue(struct io_kiocb *req)
2253 req->io_task_work.func = io_queue_async_work;
2254 io_req_task_work_add(req);
2257 static inline void io_queue_next(struct io_kiocb *req)
2259 struct io_kiocb *nxt = io_req_find_next(req);
2262 io_req_task_queue(nxt);
2265 static void io_free_req(struct io_kiocb *req)
2271 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2277 struct task_struct *task;
2282 static inline void io_init_req_batch(struct req_batch *rb)
2289 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2290 struct req_batch *rb)
2293 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2295 io_put_task(rb->task, rb->task_refs);
2298 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2299 struct io_submit_state *state)
2302 io_dismantle_req(req);
2304 if (req->task != rb->task) {
2306 io_put_task(rb->task, rb->task_refs);
2307 rb->task = req->task;
2313 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2314 state->reqs[state->free_reqs++] = req;
2316 list_add(&req->inflight_entry, &state->free_list);
2319 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2320 __must_hold(&ctx->uring_lock)
2322 struct io_submit_state *state = &ctx->submit_state;
2323 int i, nr = state->compl_nr;
2324 struct req_batch rb;
2326 spin_lock(&ctx->completion_lock);
2327 for (i = 0; i < nr; i++) {
2328 struct io_kiocb *req = state->compl_reqs[i];
2330 __io_cqring_fill_event(ctx, req->user_data, req->result,
2333 io_commit_cqring(ctx);
2334 spin_unlock(&ctx->completion_lock);
2335 io_cqring_ev_posted(ctx);
2337 io_init_req_batch(&rb);
2338 for (i = 0; i < nr; i++) {
2339 struct io_kiocb *req = state->compl_reqs[i];
2341 if (req_ref_put_and_test(req))
2342 io_req_free_batch(&rb, req, &ctx->submit_state);
2345 io_req_free_batch_finish(ctx, &rb);
2346 state->compl_nr = 0;
2350 * Drop reference to request, return next in chain (if there is one) if this
2351 * was the last reference to this request.
2353 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2355 struct io_kiocb *nxt = NULL;
2357 if (req_ref_put_and_test(req)) {
2358 nxt = io_req_find_next(req);
2364 static inline void io_put_req(struct io_kiocb *req)
2366 if (req_ref_put_and_test(req))
2370 static inline void io_put_req_deferred(struct io_kiocb *req)
2372 if (req_ref_put_and_test(req)) {
2373 req->io_task_work.func = io_free_req_work;
2374 io_req_task_work_add(req);
2378 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2380 /* See comment at the top of this file */
2382 return __io_cqring_events(ctx);
2385 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2387 struct io_rings *rings = ctx->rings;
2389 /* make sure SQ entry isn't read before tail */
2390 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2393 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2395 unsigned int cflags;
2397 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2398 cflags |= IORING_CQE_F_BUFFER;
2399 req->flags &= ~REQ_F_BUFFER_SELECTED;
2404 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2406 struct io_buffer *kbuf;
2408 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2410 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2411 return io_put_kbuf(req, kbuf);
2414 static inline bool io_run_task_work(void)
2416 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2417 __set_current_state(TASK_RUNNING);
2418 tracehook_notify_signal();
2426 * Find and free completed poll iocbs
2428 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2429 struct list_head *done)
2431 struct req_batch rb;
2432 struct io_kiocb *req;
2434 /* order with ->result store in io_complete_rw_iopoll() */
2437 io_init_req_batch(&rb);
2438 while (!list_empty(done)) {
2439 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2440 list_del(&req->inflight_entry);
2442 if (READ_ONCE(req->result) == -EAGAIN &&
2443 !(req->flags & REQ_F_DONT_REISSUE)) {
2444 req->iopoll_completed = 0;
2445 io_req_task_queue_reissue(req);
2449 __io_cqring_fill_event(ctx, req->user_data, req->result,
2450 io_put_rw_kbuf(req));
2453 if (req_ref_put_and_test(req))
2454 io_req_free_batch(&rb, req, &ctx->submit_state);
2457 io_commit_cqring(ctx);
2458 io_cqring_ev_posted_iopoll(ctx);
2459 io_req_free_batch_finish(ctx, &rb);
2462 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2465 struct io_kiocb *req, *tmp;
2470 * Only spin for completions if we don't have multiple devices hanging
2471 * off our complete list, and we're under the requested amount.
2473 spin = !ctx->poll_multi_queue && *nr_events < min;
2475 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2476 struct kiocb *kiocb = &req->rw.kiocb;
2480 * Move completed and retryable entries to our local lists.
2481 * If we find a request that requires polling, break out
2482 * and complete those lists first, if we have entries there.
2484 if (READ_ONCE(req->iopoll_completed)) {
2485 list_move_tail(&req->inflight_entry, &done);
2488 if (!list_empty(&done))
2491 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2492 if (unlikely(ret < 0))
2497 /* iopoll may have completed current req */
2498 if (READ_ONCE(req->iopoll_completed))
2499 list_move_tail(&req->inflight_entry, &done);
2502 if (!list_empty(&done))
2503 io_iopoll_complete(ctx, nr_events, &done);
2509 * We can't just wait for polled events to come to us, we have to actively
2510 * find and complete them.
2512 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2514 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2517 mutex_lock(&ctx->uring_lock);
2518 while (!list_empty(&ctx->iopoll_list)) {
2519 unsigned int nr_events = 0;
2521 io_do_iopoll(ctx, &nr_events, 0);
2523 /* let it sleep and repeat later if can't complete a request */
2527 * Ensure we allow local-to-the-cpu processing to take place,
2528 * in this case we need to ensure that we reap all events.
2529 * Also let task_work, etc. to progress by releasing the mutex
2531 if (need_resched()) {
2532 mutex_unlock(&ctx->uring_lock);
2534 mutex_lock(&ctx->uring_lock);
2537 mutex_unlock(&ctx->uring_lock);
2540 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2542 unsigned int nr_events = 0;
2546 * We disallow the app entering submit/complete with polling, but we
2547 * still need to lock the ring to prevent racing with polled issue
2548 * that got punted to a workqueue.
2550 mutex_lock(&ctx->uring_lock);
2552 * Don't enter poll loop if we already have events pending.
2553 * If we do, we can potentially be spinning for commands that
2554 * already triggered a CQE (eg in error).
2556 if (test_bit(0, &ctx->check_cq_overflow))
2557 __io_cqring_overflow_flush(ctx, false);
2558 if (io_cqring_events(ctx))
2562 * If a submit got punted to a workqueue, we can have the
2563 * application entering polling for a command before it gets
2564 * issued. That app will hold the uring_lock for the duration
2565 * of the poll right here, so we need to take a breather every
2566 * now and then to ensure that the issue has a chance to add
2567 * the poll to the issued list. Otherwise we can spin here
2568 * forever, while the workqueue is stuck trying to acquire the
2571 if (list_empty(&ctx->iopoll_list)) {
2572 u32 tail = ctx->cached_cq_tail;
2574 mutex_unlock(&ctx->uring_lock);
2576 mutex_lock(&ctx->uring_lock);
2578 /* some requests don't go through iopoll_list */
2579 if (tail != ctx->cached_cq_tail ||
2580 list_empty(&ctx->iopoll_list))
2583 ret = io_do_iopoll(ctx, &nr_events, min);
2584 } while (!ret && nr_events < min && !need_resched());
2586 mutex_unlock(&ctx->uring_lock);
2590 static void kiocb_end_write(struct io_kiocb *req)
2593 * Tell lockdep we inherited freeze protection from submission
2596 if (req->flags & REQ_F_ISREG) {
2597 struct super_block *sb = file_inode(req->file)->i_sb;
2599 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2605 static bool io_resubmit_prep(struct io_kiocb *req)
2607 struct io_async_rw *rw = req->async_data;
2610 return !io_req_prep_async(req);
2611 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2612 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2616 static bool io_rw_should_reissue(struct io_kiocb *req)
2618 umode_t mode = file_inode(req->file)->i_mode;
2619 struct io_ring_ctx *ctx = req->ctx;
2621 if (!S_ISBLK(mode) && !S_ISREG(mode))
2623 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2624 !(ctx->flags & IORING_SETUP_IOPOLL)))
2627 * If ref is dying, we might be running poll reap from the exit work.
2628 * Don't attempt to reissue from that path, just let it fail with
2631 if (percpu_ref_is_dying(&ctx->refs))
2634 * Play it safe and assume not safe to re-import and reissue if we're
2635 * not in the original thread group (or in task context).
2637 if (!same_thread_group(req->task, current) || !in_task())
2642 static bool io_resubmit_prep(struct io_kiocb *req)
2646 static bool io_rw_should_reissue(struct io_kiocb *req)
2652 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2654 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2655 kiocb_end_write(req);
2656 if (res != req->result) {
2657 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2658 io_rw_should_reissue(req)) {
2659 req->flags |= REQ_F_REISSUE;
2668 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2670 unsigned int cflags = io_put_rw_kbuf(req);
2671 long res = req->result;
2674 struct io_ring_ctx *ctx = req->ctx;
2675 struct io_submit_state *state = &ctx->submit_state;
2677 io_req_complete_state(req, res, cflags);
2678 state->compl_reqs[state->compl_nr++] = req;
2679 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2680 io_submit_flush_completions(ctx);
2682 io_req_complete_post(req, res, cflags);
2686 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2687 unsigned int issue_flags)
2689 if (__io_complete_rw_common(req, res))
2691 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2694 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2696 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2698 if (__io_complete_rw_common(req, res))
2701 req->io_task_work.func = io_req_task_complete;
2702 io_req_task_work_add(req);
2705 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2707 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2709 if (kiocb->ki_flags & IOCB_WRITE)
2710 kiocb_end_write(req);
2711 if (unlikely(res != req->result)) {
2712 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2713 io_resubmit_prep(req))) {
2715 req->flags |= REQ_F_DONT_REISSUE;
2719 WRITE_ONCE(req->result, res);
2720 /* order with io_iopoll_complete() checking ->result */
2722 WRITE_ONCE(req->iopoll_completed, 1);
2726 * After the iocb has been issued, it's safe to be found on the poll list.
2727 * Adding the kiocb to the list AFTER submission ensures that we don't
2728 * find it from a io_do_iopoll() thread before the issuer is done
2729 * accessing the kiocb cookie.
2731 static void io_iopoll_req_issued(struct io_kiocb *req)
2733 struct io_ring_ctx *ctx = req->ctx;
2734 const bool in_async = io_wq_current_is_worker();
2736 /* workqueue context doesn't hold uring_lock, grab it now */
2737 if (unlikely(in_async))
2738 mutex_lock(&ctx->uring_lock);
2741 * Track whether we have multiple files in our lists. This will impact
2742 * how we do polling eventually, not spinning if we're on potentially
2743 * different devices.
2745 if (list_empty(&ctx->iopoll_list)) {
2746 ctx->poll_multi_queue = false;
2747 } else if (!ctx->poll_multi_queue) {
2748 struct io_kiocb *list_req;
2749 unsigned int queue_num0, queue_num1;
2751 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2754 if (list_req->file != req->file) {
2755 ctx->poll_multi_queue = true;
2757 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2758 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2759 if (queue_num0 != queue_num1)
2760 ctx->poll_multi_queue = true;
2765 * For fast devices, IO may have already completed. If it has, add
2766 * it to the front so we find it first.
2768 if (READ_ONCE(req->iopoll_completed))
2769 list_add(&req->inflight_entry, &ctx->iopoll_list);
2771 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2773 if (unlikely(in_async)) {
2775 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2776 * in sq thread task context or in io worker task context. If
2777 * current task context is sq thread, we don't need to check
2778 * whether should wake up sq thread.
2780 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2781 wq_has_sleeper(&ctx->sq_data->wait))
2782 wake_up(&ctx->sq_data->wait);
2784 mutex_unlock(&ctx->uring_lock);
2788 static bool io_bdev_nowait(struct block_device *bdev)
2790 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2794 * If we tracked the file through the SCM inflight mechanism, we could support
2795 * any file. For now, just ensure that anything potentially problematic is done
2798 static bool __io_file_supports_nowait(struct file *file, int rw)
2800 umode_t mode = file_inode(file)->i_mode;
2802 if (S_ISBLK(mode)) {
2803 if (IS_ENABLED(CONFIG_BLOCK) &&
2804 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2810 if (S_ISREG(mode)) {
2811 if (IS_ENABLED(CONFIG_BLOCK) &&
2812 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2813 file->f_op != &io_uring_fops)
2818 /* any ->read/write should understand O_NONBLOCK */
2819 if (file->f_flags & O_NONBLOCK)
2822 if (!(file->f_mode & FMODE_NOWAIT))
2826 return file->f_op->read_iter != NULL;
2828 return file->f_op->write_iter != NULL;
2831 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2833 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2835 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2838 return __io_file_supports_nowait(req->file, rw);
2841 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2843 struct io_ring_ctx *ctx = req->ctx;
2844 struct kiocb *kiocb = &req->rw.kiocb;
2845 struct file *file = req->file;
2849 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2850 req->flags |= REQ_F_ISREG;
2852 kiocb->ki_pos = READ_ONCE(sqe->off);
2853 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2854 req->flags |= REQ_F_CUR_POS;
2855 kiocb->ki_pos = file->f_pos;
2857 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2858 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2859 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2863 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2864 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2865 req->flags |= REQ_F_NOWAIT;
2867 ioprio = READ_ONCE(sqe->ioprio);
2869 ret = ioprio_check_cap(ioprio);
2873 kiocb->ki_ioprio = ioprio;
2875 kiocb->ki_ioprio = get_current_ioprio();
2877 if (ctx->flags & IORING_SETUP_IOPOLL) {
2878 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2879 !kiocb->ki_filp->f_op->iopoll)
2882 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2883 kiocb->ki_complete = io_complete_rw_iopoll;
2884 req->iopoll_completed = 0;
2886 if (kiocb->ki_flags & IOCB_HIPRI)
2888 kiocb->ki_complete = io_complete_rw;
2891 if (req->opcode == IORING_OP_READ_FIXED ||
2892 req->opcode == IORING_OP_WRITE_FIXED) {
2894 io_req_set_rsrc_node(req);
2897 req->rw.addr = READ_ONCE(sqe->addr);
2898 req->rw.len = READ_ONCE(sqe->len);
2899 req->buf_index = READ_ONCE(sqe->buf_index);
2903 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2909 case -ERESTARTNOINTR:
2910 case -ERESTARTNOHAND:
2911 case -ERESTART_RESTARTBLOCK:
2913 * We can't just restart the syscall, since previously
2914 * submitted sqes may already be in progress. Just fail this
2920 kiocb->ki_complete(kiocb, ret, 0);
2924 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2925 unsigned int issue_flags)
2927 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2928 struct io_async_rw *io = req->async_data;
2929 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2931 /* add previously done IO, if any */
2932 if (io && io->bytes_done > 0) {
2934 ret = io->bytes_done;
2936 ret += io->bytes_done;
2939 if (req->flags & REQ_F_CUR_POS)
2940 req->file->f_pos = kiocb->ki_pos;
2941 if (ret >= 0 && check_reissue)
2942 __io_complete_rw(req, ret, 0, issue_flags);
2944 io_rw_done(kiocb, ret);
2946 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2947 req->flags &= ~REQ_F_REISSUE;
2948 if (io_resubmit_prep(req)) {
2949 io_req_task_queue_reissue(req);
2952 __io_req_complete(req, issue_flags, ret,
2953 io_put_rw_kbuf(req));
2958 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2959 struct io_mapped_ubuf *imu)
2961 size_t len = req->rw.len;
2962 u64 buf_end, buf_addr = req->rw.addr;
2965 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2967 /* not inside the mapped region */
2968 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2972 * May not be a start of buffer, set size appropriately
2973 * and advance us to the beginning.
2975 offset = buf_addr - imu->ubuf;
2976 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2980 * Don't use iov_iter_advance() here, as it's really slow for
2981 * using the latter parts of a big fixed buffer - it iterates
2982 * over each segment manually. We can cheat a bit here, because
2985 * 1) it's a BVEC iter, we set it up
2986 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2987 * first and last bvec
2989 * So just find our index, and adjust the iterator afterwards.
2990 * If the offset is within the first bvec (or the whole first
2991 * bvec, just use iov_iter_advance(). This makes it easier
2992 * since we can just skip the first segment, which may not
2993 * be PAGE_SIZE aligned.
2995 const struct bio_vec *bvec = imu->bvec;
2997 if (offset <= bvec->bv_len) {
2998 iov_iter_advance(iter, offset);
3000 unsigned long seg_skip;
3002 /* skip first vec */
3003 offset -= bvec->bv_len;
3004 seg_skip = 1 + (offset >> PAGE_SHIFT);
3006 iter->bvec = bvec + seg_skip;
3007 iter->nr_segs -= seg_skip;
3008 iter->count -= bvec->bv_len + offset;
3009 iter->iov_offset = offset & ~PAGE_MASK;
3016 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3018 struct io_ring_ctx *ctx = req->ctx;
3019 struct io_mapped_ubuf *imu = req->imu;
3020 u16 index, buf_index = req->buf_index;
3023 if (unlikely(buf_index >= ctx->nr_user_bufs))
3025 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3026 imu = READ_ONCE(ctx->user_bufs[index]);
3029 return __io_import_fixed(req, rw, iter, imu);
3032 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3035 mutex_unlock(&ctx->uring_lock);
3038 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3041 * "Normal" inline submissions always hold the uring_lock, since we
3042 * grab it from the system call. Same is true for the SQPOLL offload.
3043 * The only exception is when we've detached the request and issue it
3044 * from an async worker thread, grab the lock for that case.
3047 mutex_lock(&ctx->uring_lock);
3050 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3051 int bgid, struct io_buffer *kbuf,
3054 struct io_buffer *head;
3056 if (req->flags & REQ_F_BUFFER_SELECTED)
3059 io_ring_submit_lock(req->ctx, needs_lock);
3061 lockdep_assert_held(&req->ctx->uring_lock);
3063 head = xa_load(&req->ctx->io_buffers, bgid);
3065 if (!list_empty(&head->list)) {
3066 kbuf = list_last_entry(&head->list, struct io_buffer,
3068 list_del(&kbuf->list);
3071 xa_erase(&req->ctx->io_buffers, bgid);
3073 if (*len > kbuf->len)
3076 kbuf = ERR_PTR(-ENOBUFS);
3079 io_ring_submit_unlock(req->ctx, needs_lock);
3084 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3087 struct io_buffer *kbuf;
3090 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3091 bgid = req->buf_index;
3092 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3095 req->rw.addr = (u64) (unsigned long) kbuf;
3096 req->flags |= REQ_F_BUFFER_SELECTED;
3097 return u64_to_user_ptr(kbuf->addr);
3100 #ifdef CONFIG_COMPAT
3101 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3104 struct compat_iovec __user *uiov;
3105 compat_ssize_t clen;
3109 uiov = u64_to_user_ptr(req->rw.addr);
3110 if (!access_ok(uiov, sizeof(*uiov)))
3112 if (__get_user(clen, &uiov->iov_len))
3118 buf = io_rw_buffer_select(req, &len, needs_lock);
3120 return PTR_ERR(buf);
3121 iov[0].iov_base = buf;
3122 iov[0].iov_len = (compat_size_t) len;
3127 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3130 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3134 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3137 len = iov[0].iov_len;
3140 buf = io_rw_buffer_select(req, &len, needs_lock);
3142 return PTR_ERR(buf);
3143 iov[0].iov_base = buf;
3144 iov[0].iov_len = len;
3148 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3151 if (req->flags & REQ_F_BUFFER_SELECTED) {
3152 struct io_buffer *kbuf;
3154 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3155 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3156 iov[0].iov_len = kbuf->len;
3159 if (req->rw.len != 1)
3162 #ifdef CONFIG_COMPAT
3163 if (req->ctx->compat)
3164 return io_compat_import(req, iov, needs_lock);
3167 return __io_iov_buffer_select(req, iov, needs_lock);
3170 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3171 struct iov_iter *iter, bool needs_lock)
3173 void __user *buf = u64_to_user_ptr(req->rw.addr);
3174 size_t sqe_len = req->rw.len;
3175 u8 opcode = req->opcode;
3178 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3180 return io_import_fixed(req, rw, iter);
3183 /* buffer index only valid with fixed read/write, or buffer select */
3184 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3187 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3188 if (req->flags & REQ_F_BUFFER_SELECT) {
3189 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3191 return PTR_ERR(buf);
3192 req->rw.len = sqe_len;
3195 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3200 if (req->flags & REQ_F_BUFFER_SELECT) {
3201 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3203 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3208 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3212 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3214 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3218 * For files that don't have ->read_iter() and ->write_iter(), handle them
3219 * by looping over ->read() or ->write() manually.
3221 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3223 struct kiocb *kiocb = &req->rw.kiocb;
3224 struct file *file = req->file;
3228 * Don't support polled IO through this interface, and we can't
3229 * support non-blocking either. For the latter, this just causes
3230 * the kiocb to be handled from an async context.
3232 if (kiocb->ki_flags & IOCB_HIPRI)
3234 if (kiocb->ki_flags & IOCB_NOWAIT)
3237 while (iov_iter_count(iter)) {
3241 if (!iov_iter_is_bvec(iter)) {
3242 iovec = iov_iter_iovec(iter);
3244 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3245 iovec.iov_len = req->rw.len;
3249 nr = file->f_op->read(file, iovec.iov_base,
3250 iovec.iov_len, io_kiocb_ppos(kiocb));
3252 nr = file->f_op->write(file, iovec.iov_base,
3253 iovec.iov_len, io_kiocb_ppos(kiocb));
3262 if (nr != iovec.iov_len)
3266 iov_iter_advance(iter, nr);
3272 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3273 const struct iovec *fast_iov, struct iov_iter *iter)
3275 struct io_async_rw *rw = req->async_data;
3277 memcpy(&rw->iter, iter, sizeof(*iter));
3278 rw->free_iovec = iovec;
3280 /* can only be fixed buffers, no need to do anything */
3281 if (iov_iter_is_bvec(iter))
3284 unsigned iov_off = 0;
3286 rw->iter.iov = rw->fast_iov;
3287 if (iter->iov != fast_iov) {
3288 iov_off = iter->iov - fast_iov;
3289 rw->iter.iov += iov_off;
3291 if (rw->fast_iov != fast_iov)
3292 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3293 sizeof(struct iovec) * iter->nr_segs);
3295 req->flags |= REQ_F_NEED_CLEANUP;
3299 static inline int io_alloc_async_data(struct io_kiocb *req)
3301 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3302 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3303 return req->async_data == NULL;
3306 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3307 const struct iovec *fast_iov,
3308 struct iov_iter *iter, bool force)
3310 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3312 if (!req->async_data) {
3313 if (io_alloc_async_data(req)) {
3318 io_req_map_rw(req, iovec, fast_iov, iter);
3323 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3325 struct io_async_rw *iorw = req->async_data;
3326 struct iovec *iov = iorw->fast_iov;
3329 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3330 if (unlikely(ret < 0))
3333 iorw->bytes_done = 0;
3334 iorw->free_iovec = iov;
3336 req->flags |= REQ_F_NEED_CLEANUP;
3340 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3342 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3344 return io_prep_rw(req, sqe);
3348 * This is our waitqueue callback handler, registered through lock_page_async()
3349 * when we initially tried to do the IO with the iocb armed our waitqueue.
3350 * This gets called when the page is unlocked, and we generally expect that to
3351 * happen when the page IO is completed and the page is now uptodate. This will
3352 * queue a task_work based retry of the operation, attempting to copy the data
3353 * again. If the latter fails because the page was NOT uptodate, then we will
3354 * do a thread based blocking retry of the operation. That's the unexpected
3357 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3358 int sync, void *arg)
3360 struct wait_page_queue *wpq;
3361 struct io_kiocb *req = wait->private;
3362 struct wait_page_key *key = arg;
3364 wpq = container_of(wait, struct wait_page_queue, wait);
3366 if (!wake_page_match(wpq, key))
3369 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3370 list_del_init(&wait->entry);
3371 io_req_task_queue(req);
3376 * This controls whether a given IO request should be armed for async page
3377 * based retry. If we return false here, the request is handed to the async
3378 * worker threads for retry. If we're doing buffered reads on a regular file,
3379 * we prepare a private wait_page_queue entry and retry the operation. This
3380 * will either succeed because the page is now uptodate and unlocked, or it
3381 * will register a callback when the page is unlocked at IO completion. Through
3382 * that callback, io_uring uses task_work to setup a retry of the operation.
3383 * That retry will attempt the buffered read again. The retry will generally
3384 * succeed, or in rare cases where it fails, we then fall back to using the
3385 * async worker threads for a blocking retry.
3387 static bool io_rw_should_retry(struct io_kiocb *req)
3389 struct io_async_rw *rw = req->async_data;
3390 struct wait_page_queue *wait = &rw->wpq;
3391 struct kiocb *kiocb = &req->rw.kiocb;
3393 /* never retry for NOWAIT, we just complete with -EAGAIN */
3394 if (req->flags & REQ_F_NOWAIT)
3397 /* Only for buffered IO */
3398 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3402 * just use poll if we can, and don't attempt if the fs doesn't
3403 * support callback based unlocks
3405 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3408 wait->wait.func = io_async_buf_func;
3409 wait->wait.private = req;
3410 wait->wait.flags = 0;
3411 INIT_LIST_HEAD(&wait->wait.entry);
3412 kiocb->ki_flags |= IOCB_WAITQ;
3413 kiocb->ki_flags &= ~IOCB_NOWAIT;
3414 kiocb->ki_waitq = wait;
3418 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3420 if (req->file->f_op->read_iter)
3421 return call_read_iter(req->file, &req->rw.kiocb, iter);
3422 else if (req->file->f_op->read)
3423 return loop_rw_iter(READ, req, iter);
3428 static bool need_read_all(struct io_kiocb *req)
3430 return req->flags & REQ_F_ISREG ||
3431 S_ISBLK(file_inode(req->file)->i_mode);
3434 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3436 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3437 struct kiocb *kiocb = &req->rw.kiocb;
3438 struct iov_iter __iter, *iter = &__iter;
3439 struct io_async_rw *rw = req->async_data;
3440 ssize_t io_size, ret, ret2;
3441 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3447 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3451 io_size = iov_iter_count(iter);
3452 req->result = io_size;
3454 /* Ensure we clear previously set non-block flag */
3455 if (!force_nonblock)
3456 kiocb->ki_flags &= ~IOCB_NOWAIT;
3458 kiocb->ki_flags |= IOCB_NOWAIT;
3460 /* If the file doesn't support async, just async punt */
3461 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3462 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3463 return ret ?: -EAGAIN;
3466 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3467 if (unlikely(ret)) {
3472 ret = io_iter_do_read(req, iter);
3474 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3475 req->flags &= ~REQ_F_REISSUE;
3476 /* IOPOLL retry should happen for io-wq threads */
3477 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3479 /* no retry on NONBLOCK nor RWF_NOWAIT */
3480 if (req->flags & REQ_F_NOWAIT)
3482 /* some cases will consume bytes even on error returns */
3483 iov_iter_reexpand(iter, iter->count + iter->truncated);
3484 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3486 } else if (ret == -EIOCBQUEUED) {
3488 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3489 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3490 /* read all, failed, already did sync or don't want to retry */
3494 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3499 rw = req->async_data;
3500 /* now use our persistent iterator, if we aren't already */
3505 rw->bytes_done += ret;
3506 /* if we can retry, do so with the callbacks armed */
3507 if (!io_rw_should_retry(req)) {
3508 kiocb->ki_flags &= ~IOCB_WAITQ;
3513 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3514 * we get -EIOCBQUEUED, then we'll get a notification when the
3515 * desired page gets unlocked. We can also get a partial read
3516 * here, and if we do, then just retry at the new offset.
3518 ret = io_iter_do_read(req, iter);
3519 if (ret == -EIOCBQUEUED)
3521 /* we got some bytes, but not all. retry. */
3522 kiocb->ki_flags &= ~IOCB_WAITQ;
3523 } while (ret > 0 && ret < io_size);
3525 kiocb_done(kiocb, ret, issue_flags);
3527 /* it's faster to check here then delegate to kfree */
3533 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3535 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3537 return io_prep_rw(req, sqe);
3540 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3542 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3543 struct kiocb *kiocb = &req->rw.kiocb;
3544 struct iov_iter __iter, *iter = &__iter;
3545 struct io_async_rw *rw = req->async_data;
3546 ssize_t ret, ret2, io_size;
3547 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3553 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3557 io_size = iov_iter_count(iter);
3558 req->result = io_size;
3560 /* Ensure we clear previously set non-block flag */
3561 if (!force_nonblock)
3562 kiocb->ki_flags &= ~IOCB_NOWAIT;
3564 kiocb->ki_flags |= IOCB_NOWAIT;
3566 /* If the file doesn't support async, just async punt */
3567 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3570 /* file path doesn't support NOWAIT for non-direct_IO */
3571 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3572 (req->flags & REQ_F_ISREG))
3575 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3580 * Open-code file_start_write here to grab freeze protection,
3581 * which will be released by another thread in
3582 * io_complete_rw(). Fool lockdep by telling it the lock got
3583 * released so that it doesn't complain about the held lock when
3584 * we return to userspace.
3586 if (req->flags & REQ_F_ISREG) {
3587 sb_start_write(file_inode(req->file)->i_sb);
3588 __sb_writers_release(file_inode(req->file)->i_sb,
3591 kiocb->ki_flags |= IOCB_WRITE;
3593 if (req->file->f_op->write_iter)
3594 ret2 = call_write_iter(req->file, kiocb, iter);
3595 else if (req->file->f_op->write)
3596 ret2 = loop_rw_iter(WRITE, req, iter);
3600 if (req->flags & REQ_F_REISSUE) {
3601 req->flags &= ~REQ_F_REISSUE;
3606 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3607 * retry them without IOCB_NOWAIT.
3609 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3611 /* no retry on NONBLOCK nor RWF_NOWAIT */
3612 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3614 if (!force_nonblock || ret2 != -EAGAIN) {
3615 /* IOPOLL retry should happen for io-wq threads */
3616 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3619 kiocb_done(kiocb, ret2, issue_flags);
3622 /* some cases will consume bytes even on error returns */
3623 iov_iter_reexpand(iter, iter->count + iter->truncated);
3624 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3625 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3626 return ret ?: -EAGAIN;
3629 /* it's reportedly faster than delegating the null check to kfree() */
3635 static int io_renameat_prep(struct io_kiocb *req,
3636 const struct io_uring_sqe *sqe)
3638 struct io_rename *ren = &req->rename;
3639 const char __user *oldf, *newf;
3641 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3643 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3645 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3648 ren->old_dfd = READ_ONCE(sqe->fd);
3649 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3650 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3651 ren->new_dfd = READ_ONCE(sqe->len);
3652 ren->flags = READ_ONCE(sqe->rename_flags);
3654 ren->oldpath = getname(oldf);
3655 if (IS_ERR(ren->oldpath))
3656 return PTR_ERR(ren->oldpath);
3658 ren->newpath = getname(newf);
3659 if (IS_ERR(ren->newpath)) {
3660 putname(ren->oldpath);
3661 return PTR_ERR(ren->newpath);
3664 req->flags |= REQ_F_NEED_CLEANUP;
3668 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3670 struct io_rename *ren = &req->rename;
3673 if (issue_flags & IO_URING_F_NONBLOCK)
3676 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3677 ren->newpath, ren->flags);
3679 req->flags &= ~REQ_F_NEED_CLEANUP;
3682 io_req_complete(req, ret);
3686 static int io_unlinkat_prep(struct io_kiocb *req,
3687 const struct io_uring_sqe *sqe)
3689 struct io_unlink *un = &req->unlink;
3690 const char __user *fname;
3692 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3694 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3697 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3700 un->dfd = READ_ONCE(sqe->fd);
3702 un->flags = READ_ONCE(sqe->unlink_flags);
3703 if (un->flags & ~AT_REMOVEDIR)
3706 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3707 un->filename = getname(fname);
3708 if (IS_ERR(un->filename))
3709 return PTR_ERR(un->filename);
3711 req->flags |= REQ_F_NEED_CLEANUP;
3715 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3717 struct io_unlink *un = &req->unlink;
3720 if (issue_flags & IO_URING_F_NONBLOCK)
3723 if (un->flags & AT_REMOVEDIR)
3724 ret = do_rmdir(un->dfd, un->filename);
3726 ret = do_unlinkat(un->dfd, un->filename);
3728 req->flags &= ~REQ_F_NEED_CLEANUP;
3731 io_req_complete(req, ret);
3735 static int io_mkdirat_prep(struct io_kiocb *req,
3736 const struct io_uring_sqe *sqe)
3738 struct io_mkdir *mkd = &req->mkdir;
3739 const char __user *fname;
3741 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3743 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3746 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3749 mkd->dfd = READ_ONCE(sqe->fd);
3750 mkd->mode = READ_ONCE(sqe->len);
3752 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3753 mkd->filename = getname(fname);
3754 if (IS_ERR(mkd->filename))
3755 return PTR_ERR(mkd->filename);
3757 req->flags |= REQ_F_NEED_CLEANUP;
3761 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3763 struct io_mkdir *mkd = &req->mkdir;
3766 if (issue_flags & IO_URING_F_NONBLOCK)
3769 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3771 req->flags &= ~REQ_F_NEED_CLEANUP;
3774 io_req_complete(req, ret);
3778 static int io_symlinkat_prep(struct io_kiocb *req,
3779 const struct io_uring_sqe *sqe)
3781 struct io_symlink *sl = &req->symlink;
3782 const char __user *oldpath, *newpath;
3784 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3786 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3789 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3792 sl->new_dfd = READ_ONCE(sqe->fd);
3793 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3794 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3796 sl->oldpath = getname(oldpath);
3797 if (IS_ERR(sl->oldpath))
3798 return PTR_ERR(sl->oldpath);
3800 sl->newpath = getname(newpath);
3801 if (IS_ERR(sl->newpath)) {
3802 putname(sl->oldpath);
3803 return PTR_ERR(sl->newpath);
3806 req->flags |= REQ_F_NEED_CLEANUP;
3810 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3812 struct io_symlink *sl = &req->symlink;
3815 if (issue_flags & IO_URING_F_NONBLOCK)
3818 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3820 req->flags &= ~REQ_F_NEED_CLEANUP;
3823 io_req_complete(req, ret);
3827 static int io_linkat_prep(struct io_kiocb *req,
3828 const struct io_uring_sqe *sqe)
3830 struct io_hardlink *lnk = &req->hardlink;
3831 const char __user *oldf, *newf;
3833 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3835 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3837 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3840 lnk->old_dfd = READ_ONCE(sqe->fd);
3841 lnk->new_dfd = READ_ONCE(sqe->len);
3842 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3843 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3844 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3846 lnk->oldpath = getname(oldf);
3847 if (IS_ERR(lnk->oldpath))
3848 return PTR_ERR(lnk->oldpath);
3850 lnk->newpath = getname(newf);
3851 if (IS_ERR(lnk->newpath)) {
3852 putname(lnk->oldpath);
3853 return PTR_ERR(lnk->newpath);
3856 req->flags |= REQ_F_NEED_CLEANUP;
3860 static int io_linkat(struct io_kiocb *req, int issue_flags)
3862 struct io_hardlink *lnk = &req->hardlink;
3865 if (issue_flags & IO_URING_F_NONBLOCK)
3868 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3869 lnk->newpath, lnk->flags);
3871 req->flags &= ~REQ_F_NEED_CLEANUP;
3874 io_req_complete(req, ret);
3878 static int io_shutdown_prep(struct io_kiocb *req,
3879 const struct io_uring_sqe *sqe)
3881 #if defined(CONFIG_NET)
3882 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3884 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3885 sqe->buf_index || sqe->splice_fd_in))
3888 req->shutdown.how = READ_ONCE(sqe->len);
3895 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3897 #if defined(CONFIG_NET)
3898 struct socket *sock;
3901 if (issue_flags & IO_URING_F_NONBLOCK)
3904 sock = sock_from_file(req->file);
3905 if (unlikely(!sock))
3908 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3911 io_req_complete(req, ret);
3918 static int __io_splice_prep(struct io_kiocb *req,
3919 const struct io_uring_sqe *sqe)
3921 struct io_splice *sp = &req->splice;
3922 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3924 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3928 sp->len = READ_ONCE(sqe->len);
3929 sp->flags = READ_ONCE(sqe->splice_flags);
3931 if (unlikely(sp->flags & ~valid_flags))
3934 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3935 (sp->flags & SPLICE_F_FD_IN_FIXED));
3938 req->flags |= REQ_F_NEED_CLEANUP;
3942 static int io_tee_prep(struct io_kiocb *req,
3943 const struct io_uring_sqe *sqe)
3945 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3947 return __io_splice_prep(req, sqe);
3950 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3952 struct io_splice *sp = &req->splice;
3953 struct file *in = sp->file_in;
3954 struct file *out = sp->file_out;
3955 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3958 if (issue_flags & IO_URING_F_NONBLOCK)
3961 ret = do_tee(in, out, sp->len, flags);
3963 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3965 req->flags &= ~REQ_F_NEED_CLEANUP;
3969 io_req_complete(req, ret);
3973 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3975 struct io_splice *sp = &req->splice;
3977 sp->off_in = READ_ONCE(sqe->splice_off_in);
3978 sp->off_out = READ_ONCE(sqe->off);
3979 return __io_splice_prep(req, sqe);
3982 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3984 struct io_splice *sp = &req->splice;
3985 struct file *in = sp->file_in;
3986 struct file *out = sp->file_out;
3987 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3988 loff_t *poff_in, *poff_out;
3991 if (issue_flags & IO_URING_F_NONBLOCK)
3994 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3995 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3998 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4000 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4002 req->flags &= ~REQ_F_NEED_CLEANUP;
4006 io_req_complete(req, ret);
4011 * IORING_OP_NOP just posts a completion event, nothing else.
4013 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4015 struct io_ring_ctx *ctx = req->ctx;
4017 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4020 __io_req_complete(req, issue_flags, 0, 0);
4024 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4026 struct io_ring_ctx *ctx = req->ctx;
4031 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4033 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4037 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4038 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4041 req->sync.off = READ_ONCE(sqe->off);
4042 req->sync.len = READ_ONCE(sqe->len);
4046 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4048 loff_t end = req->sync.off + req->sync.len;
4051 /* fsync always requires a blocking context */
4052 if (issue_flags & IO_URING_F_NONBLOCK)
4055 ret = vfs_fsync_range(req->file, req->sync.off,
4056 end > 0 ? end : LLONG_MAX,
4057 req->sync.flags & IORING_FSYNC_DATASYNC);
4060 io_req_complete(req, ret);
4064 static int io_fallocate_prep(struct io_kiocb *req,
4065 const struct io_uring_sqe *sqe)
4067 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4070 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4073 req->sync.off = READ_ONCE(sqe->off);
4074 req->sync.len = READ_ONCE(sqe->addr);
4075 req->sync.mode = READ_ONCE(sqe->len);
4079 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4083 /* fallocate always requiring blocking context */
4084 if (issue_flags & IO_URING_F_NONBLOCK)
4086 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4090 io_req_complete(req, ret);
4094 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4096 const char __user *fname;
4099 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4101 if (unlikely(sqe->ioprio || sqe->buf_index))
4103 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4106 /* open.how should be already initialised */
4107 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4108 req->open.how.flags |= O_LARGEFILE;
4110 req->open.dfd = READ_ONCE(sqe->fd);
4111 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4112 req->open.filename = getname(fname);
4113 if (IS_ERR(req->open.filename)) {
4114 ret = PTR_ERR(req->open.filename);
4115 req->open.filename = NULL;
4119 req->open.file_slot = READ_ONCE(sqe->file_index);
4120 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4123 req->open.nofile = rlimit(RLIMIT_NOFILE);
4124 req->flags |= REQ_F_NEED_CLEANUP;
4128 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4130 u64 mode = READ_ONCE(sqe->len);
4131 u64 flags = READ_ONCE(sqe->open_flags);
4133 req->open.how = build_open_how(flags, mode);
4134 return __io_openat_prep(req, sqe);
4137 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4139 struct open_how __user *how;
4143 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4144 len = READ_ONCE(sqe->len);
4145 if (len < OPEN_HOW_SIZE_VER0)
4148 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4153 return __io_openat_prep(req, sqe);
4156 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4158 struct open_flags op;
4160 bool resolve_nonblock, nonblock_set;
4161 bool fixed = !!req->open.file_slot;
4164 ret = build_open_flags(&req->open.how, &op);
4167 nonblock_set = op.open_flag & O_NONBLOCK;
4168 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4169 if (issue_flags & IO_URING_F_NONBLOCK) {
4171 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4172 * it'll always -EAGAIN
4174 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4176 op.lookup_flags |= LOOKUP_CACHED;
4177 op.open_flag |= O_NONBLOCK;
4181 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4186 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4189 * We could hang on to this 'fd' on retrying, but seems like
4190 * marginal gain for something that is now known to be a slower
4191 * path. So just put it, and we'll get a new one when we retry.
4196 ret = PTR_ERR(file);
4197 /* only retry if RESOLVE_CACHED wasn't already set by application */
4198 if (ret == -EAGAIN &&
4199 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4204 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4205 file->f_flags &= ~O_NONBLOCK;
4206 fsnotify_open(file);
4209 fd_install(ret, file);
4211 ret = io_install_fixed_file(req, file, issue_flags,
4212 req->open.file_slot - 1);
4214 putname(req->open.filename);
4215 req->flags &= ~REQ_F_NEED_CLEANUP;
4218 __io_req_complete(req, issue_flags, ret, 0);
4222 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4224 return io_openat2(req, issue_flags);
4227 static int io_remove_buffers_prep(struct io_kiocb *req,
4228 const struct io_uring_sqe *sqe)
4230 struct io_provide_buf *p = &req->pbuf;
4233 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4237 tmp = READ_ONCE(sqe->fd);
4238 if (!tmp || tmp > USHRT_MAX)
4241 memset(p, 0, sizeof(*p));
4243 p->bgid = READ_ONCE(sqe->buf_group);
4247 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4248 int bgid, unsigned nbufs)
4252 /* shouldn't happen */
4256 /* the head kbuf is the list itself */
4257 while (!list_empty(&buf->list)) {
4258 struct io_buffer *nxt;
4260 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4261 list_del(&nxt->list);
4268 xa_erase(&ctx->io_buffers, bgid);
4273 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4275 struct io_provide_buf *p = &req->pbuf;
4276 struct io_ring_ctx *ctx = req->ctx;
4277 struct io_buffer *head;
4279 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4281 io_ring_submit_lock(ctx, !force_nonblock);
4283 lockdep_assert_held(&ctx->uring_lock);
4286 head = xa_load(&ctx->io_buffers, p->bgid);
4288 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4292 /* complete before unlock, IOPOLL may need the lock */
4293 __io_req_complete(req, issue_flags, ret, 0);
4294 io_ring_submit_unlock(ctx, !force_nonblock);
4298 static int io_provide_buffers_prep(struct io_kiocb *req,
4299 const struct io_uring_sqe *sqe)
4301 unsigned long size, tmp_check;
4302 struct io_provide_buf *p = &req->pbuf;
4305 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4308 tmp = READ_ONCE(sqe->fd);
4309 if (!tmp || tmp > USHRT_MAX)
4312 p->addr = READ_ONCE(sqe->addr);
4313 p->len = READ_ONCE(sqe->len);
4315 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4318 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4321 size = (unsigned long)p->len * p->nbufs;
4322 if (!access_ok(u64_to_user_ptr(p->addr), size))
4325 p->bgid = READ_ONCE(sqe->buf_group);
4326 tmp = READ_ONCE(sqe->off);
4327 if (tmp > USHRT_MAX)
4333 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4335 struct io_buffer *buf;
4336 u64 addr = pbuf->addr;
4337 int i, bid = pbuf->bid;
4339 for (i = 0; i < pbuf->nbufs; i++) {
4340 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4345 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4350 INIT_LIST_HEAD(&buf->list);
4353 list_add_tail(&buf->list, &(*head)->list);
4357 return i ? i : -ENOMEM;
4360 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4362 struct io_provide_buf *p = &req->pbuf;
4363 struct io_ring_ctx *ctx = req->ctx;
4364 struct io_buffer *head, *list;
4366 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4368 io_ring_submit_lock(ctx, !force_nonblock);
4370 lockdep_assert_held(&ctx->uring_lock);
4372 list = head = xa_load(&ctx->io_buffers, p->bgid);
4374 ret = io_add_buffers(p, &head);
4375 if (ret >= 0 && !list) {
4376 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4378 __io_remove_buffers(ctx, head, p->bgid, -1U);
4382 /* complete before unlock, IOPOLL may need the lock */
4383 __io_req_complete(req, issue_flags, ret, 0);
4384 io_ring_submit_unlock(ctx, !force_nonblock);
4388 static int io_epoll_ctl_prep(struct io_kiocb *req,
4389 const struct io_uring_sqe *sqe)
4391 #if defined(CONFIG_EPOLL)
4392 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4394 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4397 req->epoll.epfd = READ_ONCE(sqe->fd);
4398 req->epoll.op = READ_ONCE(sqe->len);
4399 req->epoll.fd = READ_ONCE(sqe->off);
4401 if (ep_op_has_event(req->epoll.op)) {
4402 struct epoll_event __user *ev;
4404 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4405 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4415 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4417 #if defined(CONFIG_EPOLL)
4418 struct io_epoll *ie = &req->epoll;
4420 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4422 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4423 if (force_nonblock && ret == -EAGAIN)
4428 __io_req_complete(req, issue_flags, ret, 0);
4435 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4437 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4438 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4440 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4443 req->madvise.addr = READ_ONCE(sqe->addr);
4444 req->madvise.len = READ_ONCE(sqe->len);
4445 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4452 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4454 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4455 struct io_madvise *ma = &req->madvise;
4458 if (issue_flags & IO_URING_F_NONBLOCK)
4461 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4464 io_req_complete(req, ret);
4471 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4473 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4475 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4478 req->fadvise.offset = READ_ONCE(sqe->off);
4479 req->fadvise.len = READ_ONCE(sqe->len);
4480 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4484 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4486 struct io_fadvise *fa = &req->fadvise;
4489 if (issue_flags & IO_URING_F_NONBLOCK) {
4490 switch (fa->advice) {
4491 case POSIX_FADV_NORMAL:
4492 case POSIX_FADV_RANDOM:
4493 case POSIX_FADV_SEQUENTIAL:
4500 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4503 __io_req_complete(req, issue_flags, ret, 0);
4507 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4509 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4511 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4513 if (req->flags & REQ_F_FIXED_FILE)
4516 req->statx.dfd = READ_ONCE(sqe->fd);
4517 req->statx.mask = READ_ONCE(sqe->len);
4518 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4519 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4520 req->statx.flags = READ_ONCE(sqe->statx_flags);
4525 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4527 struct io_statx *ctx = &req->statx;
4530 if (issue_flags & IO_URING_F_NONBLOCK)
4533 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4538 io_req_complete(req, ret);
4542 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4544 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4546 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4547 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4549 if (req->flags & REQ_F_FIXED_FILE)
4552 req->close.fd = READ_ONCE(sqe->fd);
4556 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4558 struct files_struct *files = current->files;
4559 struct io_close *close = &req->close;
4560 struct fdtable *fdt;
4561 struct file *file = NULL;
4564 spin_lock(&files->file_lock);
4565 fdt = files_fdtable(files);
4566 if (close->fd >= fdt->max_fds) {
4567 spin_unlock(&files->file_lock);
4570 file = fdt->fd[close->fd];
4571 if (!file || file->f_op == &io_uring_fops) {
4572 spin_unlock(&files->file_lock);
4577 /* if the file has a flush method, be safe and punt to async */
4578 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4579 spin_unlock(&files->file_lock);
4583 ret = __close_fd_get_file(close->fd, &file);
4584 spin_unlock(&files->file_lock);
4591 /* No ->flush() or already async, safely close from here */
4592 ret = filp_close(file, current->files);
4598 __io_req_complete(req, issue_flags, ret, 0);
4602 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4604 struct io_ring_ctx *ctx = req->ctx;
4606 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4608 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4612 req->sync.off = READ_ONCE(sqe->off);
4613 req->sync.len = READ_ONCE(sqe->len);
4614 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4618 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4622 /* sync_file_range always requires a blocking context */
4623 if (issue_flags & IO_URING_F_NONBLOCK)
4626 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4630 io_req_complete(req, ret);
4634 #if defined(CONFIG_NET)
4635 static int io_setup_async_msg(struct io_kiocb *req,
4636 struct io_async_msghdr *kmsg)
4638 struct io_async_msghdr *async_msg = req->async_data;
4642 if (io_alloc_async_data(req)) {
4643 kfree(kmsg->free_iov);
4646 async_msg = req->async_data;
4647 req->flags |= REQ_F_NEED_CLEANUP;
4648 memcpy(async_msg, kmsg, sizeof(*kmsg));
4649 async_msg->msg.msg_name = &async_msg->addr;
4650 /* if were using fast_iov, set it to the new one */
4651 if (!async_msg->free_iov)
4652 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4657 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4658 struct io_async_msghdr *iomsg)
4660 iomsg->msg.msg_name = &iomsg->addr;
4661 iomsg->free_iov = iomsg->fast_iov;
4662 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4663 req->sr_msg.msg_flags, &iomsg->free_iov);
4666 static int io_sendmsg_prep_async(struct io_kiocb *req)
4670 ret = io_sendmsg_copy_hdr(req, req->async_data);
4672 req->flags |= REQ_F_NEED_CLEANUP;
4676 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4678 struct io_sr_msg *sr = &req->sr_msg;
4680 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4683 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4684 sr->len = READ_ONCE(sqe->len);
4685 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4686 if (sr->msg_flags & MSG_DONTWAIT)
4687 req->flags |= REQ_F_NOWAIT;
4689 #ifdef CONFIG_COMPAT
4690 if (req->ctx->compat)
4691 sr->msg_flags |= MSG_CMSG_COMPAT;
4696 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4698 struct io_async_msghdr iomsg, *kmsg;
4699 struct socket *sock;
4704 sock = sock_from_file(req->file);
4705 if (unlikely(!sock))
4708 kmsg = req->async_data;
4710 ret = io_sendmsg_copy_hdr(req, &iomsg);
4716 flags = req->sr_msg.msg_flags;
4717 if (issue_flags & IO_URING_F_NONBLOCK)
4718 flags |= MSG_DONTWAIT;
4719 if (flags & MSG_WAITALL)
4720 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4722 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4723 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4724 return io_setup_async_msg(req, kmsg);
4725 if (ret == -ERESTARTSYS)
4728 /* fast path, check for non-NULL to avoid function call */
4730 kfree(kmsg->free_iov);
4731 req->flags &= ~REQ_F_NEED_CLEANUP;
4734 __io_req_complete(req, issue_flags, ret, 0);
4738 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4740 struct io_sr_msg *sr = &req->sr_msg;
4743 struct socket *sock;
4748 sock = sock_from_file(req->file);
4749 if (unlikely(!sock))
4752 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4756 msg.msg_name = NULL;
4757 msg.msg_control = NULL;
4758 msg.msg_controllen = 0;
4759 msg.msg_namelen = 0;
4761 flags = req->sr_msg.msg_flags;
4762 if (issue_flags & IO_URING_F_NONBLOCK)
4763 flags |= MSG_DONTWAIT;
4764 if (flags & MSG_WAITALL)
4765 min_ret = iov_iter_count(&msg.msg_iter);
4767 msg.msg_flags = flags;
4768 ret = sock_sendmsg(sock, &msg);
4769 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4771 if (ret == -ERESTARTSYS)
4776 __io_req_complete(req, issue_flags, ret, 0);
4780 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4781 struct io_async_msghdr *iomsg)
4783 struct io_sr_msg *sr = &req->sr_msg;
4784 struct iovec __user *uiov;
4788 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4789 &iomsg->uaddr, &uiov, &iov_len);
4793 if (req->flags & REQ_F_BUFFER_SELECT) {
4796 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4798 sr->len = iomsg->fast_iov[0].iov_len;
4799 iomsg->free_iov = NULL;
4801 iomsg->free_iov = iomsg->fast_iov;
4802 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4803 &iomsg->free_iov, &iomsg->msg.msg_iter,
4812 #ifdef CONFIG_COMPAT
4813 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4814 struct io_async_msghdr *iomsg)
4816 struct io_sr_msg *sr = &req->sr_msg;
4817 struct compat_iovec __user *uiov;
4822 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4827 uiov = compat_ptr(ptr);
4828 if (req->flags & REQ_F_BUFFER_SELECT) {
4829 compat_ssize_t clen;
4833 if (!access_ok(uiov, sizeof(*uiov)))
4835 if (__get_user(clen, &uiov->iov_len))
4840 iomsg->free_iov = NULL;
4842 iomsg->free_iov = iomsg->fast_iov;
4843 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4844 UIO_FASTIOV, &iomsg->free_iov,
4845 &iomsg->msg.msg_iter, true);
4854 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4855 struct io_async_msghdr *iomsg)
4857 iomsg->msg.msg_name = &iomsg->addr;
4859 #ifdef CONFIG_COMPAT
4860 if (req->ctx->compat)
4861 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4864 return __io_recvmsg_copy_hdr(req, iomsg);
4867 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4870 struct io_sr_msg *sr = &req->sr_msg;
4871 struct io_buffer *kbuf;
4873 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4878 req->flags |= REQ_F_BUFFER_SELECTED;
4882 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4884 return io_put_kbuf(req, req->sr_msg.kbuf);
4887 static int io_recvmsg_prep_async(struct io_kiocb *req)
4891 ret = io_recvmsg_copy_hdr(req, req->async_data);
4893 req->flags |= REQ_F_NEED_CLEANUP;
4897 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4899 struct io_sr_msg *sr = &req->sr_msg;
4901 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4904 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4905 sr->len = READ_ONCE(sqe->len);
4906 sr->bgid = READ_ONCE(sqe->buf_group);
4907 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4908 if (sr->msg_flags & MSG_DONTWAIT)
4909 req->flags |= REQ_F_NOWAIT;
4911 #ifdef CONFIG_COMPAT
4912 if (req->ctx->compat)
4913 sr->msg_flags |= MSG_CMSG_COMPAT;
4918 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4920 struct io_async_msghdr iomsg, *kmsg;
4921 struct socket *sock;
4922 struct io_buffer *kbuf;
4925 int ret, cflags = 0;
4926 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4928 sock = sock_from_file(req->file);
4929 if (unlikely(!sock))
4932 kmsg = req->async_data;
4934 ret = io_recvmsg_copy_hdr(req, &iomsg);
4940 if (req->flags & REQ_F_BUFFER_SELECT) {
4941 kbuf = io_recv_buffer_select(req, !force_nonblock);
4943 return PTR_ERR(kbuf);
4944 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4945 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4946 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4947 1, req->sr_msg.len);
4950 flags = req->sr_msg.msg_flags;
4952 flags |= MSG_DONTWAIT;
4953 if (flags & MSG_WAITALL)
4954 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4956 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4957 kmsg->uaddr, flags);
4958 if (force_nonblock && ret == -EAGAIN)
4959 return io_setup_async_msg(req, kmsg);
4960 if (ret == -ERESTARTSYS)
4963 if (req->flags & REQ_F_BUFFER_SELECTED)
4964 cflags = io_put_recv_kbuf(req);
4965 /* fast path, check for non-NULL to avoid function call */
4967 kfree(kmsg->free_iov);
4968 req->flags &= ~REQ_F_NEED_CLEANUP;
4969 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4971 __io_req_complete(req, issue_flags, ret, cflags);
4975 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4977 struct io_buffer *kbuf;
4978 struct io_sr_msg *sr = &req->sr_msg;
4980 void __user *buf = sr->buf;
4981 struct socket *sock;
4985 int ret, cflags = 0;
4986 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4988 sock = sock_from_file(req->file);
4989 if (unlikely(!sock))
4992 if (req->flags & REQ_F_BUFFER_SELECT) {
4993 kbuf = io_recv_buffer_select(req, !force_nonblock);
4995 return PTR_ERR(kbuf);
4996 buf = u64_to_user_ptr(kbuf->addr);
4999 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5003 msg.msg_name = NULL;
5004 msg.msg_control = NULL;
5005 msg.msg_controllen = 0;
5006 msg.msg_namelen = 0;
5007 msg.msg_iocb = NULL;
5010 flags = req->sr_msg.msg_flags;
5012 flags |= MSG_DONTWAIT;
5013 if (flags & MSG_WAITALL)
5014 min_ret = iov_iter_count(&msg.msg_iter);
5016 ret = sock_recvmsg(sock, &msg, flags);
5017 if (force_nonblock && ret == -EAGAIN)
5019 if (ret == -ERESTARTSYS)
5022 if (req->flags & REQ_F_BUFFER_SELECTED)
5023 cflags = io_put_recv_kbuf(req);
5024 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5026 __io_req_complete(req, issue_flags, ret, cflags);
5030 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5032 struct io_accept *accept = &req->accept;
5034 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5036 if (sqe->ioprio || sqe->len || sqe->buf_index)
5039 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5040 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5041 accept->flags = READ_ONCE(sqe->accept_flags);
5042 accept->nofile = rlimit(RLIMIT_NOFILE);
5044 accept->file_slot = READ_ONCE(sqe->file_index);
5045 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5046 (accept->flags & SOCK_CLOEXEC)))
5048 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5050 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5051 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5055 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5057 struct io_accept *accept = &req->accept;
5058 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5059 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5060 bool fixed = !!accept->file_slot;
5064 if (req->file->f_flags & O_NONBLOCK)
5065 req->flags |= REQ_F_NOWAIT;
5068 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5069 if (unlikely(fd < 0))
5072 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5077 ret = PTR_ERR(file);
5078 if (ret == -EAGAIN && force_nonblock)
5080 if (ret == -ERESTARTSYS)
5083 } else if (!fixed) {
5084 fd_install(fd, file);
5087 ret = io_install_fixed_file(req, file, issue_flags,
5088 accept->file_slot - 1);
5090 __io_req_complete(req, issue_flags, ret, 0);
5094 static int io_connect_prep_async(struct io_kiocb *req)
5096 struct io_async_connect *io = req->async_data;
5097 struct io_connect *conn = &req->connect;
5099 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5102 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5104 struct io_connect *conn = &req->connect;
5106 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5108 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5112 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5113 conn->addr_len = READ_ONCE(sqe->addr2);
5117 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5119 struct io_async_connect __io, *io;
5120 unsigned file_flags;
5122 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5124 if (req->async_data) {
5125 io = req->async_data;
5127 ret = move_addr_to_kernel(req->connect.addr,
5128 req->connect.addr_len,
5135 file_flags = force_nonblock ? O_NONBLOCK : 0;
5137 ret = __sys_connect_file(req->file, &io->address,
5138 req->connect.addr_len, file_flags);
5139 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5140 if (req->async_data)
5142 if (io_alloc_async_data(req)) {
5146 memcpy(req->async_data, &__io, sizeof(__io));
5149 if (ret == -ERESTARTSYS)
5154 __io_req_complete(req, issue_flags, ret, 0);
5157 #else /* !CONFIG_NET */
5158 #define IO_NETOP_FN(op) \
5159 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5161 return -EOPNOTSUPP; \
5164 #define IO_NETOP_PREP(op) \
5166 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5168 return -EOPNOTSUPP; \
5171 #define IO_NETOP_PREP_ASYNC(op) \
5173 static int io_##op##_prep_async(struct io_kiocb *req) \
5175 return -EOPNOTSUPP; \
5178 IO_NETOP_PREP_ASYNC(sendmsg);
5179 IO_NETOP_PREP_ASYNC(recvmsg);
5180 IO_NETOP_PREP_ASYNC(connect);
5181 IO_NETOP_PREP(accept);
5184 #endif /* CONFIG_NET */
5186 struct io_poll_table {
5187 struct poll_table_struct pt;
5188 struct io_kiocb *req;
5193 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5194 __poll_t mask, io_req_tw_func_t func)
5196 /* for instances that support it check for an event match first: */
5197 if (mask && !(mask & poll->events))
5200 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5202 list_del_init(&poll->wait.entry);
5205 req->io_task_work.func = func;
5208 * If this fails, then the task is exiting. When a task exits, the
5209 * work gets canceled, so just cancel this request as well instead
5210 * of executing it. We can't safely execute it anyway, as we may not
5211 * have the needed state needed for it anyway.
5213 io_req_task_work_add(req);
5217 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5218 __acquires(&req->ctx->completion_lock)
5220 struct io_ring_ctx *ctx = req->ctx;
5222 /* req->task == current here, checking PF_EXITING is safe */
5223 if (unlikely(req->task->flags & PF_EXITING))
5224 WRITE_ONCE(poll->canceled, true);
5226 if (!req->result && !READ_ONCE(poll->canceled)) {
5227 struct poll_table_struct pt = { ._key = poll->events };
5229 req->result = vfs_poll(req->file, &pt) & poll->events;
5232 spin_lock(&ctx->completion_lock);
5233 if (!req->result && !READ_ONCE(poll->canceled)) {
5234 add_wait_queue(poll->head, &poll->wait);
5241 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5243 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5244 if (req->opcode == IORING_OP_POLL_ADD)
5245 return req->async_data;
5246 return req->apoll->double_poll;
5249 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5251 if (req->opcode == IORING_OP_POLL_ADD)
5253 return &req->apoll->poll;
5256 static void io_poll_remove_double(struct io_kiocb *req)
5257 __must_hold(&req->ctx->completion_lock)
5259 struct io_poll_iocb *poll = io_poll_get_double(req);
5261 lockdep_assert_held(&req->ctx->completion_lock);
5263 if (poll && poll->head) {
5264 struct wait_queue_head *head = poll->head;
5266 spin_lock_irq(&head->lock);
5267 list_del_init(&poll->wait.entry);
5268 if (poll->wait.private)
5271 spin_unlock_irq(&head->lock);
5275 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5276 __must_hold(&req->ctx->completion_lock)
5278 struct io_ring_ctx *ctx = req->ctx;
5279 unsigned flags = IORING_CQE_F_MORE;
5282 if (READ_ONCE(req->poll.canceled)) {
5284 req->poll.events |= EPOLLONESHOT;
5286 error = mangle_poll(mask);
5288 if (req->poll.events & EPOLLONESHOT)
5290 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5291 req->poll.done = true;
5294 if (flags & IORING_CQE_F_MORE)
5297 return !(flags & IORING_CQE_F_MORE);
5300 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5301 __must_hold(&req->ctx->completion_lock)
5305 done = __io_poll_complete(req, mask);
5306 io_commit_cqring(req->ctx);
5310 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5312 struct io_ring_ctx *ctx = req->ctx;
5313 struct io_kiocb *nxt;
5315 if (io_poll_rewait(req, &req->poll)) {
5316 spin_unlock(&ctx->completion_lock);
5320 done = __io_poll_complete(req, req->result);
5322 io_poll_remove_double(req);
5323 hash_del(&req->hash_node);
5326 add_wait_queue(req->poll.head, &req->poll.wait);
5328 io_commit_cqring(ctx);
5329 spin_unlock(&ctx->completion_lock);
5330 io_cqring_ev_posted(ctx);
5333 nxt = io_put_req_find_next(req);
5335 io_req_task_submit(nxt, locked);
5340 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5341 int sync, void *key)
5343 struct io_kiocb *req = wait->private;
5344 struct io_poll_iocb *poll = io_poll_get_single(req);
5345 __poll_t mask = key_to_poll(key);
5346 unsigned long flags;
5348 /* for instances that support it check for an event match first: */
5349 if (mask && !(mask & poll->events))
5351 if (!(poll->events & EPOLLONESHOT))
5352 return poll->wait.func(&poll->wait, mode, sync, key);
5354 list_del_init(&wait->entry);
5359 spin_lock_irqsave(&poll->head->lock, flags);
5360 done = list_empty(&poll->wait.entry);
5362 list_del_init(&poll->wait.entry);
5363 /* make sure double remove sees this as being gone */
5364 wait->private = NULL;
5365 spin_unlock_irqrestore(&poll->head->lock, flags);
5367 /* use wait func handler, so it matches the rq type */
5368 poll->wait.func(&poll->wait, mode, sync, key);
5375 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5376 wait_queue_func_t wake_func)
5380 poll->canceled = false;
5381 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5382 /* mask in events that we always want/need */
5383 poll->events = events | IO_POLL_UNMASK;
5384 INIT_LIST_HEAD(&poll->wait.entry);
5385 init_waitqueue_func_entry(&poll->wait, wake_func);
5388 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5389 struct wait_queue_head *head,
5390 struct io_poll_iocb **poll_ptr)
5392 struct io_kiocb *req = pt->req;
5395 * The file being polled uses multiple waitqueues for poll handling
5396 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5399 if (unlikely(pt->nr_entries)) {
5400 struct io_poll_iocb *poll_one = poll;
5402 /* double add on the same waitqueue head, ignore */
5403 if (poll_one->head == head)
5405 /* already have a 2nd entry, fail a third attempt */
5407 if ((*poll_ptr)->head == head)
5409 pt->error = -EINVAL;
5413 * Can't handle multishot for double wait for now, turn it
5414 * into one-shot mode.
5416 if (!(poll_one->events & EPOLLONESHOT))
5417 poll_one->events |= EPOLLONESHOT;
5418 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5420 pt->error = -ENOMEM;
5423 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5425 poll->wait.private = req;
5432 if (poll->events & EPOLLEXCLUSIVE)
5433 add_wait_queue_exclusive(head, &poll->wait);
5435 add_wait_queue(head, &poll->wait);
5438 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5439 struct poll_table_struct *p)
5441 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5442 struct async_poll *apoll = pt->req->apoll;
5444 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5447 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5449 struct async_poll *apoll = req->apoll;
5450 struct io_ring_ctx *ctx = req->ctx;
5452 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5454 if (io_poll_rewait(req, &apoll->poll)) {
5455 spin_unlock(&ctx->completion_lock);
5459 hash_del(&req->hash_node);
5460 io_poll_remove_double(req);
5461 spin_unlock(&ctx->completion_lock);
5463 if (!READ_ONCE(apoll->poll.canceled))
5464 io_req_task_submit(req, locked);
5466 io_req_complete_failed(req, -ECANCELED);
5469 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5472 struct io_kiocb *req = wait->private;
5473 struct io_poll_iocb *poll = &req->apoll->poll;
5475 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5478 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5481 static void io_poll_req_insert(struct io_kiocb *req)
5483 struct io_ring_ctx *ctx = req->ctx;
5484 struct hlist_head *list;
5486 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5487 hlist_add_head(&req->hash_node, list);
5490 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5491 struct io_poll_iocb *poll,
5492 struct io_poll_table *ipt, __poll_t mask,
5493 wait_queue_func_t wake_func)
5494 __acquires(&ctx->completion_lock)
5496 struct io_ring_ctx *ctx = req->ctx;
5497 bool cancel = false;
5499 INIT_HLIST_NODE(&req->hash_node);
5500 io_init_poll_iocb(poll, mask, wake_func);
5501 poll->file = req->file;
5502 poll->wait.private = req;
5504 ipt->pt._key = mask;
5507 ipt->nr_entries = 0;
5509 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5510 if (unlikely(!ipt->nr_entries) && !ipt->error)
5511 ipt->error = -EINVAL;
5513 spin_lock(&ctx->completion_lock);
5514 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5515 io_poll_remove_double(req);
5516 if (likely(poll->head)) {
5517 spin_lock_irq(&poll->head->lock);
5518 if (unlikely(list_empty(&poll->wait.entry))) {
5524 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5525 list_del_init(&poll->wait.entry);
5527 WRITE_ONCE(poll->canceled, true);
5528 else if (!poll->done) /* actually waiting for an event */
5529 io_poll_req_insert(req);
5530 spin_unlock_irq(&poll->head->lock);
5542 static int io_arm_poll_handler(struct io_kiocb *req)
5544 const struct io_op_def *def = &io_op_defs[req->opcode];
5545 struct io_ring_ctx *ctx = req->ctx;
5546 struct async_poll *apoll;
5547 struct io_poll_table ipt;
5548 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5551 if (!req->file || !file_can_poll(req->file))
5552 return IO_APOLL_ABORTED;
5553 if (req->flags & REQ_F_POLLED)
5554 return IO_APOLL_ABORTED;
5555 if (!def->pollin && !def->pollout)
5556 return IO_APOLL_ABORTED;
5560 mask |= POLLIN | POLLRDNORM;
5562 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5563 if ((req->opcode == IORING_OP_RECVMSG) &&
5564 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5568 mask |= POLLOUT | POLLWRNORM;
5571 /* if we can't nonblock try, then no point in arming a poll handler */
5572 if (!io_file_supports_nowait(req, rw))
5573 return IO_APOLL_ABORTED;
5575 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5576 if (unlikely(!apoll))
5577 return IO_APOLL_ABORTED;
5578 apoll->double_poll = NULL;
5580 req->flags |= REQ_F_POLLED;
5581 ipt.pt._qproc = io_async_queue_proc;
5582 io_req_set_refcount(req);
5584 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5586 spin_unlock(&ctx->completion_lock);
5587 if (ret || ipt.error)
5588 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5590 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5591 mask, apoll->poll.events);
5595 static bool __io_poll_remove_one(struct io_kiocb *req,
5596 struct io_poll_iocb *poll, bool do_cancel)
5597 __must_hold(&req->ctx->completion_lock)
5599 bool do_complete = false;
5603 spin_lock_irq(&poll->head->lock);
5605 WRITE_ONCE(poll->canceled, true);
5606 if (!list_empty(&poll->wait.entry)) {
5607 list_del_init(&poll->wait.entry);
5610 spin_unlock_irq(&poll->head->lock);
5611 hash_del(&req->hash_node);
5615 static bool io_poll_remove_one(struct io_kiocb *req)
5616 __must_hold(&req->ctx->completion_lock)
5620 io_poll_remove_double(req);
5621 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5624 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5625 io_commit_cqring(req->ctx);
5627 io_put_req_deferred(req);
5633 * Returns true if we found and killed one or more poll requests
5635 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5638 struct hlist_node *tmp;
5639 struct io_kiocb *req;
5642 spin_lock(&ctx->completion_lock);
5643 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5644 struct hlist_head *list;
5646 list = &ctx->cancel_hash[i];
5647 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5648 if (io_match_task(req, tsk, cancel_all))
5649 posted += io_poll_remove_one(req);
5652 spin_unlock(&ctx->completion_lock);
5655 io_cqring_ev_posted(ctx);
5660 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5662 __must_hold(&ctx->completion_lock)
5664 struct hlist_head *list;
5665 struct io_kiocb *req;
5667 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5668 hlist_for_each_entry(req, list, hash_node) {
5669 if (sqe_addr != req->user_data)
5671 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5678 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5680 __must_hold(&ctx->completion_lock)
5682 struct io_kiocb *req;
5684 req = io_poll_find(ctx, sqe_addr, poll_only);
5687 if (io_poll_remove_one(req))
5693 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5698 events = READ_ONCE(sqe->poll32_events);
5700 events = swahw32(events);
5702 if (!(flags & IORING_POLL_ADD_MULTI))
5703 events |= EPOLLONESHOT;
5704 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5707 static int io_poll_update_prep(struct io_kiocb *req,
5708 const struct io_uring_sqe *sqe)
5710 struct io_poll_update *upd = &req->poll_update;
5713 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5715 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5717 flags = READ_ONCE(sqe->len);
5718 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5719 IORING_POLL_ADD_MULTI))
5721 /* meaningless without update */
5722 if (flags == IORING_POLL_ADD_MULTI)
5725 upd->old_user_data = READ_ONCE(sqe->addr);
5726 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5727 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5729 upd->new_user_data = READ_ONCE(sqe->off);
5730 if (!upd->update_user_data && upd->new_user_data)
5732 if (upd->update_events)
5733 upd->events = io_poll_parse_events(sqe, flags);
5734 else if (sqe->poll32_events)
5740 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5743 struct io_kiocb *req = wait->private;
5744 struct io_poll_iocb *poll = &req->poll;
5746 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5749 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5750 struct poll_table_struct *p)
5752 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5754 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5757 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5759 struct io_poll_iocb *poll = &req->poll;
5762 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5764 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5766 flags = READ_ONCE(sqe->len);
5767 if (flags & ~IORING_POLL_ADD_MULTI)
5770 io_req_set_refcount(req);
5771 poll->events = io_poll_parse_events(sqe, flags);
5775 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5777 struct io_poll_iocb *poll = &req->poll;
5778 struct io_ring_ctx *ctx = req->ctx;
5779 struct io_poll_table ipt;
5782 ipt.pt._qproc = io_poll_queue_proc;
5784 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5787 if (mask) { /* no async, we'd stolen it */
5789 io_poll_complete(req, mask);
5791 spin_unlock(&ctx->completion_lock);
5794 io_cqring_ev_posted(ctx);
5795 if (poll->events & EPOLLONESHOT)
5801 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5803 struct io_ring_ctx *ctx = req->ctx;
5804 struct io_kiocb *preq;
5808 spin_lock(&ctx->completion_lock);
5809 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5815 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5817 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5822 * Don't allow racy completion with singleshot, as we cannot safely
5823 * update those. For multishot, if we're racing with completion, just
5824 * let completion re-add it.
5826 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5827 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5831 /* we now have a detached poll request. reissue. */
5835 spin_unlock(&ctx->completion_lock);
5837 io_req_complete(req, ret);
5840 /* only mask one event flags, keep behavior flags */
5841 if (req->poll_update.update_events) {
5842 preq->poll.events &= ~0xffff;
5843 preq->poll.events |= req->poll_update.events & 0xffff;
5844 preq->poll.events |= IO_POLL_UNMASK;
5846 if (req->poll_update.update_user_data)
5847 preq->user_data = req->poll_update.new_user_data;
5848 spin_unlock(&ctx->completion_lock);
5850 /* complete update request, we're done with it */
5851 io_req_complete(req, ret);
5854 ret = io_poll_add(preq, issue_flags);
5857 io_req_complete(preq, ret);
5863 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5866 io_req_complete_post(req, -ETIME, 0);
5869 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5871 struct io_timeout_data *data = container_of(timer,
5872 struct io_timeout_data, timer);
5873 struct io_kiocb *req = data->req;
5874 struct io_ring_ctx *ctx = req->ctx;
5875 unsigned long flags;
5877 spin_lock_irqsave(&ctx->timeout_lock, flags);
5878 list_del_init(&req->timeout.list);
5879 atomic_set(&req->ctx->cq_timeouts,
5880 atomic_read(&req->ctx->cq_timeouts) + 1);
5881 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5883 req->io_task_work.func = io_req_task_timeout;
5884 io_req_task_work_add(req);
5885 return HRTIMER_NORESTART;
5888 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5890 __must_hold(&ctx->timeout_lock)
5892 struct io_timeout_data *io;
5893 struct io_kiocb *req;
5896 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5897 found = user_data == req->user_data;
5902 return ERR_PTR(-ENOENT);
5904 io = req->async_data;
5905 if (hrtimer_try_to_cancel(&io->timer) == -1)
5906 return ERR_PTR(-EALREADY);
5907 list_del_init(&req->timeout.list);
5911 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5912 __must_hold(&ctx->completion_lock)
5913 __must_hold(&ctx->timeout_lock)
5915 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5918 return PTR_ERR(req);
5921 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5922 io_put_req_deferred(req);
5926 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5928 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5929 case IORING_TIMEOUT_BOOTTIME:
5930 return CLOCK_BOOTTIME;
5931 case IORING_TIMEOUT_REALTIME:
5932 return CLOCK_REALTIME;
5934 /* can't happen, vetted at prep time */
5938 return CLOCK_MONOTONIC;
5942 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5943 struct timespec64 *ts, enum hrtimer_mode mode)
5944 __must_hold(&ctx->timeout_lock)
5946 struct io_timeout_data *io;
5947 struct io_kiocb *req;
5950 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
5951 found = user_data == req->user_data;
5958 io = req->async_data;
5959 if (hrtimer_try_to_cancel(&io->timer) == -1)
5961 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
5962 io->timer.function = io_link_timeout_fn;
5963 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
5967 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5968 struct timespec64 *ts, enum hrtimer_mode mode)
5969 __must_hold(&ctx->timeout_lock)
5971 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5972 struct io_timeout_data *data;
5975 return PTR_ERR(req);
5977 req->timeout.off = 0; /* noseq */
5978 data = req->async_data;
5979 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5980 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
5981 data->timer.function = io_timeout_fn;
5982 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5986 static int io_timeout_remove_prep(struct io_kiocb *req,
5987 const struct io_uring_sqe *sqe)
5989 struct io_timeout_rem *tr = &req->timeout_rem;
5991 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5993 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5995 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
5998 tr->ltimeout = false;
5999 tr->addr = READ_ONCE(sqe->addr);
6000 tr->flags = READ_ONCE(sqe->timeout_flags);
6001 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6002 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6004 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6005 tr->ltimeout = true;
6006 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6008 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6010 } else if (tr->flags) {
6011 /* timeout removal doesn't support flags */
6018 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6020 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6025 * Remove or update an existing timeout command
6027 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6029 struct io_timeout_rem *tr = &req->timeout_rem;
6030 struct io_ring_ctx *ctx = req->ctx;
6033 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6034 spin_lock(&ctx->completion_lock);
6035 spin_lock_irq(&ctx->timeout_lock);
6036 ret = io_timeout_cancel(ctx, tr->addr);
6037 spin_unlock_irq(&ctx->timeout_lock);
6038 spin_unlock(&ctx->completion_lock);
6040 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6042 spin_lock_irq(&ctx->timeout_lock);
6044 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6046 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6047 spin_unlock_irq(&ctx->timeout_lock);
6052 io_req_complete_post(req, ret, 0);
6056 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6057 bool is_timeout_link)
6059 struct io_timeout_data *data;
6061 u32 off = READ_ONCE(sqe->off);
6063 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6065 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6068 if (off && is_timeout_link)
6070 flags = READ_ONCE(sqe->timeout_flags);
6071 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6073 /* more than one clock specified is invalid, obviously */
6074 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6077 INIT_LIST_HEAD(&req->timeout.list);
6078 req->timeout.off = off;
6079 if (unlikely(off && !req->ctx->off_timeout_used))
6080 req->ctx->off_timeout_used = true;
6082 if (!req->async_data && io_alloc_async_data(req))
6085 data = req->async_data;
6087 data->flags = flags;
6089 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6092 data->mode = io_translate_timeout_mode(flags);
6093 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6095 if (is_timeout_link) {
6096 struct io_submit_link *link = &req->ctx->submit_state.link;
6100 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6102 req->timeout.head = link->last;
6103 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6108 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6110 struct io_ring_ctx *ctx = req->ctx;
6111 struct io_timeout_data *data = req->async_data;
6112 struct list_head *entry;
6113 u32 tail, off = req->timeout.off;
6115 spin_lock_irq(&ctx->timeout_lock);
6118 * sqe->off holds how many events that need to occur for this
6119 * timeout event to be satisfied. If it isn't set, then this is
6120 * a pure timeout request, sequence isn't used.
6122 if (io_is_timeout_noseq(req)) {
6123 entry = ctx->timeout_list.prev;
6127 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6128 req->timeout.target_seq = tail + off;
6130 /* Update the last seq here in case io_flush_timeouts() hasn't.
6131 * This is safe because ->completion_lock is held, and submissions
6132 * and completions are never mixed in the same ->completion_lock section.
6134 ctx->cq_last_tm_flush = tail;
6137 * Insertion sort, ensuring the first entry in the list is always
6138 * the one we need first.
6140 list_for_each_prev(entry, &ctx->timeout_list) {
6141 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6144 if (io_is_timeout_noseq(nxt))
6146 /* nxt.seq is behind @tail, otherwise would've been completed */
6147 if (off >= nxt->timeout.target_seq - tail)
6151 list_add(&req->timeout.list, entry);
6152 data->timer.function = io_timeout_fn;
6153 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6154 spin_unlock_irq(&ctx->timeout_lock);
6158 struct io_cancel_data {
6159 struct io_ring_ctx *ctx;
6163 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6165 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6166 struct io_cancel_data *cd = data;
6168 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6171 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6172 struct io_ring_ctx *ctx)
6174 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6175 enum io_wq_cancel cancel_ret;
6178 if (!tctx || !tctx->io_wq)
6181 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6182 switch (cancel_ret) {
6183 case IO_WQ_CANCEL_OK:
6186 case IO_WQ_CANCEL_RUNNING:
6189 case IO_WQ_CANCEL_NOTFOUND:
6197 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6199 struct io_ring_ctx *ctx = req->ctx;
6202 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6204 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6208 spin_lock(&ctx->completion_lock);
6209 spin_lock_irq(&ctx->timeout_lock);
6210 ret = io_timeout_cancel(ctx, sqe_addr);
6211 spin_unlock_irq(&ctx->timeout_lock);
6214 ret = io_poll_cancel(ctx, sqe_addr, false);
6216 spin_unlock(&ctx->completion_lock);
6220 static int io_async_cancel_prep(struct io_kiocb *req,
6221 const struct io_uring_sqe *sqe)
6223 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6225 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6227 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6231 req->cancel.addr = READ_ONCE(sqe->addr);
6235 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6237 struct io_ring_ctx *ctx = req->ctx;
6238 u64 sqe_addr = req->cancel.addr;
6239 struct io_tctx_node *node;
6242 ret = io_try_cancel_userdata(req, sqe_addr);
6246 /* slow path, try all io-wq's */
6247 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6249 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6250 struct io_uring_task *tctx = node->task->io_uring;
6252 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6256 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6260 io_req_complete_post(req, ret, 0);
6264 static int io_rsrc_update_prep(struct io_kiocb *req,
6265 const struct io_uring_sqe *sqe)
6267 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6269 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6272 req->rsrc_update.offset = READ_ONCE(sqe->off);
6273 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6274 if (!req->rsrc_update.nr_args)
6276 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6280 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6282 struct io_ring_ctx *ctx = req->ctx;
6283 struct io_uring_rsrc_update2 up;
6286 if (issue_flags & IO_URING_F_NONBLOCK)
6289 up.offset = req->rsrc_update.offset;
6290 up.data = req->rsrc_update.arg;
6295 mutex_lock(&ctx->uring_lock);
6296 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6297 &up, req->rsrc_update.nr_args);
6298 mutex_unlock(&ctx->uring_lock);
6302 __io_req_complete(req, issue_flags, ret, 0);
6306 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6308 switch (req->opcode) {
6311 case IORING_OP_READV:
6312 case IORING_OP_READ_FIXED:
6313 case IORING_OP_READ:
6314 return io_read_prep(req, sqe);
6315 case IORING_OP_WRITEV:
6316 case IORING_OP_WRITE_FIXED:
6317 case IORING_OP_WRITE:
6318 return io_write_prep(req, sqe);
6319 case IORING_OP_POLL_ADD:
6320 return io_poll_add_prep(req, sqe);
6321 case IORING_OP_POLL_REMOVE:
6322 return io_poll_update_prep(req, sqe);
6323 case IORING_OP_FSYNC:
6324 return io_fsync_prep(req, sqe);
6325 case IORING_OP_SYNC_FILE_RANGE:
6326 return io_sfr_prep(req, sqe);
6327 case IORING_OP_SENDMSG:
6328 case IORING_OP_SEND:
6329 return io_sendmsg_prep(req, sqe);
6330 case IORING_OP_RECVMSG:
6331 case IORING_OP_RECV:
6332 return io_recvmsg_prep(req, sqe);
6333 case IORING_OP_CONNECT:
6334 return io_connect_prep(req, sqe);
6335 case IORING_OP_TIMEOUT:
6336 return io_timeout_prep(req, sqe, false);
6337 case IORING_OP_TIMEOUT_REMOVE:
6338 return io_timeout_remove_prep(req, sqe);
6339 case IORING_OP_ASYNC_CANCEL:
6340 return io_async_cancel_prep(req, sqe);
6341 case IORING_OP_LINK_TIMEOUT:
6342 return io_timeout_prep(req, sqe, true);
6343 case IORING_OP_ACCEPT:
6344 return io_accept_prep(req, sqe);
6345 case IORING_OP_FALLOCATE:
6346 return io_fallocate_prep(req, sqe);
6347 case IORING_OP_OPENAT:
6348 return io_openat_prep(req, sqe);
6349 case IORING_OP_CLOSE:
6350 return io_close_prep(req, sqe);
6351 case IORING_OP_FILES_UPDATE:
6352 return io_rsrc_update_prep(req, sqe);
6353 case IORING_OP_STATX:
6354 return io_statx_prep(req, sqe);
6355 case IORING_OP_FADVISE:
6356 return io_fadvise_prep(req, sqe);
6357 case IORING_OP_MADVISE:
6358 return io_madvise_prep(req, sqe);
6359 case IORING_OP_OPENAT2:
6360 return io_openat2_prep(req, sqe);
6361 case IORING_OP_EPOLL_CTL:
6362 return io_epoll_ctl_prep(req, sqe);
6363 case IORING_OP_SPLICE:
6364 return io_splice_prep(req, sqe);
6365 case IORING_OP_PROVIDE_BUFFERS:
6366 return io_provide_buffers_prep(req, sqe);
6367 case IORING_OP_REMOVE_BUFFERS:
6368 return io_remove_buffers_prep(req, sqe);
6370 return io_tee_prep(req, sqe);
6371 case IORING_OP_SHUTDOWN:
6372 return io_shutdown_prep(req, sqe);
6373 case IORING_OP_RENAMEAT:
6374 return io_renameat_prep(req, sqe);
6375 case IORING_OP_UNLINKAT:
6376 return io_unlinkat_prep(req, sqe);
6377 case IORING_OP_MKDIRAT:
6378 return io_mkdirat_prep(req, sqe);
6379 case IORING_OP_SYMLINKAT:
6380 return io_symlinkat_prep(req, sqe);
6381 case IORING_OP_LINKAT:
6382 return io_linkat_prep(req, sqe);
6385 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6390 static int io_req_prep_async(struct io_kiocb *req)
6392 if (!io_op_defs[req->opcode].needs_async_setup)
6394 if (WARN_ON_ONCE(req->async_data))
6396 if (io_alloc_async_data(req))
6399 switch (req->opcode) {
6400 case IORING_OP_READV:
6401 return io_rw_prep_async(req, READ);
6402 case IORING_OP_WRITEV:
6403 return io_rw_prep_async(req, WRITE);
6404 case IORING_OP_SENDMSG:
6405 return io_sendmsg_prep_async(req);
6406 case IORING_OP_RECVMSG:
6407 return io_recvmsg_prep_async(req);
6408 case IORING_OP_CONNECT:
6409 return io_connect_prep_async(req);
6411 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6416 static u32 io_get_sequence(struct io_kiocb *req)
6418 u32 seq = req->ctx->cached_sq_head;
6420 /* need original cached_sq_head, but it was increased for each req */
6421 io_for_each_link(req, req)
6426 static bool io_drain_req(struct io_kiocb *req)
6428 struct io_kiocb *pos;
6429 struct io_ring_ctx *ctx = req->ctx;
6430 struct io_defer_entry *de;
6434 if (req->flags & REQ_F_FAIL) {
6435 io_req_complete_fail_submit(req);
6440 * If we need to drain a request in the middle of a link, drain the
6441 * head request and the next request/link after the current link.
6442 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6443 * maintained for every request of our link.
6445 if (ctx->drain_next) {
6446 req->flags |= REQ_F_IO_DRAIN;
6447 ctx->drain_next = false;
6449 /* not interested in head, start from the first linked */
6450 io_for_each_link(pos, req->link) {
6451 if (pos->flags & REQ_F_IO_DRAIN) {
6452 ctx->drain_next = true;
6453 req->flags |= REQ_F_IO_DRAIN;
6458 /* Still need defer if there is pending req in defer list. */
6459 if (likely(list_empty_careful(&ctx->defer_list) &&
6460 !(req->flags & REQ_F_IO_DRAIN))) {
6461 ctx->drain_active = false;
6465 seq = io_get_sequence(req);
6466 /* Still a chance to pass the sequence check */
6467 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6470 ret = io_req_prep_async(req);
6473 io_prep_async_link(req);
6474 de = kmalloc(sizeof(*de), GFP_KERNEL);
6478 io_req_complete_failed(req, ret);
6482 spin_lock(&ctx->completion_lock);
6483 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6484 spin_unlock(&ctx->completion_lock);
6486 io_queue_async_work(req, NULL);
6490 trace_io_uring_defer(ctx, req, req->user_data);
6493 list_add_tail(&de->list, &ctx->defer_list);
6494 spin_unlock(&ctx->completion_lock);
6498 static void io_clean_op(struct io_kiocb *req)
6500 if (req->flags & REQ_F_BUFFER_SELECTED) {
6501 switch (req->opcode) {
6502 case IORING_OP_READV:
6503 case IORING_OP_READ_FIXED:
6504 case IORING_OP_READ:
6505 kfree((void *)(unsigned long)req->rw.addr);
6507 case IORING_OP_RECVMSG:
6508 case IORING_OP_RECV:
6509 kfree(req->sr_msg.kbuf);
6514 if (req->flags & REQ_F_NEED_CLEANUP) {
6515 switch (req->opcode) {
6516 case IORING_OP_READV:
6517 case IORING_OP_READ_FIXED:
6518 case IORING_OP_READ:
6519 case IORING_OP_WRITEV:
6520 case IORING_OP_WRITE_FIXED:
6521 case IORING_OP_WRITE: {
6522 struct io_async_rw *io = req->async_data;
6524 kfree(io->free_iovec);
6527 case IORING_OP_RECVMSG:
6528 case IORING_OP_SENDMSG: {
6529 struct io_async_msghdr *io = req->async_data;
6531 kfree(io->free_iov);
6534 case IORING_OP_SPLICE:
6536 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6537 io_put_file(req->splice.file_in);
6539 case IORING_OP_OPENAT:
6540 case IORING_OP_OPENAT2:
6541 if (req->open.filename)
6542 putname(req->open.filename);
6544 case IORING_OP_RENAMEAT:
6545 putname(req->rename.oldpath);
6546 putname(req->rename.newpath);
6548 case IORING_OP_UNLINKAT:
6549 putname(req->unlink.filename);
6551 case IORING_OP_MKDIRAT:
6552 putname(req->mkdir.filename);
6554 case IORING_OP_SYMLINKAT:
6555 putname(req->symlink.oldpath);
6556 putname(req->symlink.newpath);
6558 case IORING_OP_LINKAT:
6559 putname(req->hardlink.oldpath);
6560 putname(req->hardlink.newpath);
6564 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6565 kfree(req->apoll->double_poll);
6569 if (req->flags & REQ_F_INFLIGHT) {
6570 struct io_uring_task *tctx = req->task->io_uring;
6572 atomic_dec(&tctx->inflight_tracked);
6574 if (req->flags & REQ_F_CREDS)
6575 put_cred(req->creds);
6577 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6580 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6582 struct io_ring_ctx *ctx = req->ctx;
6583 const struct cred *creds = NULL;
6586 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6587 creds = override_creds(req->creds);
6589 switch (req->opcode) {
6591 ret = io_nop(req, issue_flags);
6593 case IORING_OP_READV:
6594 case IORING_OP_READ_FIXED:
6595 case IORING_OP_READ:
6596 ret = io_read(req, issue_flags);
6598 case IORING_OP_WRITEV:
6599 case IORING_OP_WRITE_FIXED:
6600 case IORING_OP_WRITE:
6601 ret = io_write(req, issue_flags);
6603 case IORING_OP_FSYNC:
6604 ret = io_fsync(req, issue_flags);
6606 case IORING_OP_POLL_ADD:
6607 ret = io_poll_add(req, issue_flags);
6609 case IORING_OP_POLL_REMOVE:
6610 ret = io_poll_update(req, issue_flags);
6612 case IORING_OP_SYNC_FILE_RANGE:
6613 ret = io_sync_file_range(req, issue_flags);
6615 case IORING_OP_SENDMSG:
6616 ret = io_sendmsg(req, issue_flags);
6618 case IORING_OP_SEND:
6619 ret = io_send(req, issue_flags);
6621 case IORING_OP_RECVMSG:
6622 ret = io_recvmsg(req, issue_flags);
6624 case IORING_OP_RECV:
6625 ret = io_recv(req, issue_flags);
6627 case IORING_OP_TIMEOUT:
6628 ret = io_timeout(req, issue_flags);
6630 case IORING_OP_TIMEOUT_REMOVE:
6631 ret = io_timeout_remove(req, issue_flags);
6633 case IORING_OP_ACCEPT:
6634 ret = io_accept(req, issue_flags);
6636 case IORING_OP_CONNECT:
6637 ret = io_connect(req, issue_flags);
6639 case IORING_OP_ASYNC_CANCEL:
6640 ret = io_async_cancel(req, issue_flags);
6642 case IORING_OP_FALLOCATE:
6643 ret = io_fallocate(req, issue_flags);
6645 case IORING_OP_OPENAT:
6646 ret = io_openat(req, issue_flags);
6648 case IORING_OP_CLOSE:
6649 ret = io_close(req, issue_flags);
6651 case IORING_OP_FILES_UPDATE:
6652 ret = io_files_update(req, issue_flags);
6654 case IORING_OP_STATX:
6655 ret = io_statx(req, issue_flags);
6657 case IORING_OP_FADVISE:
6658 ret = io_fadvise(req, issue_flags);
6660 case IORING_OP_MADVISE:
6661 ret = io_madvise(req, issue_flags);
6663 case IORING_OP_OPENAT2:
6664 ret = io_openat2(req, issue_flags);
6666 case IORING_OP_EPOLL_CTL:
6667 ret = io_epoll_ctl(req, issue_flags);
6669 case IORING_OP_SPLICE:
6670 ret = io_splice(req, issue_flags);
6672 case IORING_OP_PROVIDE_BUFFERS:
6673 ret = io_provide_buffers(req, issue_flags);
6675 case IORING_OP_REMOVE_BUFFERS:
6676 ret = io_remove_buffers(req, issue_flags);
6679 ret = io_tee(req, issue_flags);
6681 case IORING_OP_SHUTDOWN:
6682 ret = io_shutdown(req, issue_flags);
6684 case IORING_OP_RENAMEAT:
6685 ret = io_renameat(req, issue_flags);
6687 case IORING_OP_UNLINKAT:
6688 ret = io_unlinkat(req, issue_flags);
6690 case IORING_OP_MKDIRAT:
6691 ret = io_mkdirat(req, issue_flags);
6693 case IORING_OP_SYMLINKAT:
6694 ret = io_symlinkat(req, issue_flags);
6696 case IORING_OP_LINKAT:
6697 ret = io_linkat(req, issue_flags);
6705 revert_creds(creds);
6708 /* If the op doesn't have a file, we're not polling for it */
6709 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6710 io_iopoll_req_issued(req);
6715 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6717 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6719 req = io_put_req_find_next(req);
6720 return req ? &req->work : NULL;
6723 static void io_wq_submit_work(struct io_wq_work *work)
6725 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6726 struct io_kiocb *timeout;
6729 /* one will be dropped by ->io_free_work() after returning to io-wq */
6730 if (!(req->flags & REQ_F_REFCOUNT))
6731 __io_req_set_refcount(req, 2);
6735 timeout = io_prep_linked_timeout(req);
6737 io_queue_linked_timeout(timeout);
6739 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6740 if (work->flags & IO_WQ_WORK_CANCEL)
6745 ret = io_issue_sqe(req, 0);
6747 * We can get EAGAIN for polled IO even though we're
6748 * forcing a sync submission from here, since we can't
6749 * wait for request slots on the block side.
6757 /* avoid locking problems by failing it from a clean context */
6759 io_req_task_queue_fail(req, ret);
6762 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6765 return &table->files[i];
6768 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6771 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6773 return (struct file *) (slot->file_ptr & FFS_MASK);
6776 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6778 unsigned long file_ptr = (unsigned long) file;
6780 if (__io_file_supports_nowait(file, READ))
6781 file_ptr |= FFS_ASYNC_READ;
6782 if (__io_file_supports_nowait(file, WRITE))
6783 file_ptr |= FFS_ASYNC_WRITE;
6784 if (S_ISREG(file_inode(file)->i_mode))
6785 file_ptr |= FFS_ISREG;
6786 file_slot->file_ptr = file_ptr;
6789 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6790 struct io_kiocb *req, int fd)
6793 unsigned long file_ptr;
6795 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6797 fd = array_index_nospec(fd, ctx->nr_user_files);
6798 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6799 file = (struct file *) (file_ptr & FFS_MASK);
6800 file_ptr &= ~FFS_MASK;
6801 /* mask in overlapping REQ_F and FFS bits */
6802 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6803 io_req_set_rsrc_node(req);
6807 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6808 struct io_kiocb *req, int fd)
6810 struct file *file = fget(fd);
6812 trace_io_uring_file_get(ctx, fd);
6814 /* we don't allow fixed io_uring files */
6815 if (file && unlikely(file->f_op == &io_uring_fops))
6816 io_req_track_inflight(req);
6820 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6821 struct io_kiocb *req, int fd, bool fixed)
6824 return io_file_get_fixed(ctx, req, fd);
6826 return io_file_get_normal(ctx, req, fd);
6829 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6831 struct io_kiocb *prev = req->timeout.prev;
6835 ret = io_try_cancel_userdata(req, prev->user_data);
6836 io_req_complete_post(req, ret ?: -ETIME, 0);
6839 io_req_complete_post(req, -ETIME, 0);
6843 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6845 struct io_timeout_data *data = container_of(timer,
6846 struct io_timeout_data, timer);
6847 struct io_kiocb *prev, *req = data->req;
6848 struct io_ring_ctx *ctx = req->ctx;
6849 unsigned long flags;
6851 spin_lock_irqsave(&ctx->timeout_lock, flags);
6852 prev = req->timeout.head;
6853 req->timeout.head = NULL;
6856 * We don't expect the list to be empty, that will only happen if we
6857 * race with the completion of the linked work.
6860 io_remove_next_linked(prev);
6861 if (!req_ref_inc_not_zero(prev))
6864 list_del(&req->timeout.list);
6865 req->timeout.prev = prev;
6866 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6868 req->io_task_work.func = io_req_task_link_timeout;
6869 io_req_task_work_add(req);
6870 return HRTIMER_NORESTART;
6873 static void io_queue_linked_timeout(struct io_kiocb *req)
6875 struct io_ring_ctx *ctx = req->ctx;
6877 spin_lock_irq(&ctx->timeout_lock);
6879 * If the back reference is NULL, then our linked request finished
6880 * before we got a chance to setup the timer
6882 if (req->timeout.head) {
6883 struct io_timeout_data *data = req->async_data;
6885 data->timer.function = io_link_timeout_fn;
6886 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6888 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6890 spin_unlock_irq(&ctx->timeout_lock);
6891 /* drop submission reference */
6895 static void __io_queue_sqe(struct io_kiocb *req)
6896 __must_hold(&req->ctx->uring_lock)
6898 struct io_kiocb *linked_timeout;
6902 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6905 * We async punt it if the file wasn't marked NOWAIT, or if the file
6906 * doesn't support non-blocking read/write attempts
6909 if (req->flags & REQ_F_COMPLETE_INLINE) {
6910 struct io_ring_ctx *ctx = req->ctx;
6911 struct io_submit_state *state = &ctx->submit_state;
6913 state->compl_reqs[state->compl_nr++] = req;
6914 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6915 io_submit_flush_completions(ctx);
6919 linked_timeout = io_prep_linked_timeout(req);
6921 io_queue_linked_timeout(linked_timeout);
6922 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6923 linked_timeout = io_prep_linked_timeout(req);
6925 switch (io_arm_poll_handler(req)) {
6926 case IO_APOLL_READY:
6928 io_unprep_linked_timeout(req);
6930 case IO_APOLL_ABORTED:
6932 * Queued up for async execution, worker will release
6933 * submit reference when the iocb is actually submitted.
6935 io_queue_async_work(req, NULL);
6940 io_queue_linked_timeout(linked_timeout);
6942 io_req_complete_failed(req, ret);
6946 static inline void io_queue_sqe(struct io_kiocb *req)
6947 __must_hold(&req->ctx->uring_lock)
6949 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6952 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
6953 __io_queue_sqe(req);
6954 } else if (req->flags & REQ_F_FAIL) {
6955 io_req_complete_fail_submit(req);
6957 int ret = io_req_prep_async(req);
6960 io_req_complete_failed(req, ret);
6962 io_queue_async_work(req, NULL);
6967 * Check SQE restrictions (opcode and flags).
6969 * Returns 'true' if SQE is allowed, 'false' otherwise.
6971 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6972 struct io_kiocb *req,
6973 unsigned int sqe_flags)
6975 if (likely(!ctx->restricted))
6978 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6981 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6982 ctx->restrictions.sqe_flags_required)
6985 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6986 ctx->restrictions.sqe_flags_required))
6992 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6993 const struct io_uring_sqe *sqe)
6994 __must_hold(&ctx->uring_lock)
6996 struct io_submit_state *state;
6997 unsigned int sqe_flags;
6998 int personality, ret = 0;
7000 /* req is partially pre-initialised, see io_preinit_req() */
7001 req->opcode = READ_ONCE(sqe->opcode);
7002 /* same numerical values with corresponding REQ_F_*, safe to copy */
7003 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7004 req->user_data = READ_ONCE(sqe->user_data);
7006 req->fixed_rsrc_refs = NULL;
7007 req->task = current;
7009 /* enforce forwards compatibility on users */
7010 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7012 if (unlikely(req->opcode >= IORING_OP_LAST))
7014 if (!io_check_restriction(ctx, req, sqe_flags))
7017 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7018 !io_op_defs[req->opcode].buffer_select)
7020 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7021 ctx->drain_active = true;
7023 personality = READ_ONCE(sqe->personality);
7025 req->creds = xa_load(&ctx->personalities, personality);
7028 get_cred(req->creds);
7029 req->flags |= REQ_F_CREDS;
7031 state = &ctx->submit_state;
7034 * Plug now if we have more than 1 IO left after this, and the target
7035 * is potentially a read/write to block based storage.
7037 if (!state->plug_started && state->ios_left > 1 &&
7038 io_op_defs[req->opcode].plug) {
7039 blk_start_plug(&state->plug);
7040 state->plug_started = true;
7043 if (io_op_defs[req->opcode].needs_file) {
7044 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7045 (sqe_flags & IOSQE_FIXED_FILE));
7046 if (unlikely(!req->file))
7054 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7055 const struct io_uring_sqe *sqe)
7056 __must_hold(&ctx->uring_lock)
7058 struct io_submit_link *link = &ctx->submit_state.link;
7061 ret = io_init_req(ctx, req, sqe);
7062 if (unlikely(ret)) {
7064 /* fail even hard links since we don't submit */
7067 * we can judge a link req is failed or cancelled by if
7068 * REQ_F_FAIL is set, but the head is an exception since
7069 * it may be set REQ_F_FAIL because of other req's failure
7070 * so let's leverage req->result to distinguish if a head
7071 * is set REQ_F_FAIL because of its failure or other req's
7072 * failure so that we can set the correct ret code for it.
7073 * init result here to avoid affecting the normal path.
7075 if (!(link->head->flags & REQ_F_FAIL))
7076 req_fail_link_node(link->head, -ECANCELED);
7077 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7079 * the current req is a normal req, we should return
7080 * error and thus break the submittion loop.
7082 io_req_complete_failed(req, ret);
7085 req_fail_link_node(req, ret);
7087 ret = io_req_prep(req, sqe);
7092 /* don't need @sqe from now on */
7093 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7095 ctx->flags & IORING_SETUP_SQPOLL);
7098 * If we already have a head request, queue this one for async
7099 * submittal once the head completes. If we don't have a head but
7100 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7101 * submitted sync once the chain is complete. If none of those
7102 * conditions are true (normal request), then just queue it.
7105 struct io_kiocb *head = link->head;
7107 if (!(req->flags & REQ_F_FAIL)) {
7108 ret = io_req_prep_async(req);
7109 if (unlikely(ret)) {
7110 req_fail_link_node(req, ret);
7111 if (!(head->flags & REQ_F_FAIL))
7112 req_fail_link_node(head, -ECANCELED);
7115 trace_io_uring_link(ctx, req, head);
7116 link->last->link = req;
7119 /* last request of a link, enqueue the link */
7120 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7125 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7137 * Batched submission is done, ensure local IO is flushed out.
7139 static void io_submit_state_end(struct io_submit_state *state,
7140 struct io_ring_ctx *ctx)
7142 if (state->link.head)
7143 io_queue_sqe(state->link.head);
7144 if (state->compl_nr)
7145 io_submit_flush_completions(ctx);
7146 if (state->plug_started)
7147 blk_finish_plug(&state->plug);
7151 * Start submission side cache.
7153 static void io_submit_state_start(struct io_submit_state *state,
7154 unsigned int max_ios)
7156 state->plug_started = false;
7157 state->ios_left = max_ios;
7158 /* set only head, no need to init link_last in advance */
7159 state->link.head = NULL;
7162 static void io_commit_sqring(struct io_ring_ctx *ctx)
7164 struct io_rings *rings = ctx->rings;
7167 * Ensure any loads from the SQEs are done at this point,
7168 * since once we write the new head, the application could
7169 * write new data to them.
7171 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7175 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7176 * that is mapped by userspace. This means that care needs to be taken to
7177 * ensure that reads are stable, as we cannot rely on userspace always
7178 * being a good citizen. If members of the sqe are validated and then later
7179 * used, it's important that those reads are done through READ_ONCE() to
7180 * prevent a re-load down the line.
7182 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7184 unsigned head, mask = ctx->sq_entries - 1;
7185 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7188 * The cached sq head (or cq tail) serves two purposes:
7190 * 1) allows us to batch the cost of updating the user visible
7192 * 2) allows the kernel side to track the head on its own, even
7193 * though the application is the one updating it.
7195 head = READ_ONCE(ctx->sq_array[sq_idx]);
7196 if (likely(head < ctx->sq_entries))
7197 return &ctx->sq_sqes[head];
7199 /* drop invalid entries */
7201 WRITE_ONCE(ctx->rings->sq_dropped,
7202 READ_ONCE(ctx->rings->sq_dropped) + 1);
7206 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7207 __must_hold(&ctx->uring_lock)
7211 /* make sure SQ entry isn't read before tail */
7212 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7213 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7215 io_get_task_refs(nr);
7217 io_submit_state_start(&ctx->submit_state, nr);
7218 while (submitted < nr) {
7219 const struct io_uring_sqe *sqe;
7220 struct io_kiocb *req;
7222 req = io_alloc_req(ctx);
7223 if (unlikely(!req)) {
7225 submitted = -EAGAIN;
7228 sqe = io_get_sqe(ctx);
7229 if (unlikely(!sqe)) {
7230 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7233 /* will complete beyond this point, count as submitted */
7235 if (io_submit_sqe(ctx, req, sqe))
7239 if (unlikely(submitted != nr)) {
7240 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7241 int unused = nr - ref_used;
7243 current->io_uring->cached_refs += unused;
7244 percpu_ref_put_many(&ctx->refs, unused);
7247 io_submit_state_end(&ctx->submit_state, ctx);
7248 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7249 io_commit_sqring(ctx);
7254 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7256 return READ_ONCE(sqd->state);
7259 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7261 /* Tell userspace we may need a wakeup call */
7262 spin_lock(&ctx->completion_lock);
7263 WRITE_ONCE(ctx->rings->sq_flags,
7264 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7265 spin_unlock(&ctx->completion_lock);
7268 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7270 spin_lock(&ctx->completion_lock);
7271 WRITE_ONCE(ctx->rings->sq_flags,
7272 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7273 spin_unlock(&ctx->completion_lock);
7276 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7278 unsigned int to_submit;
7281 to_submit = io_sqring_entries(ctx);
7282 /* if we're handling multiple rings, cap submit size for fairness */
7283 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7284 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7286 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7287 unsigned nr_events = 0;
7288 const struct cred *creds = NULL;
7290 if (ctx->sq_creds != current_cred())
7291 creds = override_creds(ctx->sq_creds);
7293 mutex_lock(&ctx->uring_lock);
7294 if (!list_empty(&ctx->iopoll_list))
7295 io_do_iopoll(ctx, &nr_events, 0);
7298 * Don't submit if refs are dying, good for io_uring_register(),
7299 * but also it is relied upon by io_ring_exit_work()
7301 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7302 !(ctx->flags & IORING_SETUP_R_DISABLED))
7303 ret = io_submit_sqes(ctx, to_submit);
7304 mutex_unlock(&ctx->uring_lock);
7306 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7307 wake_up(&ctx->sqo_sq_wait);
7309 revert_creds(creds);
7315 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7317 struct io_ring_ctx *ctx;
7318 unsigned sq_thread_idle = 0;
7320 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7321 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7322 sqd->sq_thread_idle = sq_thread_idle;
7325 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7327 bool did_sig = false;
7328 struct ksignal ksig;
7330 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7331 signal_pending(current)) {
7332 mutex_unlock(&sqd->lock);
7333 if (signal_pending(current))
7334 did_sig = get_signal(&ksig);
7336 mutex_lock(&sqd->lock);
7338 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7341 static int io_sq_thread(void *data)
7343 struct io_sq_data *sqd = data;
7344 struct io_ring_ctx *ctx;
7345 unsigned long timeout = 0;
7346 char buf[TASK_COMM_LEN];
7349 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7350 set_task_comm(current, buf);
7352 if (sqd->sq_cpu != -1)
7353 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7355 set_cpus_allowed_ptr(current, cpu_online_mask);
7356 current->flags |= PF_NO_SETAFFINITY;
7358 mutex_lock(&sqd->lock);
7360 bool cap_entries, sqt_spin = false;
7362 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7363 if (io_sqd_handle_event(sqd))
7365 timeout = jiffies + sqd->sq_thread_idle;
7368 cap_entries = !list_is_singular(&sqd->ctx_list);
7369 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7370 int ret = __io_sq_thread(ctx, cap_entries);
7372 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7375 if (io_run_task_work())
7378 if (sqt_spin || !time_after(jiffies, timeout)) {
7381 timeout = jiffies + sqd->sq_thread_idle;
7385 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7386 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7387 bool needs_sched = true;
7389 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7390 io_ring_set_wakeup_flag(ctx);
7392 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7393 !list_empty_careful(&ctx->iopoll_list)) {
7394 needs_sched = false;
7397 if (io_sqring_entries(ctx)) {
7398 needs_sched = false;
7404 mutex_unlock(&sqd->lock);
7406 mutex_lock(&sqd->lock);
7408 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7409 io_ring_clear_wakeup_flag(ctx);
7412 finish_wait(&sqd->wait, &wait);
7413 timeout = jiffies + sqd->sq_thread_idle;
7416 io_uring_cancel_generic(true, sqd);
7418 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7419 io_ring_set_wakeup_flag(ctx);
7421 mutex_unlock(&sqd->lock);
7423 complete(&sqd->exited);
7427 struct io_wait_queue {
7428 struct wait_queue_entry wq;
7429 struct io_ring_ctx *ctx;
7431 unsigned nr_timeouts;
7434 static inline bool io_should_wake(struct io_wait_queue *iowq)
7436 struct io_ring_ctx *ctx = iowq->ctx;
7437 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7440 * Wake up if we have enough events, or if a timeout occurred since we
7441 * started waiting. For timeouts, we always want to return to userspace,
7442 * regardless of event count.
7444 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7447 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7448 int wake_flags, void *key)
7450 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7454 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7455 * the task, and the next invocation will do it.
7457 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7458 return autoremove_wake_function(curr, mode, wake_flags, key);
7462 static int io_run_task_work_sig(void)
7464 if (io_run_task_work())
7466 if (!signal_pending(current))
7468 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7469 return -ERESTARTSYS;
7473 /* when returns >0, the caller should retry */
7474 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7475 struct io_wait_queue *iowq,
7476 signed long *timeout)
7480 /* make sure we run task_work before checking for signals */
7481 ret = io_run_task_work_sig();
7482 if (ret || io_should_wake(iowq))
7484 /* let the caller flush overflows, retry */
7485 if (test_bit(0, &ctx->check_cq_overflow))
7488 *timeout = schedule_timeout(*timeout);
7489 return !*timeout ? -ETIME : 1;
7493 * Wait until events become available, if we don't already have some. The
7494 * application must reap them itself, as they reside on the shared cq ring.
7496 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7497 const sigset_t __user *sig, size_t sigsz,
7498 struct __kernel_timespec __user *uts)
7500 struct io_wait_queue iowq;
7501 struct io_rings *rings = ctx->rings;
7502 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7506 io_cqring_overflow_flush(ctx);
7507 if (io_cqring_events(ctx) >= min_events)
7509 if (!io_run_task_work())
7514 #ifdef CONFIG_COMPAT
7515 if (in_compat_syscall())
7516 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7520 ret = set_user_sigmask(sig, sigsz);
7527 struct timespec64 ts;
7529 if (get_timespec64(&ts, uts))
7531 timeout = timespec64_to_jiffies(&ts);
7534 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7535 iowq.wq.private = current;
7536 INIT_LIST_HEAD(&iowq.wq.entry);
7538 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7539 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7541 trace_io_uring_cqring_wait(ctx, min_events);
7543 /* if we can't even flush overflow, don't wait for more */
7544 if (!io_cqring_overflow_flush(ctx)) {
7548 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7549 TASK_INTERRUPTIBLE);
7550 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7551 finish_wait(&ctx->cq_wait, &iowq.wq);
7555 restore_saved_sigmask_unless(ret == -EINTR);
7557 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7560 static void io_free_page_table(void **table, size_t size)
7562 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7564 for (i = 0; i < nr_tables; i++)
7569 static void **io_alloc_page_table(size_t size)
7571 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7572 size_t init_size = size;
7575 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7579 for (i = 0; i < nr_tables; i++) {
7580 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7582 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7584 io_free_page_table(table, init_size);
7592 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7594 percpu_ref_exit(&ref_node->refs);
7598 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7600 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7601 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7602 unsigned long flags;
7603 bool first_add = false;
7605 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7608 while (!list_empty(&ctx->rsrc_ref_list)) {
7609 node = list_first_entry(&ctx->rsrc_ref_list,
7610 struct io_rsrc_node, node);
7611 /* recycle ref nodes in order */
7614 list_del(&node->node);
7615 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7617 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7620 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7623 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7625 struct io_rsrc_node *ref_node;
7627 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7631 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7636 INIT_LIST_HEAD(&ref_node->node);
7637 INIT_LIST_HEAD(&ref_node->rsrc_list);
7638 ref_node->done = false;
7642 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7643 struct io_rsrc_data *data_to_kill)
7645 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7646 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7649 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7651 rsrc_node->rsrc_data = data_to_kill;
7652 spin_lock_irq(&ctx->rsrc_ref_lock);
7653 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7654 spin_unlock_irq(&ctx->rsrc_ref_lock);
7656 atomic_inc(&data_to_kill->refs);
7657 percpu_ref_kill(&rsrc_node->refs);
7658 ctx->rsrc_node = NULL;
7661 if (!ctx->rsrc_node) {
7662 ctx->rsrc_node = ctx->rsrc_backup_node;
7663 ctx->rsrc_backup_node = NULL;
7667 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7669 if (ctx->rsrc_backup_node)
7671 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7672 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7675 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7679 /* As we may drop ->uring_lock, other task may have started quiesce */
7683 data->quiesce = true;
7685 ret = io_rsrc_node_switch_start(ctx);
7688 io_rsrc_node_switch(ctx, data);
7690 /* kill initial ref, already quiesced if zero */
7691 if (atomic_dec_and_test(&data->refs))
7693 mutex_unlock(&ctx->uring_lock);
7694 flush_delayed_work(&ctx->rsrc_put_work);
7695 ret = wait_for_completion_interruptible(&data->done);
7697 mutex_lock(&ctx->uring_lock);
7701 atomic_inc(&data->refs);
7702 /* wait for all works potentially completing data->done */
7703 flush_delayed_work(&ctx->rsrc_put_work);
7704 reinit_completion(&data->done);
7706 ret = io_run_task_work_sig();
7707 mutex_lock(&ctx->uring_lock);
7709 data->quiesce = false;
7714 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7716 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7717 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7719 return &data->tags[table_idx][off];
7722 static void io_rsrc_data_free(struct io_rsrc_data *data)
7724 size_t size = data->nr * sizeof(data->tags[0][0]);
7727 io_free_page_table((void **)data->tags, size);
7731 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7732 u64 __user *utags, unsigned nr,
7733 struct io_rsrc_data **pdata)
7735 struct io_rsrc_data *data;
7739 data = kzalloc(sizeof(*data), GFP_KERNEL);
7742 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7750 data->do_put = do_put;
7753 for (i = 0; i < nr; i++) {
7754 u64 *tag_slot = io_get_tag_slot(data, i);
7756 if (copy_from_user(tag_slot, &utags[i],
7762 atomic_set(&data->refs, 1);
7763 init_completion(&data->done);
7767 io_rsrc_data_free(data);
7771 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7773 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7774 GFP_KERNEL_ACCOUNT);
7775 return !!table->files;
7778 static void io_free_file_tables(struct io_file_table *table)
7780 kvfree(table->files);
7781 table->files = NULL;
7784 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7786 #if defined(CONFIG_UNIX)
7787 if (ctx->ring_sock) {
7788 struct sock *sock = ctx->ring_sock->sk;
7789 struct sk_buff *skb;
7791 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7797 for (i = 0; i < ctx->nr_user_files; i++) {
7800 file = io_file_from_index(ctx, i);
7805 io_free_file_tables(&ctx->file_table);
7806 io_rsrc_data_free(ctx->file_data);
7807 ctx->file_data = NULL;
7808 ctx->nr_user_files = 0;
7811 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7815 if (!ctx->file_data)
7817 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7819 __io_sqe_files_unregister(ctx);
7823 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7824 __releases(&sqd->lock)
7826 WARN_ON_ONCE(sqd->thread == current);
7829 * Do the dance but not conditional clear_bit() because it'd race with
7830 * other threads incrementing park_pending and setting the bit.
7832 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7833 if (atomic_dec_return(&sqd->park_pending))
7834 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7835 mutex_unlock(&sqd->lock);
7838 static void io_sq_thread_park(struct io_sq_data *sqd)
7839 __acquires(&sqd->lock)
7841 WARN_ON_ONCE(sqd->thread == current);
7843 atomic_inc(&sqd->park_pending);
7844 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7845 mutex_lock(&sqd->lock);
7847 wake_up_process(sqd->thread);
7850 static void io_sq_thread_stop(struct io_sq_data *sqd)
7852 WARN_ON_ONCE(sqd->thread == current);
7853 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7855 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7856 mutex_lock(&sqd->lock);
7858 wake_up_process(sqd->thread);
7859 mutex_unlock(&sqd->lock);
7860 wait_for_completion(&sqd->exited);
7863 static void io_put_sq_data(struct io_sq_data *sqd)
7865 if (refcount_dec_and_test(&sqd->refs)) {
7866 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7868 io_sq_thread_stop(sqd);
7873 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7875 struct io_sq_data *sqd = ctx->sq_data;
7878 io_sq_thread_park(sqd);
7879 list_del_init(&ctx->sqd_list);
7880 io_sqd_update_thread_idle(sqd);
7881 io_sq_thread_unpark(sqd);
7883 io_put_sq_data(sqd);
7884 ctx->sq_data = NULL;
7888 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7890 struct io_ring_ctx *ctx_attach;
7891 struct io_sq_data *sqd;
7894 f = fdget(p->wq_fd);
7896 return ERR_PTR(-ENXIO);
7897 if (f.file->f_op != &io_uring_fops) {
7899 return ERR_PTR(-EINVAL);
7902 ctx_attach = f.file->private_data;
7903 sqd = ctx_attach->sq_data;
7906 return ERR_PTR(-EINVAL);
7908 if (sqd->task_tgid != current->tgid) {
7910 return ERR_PTR(-EPERM);
7913 refcount_inc(&sqd->refs);
7918 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7921 struct io_sq_data *sqd;
7924 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7925 sqd = io_attach_sq_data(p);
7930 /* fall through for EPERM case, setup new sqd/task */
7931 if (PTR_ERR(sqd) != -EPERM)
7935 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7937 return ERR_PTR(-ENOMEM);
7939 atomic_set(&sqd->park_pending, 0);
7940 refcount_set(&sqd->refs, 1);
7941 INIT_LIST_HEAD(&sqd->ctx_list);
7942 mutex_init(&sqd->lock);
7943 init_waitqueue_head(&sqd->wait);
7944 init_completion(&sqd->exited);
7948 #if defined(CONFIG_UNIX)
7950 * Ensure the UNIX gc is aware of our file set, so we are certain that
7951 * the io_uring can be safely unregistered on process exit, even if we have
7952 * loops in the file referencing.
7954 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7956 struct sock *sk = ctx->ring_sock->sk;
7957 struct scm_fp_list *fpl;
7958 struct sk_buff *skb;
7961 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7965 skb = alloc_skb(0, GFP_KERNEL);
7974 fpl->user = get_uid(current_user());
7975 for (i = 0; i < nr; i++) {
7976 struct file *file = io_file_from_index(ctx, i + offset);
7980 fpl->fp[nr_files] = get_file(file);
7981 unix_inflight(fpl->user, fpl->fp[nr_files]);
7986 fpl->max = SCM_MAX_FD;
7987 fpl->count = nr_files;
7988 UNIXCB(skb).fp = fpl;
7989 skb->destructor = unix_destruct_scm;
7990 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7991 skb_queue_head(&sk->sk_receive_queue, skb);
7993 for (i = 0; i < nr_files; i++)
8004 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8005 * causes regular reference counting to break down. We rely on the UNIX
8006 * garbage collection to take care of this problem for us.
8008 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8010 unsigned left, total;
8014 left = ctx->nr_user_files;
8016 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8018 ret = __io_sqe_files_scm(ctx, this_files, total);
8022 total += this_files;
8028 while (total < ctx->nr_user_files) {
8029 struct file *file = io_file_from_index(ctx, total);
8039 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8045 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8047 struct file *file = prsrc->file;
8048 #if defined(CONFIG_UNIX)
8049 struct sock *sock = ctx->ring_sock->sk;
8050 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8051 struct sk_buff *skb;
8054 __skb_queue_head_init(&list);
8057 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8058 * remove this entry and rearrange the file array.
8060 skb = skb_dequeue(head);
8062 struct scm_fp_list *fp;
8064 fp = UNIXCB(skb).fp;
8065 for (i = 0; i < fp->count; i++) {
8068 if (fp->fp[i] != file)
8071 unix_notinflight(fp->user, fp->fp[i]);
8072 left = fp->count - 1 - i;
8074 memmove(&fp->fp[i], &fp->fp[i + 1],
8075 left * sizeof(struct file *));
8082 __skb_queue_tail(&list, skb);
8092 __skb_queue_tail(&list, skb);
8094 skb = skb_dequeue(head);
8097 if (skb_peek(&list)) {
8098 spin_lock_irq(&head->lock);
8099 while ((skb = __skb_dequeue(&list)) != NULL)
8100 __skb_queue_tail(head, skb);
8101 spin_unlock_irq(&head->lock);
8108 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8110 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8111 struct io_ring_ctx *ctx = rsrc_data->ctx;
8112 struct io_rsrc_put *prsrc, *tmp;
8114 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8115 list_del(&prsrc->list);
8118 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8120 io_ring_submit_lock(ctx, lock_ring);
8121 spin_lock(&ctx->completion_lock);
8122 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8124 io_commit_cqring(ctx);
8125 spin_unlock(&ctx->completion_lock);
8126 io_cqring_ev_posted(ctx);
8127 io_ring_submit_unlock(ctx, lock_ring);
8130 rsrc_data->do_put(ctx, prsrc);
8134 io_rsrc_node_destroy(ref_node);
8135 if (atomic_dec_and_test(&rsrc_data->refs))
8136 complete(&rsrc_data->done);
8139 static void io_rsrc_put_work(struct work_struct *work)
8141 struct io_ring_ctx *ctx;
8142 struct llist_node *node;
8144 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8145 node = llist_del_all(&ctx->rsrc_put_llist);
8148 struct io_rsrc_node *ref_node;
8149 struct llist_node *next = node->next;
8151 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8152 __io_rsrc_put_work(ref_node);
8157 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8158 unsigned nr_args, u64 __user *tags)
8160 __s32 __user *fds = (__s32 __user *) arg;
8169 if (nr_args > IORING_MAX_FIXED_FILES)
8171 if (nr_args > rlimit(RLIMIT_NOFILE))
8173 ret = io_rsrc_node_switch_start(ctx);
8176 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8182 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8185 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8186 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8190 /* allow sparse sets */
8193 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8200 if (unlikely(!file))
8204 * Don't allow io_uring instances to be registered. If UNIX
8205 * isn't enabled, then this causes a reference cycle and this
8206 * instance can never get freed. If UNIX is enabled we'll
8207 * handle it just fine, but there's still no point in allowing
8208 * a ring fd as it doesn't support regular read/write anyway.
8210 if (file->f_op == &io_uring_fops) {
8214 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8217 ret = io_sqe_files_scm(ctx);
8219 __io_sqe_files_unregister(ctx);
8223 io_rsrc_node_switch(ctx, NULL);
8226 for (i = 0; i < ctx->nr_user_files; i++) {
8227 file = io_file_from_index(ctx, i);
8231 io_free_file_tables(&ctx->file_table);
8232 ctx->nr_user_files = 0;
8234 io_rsrc_data_free(ctx->file_data);
8235 ctx->file_data = NULL;
8239 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8242 #if defined(CONFIG_UNIX)
8243 struct sock *sock = ctx->ring_sock->sk;
8244 struct sk_buff_head *head = &sock->sk_receive_queue;
8245 struct sk_buff *skb;
8248 * See if we can merge this file into an existing skb SCM_RIGHTS
8249 * file set. If there's no room, fall back to allocating a new skb
8250 * and filling it in.
8252 spin_lock_irq(&head->lock);
8253 skb = skb_peek(head);
8255 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8257 if (fpl->count < SCM_MAX_FD) {
8258 __skb_unlink(skb, head);
8259 spin_unlock_irq(&head->lock);
8260 fpl->fp[fpl->count] = get_file(file);
8261 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8263 spin_lock_irq(&head->lock);
8264 __skb_queue_head(head, skb);
8269 spin_unlock_irq(&head->lock);
8276 return __io_sqe_files_scm(ctx, 1, index);
8282 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8283 unsigned int issue_flags, u32 slot_index)
8285 struct io_ring_ctx *ctx = req->ctx;
8286 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8287 struct io_fixed_file *file_slot;
8290 io_ring_submit_lock(ctx, !force_nonblock);
8291 if (file->f_op == &io_uring_fops)
8294 if (!ctx->file_data)
8297 if (slot_index >= ctx->nr_user_files)
8300 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8301 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8303 if (file_slot->file_ptr)
8306 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8307 io_fixed_file_set(file_slot, file);
8308 ret = io_sqe_file_register(ctx, file, slot_index);
8310 file_slot->file_ptr = 0;
8316 io_ring_submit_unlock(ctx, !force_nonblock);
8322 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8323 struct io_rsrc_node *node, void *rsrc)
8325 struct io_rsrc_put *prsrc;
8327 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8331 prsrc->tag = *io_get_tag_slot(data, idx);
8333 list_add(&prsrc->list, &node->rsrc_list);
8337 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8338 struct io_uring_rsrc_update2 *up,
8341 u64 __user *tags = u64_to_user_ptr(up->tags);
8342 __s32 __user *fds = u64_to_user_ptr(up->data);
8343 struct io_rsrc_data *data = ctx->file_data;
8344 struct io_fixed_file *file_slot;
8348 bool needs_switch = false;
8350 if (!ctx->file_data)
8352 if (up->offset + nr_args > ctx->nr_user_files)
8355 for (done = 0; done < nr_args; done++) {
8358 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8359 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8363 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8367 if (fd == IORING_REGISTER_FILES_SKIP)
8370 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8371 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8373 if (file_slot->file_ptr) {
8374 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8375 err = io_queue_rsrc_removal(data, up->offset + done,
8376 ctx->rsrc_node, file);
8379 file_slot->file_ptr = 0;
8380 needs_switch = true;
8389 * Don't allow io_uring instances to be registered. If
8390 * UNIX isn't enabled, then this causes a reference
8391 * cycle and this instance can never get freed. If UNIX
8392 * is enabled we'll handle it just fine, but there's
8393 * still no point in allowing a ring fd as it doesn't
8394 * support regular read/write anyway.
8396 if (file->f_op == &io_uring_fops) {
8401 *io_get_tag_slot(data, up->offset + done) = tag;
8402 io_fixed_file_set(file_slot, file);
8403 err = io_sqe_file_register(ctx, file, i);
8405 file_slot->file_ptr = 0;
8413 io_rsrc_node_switch(ctx, data);
8414 return done ? done : err;
8417 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8418 struct task_struct *task)
8420 struct io_wq_hash *hash;
8421 struct io_wq_data data;
8422 unsigned int concurrency;
8424 mutex_lock(&ctx->uring_lock);
8425 hash = ctx->hash_map;
8427 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8429 mutex_unlock(&ctx->uring_lock);
8430 return ERR_PTR(-ENOMEM);
8432 refcount_set(&hash->refs, 1);
8433 init_waitqueue_head(&hash->wait);
8434 ctx->hash_map = hash;
8436 mutex_unlock(&ctx->uring_lock);
8440 data.free_work = io_wq_free_work;
8441 data.do_work = io_wq_submit_work;
8443 /* Do QD, or 4 * CPUS, whatever is smallest */
8444 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8446 return io_wq_create(concurrency, &data);
8449 static int io_uring_alloc_task_context(struct task_struct *task,
8450 struct io_ring_ctx *ctx)
8452 struct io_uring_task *tctx;
8455 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8456 if (unlikely(!tctx))
8459 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8460 if (unlikely(ret)) {
8465 tctx->io_wq = io_init_wq_offload(ctx, task);
8466 if (IS_ERR(tctx->io_wq)) {
8467 ret = PTR_ERR(tctx->io_wq);
8468 percpu_counter_destroy(&tctx->inflight);
8474 init_waitqueue_head(&tctx->wait);
8475 atomic_set(&tctx->in_idle, 0);
8476 atomic_set(&tctx->inflight_tracked, 0);
8477 task->io_uring = tctx;
8478 spin_lock_init(&tctx->task_lock);
8479 INIT_WQ_LIST(&tctx->task_list);
8480 init_task_work(&tctx->task_work, tctx_task_work);
8484 void __io_uring_free(struct task_struct *tsk)
8486 struct io_uring_task *tctx = tsk->io_uring;
8488 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8489 WARN_ON_ONCE(tctx->io_wq);
8490 WARN_ON_ONCE(tctx->cached_refs);
8492 percpu_counter_destroy(&tctx->inflight);
8494 tsk->io_uring = NULL;
8497 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8498 struct io_uring_params *p)
8502 /* Retain compatibility with failing for an invalid attach attempt */
8503 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8504 IORING_SETUP_ATTACH_WQ) {
8507 f = fdget(p->wq_fd);
8510 if (f.file->f_op != &io_uring_fops) {
8516 if (ctx->flags & IORING_SETUP_SQPOLL) {
8517 struct task_struct *tsk;
8518 struct io_sq_data *sqd;
8521 sqd = io_get_sq_data(p, &attached);
8527 ctx->sq_creds = get_current_cred();
8529 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8530 if (!ctx->sq_thread_idle)
8531 ctx->sq_thread_idle = HZ;
8533 io_sq_thread_park(sqd);
8534 list_add(&ctx->sqd_list, &sqd->ctx_list);
8535 io_sqd_update_thread_idle(sqd);
8536 /* don't attach to a dying SQPOLL thread, would be racy */
8537 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8538 io_sq_thread_unpark(sqd);
8545 if (p->flags & IORING_SETUP_SQ_AFF) {
8546 int cpu = p->sq_thread_cpu;
8549 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8556 sqd->task_pid = current->pid;
8557 sqd->task_tgid = current->tgid;
8558 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8565 ret = io_uring_alloc_task_context(tsk, ctx);
8566 wake_up_new_task(tsk);
8569 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8570 /* Can't have SQ_AFF without SQPOLL */
8577 complete(&ctx->sq_data->exited);
8579 io_sq_thread_finish(ctx);
8583 static inline void __io_unaccount_mem(struct user_struct *user,
8584 unsigned long nr_pages)
8586 atomic_long_sub(nr_pages, &user->locked_vm);
8589 static inline int __io_account_mem(struct user_struct *user,
8590 unsigned long nr_pages)
8592 unsigned long page_limit, cur_pages, new_pages;
8594 /* Don't allow more pages than we can safely lock */
8595 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8598 cur_pages = atomic_long_read(&user->locked_vm);
8599 new_pages = cur_pages + nr_pages;
8600 if (new_pages > page_limit)
8602 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8603 new_pages) != cur_pages);
8608 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8611 __io_unaccount_mem(ctx->user, nr_pages);
8613 if (ctx->mm_account)
8614 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8617 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8622 ret = __io_account_mem(ctx->user, nr_pages);
8627 if (ctx->mm_account)
8628 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8633 static void io_mem_free(void *ptr)
8640 page = virt_to_head_page(ptr);
8641 if (put_page_testzero(page))
8642 free_compound_page(page);
8645 static void *io_mem_alloc(size_t size)
8647 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8648 __GFP_NORETRY | __GFP_ACCOUNT;
8650 return (void *) __get_free_pages(gfp_flags, get_order(size));
8653 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8656 struct io_rings *rings;
8657 size_t off, sq_array_size;
8659 off = struct_size(rings, cqes, cq_entries);
8660 if (off == SIZE_MAX)
8664 off = ALIGN(off, SMP_CACHE_BYTES);
8672 sq_array_size = array_size(sizeof(u32), sq_entries);
8673 if (sq_array_size == SIZE_MAX)
8676 if (check_add_overflow(off, sq_array_size, &off))
8682 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8684 struct io_mapped_ubuf *imu = *slot;
8687 if (imu != ctx->dummy_ubuf) {
8688 for (i = 0; i < imu->nr_bvecs; i++)
8689 unpin_user_page(imu->bvec[i].bv_page);
8690 if (imu->acct_pages)
8691 io_unaccount_mem(ctx, imu->acct_pages);
8697 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8699 io_buffer_unmap(ctx, &prsrc->buf);
8703 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8707 for (i = 0; i < ctx->nr_user_bufs; i++)
8708 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8709 kfree(ctx->user_bufs);
8710 io_rsrc_data_free(ctx->buf_data);
8711 ctx->user_bufs = NULL;
8712 ctx->buf_data = NULL;
8713 ctx->nr_user_bufs = 0;
8716 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8723 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8725 __io_sqe_buffers_unregister(ctx);
8729 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8730 void __user *arg, unsigned index)
8732 struct iovec __user *src;
8734 #ifdef CONFIG_COMPAT
8736 struct compat_iovec __user *ciovs;
8737 struct compat_iovec ciov;
8739 ciovs = (struct compat_iovec __user *) arg;
8740 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8743 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8744 dst->iov_len = ciov.iov_len;
8748 src = (struct iovec __user *) arg;
8749 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8755 * Not super efficient, but this is just a registration time. And we do cache
8756 * the last compound head, so generally we'll only do a full search if we don't
8759 * We check if the given compound head page has already been accounted, to
8760 * avoid double accounting it. This allows us to account the full size of the
8761 * page, not just the constituent pages of a huge page.
8763 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8764 int nr_pages, struct page *hpage)
8768 /* check current page array */
8769 for (i = 0; i < nr_pages; i++) {
8770 if (!PageCompound(pages[i]))
8772 if (compound_head(pages[i]) == hpage)
8776 /* check previously registered pages */
8777 for (i = 0; i < ctx->nr_user_bufs; i++) {
8778 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8780 for (j = 0; j < imu->nr_bvecs; j++) {
8781 if (!PageCompound(imu->bvec[j].bv_page))
8783 if (compound_head(imu->bvec[j].bv_page) == hpage)
8791 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8792 int nr_pages, struct io_mapped_ubuf *imu,
8793 struct page **last_hpage)
8797 imu->acct_pages = 0;
8798 for (i = 0; i < nr_pages; i++) {
8799 if (!PageCompound(pages[i])) {
8804 hpage = compound_head(pages[i]);
8805 if (hpage == *last_hpage)
8807 *last_hpage = hpage;
8808 if (headpage_already_acct(ctx, pages, i, hpage))
8810 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8814 if (!imu->acct_pages)
8817 ret = io_account_mem(ctx, imu->acct_pages);
8819 imu->acct_pages = 0;
8823 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8824 struct io_mapped_ubuf **pimu,
8825 struct page **last_hpage)
8827 struct io_mapped_ubuf *imu = NULL;
8828 struct vm_area_struct **vmas = NULL;
8829 struct page **pages = NULL;
8830 unsigned long off, start, end, ubuf;
8832 int ret, pret, nr_pages, i;
8834 if (!iov->iov_base) {
8835 *pimu = ctx->dummy_ubuf;
8839 ubuf = (unsigned long) iov->iov_base;
8840 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8841 start = ubuf >> PAGE_SHIFT;
8842 nr_pages = end - start;
8847 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8851 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8856 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8861 mmap_read_lock(current->mm);
8862 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8864 if (pret == nr_pages) {
8865 /* don't support file backed memory */
8866 for (i = 0; i < nr_pages; i++) {
8867 struct vm_area_struct *vma = vmas[i];
8869 if (vma_is_shmem(vma))
8872 !is_file_hugepages(vma->vm_file)) {
8878 ret = pret < 0 ? pret : -EFAULT;
8880 mmap_read_unlock(current->mm);
8883 * if we did partial map, or found file backed vmas,
8884 * release any pages we did get
8887 unpin_user_pages(pages, pret);
8891 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8893 unpin_user_pages(pages, pret);
8897 off = ubuf & ~PAGE_MASK;
8898 size = iov->iov_len;
8899 for (i = 0; i < nr_pages; i++) {
8902 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8903 imu->bvec[i].bv_page = pages[i];
8904 imu->bvec[i].bv_len = vec_len;
8905 imu->bvec[i].bv_offset = off;
8909 /* store original address for later verification */
8911 imu->ubuf_end = ubuf + iov->iov_len;
8912 imu->nr_bvecs = nr_pages;
8923 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8925 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8926 return ctx->user_bufs ? 0 : -ENOMEM;
8929 static int io_buffer_validate(struct iovec *iov)
8931 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8934 * Don't impose further limits on the size and buffer
8935 * constraints here, we'll -EINVAL later when IO is
8936 * submitted if they are wrong.
8939 return iov->iov_len ? -EFAULT : 0;
8943 /* arbitrary limit, but we need something */
8944 if (iov->iov_len > SZ_1G)
8947 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8953 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8954 unsigned int nr_args, u64 __user *tags)
8956 struct page *last_hpage = NULL;
8957 struct io_rsrc_data *data;
8963 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8965 ret = io_rsrc_node_switch_start(ctx);
8968 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8971 ret = io_buffers_map_alloc(ctx, nr_args);
8973 io_rsrc_data_free(data);
8977 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8978 ret = io_copy_iov(ctx, &iov, arg, i);
8981 ret = io_buffer_validate(&iov);
8984 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8989 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8995 WARN_ON_ONCE(ctx->buf_data);
8997 ctx->buf_data = data;
8999 __io_sqe_buffers_unregister(ctx);
9001 io_rsrc_node_switch(ctx, NULL);
9005 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9006 struct io_uring_rsrc_update2 *up,
9007 unsigned int nr_args)
9009 u64 __user *tags = u64_to_user_ptr(up->tags);
9010 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9011 struct page *last_hpage = NULL;
9012 bool needs_switch = false;
9018 if (up->offset + nr_args > ctx->nr_user_bufs)
9021 for (done = 0; done < nr_args; done++) {
9022 struct io_mapped_ubuf *imu;
9023 int offset = up->offset + done;
9026 err = io_copy_iov(ctx, &iov, iovs, done);
9029 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9033 err = io_buffer_validate(&iov);
9036 if (!iov.iov_base && tag) {
9040 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9044 i = array_index_nospec(offset, ctx->nr_user_bufs);
9045 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9046 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9047 ctx->rsrc_node, ctx->user_bufs[i]);
9048 if (unlikely(err)) {
9049 io_buffer_unmap(ctx, &imu);
9052 ctx->user_bufs[i] = NULL;
9053 needs_switch = true;
9056 ctx->user_bufs[i] = imu;
9057 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9061 io_rsrc_node_switch(ctx, ctx->buf_data);
9062 return done ? done : err;
9065 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9067 __s32 __user *fds = arg;
9073 if (copy_from_user(&fd, fds, sizeof(*fds)))
9076 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9077 if (IS_ERR(ctx->cq_ev_fd)) {
9078 int ret = PTR_ERR(ctx->cq_ev_fd);
9080 ctx->cq_ev_fd = NULL;
9087 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9089 if (ctx->cq_ev_fd) {
9090 eventfd_ctx_put(ctx->cq_ev_fd);
9091 ctx->cq_ev_fd = NULL;
9098 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9100 struct io_buffer *buf;
9101 unsigned long index;
9103 xa_for_each(&ctx->io_buffers, index, buf)
9104 __io_remove_buffers(ctx, buf, index, -1U);
9107 static void io_req_cache_free(struct list_head *list)
9109 struct io_kiocb *req, *nxt;
9111 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9112 list_del(&req->inflight_entry);
9113 kmem_cache_free(req_cachep, req);
9117 static void io_req_caches_free(struct io_ring_ctx *ctx)
9119 struct io_submit_state *state = &ctx->submit_state;
9121 mutex_lock(&ctx->uring_lock);
9123 if (state->free_reqs) {
9124 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9125 state->free_reqs = 0;
9128 io_flush_cached_locked_reqs(ctx, state);
9129 io_req_cache_free(&state->free_list);
9130 mutex_unlock(&ctx->uring_lock);
9133 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9135 if (data && !atomic_dec_and_test(&data->refs))
9136 wait_for_completion(&data->done);
9139 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9141 io_sq_thread_finish(ctx);
9143 if (ctx->mm_account) {
9144 mmdrop(ctx->mm_account);
9145 ctx->mm_account = NULL;
9148 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9149 io_wait_rsrc_data(ctx->buf_data);
9150 io_wait_rsrc_data(ctx->file_data);
9152 mutex_lock(&ctx->uring_lock);
9154 __io_sqe_buffers_unregister(ctx);
9156 __io_sqe_files_unregister(ctx);
9158 __io_cqring_overflow_flush(ctx, true);
9159 mutex_unlock(&ctx->uring_lock);
9160 io_eventfd_unregister(ctx);
9161 io_destroy_buffers(ctx);
9163 put_cred(ctx->sq_creds);
9165 /* there are no registered resources left, nobody uses it */
9167 io_rsrc_node_destroy(ctx->rsrc_node);
9168 if (ctx->rsrc_backup_node)
9169 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9170 flush_delayed_work(&ctx->rsrc_put_work);
9172 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9173 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9175 #if defined(CONFIG_UNIX)
9176 if (ctx->ring_sock) {
9177 ctx->ring_sock->file = NULL; /* so that iput() is called */
9178 sock_release(ctx->ring_sock);
9181 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9183 io_mem_free(ctx->rings);
9184 io_mem_free(ctx->sq_sqes);
9186 percpu_ref_exit(&ctx->refs);
9187 free_uid(ctx->user);
9188 io_req_caches_free(ctx);
9190 io_wq_put_hash(ctx->hash_map);
9191 kfree(ctx->cancel_hash);
9192 kfree(ctx->dummy_ubuf);
9196 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9198 struct io_ring_ctx *ctx = file->private_data;
9201 poll_wait(file, &ctx->poll_wait, wait);
9203 * synchronizes with barrier from wq_has_sleeper call in
9207 if (!io_sqring_full(ctx))
9208 mask |= EPOLLOUT | EPOLLWRNORM;
9211 * Don't flush cqring overflow list here, just do a simple check.
9212 * Otherwise there could possible be ABBA deadlock:
9215 * lock(&ctx->uring_lock);
9217 * lock(&ctx->uring_lock);
9220 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9221 * pushs them to do the flush.
9223 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9224 mask |= EPOLLIN | EPOLLRDNORM;
9229 static int io_uring_fasync(int fd, struct file *file, int on)
9231 struct io_ring_ctx *ctx = file->private_data;
9233 return fasync_helper(fd, file, on, &ctx->cq_fasync);
9236 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9238 const struct cred *creds;
9240 creds = xa_erase(&ctx->personalities, id);
9249 struct io_tctx_exit {
9250 struct callback_head task_work;
9251 struct completion completion;
9252 struct io_ring_ctx *ctx;
9255 static void io_tctx_exit_cb(struct callback_head *cb)
9257 struct io_uring_task *tctx = current->io_uring;
9258 struct io_tctx_exit *work;
9260 work = container_of(cb, struct io_tctx_exit, task_work);
9262 * When @in_idle, we're in cancellation and it's racy to remove the
9263 * node. It'll be removed by the end of cancellation, just ignore it.
9265 if (!atomic_read(&tctx->in_idle))
9266 io_uring_del_tctx_node((unsigned long)work->ctx);
9267 complete(&work->completion);
9270 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9272 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9274 return req->ctx == data;
9277 static void io_ring_exit_work(struct work_struct *work)
9279 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9280 unsigned long timeout = jiffies + HZ * 60 * 5;
9281 unsigned long interval = HZ / 20;
9282 struct io_tctx_exit exit;
9283 struct io_tctx_node *node;
9287 * If we're doing polled IO and end up having requests being
9288 * submitted async (out-of-line), then completions can come in while
9289 * we're waiting for refs to drop. We need to reap these manually,
9290 * as nobody else will be looking for them.
9293 io_uring_try_cancel_requests(ctx, NULL, true);
9295 struct io_sq_data *sqd = ctx->sq_data;
9296 struct task_struct *tsk;
9298 io_sq_thread_park(sqd);
9300 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9301 io_wq_cancel_cb(tsk->io_uring->io_wq,
9302 io_cancel_ctx_cb, ctx, true);
9303 io_sq_thread_unpark(sqd);
9306 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9307 /* there is little hope left, don't run it too often */
9310 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9312 init_completion(&exit.completion);
9313 init_task_work(&exit.task_work, io_tctx_exit_cb);
9316 * Some may use context even when all refs and requests have been put,
9317 * and they are free to do so while still holding uring_lock or
9318 * completion_lock, see io_req_task_submit(). Apart from other work,
9319 * this lock/unlock section also waits them to finish.
9321 mutex_lock(&ctx->uring_lock);
9322 while (!list_empty(&ctx->tctx_list)) {
9323 WARN_ON_ONCE(time_after(jiffies, timeout));
9325 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9327 /* don't spin on a single task if cancellation failed */
9328 list_rotate_left(&ctx->tctx_list);
9329 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9330 if (WARN_ON_ONCE(ret))
9332 wake_up_process(node->task);
9334 mutex_unlock(&ctx->uring_lock);
9335 wait_for_completion(&exit.completion);
9336 mutex_lock(&ctx->uring_lock);
9338 mutex_unlock(&ctx->uring_lock);
9339 spin_lock(&ctx->completion_lock);
9340 spin_unlock(&ctx->completion_lock);
9342 io_ring_ctx_free(ctx);
9345 /* Returns true if we found and killed one or more timeouts */
9346 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9349 struct io_kiocb *req, *tmp;
9352 spin_lock(&ctx->completion_lock);
9353 spin_lock_irq(&ctx->timeout_lock);
9354 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9355 if (io_match_task(req, tsk, cancel_all)) {
9356 io_kill_timeout(req, -ECANCELED);
9360 spin_unlock_irq(&ctx->timeout_lock);
9362 io_commit_cqring(ctx);
9363 spin_unlock(&ctx->completion_lock);
9365 io_cqring_ev_posted(ctx);
9366 return canceled != 0;
9369 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9371 unsigned long index;
9372 struct creds *creds;
9374 mutex_lock(&ctx->uring_lock);
9375 percpu_ref_kill(&ctx->refs);
9377 __io_cqring_overflow_flush(ctx, true);
9378 xa_for_each(&ctx->personalities, index, creds)
9379 io_unregister_personality(ctx, index);
9380 mutex_unlock(&ctx->uring_lock);
9382 io_kill_timeouts(ctx, NULL, true);
9383 io_poll_remove_all(ctx, NULL, true);
9385 /* if we failed setting up the ctx, we might not have any rings */
9386 io_iopoll_try_reap_events(ctx);
9388 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9390 * Use system_unbound_wq to avoid spawning tons of event kworkers
9391 * if we're exiting a ton of rings at the same time. It just adds
9392 * noise and overhead, there's no discernable change in runtime
9393 * over using system_wq.
9395 queue_work(system_unbound_wq, &ctx->exit_work);
9398 static int io_uring_release(struct inode *inode, struct file *file)
9400 struct io_ring_ctx *ctx = file->private_data;
9402 file->private_data = NULL;
9403 io_ring_ctx_wait_and_kill(ctx);
9407 struct io_task_cancel {
9408 struct task_struct *task;
9412 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9414 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9415 struct io_task_cancel *cancel = data;
9418 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9419 struct io_ring_ctx *ctx = req->ctx;
9421 /* protect against races with linked timeouts */
9422 spin_lock(&ctx->completion_lock);
9423 ret = io_match_task(req, cancel->task, cancel->all);
9424 spin_unlock(&ctx->completion_lock);
9426 ret = io_match_task(req, cancel->task, cancel->all);
9431 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9432 struct task_struct *task, bool cancel_all)
9434 struct io_defer_entry *de;
9437 spin_lock(&ctx->completion_lock);
9438 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9439 if (io_match_task(de->req, task, cancel_all)) {
9440 list_cut_position(&list, &ctx->defer_list, &de->list);
9444 spin_unlock(&ctx->completion_lock);
9445 if (list_empty(&list))
9448 while (!list_empty(&list)) {
9449 de = list_first_entry(&list, struct io_defer_entry, list);
9450 list_del_init(&de->list);
9451 io_req_complete_failed(de->req, -ECANCELED);
9457 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9459 struct io_tctx_node *node;
9460 enum io_wq_cancel cret;
9463 mutex_lock(&ctx->uring_lock);
9464 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9465 struct io_uring_task *tctx = node->task->io_uring;
9468 * io_wq will stay alive while we hold uring_lock, because it's
9469 * killed after ctx nodes, which requires to take the lock.
9471 if (!tctx || !tctx->io_wq)
9473 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9474 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9476 mutex_unlock(&ctx->uring_lock);
9481 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9482 struct task_struct *task,
9485 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9486 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9489 enum io_wq_cancel cret;
9493 ret |= io_uring_try_cancel_iowq(ctx);
9494 } else if (tctx && tctx->io_wq) {
9496 * Cancels requests of all rings, not only @ctx, but
9497 * it's fine as the task is in exit/exec.
9499 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9501 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9504 /* SQPOLL thread does its own polling */
9505 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9506 (ctx->sq_data && ctx->sq_data->thread == current)) {
9507 while (!list_empty_careful(&ctx->iopoll_list)) {
9508 io_iopoll_try_reap_events(ctx);
9513 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9514 ret |= io_poll_remove_all(ctx, task, cancel_all);
9515 ret |= io_kill_timeouts(ctx, task, cancel_all);
9517 ret |= io_run_task_work();
9524 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9526 struct io_uring_task *tctx = current->io_uring;
9527 struct io_tctx_node *node;
9530 if (unlikely(!tctx)) {
9531 ret = io_uring_alloc_task_context(current, ctx);
9534 tctx = current->io_uring;
9536 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9537 node = kmalloc(sizeof(*node), GFP_KERNEL);
9541 node->task = current;
9543 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9550 mutex_lock(&ctx->uring_lock);
9551 list_add(&node->ctx_node, &ctx->tctx_list);
9552 mutex_unlock(&ctx->uring_lock);
9559 * Note that this task has used io_uring. We use it for cancelation purposes.
9561 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9563 struct io_uring_task *tctx = current->io_uring;
9565 if (likely(tctx && tctx->last == ctx))
9567 return __io_uring_add_tctx_node(ctx);
9571 * Remove this io_uring_file -> task mapping.
9573 static void io_uring_del_tctx_node(unsigned long index)
9575 struct io_uring_task *tctx = current->io_uring;
9576 struct io_tctx_node *node;
9580 node = xa_erase(&tctx->xa, index);
9584 WARN_ON_ONCE(current != node->task);
9585 WARN_ON_ONCE(list_empty(&node->ctx_node));
9587 mutex_lock(&node->ctx->uring_lock);
9588 list_del(&node->ctx_node);
9589 mutex_unlock(&node->ctx->uring_lock);
9591 if (tctx->last == node->ctx)
9596 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9598 struct io_wq *wq = tctx->io_wq;
9599 struct io_tctx_node *node;
9600 unsigned long index;
9602 xa_for_each(&tctx->xa, index, node)
9603 io_uring_del_tctx_node(index);
9606 * Must be after io_uring_del_task_file() (removes nodes under
9607 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9609 io_wq_put_and_exit(wq);
9614 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9617 return atomic_read(&tctx->inflight_tracked);
9618 return percpu_counter_sum(&tctx->inflight);
9621 static void io_uring_drop_tctx_refs(struct task_struct *task)
9623 struct io_uring_task *tctx = task->io_uring;
9624 unsigned int refs = tctx->cached_refs;
9627 tctx->cached_refs = 0;
9628 percpu_counter_sub(&tctx->inflight, refs);
9629 put_task_struct_many(task, refs);
9634 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9635 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9637 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9639 struct io_uring_task *tctx = current->io_uring;
9640 struct io_ring_ctx *ctx;
9644 WARN_ON_ONCE(sqd && sqd->thread != current);
9646 if (!current->io_uring)
9649 io_wq_exit_start(tctx->io_wq);
9651 atomic_inc(&tctx->in_idle);
9653 io_uring_drop_tctx_refs(current);
9654 /* read completions before cancelations */
9655 inflight = tctx_inflight(tctx, !cancel_all);
9660 struct io_tctx_node *node;
9661 unsigned long index;
9663 xa_for_each(&tctx->xa, index, node) {
9664 /* sqpoll task will cancel all its requests */
9665 if (node->ctx->sq_data)
9667 io_uring_try_cancel_requests(node->ctx, current,
9671 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9672 io_uring_try_cancel_requests(ctx, current,
9676 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9677 io_uring_drop_tctx_refs(current);
9679 * If we've seen completions, retry without waiting. This
9680 * avoids a race where a completion comes in before we did
9681 * prepare_to_wait().
9683 if (inflight == tctx_inflight(tctx, !cancel_all))
9685 finish_wait(&tctx->wait, &wait);
9687 atomic_dec(&tctx->in_idle);
9689 io_uring_clean_tctx(tctx);
9691 /* for exec all current's requests should be gone, kill tctx */
9692 __io_uring_free(current);
9696 void __io_uring_cancel(bool cancel_all)
9698 io_uring_cancel_generic(cancel_all, NULL);
9701 static void *io_uring_validate_mmap_request(struct file *file,
9702 loff_t pgoff, size_t sz)
9704 struct io_ring_ctx *ctx = file->private_data;
9705 loff_t offset = pgoff << PAGE_SHIFT;
9710 case IORING_OFF_SQ_RING:
9711 case IORING_OFF_CQ_RING:
9714 case IORING_OFF_SQES:
9718 return ERR_PTR(-EINVAL);
9721 page = virt_to_head_page(ptr);
9722 if (sz > page_size(page))
9723 return ERR_PTR(-EINVAL);
9730 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9732 size_t sz = vma->vm_end - vma->vm_start;
9736 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9738 return PTR_ERR(ptr);
9740 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9741 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9744 #else /* !CONFIG_MMU */
9746 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9748 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9751 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9753 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9756 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9757 unsigned long addr, unsigned long len,
9758 unsigned long pgoff, unsigned long flags)
9762 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9764 return PTR_ERR(ptr);
9766 return (unsigned long) ptr;
9769 #endif /* !CONFIG_MMU */
9771 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9776 if (!io_sqring_full(ctx))
9778 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9780 if (!io_sqring_full(ctx))
9783 } while (!signal_pending(current));
9785 finish_wait(&ctx->sqo_sq_wait, &wait);
9789 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9790 struct __kernel_timespec __user **ts,
9791 const sigset_t __user **sig)
9793 struct io_uring_getevents_arg arg;
9796 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9797 * is just a pointer to the sigset_t.
9799 if (!(flags & IORING_ENTER_EXT_ARG)) {
9800 *sig = (const sigset_t __user *) argp;
9806 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9807 * timespec and sigset_t pointers if good.
9809 if (*argsz != sizeof(arg))
9811 if (copy_from_user(&arg, argp, sizeof(arg)))
9813 *sig = u64_to_user_ptr(arg.sigmask);
9814 *argsz = arg.sigmask_sz;
9815 *ts = u64_to_user_ptr(arg.ts);
9819 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9820 u32, min_complete, u32, flags, const void __user *, argp,
9823 struct io_ring_ctx *ctx;
9830 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9831 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9835 if (unlikely(!f.file))
9839 if (unlikely(f.file->f_op != &io_uring_fops))
9843 ctx = f.file->private_data;
9844 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9848 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9852 * For SQ polling, the thread will do all submissions and completions.
9853 * Just return the requested submit count, and wake the thread if
9857 if (ctx->flags & IORING_SETUP_SQPOLL) {
9858 io_cqring_overflow_flush(ctx);
9860 if (unlikely(ctx->sq_data->thread == NULL)) {
9864 if (flags & IORING_ENTER_SQ_WAKEUP)
9865 wake_up(&ctx->sq_data->wait);
9866 if (flags & IORING_ENTER_SQ_WAIT) {
9867 ret = io_sqpoll_wait_sq(ctx);
9871 submitted = to_submit;
9872 } else if (to_submit) {
9873 ret = io_uring_add_tctx_node(ctx);
9876 mutex_lock(&ctx->uring_lock);
9877 submitted = io_submit_sqes(ctx, to_submit);
9878 mutex_unlock(&ctx->uring_lock);
9880 if (submitted != to_submit)
9883 if (flags & IORING_ENTER_GETEVENTS) {
9884 const sigset_t __user *sig;
9885 struct __kernel_timespec __user *ts;
9887 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9891 min_complete = min(min_complete, ctx->cq_entries);
9894 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9895 * space applications don't need to do io completion events
9896 * polling again, they can rely on io_sq_thread to do polling
9897 * work, which can reduce cpu usage and uring_lock contention.
9899 if (ctx->flags & IORING_SETUP_IOPOLL &&
9900 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9901 ret = io_iopoll_check(ctx, min_complete);
9903 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9908 percpu_ref_put(&ctx->refs);
9911 return submitted ? submitted : ret;
9914 #ifdef CONFIG_PROC_FS
9915 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9916 const struct cred *cred)
9918 struct user_namespace *uns = seq_user_ns(m);
9919 struct group_info *gi;
9924 seq_printf(m, "%5d\n", id);
9925 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9926 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9927 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9928 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9929 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9930 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9931 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9932 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9933 seq_puts(m, "\n\tGroups:\t");
9934 gi = cred->group_info;
9935 for (g = 0; g < gi->ngroups; g++) {
9936 seq_put_decimal_ull(m, g ? " " : "",
9937 from_kgid_munged(uns, gi->gid[g]));
9939 seq_puts(m, "\n\tCapEff:\t");
9940 cap = cred->cap_effective;
9941 CAP_FOR_EACH_U32(__capi)
9942 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9947 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9949 struct io_sq_data *sq = NULL;
9954 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9955 * since fdinfo case grabs it in the opposite direction of normal use
9956 * cases. If we fail to get the lock, we just don't iterate any
9957 * structures that could be going away outside the io_uring mutex.
9959 has_lock = mutex_trylock(&ctx->uring_lock);
9961 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9967 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9968 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9969 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9970 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9971 struct file *f = io_file_from_index(ctx, i);
9974 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9976 seq_printf(m, "%5u: <none>\n", i);
9978 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9979 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9980 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9981 unsigned int len = buf->ubuf_end - buf->ubuf;
9983 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9985 if (has_lock && !xa_empty(&ctx->personalities)) {
9986 unsigned long index;
9987 const struct cred *cred;
9989 seq_printf(m, "Personalities:\n");
9990 xa_for_each(&ctx->personalities, index, cred)
9991 io_uring_show_cred(m, index, cred);
9993 seq_printf(m, "PollList:\n");
9994 spin_lock(&ctx->completion_lock);
9995 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9996 struct hlist_head *list = &ctx->cancel_hash[i];
9997 struct io_kiocb *req;
9999 hlist_for_each_entry(req, list, hash_node)
10000 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10001 req->task->task_works != NULL);
10003 spin_unlock(&ctx->completion_lock);
10005 mutex_unlock(&ctx->uring_lock);
10008 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10010 struct io_ring_ctx *ctx = f->private_data;
10012 if (percpu_ref_tryget(&ctx->refs)) {
10013 __io_uring_show_fdinfo(ctx, m);
10014 percpu_ref_put(&ctx->refs);
10019 static const struct file_operations io_uring_fops = {
10020 .release = io_uring_release,
10021 .mmap = io_uring_mmap,
10023 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10024 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10026 .poll = io_uring_poll,
10027 .fasync = io_uring_fasync,
10028 #ifdef CONFIG_PROC_FS
10029 .show_fdinfo = io_uring_show_fdinfo,
10033 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10034 struct io_uring_params *p)
10036 struct io_rings *rings;
10037 size_t size, sq_array_offset;
10039 /* make sure these are sane, as we already accounted them */
10040 ctx->sq_entries = p->sq_entries;
10041 ctx->cq_entries = p->cq_entries;
10043 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10044 if (size == SIZE_MAX)
10047 rings = io_mem_alloc(size);
10051 ctx->rings = rings;
10052 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10053 rings->sq_ring_mask = p->sq_entries - 1;
10054 rings->cq_ring_mask = p->cq_entries - 1;
10055 rings->sq_ring_entries = p->sq_entries;
10056 rings->cq_ring_entries = p->cq_entries;
10058 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10059 if (size == SIZE_MAX) {
10060 io_mem_free(ctx->rings);
10065 ctx->sq_sqes = io_mem_alloc(size);
10066 if (!ctx->sq_sqes) {
10067 io_mem_free(ctx->rings);
10075 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10079 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10083 ret = io_uring_add_tctx_node(ctx);
10088 fd_install(fd, file);
10093 * Allocate an anonymous fd, this is what constitutes the application
10094 * visible backing of an io_uring instance. The application mmaps this
10095 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10096 * we have to tie this fd to a socket for file garbage collection purposes.
10098 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10101 #if defined(CONFIG_UNIX)
10104 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10107 return ERR_PTR(ret);
10110 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10111 O_RDWR | O_CLOEXEC);
10112 #if defined(CONFIG_UNIX)
10113 if (IS_ERR(file)) {
10114 sock_release(ctx->ring_sock);
10115 ctx->ring_sock = NULL;
10117 ctx->ring_sock->file = file;
10123 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10124 struct io_uring_params __user *params)
10126 struct io_ring_ctx *ctx;
10132 if (entries > IORING_MAX_ENTRIES) {
10133 if (!(p->flags & IORING_SETUP_CLAMP))
10135 entries = IORING_MAX_ENTRIES;
10139 * Use twice as many entries for the CQ ring. It's possible for the
10140 * application to drive a higher depth than the size of the SQ ring,
10141 * since the sqes are only used at submission time. This allows for
10142 * some flexibility in overcommitting a bit. If the application has
10143 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10144 * of CQ ring entries manually.
10146 p->sq_entries = roundup_pow_of_two(entries);
10147 if (p->flags & IORING_SETUP_CQSIZE) {
10149 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10150 * to a power-of-two, if it isn't already. We do NOT impose
10151 * any cq vs sq ring sizing.
10153 if (!p->cq_entries)
10155 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10156 if (!(p->flags & IORING_SETUP_CLAMP))
10158 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10160 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10161 if (p->cq_entries < p->sq_entries)
10164 p->cq_entries = 2 * p->sq_entries;
10167 ctx = io_ring_ctx_alloc(p);
10170 ctx->compat = in_compat_syscall();
10171 if (!capable(CAP_IPC_LOCK))
10172 ctx->user = get_uid(current_user());
10175 * This is just grabbed for accounting purposes. When a process exits,
10176 * the mm is exited and dropped before the files, hence we need to hang
10177 * on to this mm purely for the purposes of being able to unaccount
10178 * memory (locked/pinned vm). It's not used for anything else.
10180 mmgrab(current->mm);
10181 ctx->mm_account = current->mm;
10183 ret = io_allocate_scq_urings(ctx, p);
10187 ret = io_sq_offload_create(ctx, p);
10190 /* always set a rsrc node */
10191 ret = io_rsrc_node_switch_start(ctx);
10194 io_rsrc_node_switch(ctx, NULL);
10196 memset(&p->sq_off, 0, sizeof(p->sq_off));
10197 p->sq_off.head = offsetof(struct io_rings, sq.head);
10198 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10199 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10200 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10201 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10202 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10203 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10205 memset(&p->cq_off, 0, sizeof(p->cq_off));
10206 p->cq_off.head = offsetof(struct io_rings, cq.head);
10207 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10208 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10209 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10210 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10211 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10212 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10214 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10215 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10216 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10217 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10218 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10219 IORING_FEAT_RSRC_TAGS;
10221 if (copy_to_user(params, p, sizeof(*p))) {
10226 file = io_uring_get_file(ctx);
10227 if (IS_ERR(file)) {
10228 ret = PTR_ERR(file);
10233 * Install ring fd as the very last thing, so we don't risk someone
10234 * having closed it before we finish setup
10236 ret = io_uring_install_fd(ctx, file);
10238 /* fput will clean it up */
10243 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10246 io_ring_ctx_wait_and_kill(ctx);
10251 * Sets up an aio uring context, and returns the fd. Applications asks for a
10252 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10253 * params structure passed in.
10255 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10257 struct io_uring_params p;
10260 if (copy_from_user(&p, params, sizeof(p)))
10262 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10267 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10268 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10269 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10270 IORING_SETUP_R_DISABLED))
10273 return io_uring_create(entries, &p, params);
10276 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10277 struct io_uring_params __user *, params)
10279 return io_uring_setup(entries, params);
10282 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10284 struct io_uring_probe *p;
10288 size = struct_size(p, ops, nr_args);
10289 if (size == SIZE_MAX)
10291 p = kzalloc(size, GFP_KERNEL);
10296 if (copy_from_user(p, arg, size))
10299 if (memchr_inv(p, 0, size))
10302 p->last_op = IORING_OP_LAST - 1;
10303 if (nr_args > IORING_OP_LAST)
10304 nr_args = IORING_OP_LAST;
10306 for (i = 0; i < nr_args; i++) {
10308 if (!io_op_defs[i].not_supported)
10309 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10314 if (copy_to_user(arg, p, size))
10321 static int io_register_personality(struct io_ring_ctx *ctx)
10323 const struct cred *creds;
10327 creds = get_current_cred();
10329 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10330 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10338 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10339 unsigned int nr_args)
10341 struct io_uring_restriction *res;
10345 /* Restrictions allowed only if rings started disabled */
10346 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10349 /* We allow only a single restrictions registration */
10350 if (ctx->restrictions.registered)
10353 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10356 size = array_size(nr_args, sizeof(*res));
10357 if (size == SIZE_MAX)
10360 res = memdup_user(arg, size);
10362 return PTR_ERR(res);
10366 for (i = 0; i < nr_args; i++) {
10367 switch (res[i].opcode) {
10368 case IORING_RESTRICTION_REGISTER_OP:
10369 if (res[i].register_op >= IORING_REGISTER_LAST) {
10374 __set_bit(res[i].register_op,
10375 ctx->restrictions.register_op);
10377 case IORING_RESTRICTION_SQE_OP:
10378 if (res[i].sqe_op >= IORING_OP_LAST) {
10383 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10385 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10386 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10388 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10389 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10398 /* Reset all restrictions if an error happened */
10400 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10402 ctx->restrictions.registered = true;
10408 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10410 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10413 if (ctx->restrictions.registered)
10414 ctx->restricted = 1;
10416 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10417 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10418 wake_up(&ctx->sq_data->wait);
10422 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10423 struct io_uring_rsrc_update2 *up,
10431 if (check_add_overflow(up->offset, nr_args, &tmp))
10433 err = io_rsrc_node_switch_start(ctx);
10438 case IORING_RSRC_FILE:
10439 return __io_sqe_files_update(ctx, up, nr_args);
10440 case IORING_RSRC_BUFFER:
10441 return __io_sqe_buffers_update(ctx, up, nr_args);
10446 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10449 struct io_uring_rsrc_update2 up;
10453 memset(&up, 0, sizeof(up));
10454 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10456 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10459 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10460 unsigned size, unsigned type)
10462 struct io_uring_rsrc_update2 up;
10464 if (size != sizeof(up))
10466 if (copy_from_user(&up, arg, sizeof(up)))
10468 if (!up.nr || up.resv)
10470 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10473 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10474 unsigned int size, unsigned int type)
10476 struct io_uring_rsrc_register rr;
10478 /* keep it extendible */
10479 if (size != sizeof(rr))
10482 memset(&rr, 0, sizeof(rr));
10483 if (copy_from_user(&rr, arg, size))
10485 if (!rr.nr || rr.resv || rr.resv2)
10489 case IORING_RSRC_FILE:
10490 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10491 rr.nr, u64_to_user_ptr(rr.tags));
10492 case IORING_RSRC_BUFFER:
10493 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10494 rr.nr, u64_to_user_ptr(rr.tags));
10499 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10502 struct io_uring_task *tctx = current->io_uring;
10503 cpumask_var_t new_mask;
10506 if (!tctx || !tctx->io_wq)
10509 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10512 cpumask_clear(new_mask);
10513 if (len > cpumask_size())
10514 len = cpumask_size();
10516 if (copy_from_user(new_mask, arg, len)) {
10517 free_cpumask_var(new_mask);
10521 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10522 free_cpumask_var(new_mask);
10526 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10528 struct io_uring_task *tctx = current->io_uring;
10530 if (!tctx || !tctx->io_wq)
10533 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10536 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10539 struct io_uring_task *tctx = NULL;
10540 struct io_sq_data *sqd = NULL;
10541 __u32 new_count[2];
10544 if (copy_from_user(new_count, arg, sizeof(new_count)))
10546 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10547 if (new_count[i] > INT_MAX)
10550 if (ctx->flags & IORING_SETUP_SQPOLL) {
10551 sqd = ctx->sq_data;
10553 mutex_lock(&sqd->lock);
10554 tctx = sqd->thread->io_uring;
10557 tctx = current->io_uring;
10561 if (!tctx || !tctx->io_wq)
10564 ret = io_wq_max_workers(tctx->io_wq, new_count);
10569 mutex_unlock(&sqd->lock);
10571 if (copy_to_user(arg, new_count, sizeof(new_count)))
10577 mutex_unlock(&sqd->lock);
10581 static bool io_register_op_must_quiesce(int op)
10584 case IORING_REGISTER_BUFFERS:
10585 case IORING_UNREGISTER_BUFFERS:
10586 case IORING_REGISTER_FILES:
10587 case IORING_UNREGISTER_FILES:
10588 case IORING_REGISTER_FILES_UPDATE:
10589 case IORING_REGISTER_PROBE:
10590 case IORING_REGISTER_PERSONALITY:
10591 case IORING_UNREGISTER_PERSONALITY:
10592 case IORING_REGISTER_FILES2:
10593 case IORING_REGISTER_FILES_UPDATE2:
10594 case IORING_REGISTER_BUFFERS2:
10595 case IORING_REGISTER_BUFFERS_UPDATE:
10596 case IORING_REGISTER_IOWQ_AFF:
10597 case IORING_UNREGISTER_IOWQ_AFF:
10598 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10605 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10609 percpu_ref_kill(&ctx->refs);
10612 * Drop uring mutex before waiting for references to exit. If another
10613 * thread is currently inside io_uring_enter() it might need to grab the
10614 * uring_lock to make progress. If we hold it here across the drain
10615 * wait, then we can deadlock. It's safe to drop the mutex here, since
10616 * no new references will come in after we've killed the percpu ref.
10618 mutex_unlock(&ctx->uring_lock);
10620 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10623 ret = io_run_task_work_sig();
10624 } while (ret >= 0);
10625 mutex_lock(&ctx->uring_lock);
10628 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10632 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10633 void __user *arg, unsigned nr_args)
10634 __releases(ctx->uring_lock)
10635 __acquires(ctx->uring_lock)
10640 * We're inside the ring mutex, if the ref is already dying, then
10641 * someone else killed the ctx or is already going through
10642 * io_uring_register().
10644 if (percpu_ref_is_dying(&ctx->refs))
10647 if (ctx->restricted) {
10648 if (opcode >= IORING_REGISTER_LAST)
10650 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10651 if (!test_bit(opcode, ctx->restrictions.register_op))
10655 if (io_register_op_must_quiesce(opcode)) {
10656 ret = io_ctx_quiesce(ctx);
10662 case IORING_REGISTER_BUFFERS:
10663 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10665 case IORING_UNREGISTER_BUFFERS:
10667 if (arg || nr_args)
10669 ret = io_sqe_buffers_unregister(ctx);
10671 case IORING_REGISTER_FILES:
10672 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10674 case IORING_UNREGISTER_FILES:
10676 if (arg || nr_args)
10678 ret = io_sqe_files_unregister(ctx);
10680 case IORING_REGISTER_FILES_UPDATE:
10681 ret = io_register_files_update(ctx, arg, nr_args);
10683 case IORING_REGISTER_EVENTFD:
10684 case IORING_REGISTER_EVENTFD_ASYNC:
10688 ret = io_eventfd_register(ctx, arg);
10691 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10692 ctx->eventfd_async = 1;
10694 ctx->eventfd_async = 0;
10696 case IORING_UNREGISTER_EVENTFD:
10698 if (arg || nr_args)
10700 ret = io_eventfd_unregister(ctx);
10702 case IORING_REGISTER_PROBE:
10704 if (!arg || nr_args > 256)
10706 ret = io_probe(ctx, arg, nr_args);
10708 case IORING_REGISTER_PERSONALITY:
10710 if (arg || nr_args)
10712 ret = io_register_personality(ctx);
10714 case IORING_UNREGISTER_PERSONALITY:
10718 ret = io_unregister_personality(ctx, nr_args);
10720 case IORING_REGISTER_ENABLE_RINGS:
10722 if (arg || nr_args)
10724 ret = io_register_enable_rings(ctx);
10726 case IORING_REGISTER_RESTRICTIONS:
10727 ret = io_register_restrictions(ctx, arg, nr_args);
10729 case IORING_REGISTER_FILES2:
10730 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10732 case IORING_REGISTER_FILES_UPDATE2:
10733 ret = io_register_rsrc_update(ctx, arg, nr_args,
10736 case IORING_REGISTER_BUFFERS2:
10737 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10739 case IORING_REGISTER_BUFFERS_UPDATE:
10740 ret = io_register_rsrc_update(ctx, arg, nr_args,
10741 IORING_RSRC_BUFFER);
10743 case IORING_REGISTER_IOWQ_AFF:
10745 if (!arg || !nr_args)
10747 ret = io_register_iowq_aff(ctx, arg, nr_args);
10749 case IORING_UNREGISTER_IOWQ_AFF:
10751 if (arg || nr_args)
10753 ret = io_unregister_iowq_aff(ctx);
10755 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10757 if (!arg || nr_args != 2)
10759 ret = io_register_iowq_max_workers(ctx, arg);
10766 if (io_register_op_must_quiesce(opcode)) {
10767 /* bring the ctx back to life */
10768 percpu_ref_reinit(&ctx->refs);
10769 reinit_completion(&ctx->ref_comp);
10774 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10775 void __user *, arg, unsigned int, nr_args)
10777 struct io_ring_ctx *ctx;
10786 if (f.file->f_op != &io_uring_fops)
10789 ctx = f.file->private_data;
10791 io_run_task_work();
10793 mutex_lock(&ctx->uring_lock);
10794 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10795 mutex_unlock(&ctx->uring_lock);
10796 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10797 ctx->cq_ev_fd != NULL, ret);
10803 static int __init io_uring_init(void)
10805 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10806 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10807 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10810 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10811 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10812 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10813 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10814 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10815 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10816 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10817 BUILD_BUG_SQE_ELEM(8, __u64, off);
10818 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10819 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10820 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10821 BUILD_BUG_SQE_ELEM(24, __u32, len);
10822 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10823 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10824 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10825 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10826 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10827 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10828 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10829 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10830 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10831 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10832 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10833 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10834 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10835 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10836 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10837 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10838 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10839 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10840 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10841 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10842 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10844 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10845 sizeof(struct io_uring_rsrc_update));
10846 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10847 sizeof(struct io_uring_rsrc_update2));
10849 /* ->buf_index is u16 */
10850 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10852 /* should fit into one byte */
10853 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10855 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10856 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10858 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10862 __initcall(io_uring_init);