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_revert(iter, io_size - iov_iter_count(iter));
3485 } else if (ret == -EIOCBQUEUED) {
3487 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3488 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3489 /* read all, failed, already did sync or don't want to retry */
3493 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3498 rw = req->async_data;
3499 /* now use our persistent iterator, if we aren't already */
3504 rw->bytes_done += ret;
3505 /* if we can retry, do so with the callbacks armed */
3506 if (!io_rw_should_retry(req)) {
3507 kiocb->ki_flags &= ~IOCB_WAITQ;
3512 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3513 * we get -EIOCBQUEUED, then we'll get a notification when the
3514 * desired page gets unlocked. We can also get a partial read
3515 * here, and if we do, then just retry at the new offset.
3517 ret = io_iter_do_read(req, iter);
3518 if (ret == -EIOCBQUEUED)
3520 /* we got some bytes, but not all. retry. */
3521 kiocb->ki_flags &= ~IOCB_WAITQ;
3522 } while (ret > 0 && ret < io_size);
3524 kiocb_done(kiocb, ret, issue_flags);
3526 /* it's faster to check here then delegate to kfree */
3532 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3534 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3536 return io_prep_rw(req, sqe);
3539 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3541 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3542 struct kiocb *kiocb = &req->rw.kiocb;
3543 struct iov_iter __iter, *iter = &__iter;
3544 struct io_async_rw *rw = req->async_data;
3545 ssize_t ret, ret2, io_size;
3546 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3552 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3556 io_size = iov_iter_count(iter);
3557 req->result = io_size;
3559 /* Ensure we clear previously set non-block flag */
3560 if (!force_nonblock)
3561 kiocb->ki_flags &= ~IOCB_NOWAIT;
3563 kiocb->ki_flags |= IOCB_NOWAIT;
3565 /* If the file doesn't support async, just async punt */
3566 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3569 /* file path doesn't support NOWAIT for non-direct_IO */
3570 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3571 (req->flags & REQ_F_ISREG))
3574 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3579 * Open-code file_start_write here to grab freeze protection,
3580 * which will be released by another thread in
3581 * io_complete_rw(). Fool lockdep by telling it the lock got
3582 * released so that it doesn't complain about the held lock when
3583 * we return to userspace.
3585 if (req->flags & REQ_F_ISREG) {
3586 sb_start_write(file_inode(req->file)->i_sb);
3587 __sb_writers_release(file_inode(req->file)->i_sb,
3590 kiocb->ki_flags |= IOCB_WRITE;
3592 if (req->file->f_op->write_iter)
3593 ret2 = call_write_iter(req->file, kiocb, iter);
3594 else if (req->file->f_op->write)
3595 ret2 = loop_rw_iter(WRITE, req, iter);
3599 if (req->flags & REQ_F_REISSUE) {
3600 req->flags &= ~REQ_F_REISSUE;
3605 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3606 * retry them without IOCB_NOWAIT.
3608 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3610 /* no retry on NONBLOCK nor RWF_NOWAIT */
3611 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3613 if (!force_nonblock || ret2 != -EAGAIN) {
3614 /* IOPOLL retry should happen for io-wq threads */
3615 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3618 kiocb_done(kiocb, ret2, issue_flags);
3621 /* some cases will consume bytes even on error returns */
3622 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3623 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3624 return ret ?: -EAGAIN;
3627 /* it's reportedly faster than delegating the null check to kfree() */
3633 static int io_renameat_prep(struct io_kiocb *req,
3634 const struct io_uring_sqe *sqe)
3636 struct io_rename *ren = &req->rename;
3637 const char __user *oldf, *newf;
3639 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3641 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3643 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3646 ren->old_dfd = READ_ONCE(sqe->fd);
3647 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3648 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3649 ren->new_dfd = READ_ONCE(sqe->len);
3650 ren->flags = READ_ONCE(sqe->rename_flags);
3652 ren->oldpath = getname(oldf);
3653 if (IS_ERR(ren->oldpath))
3654 return PTR_ERR(ren->oldpath);
3656 ren->newpath = getname(newf);
3657 if (IS_ERR(ren->newpath)) {
3658 putname(ren->oldpath);
3659 return PTR_ERR(ren->newpath);
3662 req->flags |= REQ_F_NEED_CLEANUP;
3666 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3668 struct io_rename *ren = &req->rename;
3671 if (issue_flags & IO_URING_F_NONBLOCK)
3674 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3675 ren->newpath, ren->flags);
3677 req->flags &= ~REQ_F_NEED_CLEANUP;
3680 io_req_complete(req, ret);
3684 static int io_unlinkat_prep(struct io_kiocb *req,
3685 const struct io_uring_sqe *sqe)
3687 struct io_unlink *un = &req->unlink;
3688 const char __user *fname;
3690 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3692 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3695 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3698 un->dfd = READ_ONCE(sqe->fd);
3700 un->flags = READ_ONCE(sqe->unlink_flags);
3701 if (un->flags & ~AT_REMOVEDIR)
3704 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3705 un->filename = getname(fname);
3706 if (IS_ERR(un->filename))
3707 return PTR_ERR(un->filename);
3709 req->flags |= REQ_F_NEED_CLEANUP;
3713 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3715 struct io_unlink *un = &req->unlink;
3718 if (issue_flags & IO_URING_F_NONBLOCK)
3721 if (un->flags & AT_REMOVEDIR)
3722 ret = do_rmdir(un->dfd, un->filename);
3724 ret = do_unlinkat(un->dfd, un->filename);
3726 req->flags &= ~REQ_F_NEED_CLEANUP;
3729 io_req_complete(req, ret);
3733 static int io_mkdirat_prep(struct io_kiocb *req,
3734 const struct io_uring_sqe *sqe)
3736 struct io_mkdir *mkd = &req->mkdir;
3737 const char __user *fname;
3739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3741 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3744 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3747 mkd->dfd = READ_ONCE(sqe->fd);
3748 mkd->mode = READ_ONCE(sqe->len);
3750 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3751 mkd->filename = getname(fname);
3752 if (IS_ERR(mkd->filename))
3753 return PTR_ERR(mkd->filename);
3755 req->flags |= REQ_F_NEED_CLEANUP;
3759 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3761 struct io_mkdir *mkd = &req->mkdir;
3764 if (issue_flags & IO_URING_F_NONBLOCK)
3767 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3769 req->flags &= ~REQ_F_NEED_CLEANUP;
3772 io_req_complete(req, ret);
3776 static int io_symlinkat_prep(struct io_kiocb *req,
3777 const struct io_uring_sqe *sqe)
3779 struct io_symlink *sl = &req->symlink;
3780 const char __user *oldpath, *newpath;
3782 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3784 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3787 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3790 sl->new_dfd = READ_ONCE(sqe->fd);
3791 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3792 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3794 sl->oldpath = getname(oldpath);
3795 if (IS_ERR(sl->oldpath))
3796 return PTR_ERR(sl->oldpath);
3798 sl->newpath = getname(newpath);
3799 if (IS_ERR(sl->newpath)) {
3800 putname(sl->oldpath);
3801 return PTR_ERR(sl->newpath);
3804 req->flags |= REQ_F_NEED_CLEANUP;
3808 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3810 struct io_symlink *sl = &req->symlink;
3813 if (issue_flags & IO_URING_F_NONBLOCK)
3816 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3818 req->flags &= ~REQ_F_NEED_CLEANUP;
3821 io_req_complete(req, ret);
3825 static int io_linkat_prep(struct io_kiocb *req,
3826 const struct io_uring_sqe *sqe)
3828 struct io_hardlink *lnk = &req->hardlink;
3829 const char __user *oldf, *newf;
3831 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3833 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3835 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3838 lnk->old_dfd = READ_ONCE(sqe->fd);
3839 lnk->new_dfd = READ_ONCE(sqe->len);
3840 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3841 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3842 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3844 lnk->oldpath = getname(oldf);
3845 if (IS_ERR(lnk->oldpath))
3846 return PTR_ERR(lnk->oldpath);
3848 lnk->newpath = getname(newf);
3849 if (IS_ERR(lnk->newpath)) {
3850 putname(lnk->oldpath);
3851 return PTR_ERR(lnk->newpath);
3854 req->flags |= REQ_F_NEED_CLEANUP;
3858 static int io_linkat(struct io_kiocb *req, int issue_flags)
3860 struct io_hardlink *lnk = &req->hardlink;
3863 if (issue_flags & IO_URING_F_NONBLOCK)
3866 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3867 lnk->newpath, lnk->flags);
3869 req->flags &= ~REQ_F_NEED_CLEANUP;
3872 io_req_complete(req, ret);
3876 static int io_shutdown_prep(struct io_kiocb *req,
3877 const struct io_uring_sqe *sqe)
3879 #if defined(CONFIG_NET)
3880 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3882 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3883 sqe->buf_index || sqe->splice_fd_in))
3886 req->shutdown.how = READ_ONCE(sqe->len);
3893 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3895 #if defined(CONFIG_NET)
3896 struct socket *sock;
3899 if (issue_flags & IO_URING_F_NONBLOCK)
3902 sock = sock_from_file(req->file);
3903 if (unlikely(!sock))
3906 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3909 io_req_complete(req, ret);
3916 static int __io_splice_prep(struct io_kiocb *req,
3917 const struct io_uring_sqe *sqe)
3919 struct io_splice *sp = &req->splice;
3920 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3922 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3926 sp->len = READ_ONCE(sqe->len);
3927 sp->flags = READ_ONCE(sqe->splice_flags);
3929 if (unlikely(sp->flags & ~valid_flags))
3932 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3933 (sp->flags & SPLICE_F_FD_IN_FIXED));
3936 req->flags |= REQ_F_NEED_CLEANUP;
3940 static int io_tee_prep(struct io_kiocb *req,
3941 const struct io_uring_sqe *sqe)
3943 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3945 return __io_splice_prep(req, sqe);
3948 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3950 struct io_splice *sp = &req->splice;
3951 struct file *in = sp->file_in;
3952 struct file *out = sp->file_out;
3953 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3956 if (issue_flags & IO_URING_F_NONBLOCK)
3959 ret = do_tee(in, out, sp->len, flags);
3961 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3963 req->flags &= ~REQ_F_NEED_CLEANUP;
3967 io_req_complete(req, ret);
3971 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3973 struct io_splice *sp = &req->splice;
3975 sp->off_in = READ_ONCE(sqe->splice_off_in);
3976 sp->off_out = READ_ONCE(sqe->off);
3977 return __io_splice_prep(req, sqe);
3980 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3982 struct io_splice *sp = &req->splice;
3983 struct file *in = sp->file_in;
3984 struct file *out = sp->file_out;
3985 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3986 loff_t *poff_in, *poff_out;
3989 if (issue_flags & IO_URING_F_NONBLOCK)
3992 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3993 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3996 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3998 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4000 req->flags &= ~REQ_F_NEED_CLEANUP;
4004 io_req_complete(req, ret);
4009 * IORING_OP_NOP just posts a completion event, nothing else.
4011 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4013 struct io_ring_ctx *ctx = req->ctx;
4015 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4018 __io_req_complete(req, issue_flags, 0, 0);
4022 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4024 struct io_ring_ctx *ctx = req->ctx;
4029 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4031 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4035 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4036 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4039 req->sync.off = READ_ONCE(sqe->off);
4040 req->sync.len = READ_ONCE(sqe->len);
4044 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4046 loff_t end = req->sync.off + req->sync.len;
4049 /* fsync always requires a blocking context */
4050 if (issue_flags & IO_URING_F_NONBLOCK)
4053 ret = vfs_fsync_range(req->file, req->sync.off,
4054 end > 0 ? end : LLONG_MAX,
4055 req->sync.flags & IORING_FSYNC_DATASYNC);
4058 io_req_complete(req, ret);
4062 static int io_fallocate_prep(struct io_kiocb *req,
4063 const struct io_uring_sqe *sqe)
4065 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4068 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4071 req->sync.off = READ_ONCE(sqe->off);
4072 req->sync.len = READ_ONCE(sqe->addr);
4073 req->sync.mode = READ_ONCE(sqe->len);
4077 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4081 /* fallocate always requiring blocking context */
4082 if (issue_flags & IO_URING_F_NONBLOCK)
4084 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4088 io_req_complete(req, ret);
4092 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4094 const char __user *fname;
4097 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4099 if (unlikely(sqe->ioprio || sqe->buf_index))
4101 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4104 /* open.how should be already initialised */
4105 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4106 req->open.how.flags |= O_LARGEFILE;
4108 req->open.dfd = READ_ONCE(sqe->fd);
4109 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4110 req->open.filename = getname(fname);
4111 if (IS_ERR(req->open.filename)) {
4112 ret = PTR_ERR(req->open.filename);
4113 req->open.filename = NULL;
4117 req->open.file_slot = READ_ONCE(sqe->file_index);
4118 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4121 req->open.nofile = rlimit(RLIMIT_NOFILE);
4122 req->flags |= REQ_F_NEED_CLEANUP;
4126 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4128 u64 mode = READ_ONCE(sqe->len);
4129 u64 flags = READ_ONCE(sqe->open_flags);
4131 req->open.how = build_open_how(flags, mode);
4132 return __io_openat_prep(req, sqe);
4135 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4137 struct open_how __user *how;
4141 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4142 len = READ_ONCE(sqe->len);
4143 if (len < OPEN_HOW_SIZE_VER0)
4146 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4151 return __io_openat_prep(req, sqe);
4154 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4156 struct open_flags op;
4158 bool resolve_nonblock, nonblock_set;
4159 bool fixed = !!req->open.file_slot;
4162 ret = build_open_flags(&req->open.how, &op);
4165 nonblock_set = op.open_flag & O_NONBLOCK;
4166 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4167 if (issue_flags & IO_URING_F_NONBLOCK) {
4169 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4170 * it'll always -EAGAIN
4172 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4174 op.lookup_flags |= LOOKUP_CACHED;
4175 op.open_flag |= O_NONBLOCK;
4179 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4184 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4187 * We could hang on to this 'fd' on retrying, but seems like
4188 * marginal gain for something that is now known to be a slower
4189 * path. So just put it, and we'll get a new one when we retry.
4194 ret = PTR_ERR(file);
4195 /* only retry if RESOLVE_CACHED wasn't already set by application */
4196 if (ret == -EAGAIN &&
4197 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4202 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4203 file->f_flags &= ~O_NONBLOCK;
4204 fsnotify_open(file);
4207 fd_install(ret, file);
4209 ret = io_install_fixed_file(req, file, issue_flags,
4210 req->open.file_slot - 1);
4212 putname(req->open.filename);
4213 req->flags &= ~REQ_F_NEED_CLEANUP;
4216 __io_req_complete(req, issue_flags, ret, 0);
4220 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4222 return io_openat2(req, issue_flags);
4225 static int io_remove_buffers_prep(struct io_kiocb *req,
4226 const struct io_uring_sqe *sqe)
4228 struct io_provide_buf *p = &req->pbuf;
4231 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4235 tmp = READ_ONCE(sqe->fd);
4236 if (!tmp || tmp > USHRT_MAX)
4239 memset(p, 0, sizeof(*p));
4241 p->bgid = READ_ONCE(sqe->buf_group);
4245 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4246 int bgid, unsigned nbufs)
4250 /* shouldn't happen */
4254 /* the head kbuf is the list itself */
4255 while (!list_empty(&buf->list)) {
4256 struct io_buffer *nxt;
4258 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4259 list_del(&nxt->list);
4266 xa_erase(&ctx->io_buffers, bgid);
4271 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4273 struct io_provide_buf *p = &req->pbuf;
4274 struct io_ring_ctx *ctx = req->ctx;
4275 struct io_buffer *head;
4277 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4279 io_ring_submit_lock(ctx, !force_nonblock);
4281 lockdep_assert_held(&ctx->uring_lock);
4284 head = xa_load(&ctx->io_buffers, p->bgid);
4286 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4290 /* complete before unlock, IOPOLL may need the lock */
4291 __io_req_complete(req, issue_flags, ret, 0);
4292 io_ring_submit_unlock(ctx, !force_nonblock);
4296 static int io_provide_buffers_prep(struct io_kiocb *req,
4297 const struct io_uring_sqe *sqe)
4299 unsigned long size, tmp_check;
4300 struct io_provide_buf *p = &req->pbuf;
4303 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4306 tmp = READ_ONCE(sqe->fd);
4307 if (!tmp || tmp > USHRT_MAX)
4310 p->addr = READ_ONCE(sqe->addr);
4311 p->len = READ_ONCE(sqe->len);
4313 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4316 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4319 size = (unsigned long)p->len * p->nbufs;
4320 if (!access_ok(u64_to_user_ptr(p->addr), size))
4323 p->bgid = READ_ONCE(sqe->buf_group);
4324 tmp = READ_ONCE(sqe->off);
4325 if (tmp > USHRT_MAX)
4331 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4333 struct io_buffer *buf;
4334 u64 addr = pbuf->addr;
4335 int i, bid = pbuf->bid;
4337 for (i = 0; i < pbuf->nbufs; i++) {
4338 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4343 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4348 INIT_LIST_HEAD(&buf->list);
4351 list_add_tail(&buf->list, &(*head)->list);
4355 return i ? i : -ENOMEM;
4358 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4360 struct io_provide_buf *p = &req->pbuf;
4361 struct io_ring_ctx *ctx = req->ctx;
4362 struct io_buffer *head, *list;
4364 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4366 io_ring_submit_lock(ctx, !force_nonblock);
4368 lockdep_assert_held(&ctx->uring_lock);
4370 list = head = xa_load(&ctx->io_buffers, p->bgid);
4372 ret = io_add_buffers(p, &head);
4373 if (ret >= 0 && !list) {
4374 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4376 __io_remove_buffers(ctx, head, p->bgid, -1U);
4380 /* complete before unlock, IOPOLL may need the lock */
4381 __io_req_complete(req, issue_flags, ret, 0);
4382 io_ring_submit_unlock(ctx, !force_nonblock);
4386 static int io_epoll_ctl_prep(struct io_kiocb *req,
4387 const struct io_uring_sqe *sqe)
4389 #if defined(CONFIG_EPOLL)
4390 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4392 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4395 req->epoll.epfd = READ_ONCE(sqe->fd);
4396 req->epoll.op = READ_ONCE(sqe->len);
4397 req->epoll.fd = READ_ONCE(sqe->off);
4399 if (ep_op_has_event(req->epoll.op)) {
4400 struct epoll_event __user *ev;
4402 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4403 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4413 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4415 #if defined(CONFIG_EPOLL)
4416 struct io_epoll *ie = &req->epoll;
4418 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4420 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4421 if (force_nonblock && ret == -EAGAIN)
4426 __io_req_complete(req, issue_flags, ret, 0);
4433 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4435 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4436 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4438 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4441 req->madvise.addr = READ_ONCE(sqe->addr);
4442 req->madvise.len = READ_ONCE(sqe->len);
4443 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4450 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4452 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4453 struct io_madvise *ma = &req->madvise;
4456 if (issue_flags & IO_URING_F_NONBLOCK)
4459 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4462 io_req_complete(req, ret);
4469 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4471 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4473 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4476 req->fadvise.offset = READ_ONCE(sqe->off);
4477 req->fadvise.len = READ_ONCE(sqe->len);
4478 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4482 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4484 struct io_fadvise *fa = &req->fadvise;
4487 if (issue_flags & IO_URING_F_NONBLOCK) {
4488 switch (fa->advice) {
4489 case POSIX_FADV_NORMAL:
4490 case POSIX_FADV_RANDOM:
4491 case POSIX_FADV_SEQUENTIAL:
4498 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4501 __io_req_complete(req, issue_flags, ret, 0);
4505 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4507 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4509 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4511 if (req->flags & REQ_F_FIXED_FILE)
4514 req->statx.dfd = READ_ONCE(sqe->fd);
4515 req->statx.mask = READ_ONCE(sqe->len);
4516 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4517 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4518 req->statx.flags = READ_ONCE(sqe->statx_flags);
4523 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4525 struct io_statx *ctx = &req->statx;
4528 if (issue_flags & IO_URING_F_NONBLOCK)
4531 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4536 io_req_complete(req, ret);
4540 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4542 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4544 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4545 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4547 if (req->flags & REQ_F_FIXED_FILE)
4550 req->close.fd = READ_ONCE(sqe->fd);
4554 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4556 struct files_struct *files = current->files;
4557 struct io_close *close = &req->close;
4558 struct fdtable *fdt;
4559 struct file *file = NULL;
4562 spin_lock(&files->file_lock);
4563 fdt = files_fdtable(files);
4564 if (close->fd >= fdt->max_fds) {
4565 spin_unlock(&files->file_lock);
4568 file = fdt->fd[close->fd];
4569 if (!file || file->f_op == &io_uring_fops) {
4570 spin_unlock(&files->file_lock);
4575 /* if the file has a flush method, be safe and punt to async */
4576 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4577 spin_unlock(&files->file_lock);
4581 ret = __close_fd_get_file(close->fd, &file);
4582 spin_unlock(&files->file_lock);
4589 /* No ->flush() or already async, safely close from here */
4590 ret = filp_close(file, current->files);
4596 __io_req_complete(req, issue_flags, ret, 0);
4600 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4602 struct io_ring_ctx *ctx = req->ctx;
4604 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4606 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4610 req->sync.off = READ_ONCE(sqe->off);
4611 req->sync.len = READ_ONCE(sqe->len);
4612 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4616 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4620 /* sync_file_range always requires a blocking context */
4621 if (issue_flags & IO_URING_F_NONBLOCK)
4624 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4628 io_req_complete(req, ret);
4632 #if defined(CONFIG_NET)
4633 static int io_setup_async_msg(struct io_kiocb *req,
4634 struct io_async_msghdr *kmsg)
4636 struct io_async_msghdr *async_msg = req->async_data;
4640 if (io_alloc_async_data(req)) {
4641 kfree(kmsg->free_iov);
4644 async_msg = req->async_data;
4645 req->flags |= REQ_F_NEED_CLEANUP;
4646 memcpy(async_msg, kmsg, sizeof(*kmsg));
4647 async_msg->msg.msg_name = &async_msg->addr;
4648 /* if were using fast_iov, set it to the new one */
4649 if (!async_msg->free_iov)
4650 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4655 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4656 struct io_async_msghdr *iomsg)
4658 iomsg->msg.msg_name = &iomsg->addr;
4659 iomsg->free_iov = iomsg->fast_iov;
4660 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4661 req->sr_msg.msg_flags, &iomsg->free_iov);
4664 static int io_sendmsg_prep_async(struct io_kiocb *req)
4668 ret = io_sendmsg_copy_hdr(req, req->async_data);
4670 req->flags |= REQ_F_NEED_CLEANUP;
4674 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4676 struct io_sr_msg *sr = &req->sr_msg;
4678 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4681 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4682 sr->len = READ_ONCE(sqe->len);
4683 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4684 if (sr->msg_flags & MSG_DONTWAIT)
4685 req->flags |= REQ_F_NOWAIT;
4687 #ifdef CONFIG_COMPAT
4688 if (req->ctx->compat)
4689 sr->msg_flags |= MSG_CMSG_COMPAT;
4694 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4696 struct io_async_msghdr iomsg, *kmsg;
4697 struct socket *sock;
4702 sock = sock_from_file(req->file);
4703 if (unlikely(!sock))
4706 kmsg = req->async_data;
4708 ret = io_sendmsg_copy_hdr(req, &iomsg);
4714 flags = req->sr_msg.msg_flags;
4715 if (issue_flags & IO_URING_F_NONBLOCK)
4716 flags |= MSG_DONTWAIT;
4717 if (flags & MSG_WAITALL)
4718 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4720 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4721 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4722 return io_setup_async_msg(req, kmsg);
4723 if (ret == -ERESTARTSYS)
4726 /* fast path, check for non-NULL to avoid function call */
4728 kfree(kmsg->free_iov);
4729 req->flags &= ~REQ_F_NEED_CLEANUP;
4732 __io_req_complete(req, issue_flags, ret, 0);
4736 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4738 struct io_sr_msg *sr = &req->sr_msg;
4741 struct socket *sock;
4746 sock = sock_from_file(req->file);
4747 if (unlikely(!sock))
4750 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4754 msg.msg_name = NULL;
4755 msg.msg_control = NULL;
4756 msg.msg_controllen = 0;
4757 msg.msg_namelen = 0;
4759 flags = req->sr_msg.msg_flags;
4760 if (issue_flags & IO_URING_F_NONBLOCK)
4761 flags |= MSG_DONTWAIT;
4762 if (flags & MSG_WAITALL)
4763 min_ret = iov_iter_count(&msg.msg_iter);
4765 msg.msg_flags = flags;
4766 ret = sock_sendmsg(sock, &msg);
4767 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4769 if (ret == -ERESTARTSYS)
4774 __io_req_complete(req, issue_flags, ret, 0);
4778 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4779 struct io_async_msghdr *iomsg)
4781 struct io_sr_msg *sr = &req->sr_msg;
4782 struct iovec __user *uiov;
4786 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4787 &iomsg->uaddr, &uiov, &iov_len);
4791 if (req->flags & REQ_F_BUFFER_SELECT) {
4794 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4796 sr->len = iomsg->fast_iov[0].iov_len;
4797 iomsg->free_iov = NULL;
4799 iomsg->free_iov = iomsg->fast_iov;
4800 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4801 &iomsg->free_iov, &iomsg->msg.msg_iter,
4810 #ifdef CONFIG_COMPAT
4811 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4812 struct io_async_msghdr *iomsg)
4814 struct io_sr_msg *sr = &req->sr_msg;
4815 struct compat_iovec __user *uiov;
4820 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4825 uiov = compat_ptr(ptr);
4826 if (req->flags & REQ_F_BUFFER_SELECT) {
4827 compat_ssize_t clen;
4831 if (!access_ok(uiov, sizeof(*uiov)))
4833 if (__get_user(clen, &uiov->iov_len))
4838 iomsg->free_iov = NULL;
4840 iomsg->free_iov = iomsg->fast_iov;
4841 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4842 UIO_FASTIOV, &iomsg->free_iov,
4843 &iomsg->msg.msg_iter, true);
4852 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4853 struct io_async_msghdr *iomsg)
4855 iomsg->msg.msg_name = &iomsg->addr;
4857 #ifdef CONFIG_COMPAT
4858 if (req->ctx->compat)
4859 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4862 return __io_recvmsg_copy_hdr(req, iomsg);
4865 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4868 struct io_sr_msg *sr = &req->sr_msg;
4869 struct io_buffer *kbuf;
4871 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4876 req->flags |= REQ_F_BUFFER_SELECTED;
4880 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4882 return io_put_kbuf(req, req->sr_msg.kbuf);
4885 static int io_recvmsg_prep_async(struct io_kiocb *req)
4889 ret = io_recvmsg_copy_hdr(req, req->async_data);
4891 req->flags |= REQ_F_NEED_CLEANUP;
4895 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4897 struct io_sr_msg *sr = &req->sr_msg;
4899 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4902 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4903 sr->len = READ_ONCE(sqe->len);
4904 sr->bgid = READ_ONCE(sqe->buf_group);
4905 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4906 if (sr->msg_flags & MSG_DONTWAIT)
4907 req->flags |= REQ_F_NOWAIT;
4909 #ifdef CONFIG_COMPAT
4910 if (req->ctx->compat)
4911 sr->msg_flags |= MSG_CMSG_COMPAT;
4916 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4918 struct io_async_msghdr iomsg, *kmsg;
4919 struct socket *sock;
4920 struct io_buffer *kbuf;
4923 int ret, cflags = 0;
4924 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4926 sock = sock_from_file(req->file);
4927 if (unlikely(!sock))
4930 kmsg = req->async_data;
4932 ret = io_recvmsg_copy_hdr(req, &iomsg);
4938 if (req->flags & REQ_F_BUFFER_SELECT) {
4939 kbuf = io_recv_buffer_select(req, !force_nonblock);
4941 return PTR_ERR(kbuf);
4942 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4943 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4944 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4945 1, req->sr_msg.len);
4948 flags = req->sr_msg.msg_flags;
4950 flags |= MSG_DONTWAIT;
4951 if (flags & MSG_WAITALL)
4952 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4954 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4955 kmsg->uaddr, flags);
4956 if (force_nonblock && ret == -EAGAIN)
4957 return io_setup_async_msg(req, kmsg);
4958 if (ret == -ERESTARTSYS)
4961 if (req->flags & REQ_F_BUFFER_SELECTED)
4962 cflags = io_put_recv_kbuf(req);
4963 /* fast path, check for non-NULL to avoid function call */
4965 kfree(kmsg->free_iov);
4966 req->flags &= ~REQ_F_NEED_CLEANUP;
4967 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4969 __io_req_complete(req, issue_flags, ret, cflags);
4973 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4975 struct io_buffer *kbuf;
4976 struct io_sr_msg *sr = &req->sr_msg;
4978 void __user *buf = sr->buf;
4979 struct socket *sock;
4983 int ret, cflags = 0;
4984 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4986 sock = sock_from_file(req->file);
4987 if (unlikely(!sock))
4990 if (req->flags & REQ_F_BUFFER_SELECT) {
4991 kbuf = io_recv_buffer_select(req, !force_nonblock);
4993 return PTR_ERR(kbuf);
4994 buf = u64_to_user_ptr(kbuf->addr);
4997 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5001 msg.msg_name = NULL;
5002 msg.msg_control = NULL;
5003 msg.msg_controllen = 0;
5004 msg.msg_namelen = 0;
5005 msg.msg_iocb = NULL;
5008 flags = req->sr_msg.msg_flags;
5010 flags |= MSG_DONTWAIT;
5011 if (flags & MSG_WAITALL)
5012 min_ret = iov_iter_count(&msg.msg_iter);
5014 ret = sock_recvmsg(sock, &msg, flags);
5015 if (force_nonblock && ret == -EAGAIN)
5017 if (ret == -ERESTARTSYS)
5020 if (req->flags & REQ_F_BUFFER_SELECTED)
5021 cflags = io_put_recv_kbuf(req);
5022 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5024 __io_req_complete(req, issue_flags, ret, cflags);
5028 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5030 struct io_accept *accept = &req->accept;
5032 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5034 if (sqe->ioprio || sqe->len || sqe->buf_index)
5037 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5038 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5039 accept->flags = READ_ONCE(sqe->accept_flags);
5040 accept->nofile = rlimit(RLIMIT_NOFILE);
5042 accept->file_slot = READ_ONCE(sqe->file_index);
5043 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5044 (accept->flags & SOCK_CLOEXEC)))
5046 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5048 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5049 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5053 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5055 struct io_accept *accept = &req->accept;
5056 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5057 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5058 bool fixed = !!accept->file_slot;
5062 if (req->file->f_flags & O_NONBLOCK)
5063 req->flags |= REQ_F_NOWAIT;
5066 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5067 if (unlikely(fd < 0))
5070 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5075 ret = PTR_ERR(file);
5076 if (ret == -EAGAIN && force_nonblock)
5078 if (ret == -ERESTARTSYS)
5081 } else if (!fixed) {
5082 fd_install(fd, file);
5085 ret = io_install_fixed_file(req, file, issue_flags,
5086 accept->file_slot - 1);
5088 __io_req_complete(req, issue_flags, ret, 0);
5092 static int io_connect_prep_async(struct io_kiocb *req)
5094 struct io_async_connect *io = req->async_data;
5095 struct io_connect *conn = &req->connect;
5097 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5100 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5102 struct io_connect *conn = &req->connect;
5104 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5106 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5110 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5111 conn->addr_len = READ_ONCE(sqe->addr2);
5115 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5117 struct io_async_connect __io, *io;
5118 unsigned file_flags;
5120 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5122 if (req->async_data) {
5123 io = req->async_data;
5125 ret = move_addr_to_kernel(req->connect.addr,
5126 req->connect.addr_len,
5133 file_flags = force_nonblock ? O_NONBLOCK : 0;
5135 ret = __sys_connect_file(req->file, &io->address,
5136 req->connect.addr_len, file_flags);
5137 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5138 if (req->async_data)
5140 if (io_alloc_async_data(req)) {
5144 memcpy(req->async_data, &__io, sizeof(__io));
5147 if (ret == -ERESTARTSYS)
5152 __io_req_complete(req, issue_flags, ret, 0);
5155 #else /* !CONFIG_NET */
5156 #define IO_NETOP_FN(op) \
5157 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5159 return -EOPNOTSUPP; \
5162 #define IO_NETOP_PREP(op) \
5164 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5166 return -EOPNOTSUPP; \
5169 #define IO_NETOP_PREP_ASYNC(op) \
5171 static int io_##op##_prep_async(struct io_kiocb *req) \
5173 return -EOPNOTSUPP; \
5176 IO_NETOP_PREP_ASYNC(sendmsg);
5177 IO_NETOP_PREP_ASYNC(recvmsg);
5178 IO_NETOP_PREP_ASYNC(connect);
5179 IO_NETOP_PREP(accept);
5182 #endif /* CONFIG_NET */
5184 struct io_poll_table {
5185 struct poll_table_struct pt;
5186 struct io_kiocb *req;
5191 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5192 __poll_t mask, io_req_tw_func_t func)
5194 /* for instances that support it check for an event match first: */
5195 if (mask && !(mask & poll->events))
5198 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5200 list_del_init(&poll->wait.entry);
5203 req->io_task_work.func = func;
5206 * If this fails, then the task is exiting. When a task exits, the
5207 * work gets canceled, so just cancel this request as well instead
5208 * of executing it. We can't safely execute it anyway, as we may not
5209 * have the needed state needed for it anyway.
5211 io_req_task_work_add(req);
5215 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5216 __acquires(&req->ctx->completion_lock)
5218 struct io_ring_ctx *ctx = req->ctx;
5220 /* req->task == current here, checking PF_EXITING is safe */
5221 if (unlikely(req->task->flags & PF_EXITING))
5222 WRITE_ONCE(poll->canceled, true);
5224 if (!req->result && !READ_ONCE(poll->canceled)) {
5225 struct poll_table_struct pt = { ._key = poll->events };
5227 req->result = vfs_poll(req->file, &pt) & poll->events;
5230 spin_lock(&ctx->completion_lock);
5231 if (!req->result && !READ_ONCE(poll->canceled)) {
5232 add_wait_queue(poll->head, &poll->wait);
5239 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5241 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5242 if (req->opcode == IORING_OP_POLL_ADD)
5243 return req->async_data;
5244 return req->apoll->double_poll;
5247 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5249 if (req->opcode == IORING_OP_POLL_ADD)
5251 return &req->apoll->poll;
5254 static void io_poll_remove_double(struct io_kiocb *req)
5255 __must_hold(&req->ctx->completion_lock)
5257 struct io_poll_iocb *poll = io_poll_get_double(req);
5259 lockdep_assert_held(&req->ctx->completion_lock);
5261 if (poll && poll->head) {
5262 struct wait_queue_head *head = poll->head;
5264 spin_lock_irq(&head->lock);
5265 list_del_init(&poll->wait.entry);
5266 if (poll->wait.private)
5269 spin_unlock_irq(&head->lock);
5273 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5274 __must_hold(&req->ctx->completion_lock)
5276 struct io_ring_ctx *ctx = req->ctx;
5277 unsigned flags = IORING_CQE_F_MORE;
5280 if (READ_ONCE(req->poll.canceled)) {
5282 req->poll.events |= EPOLLONESHOT;
5284 error = mangle_poll(mask);
5286 if (req->poll.events & EPOLLONESHOT)
5288 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5289 req->poll.done = true;
5292 if (flags & IORING_CQE_F_MORE)
5295 return !(flags & IORING_CQE_F_MORE);
5298 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5299 __must_hold(&req->ctx->completion_lock)
5303 done = __io_poll_complete(req, mask);
5304 io_commit_cqring(req->ctx);
5308 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5310 struct io_ring_ctx *ctx = req->ctx;
5311 struct io_kiocb *nxt;
5313 if (io_poll_rewait(req, &req->poll)) {
5314 spin_unlock(&ctx->completion_lock);
5318 done = __io_poll_complete(req, req->result);
5320 io_poll_remove_double(req);
5321 hash_del(&req->hash_node);
5324 add_wait_queue(req->poll.head, &req->poll.wait);
5326 io_commit_cqring(ctx);
5327 spin_unlock(&ctx->completion_lock);
5328 io_cqring_ev_posted(ctx);
5331 nxt = io_put_req_find_next(req);
5333 io_req_task_submit(nxt, locked);
5338 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5339 int sync, void *key)
5341 struct io_kiocb *req = wait->private;
5342 struct io_poll_iocb *poll = io_poll_get_single(req);
5343 __poll_t mask = key_to_poll(key);
5344 unsigned long flags;
5346 /* for instances that support it check for an event match first: */
5347 if (mask && !(mask & poll->events))
5349 if (!(poll->events & EPOLLONESHOT))
5350 return poll->wait.func(&poll->wait, mode, sync, key);
5352 list_del_init(&wait->entry);
5357 spin_lock_irqsave(&poll->head->lock, flags);
5358 done = list_empty(&poll->wait.entry);
5360 list_del_init(&poll->wait.entry);
5361 /* make sure double remove sees this as being gone */
5362 wait->private = NULL;
5363 spin_unlock_irqrestore(&poll->head->lock, flags);
5365 /* use wait func handler, so it matches the rq type */
5366 poll->wait.func(&poll->wait, mode, sync, key);
5373 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5374 wait_queue_func_t wake_func)
5378 poll->canceled = false;
5379 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5380 /* mask in events that we always want/need */
5381 poll->events = events | IO_POLL_UNMASK;
5382 INIT_LIST_HEAD(&poll->wait.entry);
5383 init_waitqueue_func_entry(&poll->wait, wake_func);
5386 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5387 struct wait_queue_head *head,
5388 struct io_poll_iocb **poll_ptr)
5390 struct io_kiocb *req = pt->req;
5393 * The file being polled uses multiple waitqueues for poll handling
5394 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5397 if (unlikely(pt->nr_entries)) {
5398 struct io_poll_iocb *poll_one = poll;
5400 /* double add on the same waitqueue head, ignore */
5401 if (poll_one->head == head)
5403 /* already have a 2nd entry, fail a third attempt */
5405 if ((*poll_ptr)->head == head)
5407 pt->error = -EINVAL;
5411 * Can't handle multishot for double wait for now, turn it
5412 * into one-shot mode.
5414 if (!(poll_one->events & EPOLLONESHOT))
5415 poll_one->events |= EPOLLONESHOT;
5416 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5418 pt->error = -ENOMEM;
5421 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5423 poll->wait.private = req;
5430 if (poll->events & EPOLLEXCLUSIVE)
5431 add_wait_queue_exclusive(head, &poll->wait);
5433 add_wait_queue(head, &poll->wait);
5436 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5437 struct poll_table_struct *p)
5439 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5440 struct async_poll *apoll = pt->req->apoll;
5442 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5445 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5447 struct async_poll *apoll = req->apoll;
5448 struct io_ring_ctx *ctx = req->ctx;
5450 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5452 if (io_poll_rewait(req, &apoll->poll)) {
5453 spin_unlock(&ctx->completion_lock);
5457 hash_del(&req->hash_node);
5458 io_poll_remove_double(req);
5459 spin_unlock(&ctx->completion_lock);
5461 if (!READ_ONCE(apoll->poll.canceled))
5462 io_req_task_submit(req, locked);
5464 io_req_complete_failed(req, -ECANCELED);
5467 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5470 struct io_kiocb *req = wait->private;
5471 struct io_poll_iocb *poll = &req->apoll->poll;
5473 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5476 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5479 static void io_poll_req_insert(struct io_kiocb *req)
5481 struct io_ring_ctx *ctx = req->ctx;
5482 struct hlist_head *list;
5484 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5485 hlist_add_head(&req->hash_node, list);
5488 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5489 struct io_poll_iocb *poll,
5490 struct io_poll_table *ipt, __poll_t mask,
5491 wait_queue_func_t wake_func)
5492 __acquires(&ctx->completion_lock)
5494 struct io_ring_ctx *ctx = req->ctx;
5495 bool cancel = false;
5497 INIT_HLIST_NODE(&req->hash_node);
5498 io_init_poll_iocb(poll, mask, wake_func);
5499 poll->file = req->file;
5500 poll->wait.private = req;
5502 ipt->pt._key = mask;
5505 ipt->nr_entries = 0;
5507 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5508 if (unlikely(!ipt->nr_entries) && !ipt->error)
5509 ipt->error = -EINVAL;
5511 spin_lock(&ctx->completion_lock);
5512 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5513 io_poll_remove_double(req);
5514 if (likely(poll->head)) {
5515 spin_lock_irq(&poll->head->lock);
5516 if (unlikely(list_empty(&poll->wait.entry))) {
5522 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5523 list_del_init(&poll->wait.entry);
5525 WRITE_ONCE(poll->canceled, true);
5526 else if (!poll->done) /* actually waiting for an event */
5527 io_poll_req_insert(req);
5528 spin_unlock_irq(&poll->head->lock);
5540 static int io_arm_poll_handler(struct io_kiocb *req)
5542 const struct io_op_def *def = &io_op_defs[req->opcode];
5543 struct io_ring_ctx *ctx = req->ctx;
5544 struct async_poll *apoll;
5545 struct io_poll_table ipt;
5546 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5549 if (!req->file || !file_can_poll(req->file))
5550 return IO_APOLL_ABORTED;
5551 if (req->flags & REQ_F_POLLED)
5552 return IO_APOLL_ABORTED;
5553 if (!def->pollin && !def->pollout)
5554 return IO_APOLL_ABORTED;
5558 mask |= POLLIN | POLLRDNORM;
5560 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5561 if ((req->opcode == IORING_OP_RECVMSG) &&
5562 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5566 mask |= POLLOUT | POLLWRNORM;
5569 /* if we can't nonblock try, then no point in arming a poll handler */
5570 if (!io_file_supports_nowait(req, rw))
5571 return IO_APOLL_ABORTED;
5573 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5574 if (unlikely(!apoll))
5575 return IO_APOLL_ABORTED;
5576 apoll->double_poll = NULL;
5578 req->flags |= REQ_F_POLLED;
5579 ipt.pt._qproc = io_async_queue_proc;
5580 io_req_set_refcount(req);
5582 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5584 spin_unlock(&ctx->completion_lock);
5585 if (ret || ipt.error)
5586 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5588 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5589 mask, apoll->poll.events);
5593 static bool __io_poll_remove_one(struct io_kiocb *req,
5594 struct io_poll_iocb *poll, bool do_cancel)
5595 __must_hold(&req->ctx->completion_lock)
5597 bool do_complete = false;
5601 spin_lock_irq(&poll->head->lock);
5603 WRITE_ONCE(poll->canceled, true);
5604 if (!list_empty(&poll->wait.entry)) {
5605 list_del_init(&poll->wait.entry);
5608 spin_unlock_irq(&poll->head->lock);
5609 hash_del(&req->hash_node);
5613 static bool io_poll_remove_one(struct io_kiocb *req)
5614 __must_hold(&req->ctx->completion_lock)
5618 io_poll_remove_double(req);
5619 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5622 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5623 io_commit_cqring(req->ctx);
5625 io_put_req_deferred(req);
5631 * Returns true if we found and killed one or more poll requests
5633 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5636 struct hlist_node *tmp;
5637 struct io_kiocb *req;
5640 spin_lock(&ctx->completion_lock);
5641 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5642 struct hlist_head *list;
5644 list = &ctx->cancel_hash[i];
5645 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5646 if (io_match_task(req, tsk, cancel_all))
5647 posted += io_poll_remove_one(req);
5650 spin_unlock(&ctx->completion_lock);
5653 io_cqring_ev_posted(ctx);
5658 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5660 __must_hold(&ctx->completion_lock)
5662 struct hlist_head *list;
5663 struct io_kiocb *req;
5665 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5666 hlist_for_each_entry(req, list, hash_node) {
5667 if (sqe_addr != req->user_data)
5669 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5676 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5678 __must_hold(&ctx->completion_lock)
5680 struct io_kiocb *req;
5682 req = io_poll_find(ctx, sqe_addr, poll_only);
5685 if (io_poll_remove_one(req))
5691 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5696 events = READ_ONCE(sqe->poll32_events);
5698 events = swahw32(events);
5700 if (!(flags & IORING_POLL_ADD_MULTI))
5701 events |= EPOLLONESHOT;
5702 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5705 static int io_poll_update_prep(struct io_kiocb *req,
5706 const struct io_uring_sqe *sqe)
5708 struct io_poll_update *upd = &req->poll_update;
5711 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5713 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5715 flags = READ_ONCE(sqe->len);
5716 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5717 IORING_POLL_ADD_MULTI))
5719 /* meaningless without update */
5720 if (flags == IORING_POLL_ADD_MULTI)
5723 upd->old_user_data = READ_ONCE(sqe->addr);
5724 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5725 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5727 upd->new_user_data = READ_ONCE(sqe->off);
5728 if (!upd->update_user_data && upd->new_user_data)
5730 if (upd->update_events)
5731 upd->events = io_poll_parse_events(sqe, flags);
5732 else if (sqe->poll32_events)
5738 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5741 struct io_kiocb *req = wait->private;
5742 struct io_poll_iocb *poll = &req->poll;
5744 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5747 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5748 struct poll_table_struct *p)
5750 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5752 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5755 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5757 struct io_poll_iocb *poll = &req->poll;
5760 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5762 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5764 flags = READ_ONCE(sqe->len);
5765 if (flags & ~IORING_POLL_ADD_MULTI)
5768 io_req_set_refcount(req);
5769 poll->events = io_poll_parse_events(sqe, flags);
5773 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5775 struct io_poll_iocb *poll = &req->poll;
5776 struct io_ring_ctx *ctx = req->ctx;
5777 struct io_poll_table ipt;
5780 ipt.pt._qproc = io_poll_queue_proc;
5782 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5785 if (mask) { /* no async, we'd stolen it */
5787 io_poll_complete(req, mask);
5789 spin_unlock(&ctx->completion_lock);
5792 io_cqring_ev_posted(ctx);
5793 if (poll->events & EPOLLONESHOT)
5799 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5801 struct io_ring_ctx *ctx = req->ctx;
5802 struct io_kiocb *preq;
5806 spin_lock(&ctx->completion_lock);
5807 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5813 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5815 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5820 * Don't allow racy completion with singleshot, as we cannot safely
5821 * update those. For multishot, if we're racing with completion, just
5822 * let completion re-add it.
5824 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5825 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5829 /* we now have a detached poll request. reissue. */
5833 spin_unlock(&ctx->completion_lock);
5835 io_req_complete(req, ret);
5838 /* only mask one event flags, keep behavior flags */
5839 if (req->poll_update.update_events) {
5840 preq->poll.events &= ~0xffff;
5841 preq->poll.events |= req->poll_update.events & 0xffff;
5842 preq->poll.events |= IO_POLL_UNMASK;
5844 if (req->poll_update.update_user_data)
5845 preq->user_data = req->poll_update.new_user_data;
5846 spin_unlock(&ctx->completion_lock);
5848 /* complete update request, we're done with it */
5849 io_req_complete(req, ret);
5852 ret = io_poll_add(preq, issue_flags);
5855 io_req_complete(preq, ret);
5861 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5864 io_req_complete_post(req, -ETIME, 0);
5867 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5869 struct io_timeout_data *data = container_of(timer,
5870 struct io_timeout_data, timer);
5871 struct io_kiocb *req = data->req;
5872 struct io_ring_ctx *ctx = req->ctx;
5873 unsigned long flags;
5875 spin_lock_irqsave(&ctx->timeout_lock, flags);
5876 list_del_init(&req->timeout.list);
5877 atomic_set(&req->ctx->cq_timeouts,
5878 atomic_read(&req->ctx->cq_timeouts) + 1);
5879 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5881 req->io_task_work.func = io_req_task_timeout;
5882 io_req_task_work_add(req);
5883 return HRTIMER_NORESTART;
5886 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5888 __must_hold(&ctx->timeout_lock)
5890 struct io_timeout_data *io;
5891 struct io_kiocb *req;
5894 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5895 found = user_data == req->user_data;
5900 return ERR_PTR(-ENOENT);
5902 io = req->async_data;
5903 if (hrtimer_try_to_cancel(&io->timer) == -1)
5904 return ERR_PTR(-EALREADY);
5905 list_del_init(&req->timeout.list);
5909 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5910 __must_hold(&ctx->completion_lock)
5911 __must_hold(&ctx->timeout_lock)
5913 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5916 return PTR_ERR(req);
5919 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5920 io_put_req_deferred(req);
5924 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5926 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5927 case IORING_TIMEOUT_BOOTTIME:
5928 return CLOCK_BOOTTIME;
5929 case IORING_TIMEOUT_REALTIME:
5930 return CLOCK_REALTIME;
5932 /* can't happen, vetted at prep time */
5936 return CLOCK_MONOTONIC;
5940 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5941 struct timespec64 *ts, enum hrtimer_mode mode)
5942 __must_hold(&ctx->timeout_lock)
5944 struct io_timeout_data *io;
5945 struct io_kiocb *req;
5948 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
5949 found = user_data == req->user_data;
5956 io = req->async_data;
5957 if (hrtimer_try_to_cancel(&io->timer) == -1)
5959 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
5960 io->timer.function = io_link_timeout_fn;
5961 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
5965 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5966 struct timespec64 *ts, enum hrtimer_mode mode)
5967 __must_hold(&ctx->timeout_lock)
5969 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5970 struct io_timeout_data *data;
5973 return PTR_ERR(req);
5975 req->timeout.off = 0; /* noseq */
5976 data = req->async_data;
5977 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5978 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
5979 data->timer.function = io_timeout_fn;
5980 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5984 static int io_timeout_remove_prep(struct io_kiocb *req,
5985 const struct io_uring_sqe *sqe)
5987 struct io_timeout_rem *tr = &req->timeout_rem;
5989 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5991 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5993 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
5996 tr->ltimeout = false;
5997 tr->addr = READ_ONCE(sqe->addr);
5998 tr->flags = READ_ONCE(sqe->timeout_flags);
5999 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6000 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6002 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6003 tr->ltimeout = true;
6004 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6006 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6008 } else if (tr->flags) {
6009 /* timeout removal doesn't support flags */
6016 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6018 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6023 * Remove or update an existing timeout command
6025 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6027 struct io_timeout_rem *tr = &req->timeout_rem;
6028 struct io_ring_ctx *ctx = req->ctx;
6031 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6032 spin_lock(&ctx->completion_lock);
6033 spin_lock_irq(&ctx->timeout_lock);
6034 ret = io_timeout_cancel(ctx, tr->addr);
6035 spin_unlock_irq(&ctx->timeout_lock);
6036 spin_unlock(&ctx->completion_lock);
6038 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6040 spin_lock_irq(&ctx->timeout_lock);
6042 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6044 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6045 spin_unlock_irq(&ctx->timeout_lock);
6050 io_req_complete_post(req, ret, 0);
6054 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6055 bool is_timeout_link)
6057 struct io_timeout_data *data;
6059 u32 off = READ_ONCE(sqe->off);
6061 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6063 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6066 if (off && is_timeout_link)
6068 flags = READ_ONCE(sqe->timeout_flags);
6069 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6071 /* more than one clock specified is invalid, obviously */
6072 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6075 INIT_LIST_HEAD(&req->timeout.list);
6076 req->timeout.off = off;
6077 if (unlikely(off && !req->ctx->off_timeout_used))
6078 req->ctx->off_timeout_used = true;
6080 if (!req->async_data && io_alloc_async_data(req))
6083 data = req->async_data;
6085 data->flags = flags;
6087 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6090 data->mode = io_translate_timeout_mode(flags);
6091 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6093 if (is_timeout_link) {
6094 struct io_submit_link *link = &req->ctx->submit_state.link;
6098 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6100 req->timeout.head = link->last;
6101 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6106 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6108 struct io_ring_ctx *ctx = req->ctx;
6109 struct io_timeout_data *data = req->async_data;
6110 struct list_head *entry;
6111 u32 tail, off = req->timeout.off;
6113 spin_lock_irq(&ctx->timeout_lock);
6116 * sqe->off holds how many events that need to occur for this
6117 * timeout event to be satisfied. If it isn't set, then this is
6118 * a pure timeout request, sequence isn't used.
6120 if (io_is_timeout_noseq(req)) {
6121 entry = ctx->timeout_list.prev;
6125 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6126 req->timeout.target_seq = tail + off;
6128 /* Update the last seq here in case io_flush_timeouts() hasn't.
6129 * This is safe because ->completion_lock is held, and submissions
6130 * and completions are never mixed in the same ->completion_lock section.
6132 ctx->cq_last_tm_flush = tail;
6135 * Insertion sort, ensuring the first entry in the list is always
6136 * the one we need first.
6138 list_for_each_prev(entry, &ctx->timeout_list) {
6139 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6142 if (io_is_timeout_noseq(nxt))
6144 /* nxt.seq is behind @tail, otherwise would've been completed */
6145 if (off >= nxt->timeout.target_seq - tail)
6149 list_add(&req->timeout.list, entry);
6150 data->timer.function = io_timeout_fn;
6151 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6152 spin_unlock_irq(&ctx->timeout_lock);
6156 struct io_cancel_data {
6157 struct io_ring_ctx *ctx;
6161 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6163 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6164 struct io_cancel_data *cd = data;
6166 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6169 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6170 struct io_ring_ctx *ctx)
6172 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6173 enum io_wq_cancel cancel_ret;
6176 if (!tctx || !tctx->io_wq)
6179 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6180 switch (cancel_ret) {
6181 case IO_WQ_CANCEL_OK:
6184 case IO_WQ_CANCEL_RUNNING:
6187 case IO_WQ_CANCEL_NOTFOUND:
6195 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6197 struct io_ring_ctx *ctx = req->ctx;
6200 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6202 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6206 spin_lock(&ctx->completion_lock);
6207 spin_lock_irq(&ctx->timeout_lock);
6208 ret = io_timeout_cancel(ctx, sqe_addr);
6209 spin_unlock_irq(&ctx->timeout_lock);
6212 ret = io_poll_cancel(ctx, sqe_addr, false);
6214 spin_unlock(&ctx->completion_lock);
6218 static int io_async_cancel_prep(struct io_kiocb *req,
6219 const struct io_uring_sqe *sqe)
6221 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6223 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6225 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6229 req->cancel.addr = READ_ONCE(sqe->addr);
6233 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6235 struct io_ring_ctx *ctx = req->ctx;
6236 u64 sqe_addr = req->cancel.addr;
6237 struct io_tctx_node *node;
6240 ret = io_try_cancel_userdata(req, sqe_addr);
6244 /* slow path, try all io-wq's */
6245 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6247 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6248 struct io_uring_task *tctx = node->task->io_uring;
6250 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6254 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6258 io_req_complete_post(req, ret, 0);
6262 static int io_rsrc_update_prep(struct io_kiocb *req,
6263 const struct io_uring_sqe *sqe)
6265 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6267 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6270 req->rsrc_update.offset = READ_ONCE(sqe->off);
6271 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6272 if (!req->rsrc_update.nr_args)
6274 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6278 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6280 struct io_ring_ctx *ctx = req->ctx;
6281 struct io_uring_rsrc_update2 up;
6284 if (issue_flags & IO_URING_F_NONBLOCK)
6287 up.offset = req->rsrc_update.offset;
6288 up.data = req->rsrc_update.arg;
6293 mutex_lock(&ctx->uring_lock);
6294 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6295 &up, req->rsrc_update.nr_args);
6296 mutex_unlock(&ctx->uring_lock);
6300 __io_req_complete(req, issue_flags, ret, 0);
6304 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6306 switch (req->opcode) {
6309 case IORING_OP_READV:
6310 case IORING_OP_READ_FIXED:
6311 case IORING_OP_READ:
6312 return io_read_prep(req, sqe);
6313 case IORING_OP_WRITEV:
6314 case IORING_OP_WRITE_FIXED:
6315 case IORING_OP_WRITE:
6316 return io_write_prep(req, sqe);
6317 case IORING_OP_POLL_ADD:
6318 return io_poll_add_prep(req, sqe);
6319 case IORING_OP_POLL_REMOVE:
6320 return io_poll_update_prep(req, sqe);
6321 case IORING_OP_FSYNC:
6322 return io_fsync_prep(req, sqe);
6323 case IORING_OP_SYNC_FILE_RANGE:
6324 return io_sfr_prep(req, sqe);
6325 case IORING_OP_SENDMSG:
6326 case IORING_OP_SEND:
6327 return io_sendmsg_prep(req, sqe);
6328 case IORING_OP_RECVMSG:
6329 case IORING_OP_RECV:
6330 return io_recvmsg_prep(req, sqe);
6331 case IORING_OP_CONNECT:
6332 return io_connect_prep(req, sqe);
6333 case IORING_OP_TIMEOUT:
6334 return io_timeout_prep(req, sqe, false);
6335 case IORING_OP_TIMEOUT_REMOVE:
6336 return io_timeout_remove_prep(req, sqe);
6337 case IORING_OP_ASYNC_CANCEL:
6338 return io_async_cancel_prep(req, sqe);
6339 case IORING_OP_LINK_TIMEOUT:
6340 return io_timeout_prep(req, sqe, true);
6341 case IORING_OP_ACCEPT:
6342 return io_accept_prep(req, sqe);
6343 case IORING_OP_FALLOCATE:
6344 return io_fallocate_prep(req, sqe);
6345 case IORING_OP_OPENAT:
6346 return io_openat_prep(req, sqe);
6347 case IORING_OP_CLOSE:
6348 return io_close_prep(req, sqe);
6349 case IORING_OP_FILES_UPDATE:
6350 return io_rsrc_update_prep(req, sqe);
6351 case IORING_OP_STATX:
6352 return io_statx_prep(req, sqe);
6353 case IORING_OP_FADVISE:
6354 return io_fadvise_prep(req, sqe);
6355 case IORING_OP_MADVISE:
6356 return io_madvise_prep(req, sqe);
6357 case IORING_OP_OPENAT2:
6358 return io_openat2_prep(req, sqe);
6359 case IORING_OP_EPOLL_CTL:
6360 return io_epoll_ctl_prep(req, sqe);
6361 case IORING_OP_SPLICE:
6362 return io_splice_prep(req, sqe);
6363 case IORING_OP_PROVIDE_BUFFERS:
6364 return io_provide_buffers_prep(req, sqe);
6365 case IORING_OP_REMOVE_BUFFERS:
6366 return io_remove_buffers_prep(req, sqe);
6368 return io_tee_prep(req, sqe);
6369 case IORING_OP_SHUTDOWN:
6370 return io_shutdown_prep(req, sqe);
6371 case IORING_OP_RENAMEAT:
6372 return io_renameat_prep(req, sqe);
6373 case IORING_OP_UNLINKAT:
6374 return io_unlinkat_prep(req, sqe);
6375 case IORING_OP_MKDIRAT:
6376 return io_mkdirat_prep(req, sqe);
6377 case IORING_OP_SYMLINKAT:
6378 return io_symlinkat_prep(req, sqe);
6379 case IORING_OP_LINKAT:
6380 return io_linkat_prep(req, sqe);
6383 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6388 static int io_req_prep_async(struct io_kiocb *req)
6390 if (!io_op_defs[req->opcode].needs_async_setup)
6392 if (WARN_ON_ONCE(req->async_data))
6394 if (io_alloc_async_data(req))
6397 switch (req->opcode) {
6398 case IORING_OP_READV:
6399 return io_rw_prep_async(req, READ);
6400 case IORING_OP_WRITEV:
6401 return io_rw_prep_async(req, WRITE);
6402 case IORING_OP_SENDMSG:
6403 return io_sendmsg_prep_async(req);
6404 case IORING_OP_RECVMSG:
6405 return io_recvmsg_prep_async(req);
6406 case IORING_OP_CONNECT:
6407 return io_connect_prep_async(req);
6409 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6414 static u32 io_get_sequence(struct io_kiocb *req)
6416 u32 seq = req->ctx->cached_sq_head;
6418 /* need original cached_sq_head, but it was increased for each req */
6419 io_for_each_link(req, req)
6424 static bool io_drain_req(struct io_kiocb *req)
6426 struct io_kiocb *pos;
6427 struct io_ring_ctx *ctx = req->ctx;
6428 struct io_defer_entry *de;
6432 if (req->flags & REQ_F_FAIL) {
6433 io_req_complete_fail_submit(req);
6438 * If we need to drain a request in the middle of a link, drain the
6439 * head request and the next request/link after the current link.
6440 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6441 * maintained for every request of our link.
6443 if (ctx->drain_next) {
6444 req->flags |= REQ_F_IO_DRAIN;
6445 ctx->drain_next = false;
6447 /* not interested in head, start from the first linked */
6448 io_for_each_link(pos, req->link) {
6449 if (pos->flags & REQ_F_IO_DRAIN) {
6450 ctx->drain_next = true;
6451 req->flags |= REQ_F_IO_DRAIN;
6456 /* Still need defer if there is pending req in defer list. */
6457 if (likely(list_empty_careful(&ctx->defer_list) &&
6458 !(req->flags & REQ_F_IO_DRAIN))) {
6459 ctx->drain_active = false;
6463 seq = io_get_sequence(req);
6464 /* Still a chance to pass the sequence check */
6465 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6468 ret = io_req_prep_async(req);
6471 io_prep_async_link(req);
6472 de = kmalloc(sizeof(*de), GFP_KERNEL);
6476 io_req_complete_failed(req, ret);
6480 spin_lock(&ctx->completion_lock);
6481 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6482 spin_unlock(&ctx->completion_lock);
6484 io_queue_async_work(req, NULL);
6488 trace_io_uring_defer(ctx, req, req->user_data);
6491 list_add_tail(&de->list, &ctx->defer_list);
6492 spin_unlock(&ctx->completion_lock);
6496 static void io_clean_op(struct io_kiocb *req)
6498 if (req->flags & REQ_F_BUFFER_SELECTED) {
6499 switch (req->opcode) {
6500 case IORING_OP_READV:
6501 case IORING_OP_READ_FIXED:
6502 case IORING_OP_READ:
6503 kfree((void *)(unsigned long)req->rw.addr);
6505 case IORING_OP_RECVMSG:
6506 case IORING_OP_RECV:
6507 kfree(req->sr_msg.kbuf);
6512 if (req->flags & REQ_F_NEED_CLEANUP) {
6513 switch (req->opcode) {
6514 case IORING_OP_READV:
6515 case IORING_OP_READ_FIXED:
6516 case IORING_OP_READ:
6517 case IORING_OP_WRITEV:
6518 case IORING_OP_WRITE_FIXED:
6519 case IORING_OP_WRITE: {
6520 struct io_async_rw *io = req->async_data;
6522 kfree(io->free_iovec);
6525 case IORING_OP_RECVMSG:
6526 case IORING_OP_SENDMSG: {
6527 struct io_async_msghdr *io = req->async_data;
6529 kfree(io->free_iov);
6532 case IORING_OP_SPLICE:
6534 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6535 io_put_file(req->splice.file_in);
6537 case IORING_OP_OPENAT:
6538 case IORING_OP_OPENAT2:
6539 if (req->open.filename)
6540 putname(req->open.filename);
6542 case IORING_OP_RENAMEAT:
6543 putname(req->rename.oldpath);
6544 putname(req->rename.newpath);
6546 case IORING_OP_UNLINKAT:
6547 putname(req->unlink.filename);
6549 case IORING_OP_MKDIRAT:
6550 putname(req->mkdir.filename);
6552 case IORING_OP_SYMLINKAT:
6553 putname(req->symlink.oldpath);
6554 putname(req->symlink.newpath);
6556 case IORING_OP_LINKAT:
6557 putname(req->hardlink.oldpath);
6558 putname(req->hardlink.newpath);
6562 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6563 kfree(req->apoll->double_poll);
6567 if (req->flags & REQ_F_INFLIGHT) {
6568 struct io_uring_task *tctx = req->task->io_uring;
6570 atomic_dec(&tctx->inflight_tracked);
6572 if (req->flags & REQ_F_CREDS)
6573 put_cred(req->creds);
6575 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6578 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6580 struct io_ring_ctx *ctx = req->ctx;
6581 const struct cred *creds = NULL;
6584 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6585 creds = override_creds(req->creds);
6587 switch (req->opcode) {
6589 ret = io_nop(req, issue_flags);
6591 case IORING_OP_READV:
6592 case IORING_OP_READ_FIXED:
6593 case IORING_OP_READ:
6594 ret = io_read(req, issue_flags);
6596 case IORING_OP_WRITEV:
6597 case IORING_OP_WRITE_FIXED:
6598 case IORING_OP_WRITE:
6599 ret = io_write(req, issue_flags);
6601 case IORING_OP_FSYNC:
6602 ret = io_fsync(req, issue_flags);
6604 case IORING_OP_POLL_ADD:
6605 ret = io_poll_add(req, issue_flags);
6607 case IORING_OP_POLL_REMOVE:
6608 ret = io_poll_update(req, issue_flags);
6610 case IORING_OP_SYNC_FILE_RANGE:
6611 ret = io_sync_file_range(req, issue_flags);
6613 case IORING_OP_SENDMSG:
6614 ret = io_sendmsg(req, issue_flags);
6616 case IORING_OP_SEND:
6617 ret = io_send(req, issue_flags);
6619 case IORING_OP_RECVMSG:
6620 ret = io_recvmsg(req, issue_flags);
6622 case IORING_OP_RECV:
6623 ret = io_recv(req, issue_flags);
6625 case IORING_OP_TIMEOUT:
6626 ret = io_timeout(req, issue_flags);
6628 case IORING_OP_TIMEOUT_REMOVE:
6629 ret = io_timeout_remove(req, issue_flags);
6631 case IORING_OP_ACCEPT:
6632 ret = io_accept(req, issue_flags);
6634 case IORING_OP_CONNECT:
6635 ret = io_connect(req, issue_flags);
6637 case IORING_OP_ASYNC_CANCEL:
6638 ret = io_async_cancel(req, issue_flags);
6640 case IORING_OP_FALLOCATE:
6641 ret = io_fallocate(req, issue_flags);
6643 case IORING_OP_OPENAT:
6644 ret = io_openat(req, issue_flags);
6646 case IORING_OP_CLOSE:
6647 ret = io_close(req, issue_flags);
6649 case IORING_OP_FILES_UPDATE:
6650 ret = io_files_update(req, issue_flags);
6652 case IORING_OP_STATX:
6653 ret = io_statx(req, issue_flags);
6655 case IORING_OP_FADVISE:
6656 ret = io_fadvise(req, issue_flags);
6658 case IORING_OP_MADVISE:
6659 ret = io_madvise(req, issue_flags);
6661 case IORING_OP_OPENAT2:
6662 ret = io_openat2(req, issue_flags);
6664 case IORING_OP_EPOLL_CTL:
6665 ret = io_epoll_ctl(req, issue_flags);
6667 case IORING_OP_SPLICE:
6668 ret = io_splice(req, issue_flags);
6670 case IORING_OP_PROVIDE_BUFFERS:
6671 ret = io_provide_buffers(req, issue_flags);
6673 case IORING_OP_REMOVE_BUFFERS:
6674 ret = io_remove_buffers(req, issue_flags);
6677 ret = io_tee(req, issue_flags);
6679 case IORING_OP_SHUTDOWN:
6680 ret = io_shutdown(req, issue_flags);
6682 case IORING_OP_RENAMEAT:
6683 ret = io_renameat(req, issue_flags);
6685 case IORING_OP_UNLINKAT:
6686 ret = io_unlinkat(req, issue_flags);
6688 case IORING_OP_MKDIRAT:
6689 ret = io_mkdirat(req, issue_flags);
6691 case IORING_OP_SYMLINKAT:
6692 ret = io_symlinkat(req, issue_flags);
6694 case IORING_OP_LINKAT:
6695 ret = io_linkat(req, issue_flags);
6703 revert_creds(creds);
6706 /* If the op doesn't have a file, we're not polling for it */
6707 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6708 io_iopoll_req_issued(req);
6713 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6715 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6717 req = io_put_req_find_next(req);
6718 return req ? &req->work : NULL;
6721 static void io_wq_submit_work(struct io_wq_work *work)
6723 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6724 struct io_kiocb *timeout;
6727 /* one will be dropped by ->io_free_work() after returning to io-wq */
6728 if (!(req->flags & REQ_F_REFCOUNT))
6729 __io_req_set_refcount(req, 2);
6733 timeout = io_prep_linked_timeout(req);
6735 io_queue_linked_timeout(timeout);
6737 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6738 if (work->flags & IO_WQ_WORK_CANCEL)
6743 ret = io_issue_sqe(req, 0);
6745 * We can get EAGAIN for polled IO even though we're
6746 * forcing a sync submission from here, since we can't
6747 * wait for request slots on the block side.
6755 /* avoid locking problems by failing it from a clean context */
6757 io_req_task_queue_fail(req, ret);
6760 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6763 return &table->files[i];
6766 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6769 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6771 return (struct file *) (slot->file_ptr & FFS_MASK);
6774 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6776 unsigned long file_ptr = (unsigned long) file;
6778 if (__io_file_supports_nowait(file, READ))
6779 file_ptr |= FFS_ASYNC_READ;
6780 if (__io_file_supports_nowait(file, WRITE))
6781 file_ptr |= FFS_ASYNC_WRITE;
6782 if (S_ISREG(file_inode(file)->i_mode))
6783 file_ptr |= FFS_ISREG;
6784 file_slot->file_ptr = file_ptr;
6787 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6788 struct io_kiocb *req, int fd)
6791 unsigned long file_ptr;
6793 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6795 fd = array_index_nospec(fd, ctx->nr_user_files);
6796 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6797 file = (struct file *) (file_ptr & FFS_MASK);
6798 file_ptr &= ~FFS_MASK;
6799 /* mask in overlapping REQ_F and FFS bits */
6800 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6801 io_req_set_rsrc_node(req);
6805 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6806 struct io_kiocb *req, int fd)
6808 struct file *file = fget(fd);
6810 trace_io_uring_file_get(ctx, fd);
6812 /* we don't allow fixed io_uring files */
6813 if (file && unlikely(file->f_op == &io_uring_fops))
6814 io_req_track_inflight(req);
6818 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6819 struct io_kiocb *req, int fd, bool fixed)
6822 return io_file_get_fixed(ctx, req, fd);
6824 return io_file_get_normal(ctx, req, fd);
6827 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6829 struct io_kiocb *prev = req->timeout.prev;
6833 ret = io_try_cancel_userdata(req, prev->user_data);
6834 io_req_complete_post(req, ret ?: -ETIME, 0);
6837 io_req_complete_post(req, -ETIME, 0);
6841 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6843 struct io_timeout_data *data = container_of(timer,
6844 struct io_timeout_data, timer);
6845 struct io_kiocb *prev, *req = data->req;
6846 struct io_ring_ctx *ctx = req->ctx;
6847 unsigned long flags;
6849 spin_lock_irqsave(&ctx->timeout_lock, flags);
6850 prev = req->timeout.head;
6851 req->timeout.head = NULL;
6854 * We don't expect the list to be empty, that will only happen if we
6855 * race with the completion of the linked work.
6858 io_remove_next_linked(prev);
6859 if (!req_ref_inc_not_zero(prev))
6862 list_del(&req->timeout.list);
6863 req->timeout.prev = prev;
6864 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6866 req->io_task_work.func = io_req_task_link_timeout;
6867 io_req_task_work_add(req);
6868 return HRTIMER_NORESTART;
6871 static void io_queue_linked_timeout(struct io_kiocb *req)
6873 struct io_ring_ctx *ctx = req->ctx;
6875 spin_lock_irq(&ctx->timeout_lock);
6877 * If the back reference is NULL, then our linked request finished
6878 * before we got a chance to setup the timer
6880 if (req->timeout.head) {
6881 struct io_timeout_data *data = req->async_data;
6883 data->timer.function = io_link_timeout_fn;
6884 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6886 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6888 spin_unlock_irq(&ctx->timeout_lock);
6889 /* drop submission reference */
6893 static void __io_queue_sqe(struct io_kiocb *req)
6894 __must_hold(&req->ctx->uring_lock)
6896 struct io_kiocb *linked_timeout;
6900 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6903 * We async punt it if the file wasn't marked NOWAIT, or if the file
6904 * doesn't support non-blocking read/write attempts
6907 if (req->flags & REQ_F_COMPLETE_INLINE) {
6908 struct io_ring_ctx *ctx = req->ctx;
6909 struct io_submit_state *state = &ctx->submit_state;
6911 state->compl_reqs[state->compl_nr++] = req;
6912 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6913 io_submit_flush_completions(ctx);
6917 linked_timeout = io_prep_linked_timeout(req);
6919 io_queue_linked_timeout(linked_timeout);
6920 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6921 linked_timeout = io_prep_linked_timeout(req);
6923 switch (io_arm_poll_handler(req)) {
6924 case IO_APOLL_READY:
6926 io_unprep_linked_timeout(req);
6928 case IO_APOLL_ABORTED:
6930 * Queued up for async execution, worker will release
6931 * submit reference when the iocb is actually submitted.
6933 io_queue_async_work(req, NULL);
6938 io_queue_linked_timeout(linked_timeout);
6940 io_req_complete_failed(req, ret);
6944 static inline void io_queue_sqe(struct io_kiocb *req)
6945 __must_hold(&req->ctx->uring_lock)
6947 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6950 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
6951 __io_queue_sqe(req);
6952 } else if (req->flags & REQ_F_FAIL) {
6953 io_req_complete_fail_submit(req);
6955 int ret = io_req_prep_async(req);
6958 io_req_complete_failed(req, ret);
6960 io_queue_async_work(req, NULL);
6965 * Check SQE restrictions (opcode and flags).
6967 * Returns 'true' if SQE is allowed, 'false' otherwise.
6969 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6970 struct io_kiocb *req,
6971 unsigned int sqe_flags)
6973 if (likely(!ctx->restricted))
6976 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6979 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6980 ctx->restrictions.sqe_flags_required)
6983 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6984 ctx->restrictions.sqe_flags_required))
6990 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6991 const struct io_uring_sqe *sqe)
6992 __must_hold(&ctx->uring_lock)
6994 struct io_submit_state *state;
6995 unsigned int sqe_flags;
6996 int personality, ret = 0;
6998 /* req is partially pre-initialised, see io_preinit_req() */
6999 req->opcode = READ_ONCE(sqe->opcode);
7000 /* same numerical values with corresponding REQ_F_*, safe to copy */
7001 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7002 req->user_data = READ_ONCE(sqe->user_data);
7004 req->fixed_rsrc_refs = NULL;
7005 req->task = current;
7007 /* enforce forwards compatibility on users */
7008 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7010 if (unlikely(req->opcode >= IORING_OP_LAST))
7012 if (!io_check_restriction(ctx, req, sqe_flags))
7015 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7016 !io_op_defs[req->opcode].buffer_select)
7018 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7019 ctx->drain_active = true;
7021 personality = READ_ONCE(sqe->personality);
7023 req->creds = xa_load(&ctx->personalities, personality);
7026 get_cred(req->creds);
7027 req->flags |= REQ_F_CREDS;
7029 state = &ctx->submit_state;
7032 * Plug now if we have more than 1 IO left after this, and the target
7033 * is potentially a read/write to block based storage.
7035 if (!state->plug_started && state->ios_left > 1 &&
7036 io_op_defs[req->opcode].plug) {
7037 blk_start_plug(&state->plug);
7038 state->plug_started = true;
7041 if (io_op_defs[req->opcode].needs_file) {
7042 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7043 (sqe_flags & IOSQE_FIXED_FILE));
7044 if (unlikely(!req->file))
7052 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7053 const struct io_uring_sqe *sqe)
7054 __must_hold(&ctx->uring_lock)
7056 struct io_submit_link *link = &ctx->submit_state.link;
7059 ret = io_init_req(ctx, req, sqe);
7060 if (unlikely(ret)) {
7062 /* fail even hard links since we don't submit */
7065 * we can judge a link req is failed or cancelled by if
7066 * REQ_F_FAIL is set, but the head is an exception since
7067 * it may be set REQ_F_FAIL because of other req's failure
7068 * so let's leverage req->result to distinguish if a head
7069 * is set REQ_F_FAIL because of its failure or other req's
7070 * failure so that we can set the correct ret code for it.
7071 * init result here to avoid affecting the normal path.
7073 if (!(link->head->flags & REQ_F_FAIL))
7074 req_fail_link_node(link->head, -ECANCELED);
7075 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7077 * the current req is a normal req, we should return
7078 * error and thus break the submittion loop.
7080 io_req_complete_failed(req, ret);
7083 req_fail_link_node(req, ret);
7085 ret = io_req_prep(req, sqe);
7090 /* don't need @sqe from now on */
7091 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7093 ctx->flags & IORING_SETUP_SQPOLL);
7096 * If we already have a head request, queue this one for async
7097 * submittal once the head completes. If we don't have a head but
7098 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7099 * submitted sync once the chain is complete. If none of those
7100 * conditions are true (normal request), then just queue it.
7103 struct io_kiocb *head = link->head;
7105 if (!(req->flags & REQ_F_FAIL)) {
7106 ret = io_req_prep_async(req);
7107 if (unlikely(ret)) {
7108 req_fail_link_node(req, ret);
7109 if (!(head->flags & REQ_F_FAIL))
7110 req_fail_link_node(head, -ECANCELED);
7113 trace_io_uring_link(ctx, req, head);
7114 link->last->link = req;
7117 /* last request of a link, enqueue the link */
7118 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7123 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7135 * Batched submission is done, ensure local IO is flushed out.
7137 static void io_submit_state_end(struct io_submit_state *state,
7138 struct io_ring_ctx *ctx)
7140 if (state->link.head)
7141 io_queue_sqe(state->link.head);
7142 if (state->compl_nr)
7143 io_submit_flush_completions(ctx);
7144 if (state->plug_started)
7145 blk_finish_plug(&state->plug);
7149 * Start submission side cache.
7151 static void io_submit_state_start(struct io_submit_state *state,
7152 unsigned int max_ios)
7154 state->plug_started = false;
7155 state->ios_left = max_ios;
7156 /* set only head, no need to init link_last in advance */
7157 state->link.head = NULL;
7160 static void io_commit_sqring(struct io_ring_ctx *ctx)
7162 struct io_rings *rings = ctx->rings;
7165 * Ensure any loads from the SQEs are done at this point,
7166 * since once we write the new head, the application could
7167 * write new data to them.
7169 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7173 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7174 * that is mapped by userspace. This means that care needs to be taken to
7175 * ensure that reads are stable, as we cannot rely on userspace always
7176 * being a good citizen. If members of the sqe are validated and then later
7177 * used, it's important that those reads are done through READ_ONCE() to
7178 * prevent a re-load down the line.
7180 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7182 unsigned head, mask = ctx->sq_entries - 1;
7183 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7186 * The cached sq head (or cq tail) serves two purposes:
7188 * 1) allows us to batch the cost of updating the user visible
7190 * 2) allows the kernel side to track the head on its own, even
7191 * though the application is the one updating it.
7193 head = READ_ONCE(ctx->sq_array[sq_idx]);
7194 if (likely(head < ctx->sq_entries))
7195 return &ctx->sq_sqes[head];
7197 /* drop invalid entries */
7199 WRITE_ONCE(ctx->rings->sq_dropped,
7200 READ_ONCE(ctx->rings->sq_dropped) + 1);
7204 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7205 __must_hold(&ctx->uring_lock)
7209 /* make sure SQ entry isn't read before tail */
7210 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7211 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7213 io_get_task_refs(nr);
7215 io_submit_state_start(&ctx->submit_state, nr);
7216 while (submitted < nr) {
7217 const struct io_uring_sqe *sqe;
7218 struct io_kiocb *req;
7220 req = io_alloc_req(ctx);
7221 if (unlikely(!req)) {
7223 submitted = -EAGAIN;
7226 sqe = io_get_sqe(ctx);
7227 if (unlikely(!sqe)) {
7228 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7231 /* will complete beyond this point, count as submitted */
7233 if (io_submit_sqe(ctx, req, sqe))
7237 if (unlikely(submitted != nr)) {
7238 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7239 int unused = nr - ref_used;
7241 current->io_uring->cached_refs += unused;
7242 percpu_ref_put_many(&ctx->refs, unused);
7245 io_submit_state_end(&ctx->submit_state, ctx);
7246 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7247 io_commit_sqring(ctx);
7252 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7254 return READ_ONCE(sqd->state);
7257 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7259 /* Tell userspace we may need a wakeup call */
7260 spin_lock(&ctx->completion_lock);
7261 WRITE_ONCE(ctx->rings->sq_flags,
7262 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7263 spin_unlock(&ctx->completion_lock);
7266 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7268 spin_lock(&ctx->completion_lock);
7269 WRITE_ONCE(ctx->rings->sq_flags,
7270 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7271 spin_unlock(&ctx->completion_lock);
7274 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7276 unsigned int to_submit;
7279 to_submit = io_sqring_entries(ctx);
7280 /* if we're handling multiple rings, cap submit size for fairness */
7281 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7282 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7284 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7285 unsigned nr_events = 0;
7286 const struct cred *creds = NULL;
7288 if (ctx->sq_creds != current_cred())
7289 creds = override_creds(ctx->sq_creds);
7291 mutex_lock(&ctx->uring_lock);
7292 if (!list_empty(&ctx->iopoll_list))
7293 io_do_iopoll(ctx, &nr_events, 0);
7296 * Don't submit if refs are dying, good for io_uring_register(),
7297 * but also it is relied upon by io_ring_exit_work()
7299 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7300 !(ctx->flags & IORING_SETUP_R_DISABLED))
7301 ret = io_submit_sqes(ctx, to_submit);
7302 mutex_unlock(&ctx->uring_lock);
7304 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7305 wake_up(&ctx->sqo_sq_wait);
7307 revert_creds(creds);
7313 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7315 struct io_ring_ctx *ctx;
7316 unsigned sq_thread_idle = 0;
7318 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7319 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7320 sqd->sq_thread_idle = sq_thread_idle;
7323 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7325 bool did_sig = false;
7326 struct ksignal ksig;
7328 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7329 signal_pending(current)) {
7330 mutex_unlock(&sqd->lock);
7331 if (signal_pending(current))
7332 did_sig = get_signal(&ksig);
7334 mutex_lock(&sqd->lock);
7336 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7339 static int io_sq_thread(void *data)
7341 struct io_sq_data *sqd = data;
7342 struct io_ring_ctx *ctx;
7343 unsigned long timeout = 0;
7344 char buf[TASK_COMM_LEN];
7347 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7348 set_task_comm(current, buf);
7350 if (sqd->sq_cpu != -1)
7351 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7353 set_cpus_allowed_ptr(current, cpu_online_mask);
7354 current->flags |= PF_NO_SETAFFINITY;
7356 mutex_lock(&sqd->lock);
7358 bool cap_entries, sqt_spin = false;
7360 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7361 if (io_sqd_handle_event(sqd))
7363 timeout = jiffies + sqd->sq_thread_idle;
7366 cap_entries = !list_is_singular(&sqd->ctx_list);
7367 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7368 int ret = __io_sq_thread(ctx, cap_entries);
7370 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7373 if (io_run_task_work())
7376 if (sqt_spin || !time_after(jiffies, timeout)) {
7379 timeout = jiffies + sqd->sq_thread_idle;
7383 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7384 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7385 bool needs_sched = true;
7387 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7388 io_ring_set_wakeup_flag(ctx);
7390 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7391 !list_empty_careful(&ctx->iopoll_list)) {
7392 needs_sched = false;
7395 if (io_sqring_entries(ctx)) {
7396 needs_sched = false;
7402 mutex_unlock(&sqd->lock);
7404 mutex_lock(&sqd->lock);
7406 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7407 io_ring_clear_wakeup_flag(ctx);
7410 finish_wait(&sqd->wait, &wait);
7411 timeout = jiffies + sqd->sq_thread_idle;
7414 io_uring_cancel_generic(true, sqd);
7416 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7417 io_ring_set_wakeup_flag(ctx);
7419 mutex_unlock(&sqd->lock);
7421 complete(&sqd->exited);
7425 struct io_wait_queue {
7426 struct wait_queue_entry wq;
7427 struct io_ring_ctx *ctx;
7429 unsigned nr_timeouts;
7432 static inline bool io_should_wake(struct io_wait_queue *iowq)
7434 struct io_ring_ctx *ctx = iowq->ctx;
7435 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7438 * Wake up if we have enough events, or if a timeout occurred since we
7439 * started waiting. For timeouts, we always want to return to userspace,
7440 * regardless of event count.
7442 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7445 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7446 int wake_flags, void *key)
7448 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7452 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7453 * the task, and the next invocation will do it.
7455 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7456 return autoremove_wake_function(curr, mode, wake_flags, key);
7460 static int io_run_task_work_sig(void)
7462 if (io_run_task_work())
7464 if (!signal_pending(current))
7466 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7467 return -ERESTARTSYS;
7471 /* when returns >0, the caller should retry */
7472 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7473 struct io_wait_queue *iowq,
7474 signed long *timeout)
7478 /* make sure we run task_work before checking for signals */
7479 ret = io_run_task_work_sig();
7480 if (ret || io_should_wake(iowq))
7482 /* let the caller flush overflows, retry */
7483 if (test_bit(0, &ctx->check_cq_overflow))
7486 *timeout = schedule_timeout(*timeout);
7487 return !*timeout ? -ETIME : 1;
7491 * Wait until events become available, if we don't already have some. The
7492 * application must reap them itself, as they reside on the shared cq ring.
7494 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7495 const sigset_t __user *sig, size_t sigsz,
7496 struct __kernel_timespec __user *uts)
7498 struct io_wait_queue iowq;
7499 struct io_rings *rings = ctx->rings;
7500 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7504 io_cqring_overflow_flush(ctx);
7505 if (io_cqring_events(ctx) >= min_events)
7507 if (!io_run_task_work())
7512 #ifdef CONFIG_COMPAT
7513 if (in_compat_syscall())
7514 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7518 ret = set_user_sigmask(sig, sigsz);
7525 struct timespec64 ts;
7527 if (get_timespec64(&ts, uts))
7529 timeout = timespec64_to_jiffies(&ts);
7532 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7533 iowq.wq.private = current;
7534 INIT_LIST_HEAD(&iowq.wq.entry);
7536 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7537 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7539 trace_io_uring_cqring_wait(ctx, min_events);
7541 /* if we can't even flush overflow, don't wait for more */
7542 if (!io_cqring_overflow_flush(ctx)) {
7546 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7547 TASK_INTERRUPTIBLE);
7548 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7549 finish_wait(&ctx->cq_wait, &iowq.wq);
7553 restore_saved_sigmask_unless(ret == -EINTR);
7555 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7558 static void io_free_page_table(void **table, size_t size)
7560 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7562 for (i = 0; i < nr_tables; i++)
7567 static void **io_alloc_page_table(size_t size)
7569 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7570 size_t init_size = size;
7573 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7577 for (i = 0; i < nr_tables; i++) {
7578 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7580 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7582 io_free_page_table(table, init_size);
7590 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7592 percpu_ref_exit(&ref_node->refs);
7596 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7598 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7599 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7600 unsigned long flags;
7601 bool first_add = false;
7603 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7606 while (!list_empty(&ctx->rsrc_ref_list)) {
7607 node = list_first_entry(&ctx->rsrc_ref_list,
7608 struct io_rsrc_node, node);
7609 /* recycle ref nodes in order */
7612 list_del(&node->node);
7613 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7615 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7618 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7621 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7623 struct io_rsrc_node *ref_node;
7625 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7629 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7634 INIT_LIST_HEAD(&ref_node->node);
7635 INIT_LIST_HEAD(&ref_node->rsrc_list);
7636 ref_node->done = false;
7640 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7641 struct io_rsrc_data *data_to_kill)
7643 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7644 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7647 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7649 rsrc_node->rsrc_data = data_to_kill;
7650 spin_lock_irq(&ctx->rsrc_ref_lock);
7651 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7652 spin_unlock_irq(&ctx->rsrc_ref_lock);
7654 atomic_inc(&data_to_kill->refs);
7655 percpu_ref_kill(&rsrc_node->refs);
7656 ctx->rsrc_node = NULL;
7659 if (!ctx->rsrc_node) {
7660 ctx->rsrc_node = ctx->rsrc_backup_node;
7661 ctx->rsrc_backup_node = NULL;
7665 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7667 if (ctx->rsrc_backup_node)
7669 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7670 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7673 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7677 /* As we may drop ->uring_lock, other task may have started quiesce */
7681 data->quiesce = true;
7683 ret = io_rsrc_node_switch_start(ctx);
7686 io_rsrc_node_switch(ctx, data);
7688 /* kill initial ref, already quiesced if zero */
7689 if (atomic_dec_and_test(&data->refs))
7691 mutex_unlock(&ctx->uring_lock);
7692 flush_delayed_work(&ctx->rsrc_put_work);
7693 ret = wait_for_completion_interruptible(&data->done);
7695 mutex_lock(&ctx->uring_lock);
7699 atomic_inc(&data->refs);
7700 /* wait for all works potentially completing data->done */
7701 flush_delayed_work(&ctx->rsrc_put_work);
7702 reinit_completion(&data->done);
7704 ret = io_run_task_work_sig();
7705 mutex_lock(&ctx->uring_lock);
7707 data->quiesce = false;
7712 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7714 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7715 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7717 return &data->tags[table_idx][off];
7720 static void io_rsrc_data_free(struct io_rsrc_data *data)
7722 size_t size = data->nr * sizeof(data->tags[0][0]);
7725 io_free_page_table((void **)data->tags, size);
7729 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7730 u64 __user *utags, unsigned nr,
7731 struct io_rsrc_data **pdata)
7733 struct io_rsrc_data *data;
7737 data = kzalloc(sizeof(*data), GFP_KERNEL);
7740 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7748 data->do_put = do_put;
7751 for (i = 0; i < nr; i++) {
7752 u64 *tag_slot = io_get_tag_slot(data, i);
7754 if (copy_from_user(tag_slot, &utags[i],
7760 atomic_set(&data->refs, 1);
7761 init_completion(&data->done);
7765 io_rsrc_data_free(data);
7769 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7771 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7772 GFP_KERNEL_ACCOUNT);
7773 return !!table->files;
7776 static void io_free_file_tables(struct io_file_table *table)
7778 kvfree(table->files);
7779 table->files = NULL;
7782 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7784 #if defined(CONFIG_UNIX)
7785 if (ctx->ring_sock) {
7786 struct sock *sock = ctx->ring_sock->sk;
7787 struct sk_buff *skb;
7789 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7795 for (i = 0; i < ctx->nr_user_files; i++) {
7798 file = io_file_from_index(ctx, i);
7803 io_free_file_tables(&ctx->file_table);
7804 io_rsrc_data_free(ctx->file_data);
7805 ctx->file_data = NULL;
7806 ctx->nr_user_files = 0;
7809 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7813 if (!ctx->file_data)
7815 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7817 __io_sqe_files_unregister(ctx);
7821 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7822 __releases(&sqd->lock)
7824 WARN_ON_ONCE(sqd->thread == current);
7827 * Do the dance but not conditional clear_bit() because it'd race with
7828 * other threads incrementing park_pending and setting the bit.
7830 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7831 if (atomic_dec_return(&sqd->park_pending))
7832 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7833 mutex_unlock(&sqd->lock);
7836 static void io_sq_thread_park(struct io_sq_data *sqd)
7837 __acquires(&sqd->lock)
7839 WARN_ON_ONCE(sqd->thread == current);
7841 atomic_inc(&sqd->park_pending);
7842 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7843 mutex_lock(&sqd->lock);
7845 wake_up_process(sqd->thread);
7848 static void io_sq_thread_stop(struct io_sq_data *sqd)
7850 WARN_ON_ONCE(sqd->thread == current);
7851 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7853 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7854 mutex_lock(&sqd->lock);
7856 wake_up_process(sqd->thread);
7857 mutex_unlock(&sqd->lock);
7858 wait_for_completion(&sqd->exited);
7861 static void io_put_sq_data(struct io_sq_data *sqd)
7863 if (refcount_dec_and_test(&sqd->refs)) {
7864 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7866 io_sq_thread_stop(sqd);
7871 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7873 struct io_sq_data *sqd = ctx->sq_data;
7876 io_sq_thread_park(sqd);
7877 list_del_init(&ctx->sqd_list);
7878 io_sqd_update_thread_idle(sqd);
7879 io_sq_thread_unpark(sqd);
7881 io_put_sq_data(sqd);
7882 ctx->sq_data = NULL;
7886 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7888 struct io_ring_ctx *ctx_attach;
7889 struct io_sq_data *sqd;
7892 f = fdget(p->wq_fd);
7894 return ERR_PTR(-ENXIO);
7895 if (f.file->f_op != &io_uring_fops) {
7897 return ERR_PTR(-EINVAL);
7900 ctx_attach = f.file->private_data;
7901 sqd = ctx_attach->sq_data;
7904 return ERR_PTR(-EINVAL);
7906 if (sqd->task_tgid != current->tgid) {
7908 return ERR_PTR(-EPERM);
7911 refcount_inc(&sqd->refs);
7916 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7919 struct io_sq_data *sqd;
7922 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7923 sqd = io_attach_sq_data(p);
7928 /* fall through for EPERM case, setup new sqd/task */
7929 if (PTR_ERR(sqd) != -EPERM)
7933 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7935 return ERR_PTR(-ENOMEM);
7937 atomic_set(&sqd->park_pending, 0);
7938 refcount_set(&sqd->refs, 1);
7939 INIT_LIST_HEAD(&sqd->ctx_list);
7940 mutex_init(&sqd->lock);
7941 init_waitqueue_head(&sqd->wait);
7942 init_completion(&sqd->exited);
7946 #if defined(CONFIG_UNIX)
7948 * Ensure the UNIX gc is aware of our file set, so we are certain that
7949 * the io_uring can be safely unregistered on process exit, even if we have
7950 * loops in the file referencing.
7952 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7954 struct sock *sk = ctx->ring_sock->sk;
7955 struct scm_fp_list *fpl;
7956 struct sk_buff *skb;
7959 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7963 skb = alloc_skb(0, GFP_KERNEL);
7972 fpl->user = get_uid(current_user());
7973 for (i = 0; i < nr; i++) {
7974 struct file *file = io_file_from_index(ctx, i + offset);
7978 fpl->fp[nr_files] = get_file(file);
7979 unix_inflight(fpl->user, fpl->fp[nr_files]);
7984 fpl->max = SCM_MAX_FD;
7985 fpl->count = nr_files;
7986 UNIXCB(skb).fp = fpl;
7987 skb->destructor = unix_destruct_scm;
7988 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7989 skb_queue_head(&sk->sk_receive_queue, skb);
7991 for (i = 0; i < nr_files; i++)
8002 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8003 * causes regular reference counting to break down. We rely on the UNIX
8004 * garbage collection to take care of this problem for us.
8006 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8008 unsigned left, total;
8012 left = ctx->nr_user_files;
8014 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8016 ret = __io_sqe_files_scm(ctx, this_files, total);
8020 total += this_files;
8026 while (total < ctx->nr_user_files) {
8027 struct file *file = io_file_from_index(ctx, total);
8037 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8043 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8045 struct file *file = prsrc->file;
8046 #if defined(CONFIG_UNIX)
8047 struct sock *sock = ctx->ring_sock->sk;
8048 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8049 struct sk_buff *skb;
8052 __skb_queue_head_init(&list);
8055 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8056 * remove this entry and rearrange the file array.
8058 skb = skb_dequeue(head);
8060 struct scm_fp_list *fp;
8062 fp = UNIXCB(skb).fp;
8063 for (i = 0; i < fp->count; i++) {
8066 if (fp->fp[i] != file)
8069 unix_notinflight(fp->user, fp->fp[i]);
8070 left = fp->count - 1 - i;
8072 memmove(&fp->fp[i], &fp->fp[i + 1],
8073 left * sizeof(struct file *));
8080 __skb_queue_tail(&list, skb);
8090 __skb_queue_tail(&list, skb);
8092 skb = skb_dequeue(head);
8095 if (skb_peek(&list)) {
8096 spin_lock_irq(&head->lock);
8097 while ((skb = __skb_dequeue(&list)) != NULL)
8098 __skb_queue_tail(head, skb);
8099 spin_unlock_irq(&head->lock);
8106 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8108 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8109 struct io_ring_ctx *ctx = rsrc_data->ctx;
8110 struct io_rsrc_put *prsrc, *tmp;
8112 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8113 list_del(&prsrc->list);
8116 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8118 io_ring_submit_lock(ctx, lock_ring);
8119 spin_lock(&ctx->completion_lock);
8120 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8122 io_commit_cqring(ctx);
8123 spin_unlock(&ctx->completion_lock);
8124 io_cqring_ev_posted(ctx);
8125 io_ring_submit_unlock(ctx, lock_ring);
8128 rsrc_data->do_put(ctx, prsrc);
8132 io_rsrc_node_destroy(ref_node);
8133 if (atomic_dec_and_test(&rsrc_data->refs))
8134 complete(&rsrc_data->done);
8137 static void io_rsrc_put_work(struct work_struct *work)
8139 struct io_ring_ctx *ctx;
8140 struct llist_node *node;
8142 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8143 node = llist_del_all(&ctx->rsrc_put_llist);
8146 struct io_rsrc_node *ref_node;
8147 struct llist_node *next = node->next;
8149 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8150 __io_rsrc_put_work(ref_node);
8155 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8156 unsigned nr_args, u64 __user *tags)
8158 __s32 __user *fds = (__s32 __user *) arg;
8167 if (nr_args > IORING_MAX_FIXED_FILES)
8169 if (nr_args > rlimit(RLIMIT_NOFILE))
8171 ret = io_rsrc_node_switch_start(ctx);
8174 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8180 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8183 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8184 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8188 /* allow sparse sets */
8191 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8198 if (unlikely(!file))
8202 * Don't allow io_uring instances to be registered. If UNIX
8203 * isn't enabled, then this causes a reference cycle and this
8204 * instance can never get freed. If UNIX is enabled we'll
8205 * handle it just fine, but there's still no point in allowing
8206 * a ring fd as it doesn't support regular read/write anyway.
8208 if (file->f_op == &io_uring_fops) {
8212 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8215 ret = io_sqe_files_scm(ctx);
8217 __io_sqe_files_unregister(ctx);
8221 io_rsrc_node_switch(ctx, NULL);
8224 for (i = 0; i < ctx->nr_user_files; i++) {
8225 file = io_file_from_index(ctx, i);
8229 io_free_file_tables(&ctx->file_table);
8230 ctx->nr_user_files = 0;
8232 io_rsrc_data_free(ctx->file_data);
8233 ctx->file_data = NULL;
8237 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8240 #if defined(CONFIG_UNIX)
8241 struct sock *sock = ctx->ring_sock->sk;
8242 struct sk_buff_head *head = &sock->sk_receive_queue;
8243 struct sk_buff *skb;
8246 * See if we can merge this file into an existing skb SCM_RIGHTS
8247 * file set. If there's no room, fall back to allocating a new skb
8248 * and filling it in.
8250 spin_lock_irq(&head->lock);
8251 skb = skb_peek(head);
8253 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8255 if (fpl->count < SCM_MAX_FD) {
8256 __skb_unlink(skb, head);
8257 spin_unlock_irq(&head->lock);
8258 fpl->fp[fpl->count] = get_file(file);
8259 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8261 spin_lock_irq(&head->lock);
8262 __skb_queue_head(head, skb);
8267 spin_unlock_irq(&head->lock);
8274 return __io_sqe_files_scm(ctx, 1, index);
8280 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8281 unsigned int issue_flags, u32 slot_index)
8283 struct io_ring_ctx *ctx = req->ctx;
8284 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8285 struct io_fixed_file *file_slot;
8288 io_ring_submit_lock(ctx, !force_nonblock);
8289 if (file->f_op == &io_uring_fops)
8292 if (!ctx->file_data)
8295 if (slot_index >= ctx->nr_user_files)
8298 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8299 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8301 if (file_slot->file_ptr)
8304 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8305 io_fixed_file_set(file_slot, file);
8306 ret = io_sqe_file_register(ctx, file, slot_index);
8308 file_slot->file_ptr = 0;
8314 io_ring_submit_unlock(ctx, !force_nonblock);
8320 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8321 struct io_rsrc_node *node, void *rsrc)
8323 struct io_rsrc_put *prsrc;
8325 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8329 prsrc->tag = *io_get_tag_slot(data, idx);
8331 list_add(&prsrc->list, &node->rsrc_list);
8335 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8336 struct io_uring_rsrc_update2 *up,
8339 u64 __user *tags = u64_to_user_ptr(up->tags);
8340 __s32 __user *fds = u64_to_user_ptr(up->data);
8341 struct io_rsrc_data *data = ctx->file_data;
8342 struct io_fixed_file *file_slot;
8346 bool needs_switch = false;
8348 if (!ctx->file_data)
8350 if (up->offset + nr_args > ctx->nr_user_files)
8353 for (done = 0; done < nr_args; done++) {
8356 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8357 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8361 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8365 if (fd == IORING_REGISTER_FILES_SKIP)
8368 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8369 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8371 if (file_slot->file_ptr) {
8372 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8373 err = io_queue_rsrc_removal(data, up->offset + done,
8374 ctx->rsrc_node, file);
8377 file_slot->file_ptr = 0;
8378 needs_switch = true;
8387 * Don't allow io_uring instances to be registered. If
8388 * UNIX isn't enabled, then this causes a reference
8389 * cycle and this instance can never get freed. If UNIX
8390 * is enabled we'll handle it just fine, but there's
8391 * still no point in allowing a ring fd as it doesn't
8392 * support regular read/write anyway.
8394 if (file->f_op == &io_uring_fops) {
8399 *io_get_tag_slot(data, up->offset + done) = tag;
8400 io_fixed_file_set(file_slot, file);
8401 err = io_sqe_file_register(ctx, file, i);
8403 file_slot->file_ptr = 0;
8411 io_rsrc_node_switch(ctx, data);
8412 return done ? done : err;
8415 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8416 struct task_struct *task)
8418 struct io_wq_hash *hash;
8419 struct io_wq_data data;
8420 unsigned int concurrency;
8422 mutex_lock(&ctx->uring_lock);
8423 hash = ctx->hash_map;
8425 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8427 mutex_unlock(&ctx->uring_lock);
8428 return ERR_PTR(-ENOMEM);
8430 refcount_set(&hash->refs, 1);
8431 init_waitqueue_head(&hash->wait);
8432 ctx->hash_map = hash;
8434 mutex_unlock(&ctx->uring_lock);
8438 data.free_work = io_wq_free_work;
8439 data.do_work = io_wq_submit_work;
8441 /* Do QD, or 4 * CPUS, whatever is smallest */
8442 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8444 return io_wq_create(concurrency, &data);
8447 static int io_uring_alloc_task_context(struct task_struct *task,
8448 struct io_ring_ctx *ctx)
8450 struct io_uring_task *tctx;
8453 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8454 if (unlikely(!tctx))
8457 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8458 if (unlikely(ret)) {
8463 tctx->io_wq = io_init_wq_offload(ctx, task);
8464 if (IS_ERR(tctx->io_wq)) {
8465 ret = PTR_ERR(tctx->io_wq);
8466 percpu_counter_destroy(&tctx->inflight);
8472 init_waitqueue_head(&tctx->wait);
8473 atomic_set(&tctx->in_idle, 0);
8474 atomic_set(&tctx->inflight_tracked, 0);
8475 task->io_uring = tctx;
8476 spin_lock_init(&tctx->task_lock);
8477 INIT_WQ_LIST(&tctx->task_list);
8478 init_task_work(&tctx->task_work, tctx_task_work);
8482 void __io_uring_free(struct task_struct *tsk)
8484 struct io_uring_task *tctx = tsk->io_uring;
8486 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8487 WARN_ON_ONCE(tctx->io_wq);
8488 WARN_ON_ONCE(tctx->cached_refs);
8490 percpu_counter_destroy(&tctx->inflight);
8492 tsk->io_uring = NULL;
8495 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8496 struct io_uring_params *p)
8500 /* Retain compatibility with failing for an invalid attach attempt */
8501 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8502 IORING_SETUP_ATTACH_WQ) {
8505 f = fdget(p->wq_fd);
8508 if (f.file->f_op != &io_uring_fops) {
8514 if (ctx->flags & IORING_SETUP_SQPOLL) {
8515 struct task_struct *tsk;
8516 struct io_sq_data *sqd;
8519 sqd = io_get_sq_data(p, &attached);
8525 ctx->sq_creds = get_current_cred();
8527 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8528 if (!ctx->sq_thread_idle)
8529 ctx->sq_thread_idle = HZ;
8531 io_sq_thread_park(sqd);
8532 list_add(&ctx->sqd_list, &sqd->ctx_list);
8533 io_sqd_update_thread_idle(sqd);
8534 /* don't attach to a dying SQPOLL thread, would be racy */
8535 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8536 io_sq_thread_unpark(sqd);
8543 if (p->flags & IORING_SETUP_SQ_AFF) {
8544 int cpu = p->sq_thread_cpu;
8547 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8554 sqd->task_pid = current->pid;
8555 sqd->task_tgid = current->tgid;
8556 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8563 ret = io_uring_alloc_task_context(tsk, ctx);
8564 wake_up_new_task(tsk);
8567 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8568 /* Can't have SQ_AFF without SQPOLL */
8575 complete(&ctx->sq_data->exited);
8577 io_sq_thread_finish(ctx);
8581 static inline void __io_unaccount_mem(struct user_struct *user,
8582 unsigned long nr_pages)
8584 atomic_long_sub(nr_pages, &user->locked_vm);
8587 static inline int __io_account_mem(struct user_struct *user,
8588 unsigned long nr_pages)
8590 unsigned long page_limit, cur_pages, new_pages;
8592 /* Don't allow more pages than we can safely lock */
8593 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8596 cur_pages = atomic_long_read(&user->locked_vm);
8597 new_pages = cur_pages + nr_pages;
8598 if (new_pages > page_limit)
8600 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8601 new_pages) != cur_pages);
8606 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8609 __io_unaccount_mem(ctx->user, nr_pages);
8611 if (ctx->mm_account)
8612 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8615 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8620 ret = __io_account_mem(ctx->user, nr_pages);
8625 if (ctx->mm_account)
8626 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8631 static void io_mem_free(void *ptr)
8638 page = virt_to_head_page(ptr);
8639 if (put_page_testzero(page))
8640 free_compound_page(page);
8643 static void *io_mem_alloc(size_t size)
8645 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8646 __GFP_NORETRY | __GFP_ACCOUNT;
8648 return (void *) __get_free_pages(gfp_flags, get_order(size));
8651 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8654 struct io_rings *rings;
8655 size_t off, sq_array_size;
8657 off = struct_size(rings, cqes, cq_entries);
8658 if (off == SIZE_MAX)
8662 off = ALIGN(off, SMP_CACHE_BYTES);
8670 sq_array_size = array_size(sizeof(u32), sq_entries);
8671 if (sq_array_size == SIZE_MAX)
8674 if (check_add_overflow(off, sq_array_size, &off))
8680 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8682 struct io_mapped_ubuf *imu = *slot;
8685 if (imu != ctx->dummy_ubuf) {
8686 for (i = 0; i < imu->nr_bvecs; i++)
8687 unpin_user_page(imu->bvec[i].bv_page);
8688 if (imu->acct_pages)
8689 io_unaccount_mem(ctx, imu->acct_pages);
8695 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8697 io_buffer_unmap(ctx, &prsrc->buf);
8701 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8705 for (i = 0; i < ctx->nr_user_bufs; i++)
8706 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8707 kfree(ctx->user_bufs);
8708 io_rsrc_data_free(ctx->buf_data);
8709 ctx->user_bufs = NULL;
8710 ctx->buf_data = NULL;
8711 ctx->nr_user_bufs = 0;
8714 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8721 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8723 __io_sqe_buffers_unregister(ctx);
8727 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8728 void __user *arg, unsigned index)
8730 struct iovec __user *src;
8732 #ifdef CONFIG_COMPAT
8734 struct compat_iovec __user *ciovs;
8735 struct compat_iovec ciov;
8737 ciovs = (struct compat_iovec __user *) arg;
8738 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8741 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8742 dst->iov_len = ciov.iov_len;
8746 src = (struct iovec __user *) arg;
8747 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8753 * Not super efficient, but this is just a registration time. And we do cache
8754 * the last compound head, so generally we'll only do a full search if we don't
8757 * We check if the given compound head page has already been accounted, to
8758 * avoid double accounting it. This allows us to account the full size of the
8759 * page, not just the constituent pages of a huge page.
8761 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8762 int nr_pages, struct page *hpage)
8766 /* check current page array */
8767 for (i = 0; i < nr_pages; i++) {
8768 if (!PageCompound(pages[i]))
8770 if (compound_head(pages[i]) == hpage)
8774 /* check previously registered pages */
8775 for (i = 0; i < ctx->nr_user_bufs; i++) {
8776 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8778 for (j = 0; j < imu->nr_bvecs; j++) {
8779 if (!PageCompound(imu->bvec[j].bv_page))
8781 if (compound_head(imu->bvec[j].bv_page) == hpage)
8789 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8790 int nr_pages, struct io_mapped_ubuf *imu,
8791 struct page **last_hpage)
8795 imu->acct_pages = 0;
8796 for (i = 0; i < nr_pages; i++) {
8797 if (!PageCompound(pages[i])) {
8802 hpage = compound_head(pages[i]);
8803 if (hpage == *last_hpage)
8805 *last_hpage = hpage;
8806 if (headpage_already_acct(ctx, pages, i, hpage))
8808 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8812 if (!imu->acct_pages)
8815 ret = io_account_mem(ctx, imu->acct_pages);
8817 imu->acct_pages = 0;
8821 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8822 struct io_mapped_ubuf **pimu,
8823 struct page **last_hpage)
8825 struct io_mapped_ubuf *imu = NULL;
8826 struct vm_area_struct **vmas = NULL;
8827 struct page **pages = NULL;
8828 unsigned long off, start, end, ubuf;
8830 int ret, pret, nr_pages, i;
8832 if (!iov->iov_base) {
8833 *pimu = ctx->dummy_ubuf;
8837 ubuf = (unsigned long) iov->iov_base;
8838 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8839 start = ubuf >> PAGE_SHIFT;
8840 nr_pages = end - start;
8845 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8849 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8854 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8859 mmap_read_lock(current->mm);
8860 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8862 if (pret == nr_pages) {
8863 /* don't support file backed memory */
8864 for (i = 0; i < nr_pages; i++) {
8865 struct vm_area_struct *vma = vmas[i];
8867 if (vma_is_shmem(vma))
8870 !is_file_hugepages(vma->vm_file)) {
8876 ret = pret < 0 ? pret : -EFAULT;
8878 mmap_read_unlock(current->mm);
8881 * if we did partial map, or found file backed vmas,
8882 * release any pages we did get
8885 unpin_user_pages(pages, pret);
8889 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8891 unpin_user_pages(pages, pret);
8895 off = ubuf & ~PAGE_MASK;
8896 size = iov->iov_len;
8897 for (i = 0; i < nr_pages; i++) {
8900 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8901 imu->bvec[i].bv_page = pages[i];
8902 imu->bvec[i].bv_len = vec_len;
8903 imu->bvec[i].bv_offset = off;
8907 /* store original address for later verification */
8909 imu->ubuf_end = ubuf + iov->iov_len;
8910 imu->nr_bvecs = nr_pages;
8921 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8923 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8924 return ctx->user_bufs ? 0 : -ENOMEM;
8927 static int io_buffer_validate(struct iovec *iov)
8929 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8932 * Don't impose further limits on the size and buffer
8933 * constraints here, we'll -EINVAL later when IO is
8934 * submitted if they are wrong.
8937 return iov->iov_len ? -EFAULT : 0;
8941 /* arbitrary limit, but we need something */
8942 if (iov->iov_len > SZ_1G)
8945 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8951 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8952 unsigned int nr_args, u64 __user *tags)
8954 struct page *last_hpage = NULL;
8955 struct io_rsrc_data *data;
8961 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8963 ret = io_rsrc_node_switch_start(ctx);
8966 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8969 ret = io_buffers_map_alloc(ctx, nr_args);
8971 io_rsrc_data_free(data);
8975 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8976 ret = io_copy_iov(ctx, &iov, arg, i);
8979 ret = io_buffer_validate(&iov);
8982 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8987 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8993 WARN_ON_ONCE(ctx->buf_data);
8995 ctx->buf_data = data;
8997 __io_sqe_buffers_unregister(ctx);
8999 io_rsrc_node_switch(ctx, NULL);
9003 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9004 struct io_uring_rsrc_update2 *up,
9005 unsigned int nr_args)
9007 u64 __user *tags = u64_to_user_ptr(up->tags);
9008 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9009 struct page *last_hpage = NULL;
9010 bool needs_switch = false;
9016 if (up->offset + nr_args > ctx->nr_user_bufs)
9019 for (done = 0; done < nr_args; done++) {
9020 struct io_mapped_ubuf *imu;
9021 int offset = up->offset + done;
9024 err = io_copy_iov(ctx, &iov, iovs, done);
9027 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9031 err = io_buffer_validate(&iov);
9034 if (!iov.iov_base && tag) {
9038 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9042 i = array_index_nospec(offset, ctx->nr_user_bufs);
9043 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9044 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9045 ctx->rsrc_node, ctx->user_bufs[i]);
9046 if (unlikely(err)) {
9047 io_buffer_unmap(ctx, &imu);
9050 ctx->user_bufs[i] = NULL;
9051 needs_switch = true;
9054 ctx->user_bufs[i] = imu;
9055 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9059 io_rsrc_node_switch(ctx, ctx->buf_data);
9060 return done ? done : err;
9063 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9065 __s32 __user *fds = arg;
9071 if (copy_from_user(&fd, fds, sizeof(*fds)))
9074 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9075 if (IS_ERR(ctx->cq_ev_fd)) {
9076 int ret = PTR_ERR(ctx->cq_ev_fd);
9078 ctx->cq_ev_fd = NULL;
9085 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9087 if (ctx->cq_ev_fd) {
9088 eventfd_ctx_put(ctx->cq_ev_fd);
9089 ctx->cq_ev_fd = NULL;
9096 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9098 struct io_buffer *buf;
9099 unsigned long index;
9101 xa_for_each(&ctx->io_buffers, index, buf)
9102 __io_remove_buffers(ctx, buf, index, -1U);
9105 static void io_req_cache_free(struct list_head *list)
9107 struct io_kiocb *req, *nxt;
9109 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9110 list_del(&req->inflight_entry);
9111 kmem_cache_free(req_cachep, req);
9115 static void io_req_caches_free(struct io_ring_ctx *ctx)
9117 struct io_submit_state *state = &ctx->submit_state;
9119 mutex_lock(&ctx->uring_lock);
9121 if (state->free_reqs) {
9122 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9123 state->free_reqs = 0;
9126 io_flush_cached_locked_reqs(ctx, state);
9127 io_req_cache_free(&state->free_list);
9128 mutex_unlock(&ctx->uring_lock);
9131 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9133 if (data && !atomic_dec_and_test(&data->refs))
9134 wait_for_completion(&data->done);
9137 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9139 io_sq_thread_finish(ctx);
9141 if (ctx->mm_account) {
9142 mmdrop(ctx->mm_account);
9143 ctx->mm_account = NULL;
9146 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9147 io_wait_rsrc_data(ctx->buf_data);
9148 io_wait_rsrc_data(ctx->file_data);
9150 mutex_lock(&ctx->uring_lock);
9152 __io_sqe_buffers_unregister(ctx);
9154 __io_sqe_files_unregister(ctx);
9156 __io_cqring_overflow_flush(ctx, true);
9157 mutex_unlock(&ctx->uring_lock);
9158 io_eventfd_unregister(ctx);
9159 io_destroy_buffers(ctx);
9161 put_cred(ctx->sq_creds);
9163 /* there are no registered resources left, nobody uses it */
9165 io_rsrc_node_destroy(ctx->rsrc_node);
9166 if (ctx->rsrc_backup_node)
9167 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9168 flush_delayed_work(&ctx->rsrc_put_work);
9170 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9171 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9173 #if defined(CONFIG_UNIX)
9174 if (ctx->ring_sock) {
9175 ctx->ring_sock->file = NULL; /* so that iput() is called */
9176 sock_release(ctx->ring_sock);
9179 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9181 io_mem_free(ctx->rings);
9182 io_mem_free(ctx->sq_sqes);
9184 percpu_ref_exit(&ctx->refs);
9185 free_uid(ctx->user);
9186 io_req_caches_free(ctx);
9188 io_wq_put_hash(ctx->hash_map);
9189 kfree(ctx->cancel_hash);
9190 kfree(ctx->dummy_ubuf);
9194 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9196 struct io_ring_ctx *ctx = file->private_data;
9199 poll_wait(file, &ctx->poll_wait, wait);
9201 * synchronizes with barrier from wq_has_sleeper call in
9205 if (!io_sqring_full(ctx))
9206 mask |= EPOLLOUT | EPOLLWRNORM;
9209 * Don't flush cqring overflow list here, just do a simple check.
9210 * Otherwise there could possible be ABBA deadlock:
9213 * lock(&ctx->uring_lock);
9215 * lock(&ctx->uring_lock);
9218 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9219 * pushs them to do the flush.
9221 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9222 mask |= EPOLLIN | EPOLLRDNORM;
9227 static int io_uring_fasync(int fd, struct file *file, int on)
9229 struct io_ring_ctx *ctx = file->private_data;
9231 return fasync_helper(fd, file, on, &ctx->cq_fasync);
9234 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9236 const struct cred *creds;
9238 creds = xa_erase(&ctx->personalities, id);
9247 struct io_tctx_exit {
9248 struct callback_head task_work;
9249 struct completion completion;
9250 struct io_ring_ctx *ctx;
9253 static void io_tctx_exit_cb(struct callback_head *cb)
9255 struct io_uring_task *tctx = current->io_uring;
9256 struct io_tctx_exit *work;
9258 work = container_of(cb, struct io_tctx_exit, task_work);
9260 * When @in_idle, we're in cancellation and it's racy to remove the
9261 * node. It'll be removed by the end of cancellation, just ignore it.
9263 if (!atomic_read(&tctx->in_idle))
9264 io_uring_del_tctx_node((unsigned long)work->ctx);
9265 complete(&work->completion);
9268 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9270 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9272 return req->ctx == data;
9275 static void io_ring_exit_work(struct work_struct *work)
9277 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9278 unsigned long timeout = jiffies + HZ * 60 * 5;
9279 unsigned long interval = HZ / 20;
9280 struct io_tctx_exit exit;
9281 struct io_tctx_node *node;
9285 * If we're doing polled IO and end up having requests being
9286 * submitted async (out-of-line), then completions can come in while
9287 * we're waiting for refs to drop. We need to reap these manually,
9288 * as nobody else will be looking for them.
9291 io_uring_try_cancel_requests(ctx, NULL, true);
9293 struct io_sq_data *sqd = ctx->sq_data;
9294 struct task_struct *tsk;
9296 io_sq_thread_park(sqd);
9298 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9299 io_wq_cancel_cb(tsk->io_uring->io_wq,
9300 io_cancel_ctx_cb, ctx, true);
9301 io_sq_thread_unpark(sqd);
9304 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9305 /* there is little hope left, don't run it too often */
9308 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9310 init_completion(&exit.completion);
9311 init_task_work(&exit.task_work, io_tctx_exit_cb);
9314 * Some may use context even when all refs and requests have been put,
9315 * and they are free to do so while still holding uring_lock or
9316 * completion_lock, see io_req_task_submit(). Apart from other work,
9317 * this lock/unlock section also waits them to finish.
9319 mutex_lock(&ctx->uring_lock);
9320 while (!list_empty(&ctx->tctx_list)) {
9321 WARN_ON_ONCE(time_after(jiffies, timeout));
9323 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9325 /* don't spin on a single task if cancellation failed */
9326 list_rotate_left(&ctx->tctx_list);
9327 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9328 if (WARN_ON_ONCE(ret))
9330 wake_up_process(node->task);
9332 mutex_unlock(&ctx->uring_lock);
9333 wait_for_completion(&exit.completion);
9334 mutex_lock(&ctx->uring_lock);
9336 mutex_unlock(&ctx->uring_lock);
9337 spin_lock(&ctx->completion_lock);
9338 spin_unlock(&ctx->completion_lock);
9340 io_ring_ctx_free(ctx);
9343 /* Returns true if we found and killed one or more timeouts */
9344 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9347 struct io_kiocb *req, *tmp;
9350 spin_lock(&ctx->completion_lock);
9351 spin_lock_irq(&ctx->timeout_lock);
9352 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9353 if (io_match_task(req, tsk, cancel_all)) {
9354 io_kill_timeout(req, -ECANCELED);
9358 spin_unlock_irq(&ctx->timeout_lock);
9360 io_commit_cqring(ctx);
9361 spin_unlock(&ctx->completion_lock);
9363 io_cqring_ev_posted(ctx);
9364 return canceled != 0;
9367 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9369 unsigned long index;
9370 struct creds *creds;
9372 mutex_lock(&ctx->uring_lock);
9373 percpu_ref_kill(&ctx->refs);
9375 __io_cqring_overflow_flush(ctx, true);
9376 xa_for_each(&ctx->personalities, index, creds)
9377 io_unregister_personality(ctx, index);
9378 mutex_unlock(&ctx->uring_lock);
9380 io_kill_timeouts(ctx, NULL, true);
9381 io_poll_remove_all(ctx, NULL, true);
9383 /* if we failed setting up the ctx, we might not have any rings */
9384 io_iopoll_try_reap_events(ctx);
9386 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9388 * Use system_unbound_wq to avoid spawning tons of event kworkers
9389 * if we're exiting a ton of rings at the same time. It just adds
9390 * noise and overhead, there's no discernable change in runtime
9391 * over using system_wq.
9393 queue_work(system_unbound_wq, &ctx->exit_work);
9396 static int io_uring_release(struct inode *inode, struct file *file)
9398 struct io_ring_ctx *ctx = file->private_data;
9400 file->private_data = NULL;
9401 io_ring_ctx_wait_and_kill(ctx);
9405 struct io_task_cancel {
9406 struct task_struct *task;
9410 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9412 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9413 struct io_task_cancel *cancel = data;
9416 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9417 struct io_ring_ctx *ctx = req->ctx;
9419 /* protect against races with linked timeouts */
9420 spin_lock(&ctx->completion_lock);
9421 ret = io_match_task(req, cancel->task, cancel->all);
9422 spin_unlock(&ctx->completion_lock);
9424 ret = io_match_task(req, cancel->task, cancel->all);
9429 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9430 struct task_struct *task, bool cancel_all)
9432 struct io_defer_entry *de;
9435 spin_lock(&ctx->completion_lock);
9436 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9437 if (io_match_task(de->req, task, cancel_all)) {
9438 list_cut_position(&list, &ctx->defer_list, &de->list);
9442 spin_unlock(&ctx->completion_lock);
9443 if (list_empty(&list))
9446 while (!list_empty(&list)) {
9447 de = list_first_entry(&list, struct io_defer_entry, list);
9448 list_del_init(&de->list);
9449 io_req_complete_failed(de->req, -ECANCELED);
9455 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9457 struct io_tctx_node *node;
9458 enum io_wq_cancel cret;
9461 mutex_lock(&ctx->uring_lock);
9462 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9463 struct io_uring_task *tctx = node->task->io_uring;
9466 * io_wq will stay alive while we hold uring_lock, because it's
9467 * killed after ctx nodes, which requires to take the lock.
9469 if (!tctx || !tctx->io_wq)
9471 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9472 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9474 mutex_unlock(&ctx->uring_lock);
9479 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9480 struct task_struct *task,
9483 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9484 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9487 enum io_wq_cancel cret;
9491 ret |= io_uring_try_cancel_iowq(ctx);
9492 } else if (tctx && tctx->io_wq) {
9494 * Cancels requests of all rings, not only @ctx, but
9495 * it's fine as the task is in exit/exec.
9497 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9499 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9502 /* SQPOLL thread does its own polling */
9503 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9504 (ctx->sq_data && ctx->sq_data->thread == current)) {
9505 while (!list_empty_careful(&ctx->iopoll_list)) {
9506 io_iopoll_try_reap_events(ctx);
9511 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9512 ret |= io_poll_remove_all(ctx, task, cancel_all);
9513 ret |= io_kill_timeouts(ctx, task, cancel_all);
9515 ret |= io_run_task_work();
9522 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9524 struct io_uring_task *tctx = current->io_uring;
9525 struct io_tctx_node *node;
9528 if (unlikely(!tctx)) {
9529 ret = io_uring_alloc_task_context(current, ctx);
9532 tctx = current->io_uring;
9534 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9535 node = kmalloc(sizeof(*node), GFP_KERNEL);
9539 node->task = current;
9541 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9548 mutex_lock(&ctx->uring_lock);
9549 list_add(&node->ctx_node, &ctx->tctx_list);
9550 mutex_unlock(&ctx->uring_lock);
9557 * Note that this task has used io_uring. We use it for cancelation purposes.
9559 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9561 struct io_uring_task *tctx = current->io_uring;
9563 if (likely(tctx && tctx->last == ctx))
9565 return __io_uring_add_tctx_node(ctx);
9569 * Remove this io_uring_file -> task mapping.
9571 static void io_uring_del_tctx_node(unsigned long index)
9573 struct io_uring_task *tctx = current->io_uring;
9574 struct io_tctx_node *node;
9578 node = xa_erase(&tctx->xa, index);
9582 WARN_ON_ONCE(current != node->task);
9583 WARN_ON_ONCE(list_empty(&node->ctx_node));
9585 mutex_lock(&node->ctx->uring_lock);
9586 list_del(&node->ctx_node);
9587 mutex_unlock(&node->ctx->uring_lock);
9589 if (tctx->last == node->ctx)
9594 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9596 struct io_wq *wq = tctx->io_wq;
9597 struct io_tctx_node *node;
9598 unsigned long index;
9600 xa_for_each(&tctx->xa, index, node)
9601 io_uring_del_tctx_node(index);
9604 * Must be after io_uring_del_task_file() (removes nodes under
9605 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9607 io_wq_put_and_exit(wq);
9612 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9615 return atomic_read(&tctx->inflight_tracked);
9616 return percpu_counter_sum(&tctx->inflight);
9619 static void io_uring_drop_tctx_refs(struct task_struct *task)
9621 struct io_uring_task *tctx = task->io_uring;
9622 unsigned int refs = tctx->cached_refs;
9625 tctx->cached_refs = 0;
9626 percpu_counter_sub(&tctx->inflight, refs);
9627 put_task_struct_many(task, refs);
9632 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9633 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9635 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9637 struct io_uring_task *tctx = current->io_uring;
9638 struct io_ring_ctx *ctx;
9642 WARN_ON_ONCE(sqd && sqd->thread != current);
9644 if (!current->io_uring)
9647 io_wq_exit_start(tctx->io_wq);
9649 atomic_inc(&tctx->in_idle);
9651 io_uring_drop_tctx_refs(current);
9652 /* read completions before cancelations */
9653 inflight = tctx_inflight(tctx, !cancel_all);
9658 struct io_tctx_node *node;
9659 unsigned long index;
9661 xa_for_each(&tctx->xa, index, node) {
9662 /* sqpoll task will cancel all its requests */
9663 if (node->ctx->sq_data)
9665 io_uring_try_cancel_requests(node->ctx, current,
9669 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9670 io_uring_try_cancel_requests(ctx, current,
9674 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9675 io_uring_drop_tctx_refs(current);
9677 * If we've seen completions, retry without waiting. This
9678 * avoids a race where a completion comes in before we did
9679 * prepare_to_wait().
9681 if (inflight == tctx_inflight(tctx, !cancel_all))
9683 finish_wait(&tctx->wait, &wait);
9685 atomic_dec(&tctx->in_idle);
9687 io_uring_clean_tctx(tctx);
9689 /* for exec all current's requests should be gone, kill tctx */
9690 __io_uring_free(current);
9694 void __io_uring_cancel(bool cancel_all)
9696 io_uring_cancel_generic(cancel_all, NULL);
9699 static void *io_uring_validate_mmap_request(struct file *file,
9700 loff_t pgoff, size_t sz)
9702 struct io_ring_ctx *ctx = file->private_data;
9703 loff_t offset = pgoff << PAGE_SHIFT;
9708 case IORING_OFF_SQ_RING:
9709 case IORING_OFF_CQ_RING:
9712 case IORING_OFF_SQES:
9716 return ERR_PTR(-EINVAL);
9719 page = virt_to_head_page(ptr);
9720 if (sz > page_size(page))
9721 return ERR_PTR(-EINVAL);
9728 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9730 size_t sz = vma->vm_end - vma->vm_start;
9734 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9736 return PTR_ERR(ptr);
9738 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9739 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9742 #else /* !CONFIG_MMU */
9744 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9746 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9749 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9751 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9754 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9755 unsigned long addr, unsigned long len,
9756 unsigned long pgoff, unsigned long flags)
9760 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9762 return PTR_ERR(ptr);
9764 return (unsigned long) ptr;
9767 #endif /* !CONFIG_MMU */
9769 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9774 if (!io_sqring_full(ctx))
9776 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9778 if (!io_sqring_full(ctx))
9781 } while (!signal_pending(current));
9783 finish_wait(&ctx->sqo_sq_wait, &wait);
9787 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9788 struct __kernel_timespec __user **ts,
9789 const sigset_t __user **sig)
9791 struct io_uring_getevents_arg arg;
9794 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9795 * is just a pointer to the sigset_t.
9797 if (!(flags & IORING_ENTER_EXT_ARG)) {
9798 *sig = (const sigset_t __user *) argp;
9804 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9805 * timespec and sigset_t pointers if good.
9807 if (*argsz != sizeof(arg))
9809 if (copy_from_user(&arg, argp, sizeof(arg)))
9811 *sig = u64_to_user_ptr(arg.sigmask);
9812 *argsz = arg.sigmask_sz;
9813 *ts = u64_to_user_ptr(arg.ts);
9817 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9818 u32, min_complete, u32, flags, const void __user *, argp,
9821 struct io_ring_ctx *ctx;
9828 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9829 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9833 if (unlikely(!f.file))
9837 if (unlikely(f.file->f_op != &io_uring_fops))
9841 ctx = f.file->private_data;
9842 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9846 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9850 * For SQ polling, the thread will do all submissions and completions.
9851 * Just return the requested submit count, and wake the thread if
9855 if (ctx->flags & IORING_SETUP_SQPOLL) {
9856 io_cqring_overflow_flush(ctx);
9858 if (unlikely(ctx->sq_data->thread == NULL)) {
9862 if (flags & IORING_ENTER_SQ_WAKEUP)
9863 wake_up(&ctx->sq_data->wait);
9864 if (flags & IORING_ENTER_SQ_WAIT) {
9865 ret = io_sqpoll_wait_sq(ctx);
9869 submitted = to_submit;
9870 } else if (to_submit) {
9871 ret = io_uring_add_tctx_node(ctx);
9874 mutex_lock(&ctx->uring_lock);
9875 submitted = io_submit_sqes(ctx, to_submit);
9876 mutex_unlock(&ctx->uring_lock);
9878 if (submitted != to_submit)
9881 if (flags & IORING_ENTER_GETEVENTS) {
9882 const sigset_t __user *sig;
9883 struct __kernel_timespec __user *ts;
9885 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9889 min_complete = min(min_complete, ctx->cq_entries);
9892 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9893 * space applications don't need to do io completion events
9894 * polling again, they can rely on io_sq_thread to do polling
9895 * work, which can reduce cpu usage and uring_lock contention.
9897 if (ctx->flags & IORING_SETUP_IOPOLL &&
9898 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9899 ret = io_iopoll_check(ctx, min_complete);
9901 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9906 percpu_ref_put(&ctx->refs);
9909 return submitted ? submitted : ret;
9912 #ifdef CONFIG_PROC_FS
9913 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9914 const struct cred *cred)
9916 struct user_namespace *uns = seq_user_ns(m);
9917 struct group_info *gi;
9922 seq_printf(m, "%5d\n", id);
9923 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9924 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9925 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9926 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9927 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9928 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9929 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9930 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9931 seq_puts(m, "\n\tGroups:\t");
9932 gi = cred->group_info;
9933 for (g = 0; g < gi->ngroups; g++) {
9934 seq_put_decimal_ull(m, g ? " " : "",
9935 from_kgid_munged(uns, gi->gid[g]));
9937 seq_puts(m, "\n\tCapEff:\t");
9938 cap = cred->cap_effective;
9939 CAP_FOR_EACH_U32(__capi)
9940 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9945 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9947 struct io_sq_data *sq = NULL;
9952 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9953 * since fdinfo case grabs it in the opposite direction of normal use
9954 * cases. If we fail to get the lock, we just don't iterate any
9955 * structures that could be going away outside the io_uring mutex.
9957 has_lock = mutex_trylock(&ctx->uring_lock);
9959 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9965 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9966 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9967 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9968 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9969 struct file *f = io_file_from_index(ctx, i);
9972 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9974 seq_printf(m, "%5u: <none>\n", i);
9976 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9977 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9978 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9979 unsigned int len = buf->ubuf_end - buf->ubuf;
9981 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9983 if (has_lock && !xa_empty(&ctx->personalities)) {
9984 unsigned long index;
9985 const struct cred *cred;
9987 seq_printf(m, "Personalities:\n");
9988 xa_for_each(&ctx->personalities, index, cred)
9989 io_uring_show_cred(m, index, cred);
9991 seq_printf(m, "PollList:\n");
9992 spin_lock(&ctx->completion_lock);
9993 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9994 struct hlist_head *list = &ctx->cancel_hash[i];
9995 struct io_kiocb *req;
9997 hlist_for_each_entry(req, list, hash_node)
9998 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9999 req->task->task_works != NULL);
10001 spin_unlock(&ctx->completion_lock);
10003 mutex_unlock(&ctx->uring_lock);
10006 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10008 struct io_ring_ctx *ctx = f->private_data;
10010 if (percpu_ref_tryget(&ctx->refs)) {
10011 __io_uring_show_fdinfo(ctx, m);
10012 percpu_ref_put(&ctx->refs);
10017 static const struct file_operations io_uring_fops = {
10018 .release = io_uring_release,
10019 .mmap = io_uring_mmap,
10021 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10022 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10024 .poll = io_uring_poll,
10025 .fasync = io_uring_fasync,
10026 #ifdef CONFIG_PROC_FS
10027 .show_fdinfo = io_uring_show_fdinfo,
10031 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10032 struct io_uring_params *p)
10034 struct io_rings *rings;
10035 size_t size, sq_array_offset;
10037 /* make sure these are sane, as we already accounted them */
10038 ctx->sq_entries = p->sq_entries;
10039 ctx->cq_entries = p->cq_entries;
10041 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10042 if (size == SIZE_MAX)
10045 rings = io_mem_alloc(size);
10049 ctx->rings = rings;
10050 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10051 rings->sq_ring_mask = p->sq_entries - 1;
10052 rings->cq_ring_mask = p->cq_entries - 1;
10053 rings->sq_ring_entries = p->sq_entries;
10054 rings->cq_ring_entries = p->cq_entries;
10056 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10057 if (size == SIZE_MAX) {
10058 io_mem_free(ctx->rings);
10063 ctx->sq_sqes = io_mem_alloc(size);
10064 if (!ctx->sq_sqes) {
10065 io_mem_free(ctx->rings);
10073 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10077 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10081 ret = io_uring_add_tctx_node(ctx);
10086 fd_install(fd, file);
10091 * Allocate an anonymous fd, this is what constitutes the application
10092 * visible backing of an io_uring instance. The application mmaps this
10093 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10094 * we have to tie this fd to a socket for file garbage collection purposes.
10096 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10099 #if defined(CONFIG_UNIX)
10102 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10105 return ERR_PTR(ret);
10108 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10109 O_RDWR | O_CLOEXEC);
10110 #if defined(CONFIG_UNIX)
10111 if (IS_ERR(file)) {
10112 sock_release(ctx->ring_sock);
10113 ctx->ring_sock = NULL;
10115 ctx->ring_sock->file = file;
10121 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10122 struct io_uring_params __user *params)
10124 struct io_ring_ctx *ctx;
10130 if (entries > IORING_MAX_ENTRIES) {
10131 if (!(p->flags & IORING_SETUP_CLAMP))
10133 entries = IORING_MAX_ENTRIES;
10137 * Use twice as many entries for the CQ ring. It's possible for the
10138 * application to drive a higher depth than the size of the SQ ring,
10139 * since the sqes are only used at submission time. This allows for
10140 * some flexibility in overcommitting a bit. If the application has
10141 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10142 * of CQ ring entries manually.
10144 p->sq_entries = roundup_pow_of_two(entries);
10145 if (p->flags & IORING_SETUP_CQSIZE) {
10147 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10148 * to a power-of-two, if it isn't already. We do NOT impose
10149 * any cq vs sq ring sizing.
10151 if (!p->cq_entries)
10153 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10154 if (!(p->flags & IORING_SETUP_CLAMP))
10156 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10158 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10159 if (p->cq_entries < p->sq_entries)
10162 p->cq_entries = 2 * p->sq_entries;
10165 ctx = io_ring_ctx_alloc(p);
10168 ctx->compat = in_compat_syscall();
10169 if (!capable(CAP_IPC_LOCK))
10170 ctx->user = get_uid(current_user());
10173 * This is just grabbed for accounting purposes. When a process exits,
10174 * the mm is exited and dropped before the files, hence we need to hang
10175 * on to this mm purely for the purposes of being able to unaccount
10176 * memory (locked/pinned vm). It's not used for anything else.
10178 mmgrab(current->mm);
10179 ctx->mm_account = current->mm;
10181 ret = io_allocate_scq_urings(ctx, p);
10185 ret = io_sq_offload_create(ctx, p);
10188 /* always set a rsrc node */
10189 ret = io_rsrc_node_switch_start(ctx);
10192 io_rsrc_node_switch(ctx, NULL);
10194 memset(&p->sq_off, 0, sizeof(p->sq_off));
10195 p->sq_off.head = offsetof(struct io_rings, sq.head);
10196 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10197 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10198 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10199 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10200 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10201 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10203 memset(&p->cq_off, 0, sizeof(p->cq_off));
10204 p->cq_off.head = offsetof(struct io_rings, cq.head);
10205 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10206 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10207 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10208 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10209 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10210 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10212 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10213 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10214 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10215 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10216 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10217 IORING_FEAT_RSRC_TAGS;
10219 if (copy_to_user(params, p, sizeof(*p))) {
10224 file = io_uring_get_file(ctx);
10225 if (IS_ERR(file)) {
10226 ret = PTR_ERR(file);
10231 * Install ring fd as the very last thing, so we don't risk someone
10232 * having closed it before we finish setup
10234 ret = io_uring_install_fd(ctx, file);
10236 /* fput will clean it up */
10241 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10244 io_ring_ctx_wait_and_kill(ctx);
10249 * Sets up an aio uring context, and returns the fd. Applications asks for a
10250 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10251 * params structure passed in.
10253 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10255 struct io_uring_params p;
10258 if (copy_from_user(&p, params, sizeof(p)))
10260 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10265 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10266 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10267 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10268 IORING_SETUP_R_DISABLED))
10271 return io_uring_create(entries, &p, params);
10274 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10275 struct io_uring_params __user *, params)
10277 return io_uring_setup(entries, params);
10280 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10282 struct io_uring_probe *p;
10286 size = struct_size(p, ops, nr_args);
10287 if (size == SIZE_MAX)
10289 p = kzalloc(size, GFP_KERNEL);
10294 if (copy_from_user(p, arg, size))
10297 if (memchr_inv(p, 0, size))
10300 p->last_op = IORING_OP_LAST - 1;
10301 if (nr_args > IORING_OP_LAST)
10302 nr_args = IORING_OP_LAST;
10304 for (i = 0; i < nr_args; i++) {
10306 if (!io_op_defs[i].not_supported)
10307 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10312 if (copy_to_user(arg, p, size))
10319 static int io_register_personality(struct io_ring_ctx *ctx)
10321 const struct cred *creds;
10325 creds = get_current_cred();
10327 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10328 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10336 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10337 unsigned int nr_args)
10339 struct io_uring_restriction *res;
10343 /* Restrictions allowed only if rings started disabled */
10344 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10347 /* We allow only a single restrictions registration */
10348 if (ctx->restrictions.registered)
10351 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10354 size = array_size(nr_args, sizeof(*res));
10355 if (size == SIZE_MAX)
10358 res = memdup_user(arg, size);
10360 return PTR_ERR(res);
10364 for (i = 0; i < nr_args; i++) {
10365 switch (res[i].opcode) {
10366 case IORING_RESTRICTION_REGISTER_OP:
10367 if (res[i].register_op >= IORING_REGISTER_LAST) {
10372 __set_bit(res[i].register_op,
10373 ctx->restrictions.register_op);
10375 case IORING_RESTRICTION_SQE_OP:
10376 if (res[i].sqe_op >= IORING_OP_LAST) {
10381 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10383 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10384 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10386 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10387 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10396 /* Reset all restrictions if an error happened */
10398 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10400 ctx->restrictions.registered = true;
10406 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10408 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10411 if (ctx->restrictions.registered)
10412 ctx->restricted = 1;
10414 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10415 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10416 wake_up(&ctx->sq_data->wait);
10420 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10421 struct io_uring_rsrc_update2 *up,
10429 if (check_add_overflow(up->offset, nr_args, &tmp))
10431 err = io_rsrc_node_switch_start(ctx);
10436 case IORING_RSRC_FILE:
10437 return __io_sqe_files_update(ctx, up, nr_args);
10438 case IORING_RSRC_BUFFER:
10439 return __io_sqe_buffers_update(ctx, up, nr_args);
10444 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10447 struct io_uring_rsrc_update2 up;
10451 memset(&up, 0, sizeof(up));
10452 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10454 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10457 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10458 unsigned size, unsigned type)
10460 struct io_uring_rsrc_update2 up;
10462 if (size != sizeof(up))
10464 if (copy_from_user(&up, arg, sizeof(up)))
10466 if (!up.nr || up.resv)
10468 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10471 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10472 unsigned int size, unsigned int type)
10474 struct io_uring_rsrc_register rr;
10476 /* keep it extendible */
10477 if (size != sizeof(rr))
10480 memset(&rr, 0, sizeof(rr));
10481 if (copy_from_user(&rr, arg, size))
10483 if (!rr.nr || rr.resv || rr.resv2)
10487 case IORING_RSRC_FILE:
10488 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10489 rr.nr, u64_to_user_ptr(rr.tags));
10490 case IORING_RSRC_BUFFER:
10491 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10492 rr.nr, u64_to_user_ptr(rr.tags));
10497 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10500 struct io_uring_task *tctx = current->io_uring;
10501 cpumask_var_t new_mask;
10504 if (!tctx || !tctx->io_wq)
10507 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10510 cpumask_clear(new_mask);
10511 if (len > cpumask_size())
10512 len = cpumask_size();
10514 if (copy_from_user(new_mask, arg, len)) {
10515 free_cpumask_var(new_mask);
10519 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10520 free_cpumask_var(new_mask);
10524 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10526 struct io_uring_task *tctx = current->io_uring;
10528 if (!tctx || !tctx->io_wq)
10531 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10534 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10537 struct io_uring_task *tctx = NULL;
10538 struct io_sq_data *sqd = NULL;
10539 __u32 new_count[2];
10542 if (copy_from_user(new_count, arg, sizeof(new_count)))
10544 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10545 if (new_count[i] > INT_MAX)
10548 if (ctx->flags & IORING_SETUP_SQPOLL) {
10549 sqd = ctx->sq_data;
10551 mutex_lock(&sqd->lock);
10552 tctx = sqd->thread->io_uring;
10555 tctx = current->io_uring;
10559 if (!tctx || !tctx->io_wq)
10562 ret = io_wq_max_workers(tctx->io_wq, new_count);
10567 mutex_unlock(&sqd->lock);
10569 if (copy_to_user(arg, new_count, sizeof(new_count)))
10575 mutex_unlock(&sqd->lock);
10579 static bool io_register_op_must_quiesce(int op)
10582 case IORING_REGISTER_BUFFERS:
10583 case IORING_UNREGISTER_BUFFERS:
10584 case IORING_REGISTER_FILES:
10585 case IORING_UNREGISTER_FILES:
10586 case IORING_REGISTER_FILES_UPDATE:
10587 case IORING_REGISTER_PROBE:
10588 case IORING_REGISTER_PERSONALITY:
10589 case IORING_UNREGISTER_PERSONALITY:
10590 case IORING_REGISTER_FILES2:
10591 case IORING_REGISTER_FILES_UPDATE2:
10592 case IORING_REGISTER_BUFFERS2:
10593 case IORING_REGISTER_BUFFERS_UPDATE:
10594 case IORING_REGISTER_IOWQ_AFF:
10595 case IORING_UNREGISTER_IOWQ_AFF:
10596 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10603 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10607 percpu_ref_kill(&ctx->refs);
10610 * Drop uring mutex before waiting for references to exit. If another
10611 * thread is currently inside io_uring_enter() it might need to grab the
10612 * uring_lock to make progress. If we hold it here across the drain
10613 * wait, then we can deadlock. It's safe to drop the mutex here, since
10614 * no new references will come in after we've killed the percpu ref.
10616 mutex_unlock(&ctx->uring_lock);
10618 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10621 ret = io_run_task_work_sig();
10622 } while (ret >= 0);
10623 mutex_lock(&ctx->uring_lock);
10626 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10630 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10631 void __user *arg, unsigned nr_args)
10632 __releases(ctx->uring_lock)
10633 __acquires(ctx->uring_lock)
10638 * We're inside the ring mutex, if the ref is already dying, then
10639 * someone else killed the ctx or is already going through
10640 * io_uring_register().
10642 if (percpu_ref_is_dying(&ctx->refs))
10645 if (ctx->restricted) {
10646 if (opcode >= IORING_REGISTER_LAST)
10648 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10649 if (!test_bit(opcode, ctx->restrictions.register_op))
10653 if (io_register_op_must_quiesce(opcode)) {
10654 ret = io_ctx_quiesce(ctx);
10660 case IORING_REGISTER_BUFFERS:
10661 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10663 case IORING_UNREGISTER_BUFFERS:
10665 if (arg || nr_args)
10667 ret = io_sqe_buffers_unregister(ctx);
10669 case IORING_REGISTER_FILES:
10670 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10672 case IORING_UNREGISTER_FILES:
10674 if (arg || nr_args)
10676 ret = io_sqe_files_unregister(ctx);
10678 case IORING_REGISTER_FILES_UPDATE:
10679 ret = io_register_files_update(ctx, arg, nr_args);
10681 case IORING_REGISTER_EVENTFD:
10682 case IORING_REGISTER_EVENTFD_ASYNC:
10686 ret = io_eventfd_register(ctx, arg);
10689 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10690 ctx->eventfd_async = 1;
10692 ctx->eventfd_async = 0;
10694 case IORING_UNREGISTER_EVENTFD:
10696 if (arg || nr_args)
10698 ret = io_eventfd_unregister(ctx);
10700 case IORING_REGISTER_PROBE:
10702 if (!arg || nr_args > 256)
10704 ret = io_probe(ctx, arg, nr_args);
10706 case IORING_REGISTER_PERSONALITY:
10708 if (arg || nr_args)
10710 ret = io_register_personality(ctx);
10712 case IORING_UNREGISTER_PERSONALITY:
10716 ret = io_unregister_personality(ctx, nr_args);
10718 case IORING_REGISTER_ENABLE_RINGS:
10720 if (arg || nr_args)
10722 ret = io_register_enable_rings(ctx);
10724 case IORING_REGISTER_RESTRICTIONS:
10725 ret = io_register_restrictions(ctx, arg, nr_args);
10727 case IORING_REGISTER_FILES2:
10728 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10730 case IORING_REGISTER_FILES_UPDATE2:
10731 ret = io_register_rsrc_update(ctx, arg, nr_args,
10734 case IORING_REGISTER_BUFFERS2:
10735 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10737 case IORING_REGISTER_BUFFERS_UPDATE:
10738 ret = io_register_rsrc_update(ctx, arg, nr_args,
10739 IORING_RSRC_BUFFER);
10741 case IORING_REGISTER_IOWQ_AFF:
10743 if (!arg || !nr_args)
10745 ret = io_register_iowq_aff(ctx, arg, nr_args);
10747 case IORING_UNREGISTER_IOWQ_AFF:
10749 if (arg || nr_args)
10751 ret = io_unregister_iowq_aff(ctx);
10753 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10755 if (!arg || nr_args != 2)
10757 ret = io_register_iowq_max_workers(ctx, arg);
10764 if (io_register_op_must_quiesce(opcode)) {
10765 /* bring the ctx back to life */
10766 percpu_ref_reinit(&ctx->refs);
10767 reinit_completion(&ctx->ref_comp);
10772 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10773 void __user *, arg, unsigned int, nr_args)
10775 struct io_ring_ctx *ctx;
10784 if (f.file->f_op != &io_uring_fops)
10787 ctx = f.file->private_data;
10789 io_run_task_work();
10791 mutex_lock(&ctx->uring_lock);
10792 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10793 mutex_unlock(&ctx->uring_lock);
10794 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10795 ctx->cq_ev_fd != NULL, ret);
10801 static int __init io_uring_init(void)
10803 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10804 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10805 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10808 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10809 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10810 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10811 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10812 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10813 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10814 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10815 BUILD_BUG_SQE_ELEM(8, __u64, off);
10816 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10817 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10818 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10819 BUILD_BUG_SQE_ELEM(24, __u32, len);
10820 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10821 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10822 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10823 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10824 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10825 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10826 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10827 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10828 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10829 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10830 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10831 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10832 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10833 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10834 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10835 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10836 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10837 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10838 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10839 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10840 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10842 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10843 sizeof(struct io_uring_rsrc_update));
10844 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10845 sizeof(struct io_uring_rsrc_update2));
10847 /* ->buf_index is u16 */
10848 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10850 /* should fit into one byte */
10851 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10853 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10854 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10856 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10860 __initcall(io_uring_init);