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 /* see waitqueue_active() comment */
1625 if (ctx->flags & IORING_SETUP_SQPOLL) {
1626 if (waitqueue_active(&ctx->cq_wait))
1627 wake_up_all(&ctx->cq_wait);
1629 if (io_should_trigger_evfd(ctx))
1630 eventfd_signal(ctx->cq_ev_fd, 1);
1631 if (waitqueue_active(&ctx->poll_wait)) {
1632 wake_up_interruptible(&ctx->poll_wait);
1633 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1637 /* Returns true if there are no backlogged entries after the flush */
1638 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1640 bool all_flushed, posted;
1642 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1646 spin_lock(&ctx->completion_lock);
1647 while (!list_empty(&ctx->cq_overflow_list)) {
1648 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1649 struct io_overflow_cqe *ocqe;
1653 ocqe = list_first_entry(&ctx->cq_overflow_list,
1654 struct io_overflow_cqe, list);
1656 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1658 io_account_cq_overflow(ctx);
1661 list_del(&ocqe->list);
1665 all_flushed = list_empty(&ctx->cq_overflow_list);
1667 clear_bit(0, &ctx->check_cq_overflow);
1668 WRITE_ONCE(ctx->rings->sq_flags,
1669 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1673 io_commit_cqring(ctx);
1674 spin_unlock(&ctx->completion_lock);
1676 io_cqring_ev_posted(ctx);
1680 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1684 if (test_bit(0, &ctx->check_cq_overflow)) {
1685 /* iopoll syncs against uring_lock, not completion_lock */
1686 if (ctx->flags & IORING_SETUP_IOPOLL)
1687 mutex_lock(&ctx->uring_lock);
1688 ret = __io_cqring_overflow_flush(ctx, false);
1689 if (ctx->flags & IORING_SETUP_IOPOLL)
1690 mutex_unlock(&ctx->uring_lock);
1696 /* must to be called somewhat shortly after putting a request */
1697 static inline void io_put_task(struct task_struct *task, int nr)
1699 struct io_uring_task *tctx = task->io_uring;
1701 if (likely(task == current)) {
1702 tctx->cached_refs += nr;
1704 percpu_counter_sub(&tctx->inflight, nr);
1705 if (unlikely(atomic_read(&tctx->in_idle)))
1706 wake_up(&tctx->wait);
1707 put_task_struct_many(task, nr);
1711 static void io_task_refs_refill(struct io_uring_task *tctx)
1713 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1715 percpu_counter_add(&tctx->inflight, refill);
1716 refcount_add(refill, ¤t->usage);
1717 tctx->cached_refs += refill;
1720 static inline void io_get_task_refs(int nr)
1722 struct io_uring_task *tctx = current->io_uring;
1724 tctx->cached_refs -= nr;
1725 if (unlikely(tctx->cached_refs < 0))
1726 io_task_refs_refill(tctx);
1729 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1730 long res, unsigned int cflags)
1732 struct io_overflow_cqe *ocqe;
1734 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1737 * If we're in ring overflow flush mode, or in task cancel mode,
1738 * or cannot allocate an overflow entry, then we need to drop it
1741 io_account_cq_overflow(ctx);
1744 if (list_empty(&ctx->cq_overflow_list)) {
1745 set_bit(0, &ctx->check_cq_overflow);
1746 WRITE_ONCE(ctx->rings->sq_flags,
1747 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1750 ocqe->cqe.user_data = user_data;
1751 ocqe->cqe.res = res;
1752 ocqe->cqe.flags = cflags;
1753 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1757 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1758 long res, unsigned int cflags)
1760 struct io_uring_cqe *cqe;
1762 trace_io_uring_complete(ctx, user_data, res, cflags);
1765 * If we can't get a cq entry, userspace overflowed the
1766 * submission (by quite a lot). Increment the overflow count in
1769 cqe = io_get_cqe(ctx);
1771 WRITE_ONCE(cqe->user_data, user_data);
1772 WRITE_ONCE(cqe->res, res);
1773 WRITE_ONCE(cqe->flags, cflags);
1776 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1779 /* not as hot to bloat with inlining */
1780 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1781 long res, unsigned int cflags)
1783 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1786 static void io_req_complete_post(struct io_kiocb *req, long res,
1787 unsigned int cflags)
1789 struct io_ring_ctx *ctx = req->ctx;
1791 spin_lock(&ctx->completion_lock);
1792 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1794 * If we're the last reference to this request, add to our locked
1797 if (req_ref_put_and_test(req)) {
1798 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1799 if (req->flags & IO_DISARM_MASK)
1800 io_disarm_next(req);
1802 io_req_task_queue(req->link);
1806 io_dismantle_req(req);
1807 io_put_task(req->task, 1);
1808 list_add(&req->inflight_entry, &ctx->locked_free_list);
1809 ctx->locked_free_nr++;
1811 if (!percpu_ref_tryget(&ctx->refs))
1814 io_commit_cqring(ctx);
1815 spin_unlock(&ctx->completion_lock);
1818 io_cqring_ev_posted(ctx);
1819 percpu_ref_put(&ctx->refs);
1823 static inline bool io_req_needs_clean(struct io_kiocb *req)
1825 return req->flags & IO_REQ_CLEAN_FLAGS;
1828 static void io_req_complete_state(struct io_kiocb *req, long res,
1829 unsigned int cflags)
1831 if (io_req_needs_clean(req))
1834 req->compl.cflags = cflags;
1835 req->flags |= REQ_F_COMPLETE_INLINE;
1838 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1839 long res, unsigned cflags)
1841 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1842 io_req_complete_state(req, res, cflags);
1844 io_req_complete_post(req, res, cflags);
1847 static inline void io_req_complete(struct io_kiocb *req, long res)
1849 __io_req_complete(req, 0, res, 0);
1852 static void io_req_complete_failed(struct io_kiocb *req, long res)
1855 io_req_complete_post(req, res, 0);
1858 static void io_req_complete_fail_submit(struct io_kiocb *req)
1861 * We don't submit, fail them all, for that replace hardlinks with
1862 * normal links. Extra REQ_F_LINK is tolerated.
1864 req->flags &= ~REQ_F_HARDLINK;
1865 req->flags |= REQ_F_LINK;
1866 io_req_complete_failed(req, req->result);
1870 * Don't initialise the fields below on every allocation, but do that in
1871 * advance and keep them valid across allocations.
1873 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1877 req->async_data = NULL;
1878 /* not necessary, but safer to zero */
1882 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1883 struct io_submit_state *state)
1885 spin_lock(&ctx->completion_lock);
1886 list_splice_init(&ctx->locked_free_list, &state->free_list);
1887 ctx->locked_free_nr = 0;
1888 spin_unlock(&ctx->completion_lock);
1891 /* Returns true IFF there are requests in the cache */
1892 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1894 struct io_submit_state *state = &ctx->submit_state;
1898 * If we have more than a batch's worth of requests in our IRQ side
1899 * locked cache, grab the lock and move them over to our submission
1902 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1903 io_flush_cached_locked_reqs(ctx, state);
1905 nr = state->free_reqs;
1906 while (!list_empty(&state->free_list)) {
1907 struct io_kiocb *req = list_first_entry(&state->free_list,
1908 struct io_kiocb, inflight_entry);
1910 list_del(&req->inflight_entry);
1911 state->reqs[nr++] = req;
1912 if (nr == ARRAY_SIZE(state->reqs))
1916 state->free_reqs = nr;
1921 * A request might get retired back into the request caches even before opcode
1922 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1923 * Because of that, io_alloc_req() should be called only under ->uring_lock
1924 * and with extra caution to not get a request that is still worked on.
1926 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1927 __must_hold(&ctx->uring_lock)
1929 struct io_submit_state *state = &ctx->submit_state;
1930 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1933 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1935 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1938 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1942 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1943 * retry single alloc to be on the safe side.
1945 if (unlikely(ret <= 0)) {
1946 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1947 if (!state->reqs[0])
1952 for (i = 0; i < ret; i++)
1953 io_preinit_req(state->reqs[i], ctx);
1954 state->free_reqs = ret;
1957 return state->reqs[state->free_reqs];
1960 static inline void io_put_file(struct file *file)
1966 static void io_dismantle_req(struct io_kiocb *req)
1968 unsigned int flags = req->flags;
1970 if (io_req_needs_clean(req))
1972 if (!(flags & REQ_F_FIXED_FILE))
1973 io_put_file(req->file);
1974 if (req->fixed_rsrc_refs)
1975 percpu_ref_put(req->fixed_rsrc_refs);
1976 if (req->async_data) {
1977 kfree(req->async_data);
1978 req->async_data = NULL;
1982 static void __io_free_req(struct io_kiocb *req)
1984 struct io_ring_ctx *ctx = req->ctx;
1986 io_dismantle_req(req);
1987 io_put_task(req->task, 1);
1989 spin_lock(&ctx->completion_lock);
1990 list_add(&req->inflight_entry, &ctx->locked_free_list);
1991 ctx->locked_free_nr++;
1992 spin_unlock(&ctx->completion_lock);
1994 percpu_ref_put(&ctx->refs);
1997 static inline void io_remove_next_linked(struct io_kiocb *req)
1999 struct io_kiocb *nxt = req->link;
2001 req->link = nxt->link;
2005 static bool io_kill_linked_timeout(struct io_kiocb *req)
2006 __must_hold(&req->ctx->completion_lock)
2007 __must_hold(&req->ctx->timeout_lock)
2009 struct io_kiocb *link = req->link;
2011 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2012 struct io_timeout_data *io = link->async_data;
2014 io_remove_next_linked(req);
2015 link->timeout.head = NULL;
2016 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2017 list_del(&link->timeout.list);
2018 io_cqring_fill_event(link->ctx, link->user_data,
2020 io_put_req_deferred(link);
2027 static void io_fail_links(struct io_kiocb *req)
2028 __must_hold(&req->ctx->completion_lock)
2030 struct io_kiocb *nxt, *link = req->link;
2034 long res = -ECANCELED;
2036 if (link->flags & REQ_F_FAIL)
2042 trace_io_uring_fail_link(req, link);
2043 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2044 io_put_req_deferred(link);
2049 static bool io_disarm_next(struct io_kiocb *req)
2050 __must_hold(&req->ctx->completion_lock)
2052 bool posted = false;
2054 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2055 struct io_kiocb *link = req->link;
2057 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2058 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2059 io_remove_next_linked(req);
2060 io_cqring_fill_event(link->ctx, link->user_data,
2062 io_put_req_deferred(link);
2065 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2066 struct io_ring_ctx *ctx = req->ctx;
2068 spin_lock_irq(&ctx->timeout_lock);
2069 posted = io_kill_linked_timeout(req);
2070 spin_unlock_irq(&ctx->timeout_lock);
2072 if (unlikely((req->flags & REQ_F_FAIL) &&
2073 !(req->flags & REQ_F_HARDLINK))) {
2074 posted |= (req->link != NULL);
2080 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2082 struct io_kiocb *nxt;
2085 * If LINK is set, we have dependent requests in this chain. If we
2086 * didn't fail this request, queue the first one up, moving any other
2087 * dependencies to the next request. In case of failure, fail the rest
2090 if (req->flags & IO_DISARM_MASK) {
2091 struct io_ring_ctx *ctx = req->ctx;
2094 spin_lock(&ctx->completion_lock);
2095 posted = io_disarm_next(req);
2097 io_commit_cqring(req->ctx);
2098 spin_unlock(&ctx->completion_lock);
2100 io_cqring_ev_posted(ctx);
2107 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2109 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2111 return __io_req_find_next(req);
2114 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2119 if (ctx->submit_state.compl_nr)
2120 io_submit_flush_completions(ctx);
2121 mutex_unlock(&ctx->uring_lock);
2124 percpu_ref_put(&ctx->refs);
2127 static void tctx_task_work(struct callback_head *cb)
2129 bool locked = false;
2130 struct io_ring_ctx *ctx = NULL;
2131 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2135 struct io_wq_work_node *node;
2137 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2138 io_submit_flush_completions(ctx);
2140 spin_lock_irq(&tctx->task_lock);
2141 node = tctx->task_list.first;
2142 INIT_WQ_LIST(&tctx->task_list);
2144 tctx->task_running = false;
2145 spin_unlock_irq(&tctx->task_lock);
2150 struct io_wq_work_node *next = node->next;
2151 struct io_kiocb *req = container_of(node, struct io_kiocb,
2154 if (req->ctx != ctx) {
2155 ctx_flush_and_put(ctx, &locked);
2157 /* if not contended, grab and improve batching */
2158 locked = mutex_trylock(&ctx->uring_lock);
2159 percpu_ref_get(&ctx->refs);
2161 req->io_task_work.func(req, &locked);
2168 ctx_flush_and_put(ctx, &locked);
2171 static void io_req_task_work_add(struct io_kiocb *req)
2173 struct task_struct *tsk = req->task;
2174 struct io_uring_task *tctx = tsk->io_uring;
2175 enum task_work_notify_mode notify;
2176 struct io_wq_work_node *node;
2177 unsigned long flags;
2180 WARN_ON_ONCE(!tctx);
2182 spin_lock_irqsave(&tctx->task_lock, flags);
2183 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2184 running = tctx->task_running;
2186 tctx->task_running = true;
2187 spin_unlock_irqrestore(&tctx->task_lock, flags);
2189 /* task_work already pending, we're done */
2194 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2195 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2196 * processing task_work. There's no reliable way to tell if TWA_RESUME
2199 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2200 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2201 wake_up_process(tsk);
2205 spin_lock_irqsave(&tctx->task_lock, flags);
2206 tctx->task_running = false;
2207 node = tctx->task_list.first;
2208 INIT_WQ_LIST(&tctx->task_list);
2209 spin_unlock_irqrestore(&tctx->task_lock, flags);
2212 req = container_of(node, struct io_kiocb, io_task_work.node);
2214 if (llist_add(&req->io_task_work.fallback_node,
2215 &req->ctx->fallback_llist))
2216 schedule_delayed_work(&req->ctx->fallback_work, 1);
2220 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2222 struct io_ring_ctx *ctx = req->ctx;
2224 /* not needed for normal modes, but SQPOLL depends on it */
2225 io_tw_lock(ctx, locked);
2226 io_req_complete_failed(req, req->result);
2229 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2231 struct io_ring_ctx *ctx = req->ctx;
2233 io_tw_lock(ctx, locked);
2234 /* req->task == current here, checking PF_EXITING is safe */
2235 if (likely(!(req->task->flags & PF_EXITING)))
2236 __io_queue_sqe(req);
2238 io_req_complete_failed(req, -EFAULT);
2241 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2244 req->io_task_work.func = io_req_task_cancel;
2245 io_req_task_work_add(req);
2248 static void io_req_task_queue(struct io_kiocb *req)
2250 req->io_task_work.func = io_req_task_submit;
2251 io_req_task_work_add(req);
2254 static void io_req_task_queue_reissue(struct io_kiocb *req)
2256 req->io_task_work.func = io_queue_async_work;
2257 io_req_task_work_add(req);
2260 static inline void io_queue_next(struct io_kiocb *req)
2262 struct io_kiocb *nxt = io_req_find_next(req);
2265 io_req_task_queue(nxt);
2268 static void io_free_req(struct io_kiocb *req)
2274 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2280 struct task_struct *task;
2285 static inline void io_init_req_batch(struct req_batch *rb)
2292 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2293 struct req_batch *rb)
2296 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2298 io_put_task(rb->task, rb->task_refs);
2301 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2302 struct io_submit_state *state)
2305 io_dismantle_req(req);
2307 if (req->task != rb->task) {
2309 io_put_task(rb->task, rb->task_refs);
2310 rb->task = req->task;
2316 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2317 state->reqs[state->free_reqs++] = req;
2319 list_add(&req->inflight_entry, &state->free_list);
2322 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2323 __must_hold(&ctx->uring_lock)
2325 struct io_submit_state *state = &ctx->submit_state;
2326 int i, nr = state->compl_nr;
2327 struct req_batch rb;
2329 spin_lock(&ctx->completion_lock);
2330 for (i = 0; i < nr; i++) {
2331 struct io_kiocb *req = state->compl_reqs[i];
2333 __io_cqring_fill_event(ctx, req->user_data, req->result,
2336 io_commit_cqring(ctx);
2337 spin_unlock(&ctx->completion_lock);
2338 io_cqring_ev_posted(ctx);
2340 io_init_req_batch(&rb);
2341 for (i = 0; i < nr; i++) {
2342 struct io_kiocb *req = state->compl_reqs[i];
2344 if (req_ref_put_and_test(req))
2345 io_req_free_batch(&rb, req, &ctx->submit_state);
2348 io_req_free_batch_finish(ctx, &rb);
2349 state->compl_nr = 0;
2353 * Drop reference to request, return next in chain (if there is one) if this
2354 * was the last reference to this request.
2356 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2358 struct io_kiocb *nxt = NULL;
2360 if (req_ref_put_and_test(req)) {
2361 nxt = io_req_find_next(req);
2367 static inline void io_put_req(struct io_kiocb *req)
2369 if (req_ref_put_and_test(req))
2373 static inline void io_put_req_deferred(struct io_kiocb *req)
2375 if (req_ref_put_and_test(req)) {
2376 req->io_task_work.func = io_free_req_work;
2377 io_req_task_work_add(req);
2381 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2383 /* See comment at the top of this file */
2385 return __io_cqring_events(ctx);
2388 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2390 struct io_rings *rings = ctx->rings;
2392 /* make sure SQ entry isn't read before tail */
2393 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2396 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2398 unsigned int cflags;
2400 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2401 cflags |= IORING_CQE_F_BUFFER;
2402 req->flags &= ~REQ_F_BUFFER_SELECTED;
2407 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2409 struct io_buffer *kbuf;
2411 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2413 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2414 return io_put_kbuf(req, kbuf);
2417 static inline bool io_run_task_work(void)
2419 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2420 __set_current_state(TASK_RUNNING);
2421 tracehook_notify_signal();
2429 * Find and free completed poll iocbs
2431 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2432 struct list_head *done)
2434 struct req_batch rb;
2435 struct io_kiocb *req;
2437 /* order with ->result store in io_complete_rw_iopoll() */
2440 io_init_req_batch(&rb);
2441 while (!list_empty(done)) {
2442 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2443 list_del(&req->inflight_entry);
2445 if (READ_ONCE(req->result) == -EAGAIN &&
2446 !(req->flags & REQ_F_DONT_REISSUE)) {
2447 req->iopoll_completed = 0;
2448 io_req_task_queue_reissue(req);
2452 __io_cqring_fill_event(ctx, req->user_data, req->result,
2453 io_put_rw_kbuf(req));
2456 if (req_ref_put_and_test(req))
2457 io_req_free_batch(&rb, req, &ctx->submit_state);
2460 io_commit_cqring(ctx);
2461 io_cqring_ev_posted_iopoll(ctx);
2462 io_req_free_batch_finish(ctx, &rb);
2465 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2468 struct io_kiocb *req, *tmp;
2473 * Only spin for completions if we don't have multiple devices hanging
2474 * off our complete list, and we're under the requested amount.
2476 spin = !ctx->poll_multi_queue && *nr_events < min;
2478 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2479 struct kiocb *kiocb = &req->rw.kiocb;
2483 * Move completed and retryable entries to our local lists.
2484 * If we find a request that requires polling, break out
2485 * and complete those lists first, if we have entries there.
2487 if (READ_ONCE(req->iopoll_completed)) {
2488 list_move_tail(&req->inflight_entry, &done);
2491 if (!list_empty(&done))
2494 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2495 if (unlikely(ret < 0))
2500 /* iopoll may have completed current req */
2501 if (READ_ONCE(req->iopoll_completed))
2502 list_move_tail(&req->inflight_entry, &done);
2505 if (!list_empty(&done))
2506 io_iopoll_complete(ctx, nr_events, &done);
2512 * We can't just wait for polled events to come to us, we have to actively
2513 * find and complete them.
2515 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2517 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2520 mutex_lock(&ctx->uring_lock);
2521 while (!list_empty(&ctx->iopoll_list)) {
2522 unsigned int nr_events = 0;
2524 io_do_iopoll(ctx, &nr_events, 0);
2526 /* let it sleep and repeat later if can't complete a request */
2530 * Ensure we allow local-to-the-cpu processing to take place,
2531 * in this case we need to ensure that we reap all events.
2532 * Also let task_work, etc. to progress by releasing the mutex
2534 if (need_resched()) {
2535 mutex_unlock(&ctx->uring_lock);
2537 mutex_lock(&ctx->uring_lock);
2540 mutex_unlock(&ctx->uring_lock);
2543 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2545 unsigned int nr_events = 0;
2549 * We disallow the app entering submit/complete with polling, but we
2550 * still need to lock the ring to prevent racing with polled issue
2551 * that got punted to a workqueue.
2553 mutex_lock(&ctx->uring_lock);
2555 * Don't enter poll loop if we already have events pending.
2556 * If we do, we can potentially be spinning for commands that
2557 * already triggered a CQE (eg in error).
2559 if (test_bit(0, &ctx->check_cq_overflow))
2560 __io_cqring_overflow_flush(ctx, false);
2561 if (io_cqring_events(ctx))
2565 * If a submit got punted to a workqueue, we can have the
2566 * application entering polling for a command before it gets
2567 * issued. That app will hold the uring_lock for the duration
2568 * of the poll right here, so we need to take a breather every
2569 * now and then to ensure that the issue has a chance to add
2570 * the poll to the issued list. Otherwise we can spin here
2571 * forever, while the workqueue is stuck trying to acquire the
2574 if (list_empty(&ctx->iopoll_list)) {
2575 u32 tail = ctx->cached_cq_tail;
2577 mutex_unlock(&ctx->uring_lock);
2579 mutex_lock(&ctx->uring_lock);
2581 /* some requests don't go through iopoll_list */
2582 if (tail != ctx->cached_cq_tail ||
2583 list_empty(&ctx->iopoll_list))
2586 ret = io_do_iopoll(ctx, &nr_events, min);
2587 } while (!ret && nr_events < min && !need_resched());
2589 mutex_unlock(&ctx->uring_lock);
2593 static void kiocb_end_write(struct io_kiocb *req)
2596 * Tell lockdep we inherited freeze protection from submission
2599 if (req->flags & REQ_F_ISREG) {
2600 struct super_block *sb = file_inode(req->file)->i_sb;
2602 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2608 static bool io_resubmit_prep(struct io_kiocb *req)
2610 struct io_async_rw *rw = req->async_data;
2613 return !io_req_prep_async(req);
2614 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2615 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2619 static bool io_rw_should_reissue(struct io_kiocb *req)
2621 umode_t mode = file_inode(req->file)->i_mode;
2622 struct io_ring_ctx *ctx = req->ctx;
2624 if (!S_ISBLK(mode) && !S_ISREG(mode))
2626 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2627 !(ctx->flags & IORING_SETUP_IOPOLL)))
2630 * If ref is dying, we might be running poll reap from the exit work.
2631 * Don't attempt to reissue from that path, just let it fail with
2634 if (percpu_ref_is_dying(&ctx->refs))
2637 * Play it safe and assume not safe to re-import and reissue if we're
2638 * not in the original thread group (or in task context).
2640 if (!same_thread_group(req->task, current) || !in_task())
2645 static bool io_resubmit_prep(struct io_kiocb *req)
2649 static bool io_rw_should_reissue(struct io_kiocb *req)
2655 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2657 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2658 kiocb_end_write(req);
2659 if (res != req->result) {
2660 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2661 io_rw_should_reissue(req)) {
2662 req->flags |= REQ_F_REISSUE;
2671 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2673 unsigned int cflags = io_put_rw_kbuf(req);
2674 long res = req->result;
2677 struct io_ring_ctx *ctx = req->ctx;
2678 struct io_submit_state *state = &ctx->submit_state;
2680 io_req_complete_state(req, res, cflags);
2681 state->compl_reqs[state->compl_nr++] = req;
2682 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2683 io_submit_flush_completions(ctx);
2685 io_req_complete_post(req, res, cflags);
2689 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2690 unsigned int issue_flags)
2692 if (__io_complete_rw_common(req, res))
2694 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2697 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2699 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2701 if (__io_complete_rw_common(req, res))
2704 req->io_task_work.func = io_req_task_complete;
2705 io_req_task_work_add(req);
2708 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2710 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2712 if (kiocb->ki_flags & IOCB_WRITE)
2713 kiocb_end_write(req);
2714 if (unlikely(res != req->result)) {
2715 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2716 io_resubmit_prep(req))) {
2718 req->flags |= REQ_F_DONT_REISSUE;
2722 WRITE_ONCE(req->result, res);
2723 /* order with io_iopoll_complete() checking ->result */
2725 WRITE_ONCE(req->iopoll_completed, 1);
2729 * After the iocb has been issued, it's safe to be found on the poll list.
2730 * Adding the kiocb to the list AFTER submission ensures that we don't
2731 * find it from a io_do_iopoll() thread before the issuer is done
2732 * accessing the kiocb cookie.
2734 static void io_iopoll_req_issued(struct io_kiocb *req)
2736 struct io_ring_ctx *ctx = req->ctx;
2737 const bool in_async = io_wq_current_is_worker();
2739 /* workqueue context doesn't hold uring_lock, grab it now */
2740 if (unlikely(in_async))
2741 mutex_lock(&ctx->uring_lock);
2744 * Track whether we have multiple files in our lists. This will impact
2745 * how we do polling eventually, not spinning if we're on potentially
2746 * different devices.
2748 if (list_empty(&ctx->iopoll_list)) {
2749 ctx->poll_multi_queue = false;
2750 } else if (!ctx->poll_multi_queue) {
2751 struct io_kiocb *list_req;
2752 unsigned int queue_num0, queue_num1;
2754 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2757 if (list_req->file != req->file) {
2758 ctx->poll_multi_queue = true;
2760 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2761 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2762 if (queue_num0 != queue_num1)
2763 ctx->poll_multi_queue = true;
2768 * For fast devices, IO may have already completed. If it has, add
2769 * it to the front so we find it first.
2771 if (READ_ONCE(req->iopoll_completed))
2772 list_add(&req->inflight_entry, &ctx->iopoll_list);
2774 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2776 if (unlikely(in_async)) {
2778 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2779 * in sq thread task context or in io worker task context. If
2780 * current task context is sq thread, we don't need to check
2781 * whether should wake up sq thread.
2783 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2784 wq_has_sleeper(&ctx->sq_data->wait))
2785 wake_up(&ctx->sq_data->wait);
2787 mutex_unlock(&ctx->uring_lock);
2791 static bool io_bdev_nowait(struct block_device *bdev)
2793 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2797 * If we tracked the file through the SCM inflight mechanism, we could support
2798 * any file. For now, just ensure that anything potentially problematic is done
2801 static bool __io_file_supports_nowait(struct file *file, int rw)
2803 umode_t mode = file_inode(file)->i_mode;
2805 if (S_ISBLK(mode)) {
2806 if (IS_ENABLED(CONFIG_BLOCK) &&
2807 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2813 if (S_ISREG(mode)) {
2814 if (IS_ENABLED(CONFIG_BLOCK) &&
2815 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2816 file->f_op != &io_uring_fops)
2821 /* any ->read/write should understand O_NONBLOCK */
2822 if (file->f_flags & O_NONBLOCK)
2825 if (!(file->f_mode & FMODE_NOWAIT))
2829 return file->f_op->read_iter != NULL;
2831 return file->f_op->write_iter != NULL;
2834 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2836 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2838 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2841 return __io_file_supports_nowait(req->file, rw);
2844 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2846 struct io_ring_ctx *ctx = req->ctx;
2847 struct kiocb *kiocb = &req->rw.kiocb;
2848 struct file *file = req->file;
2852 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2853 req->flags |= REQ_F_ISREG;
2855 kiocb->ki_pos = READ_ONCE(sqe->off);
2856 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2857 req->flags |= REQ_F_CUR_POS;
2858 kiocb->ki_pos = file->f_pos;
2860 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2861 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2862 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2866 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2867 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2868 req->flags |= REQ_F_NOWAIT;
2870 ioprio = READ_ONCE(sqe->ioprio);
2872 ret = ioprio_check_cap(ioprio);
2876 kiocb->ki_ioprio = ioprio;
2878 kiocb->ki_ioprio = get_current_ioprio();
2880 if (ctx->flags & IORING_SETUP_IOPOLL) {
2881 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2882 !kiocb->ki_filp->f_op->iopoll)
2885 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2886 kiocb->ki_complete = io_complete_rw_iopoll;
2887 req->iopoll_completed = 0;
2889 if (kiocb->ki_flags & IOCB_HIPRI)
2891 kiocb->ki_complete = io_complete_rw;
2894 if (req->opcode == IORING_OP_READ_FIXED ||
2895 req->opcode == IORING_OP_WRITE_FIXED) {
2897 io_req_set_rsrc_node(req);
2900 req->rw.addr = READ_ONCE(sqe->addr);
2901 req->rw.len = READ_ONCE(sqe->len);
2902 req->buf_index = READ_ONCE(sqe->buf_index);
2906 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2912 case -ERESTARTNOINTR:
2913 case -ERESTARTNOHAND:
2914 case -ERESTART_RESTARTBLOCK:
2916 * We can't just restart the syscall, since previously
2917 * submitted sqes may already be in progress. Just fail this
2923 kiocb->ki_complete(kiocb, ret, 0);
2927 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2928 unsigned int issue_flags)
2930 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2931 struct io_async_rw *io = req->async_data;
2932 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2934 /* add previously done IO, if any */
2935 if (io && io->bytes_done > 0) {
2937 ret = io->bytes_done;
2939 ret += io->bytes_done;
2942 if (req->flags & REQ_F_CUR_POS)
2943 req->file->f_pos = kiocb->ki_pos;
2944 if (ret >= 0 && check_reissue)
2945 __io_complete_rw(req, ret, 0, issue_flags);
2947 io_rw_done(kiocb, ret);
2949 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2950 req->flags &= ~REQ_F_REISSUE;
2951 if (io_resubmit_prep(req)) {
2952 io_req_task_queue_reissue(req);
2955 __io_req_complete(req, issue_flags, ret,
2956 io_put_rw_kbuf(req));
2961 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2962 struct io_mapped_ubuf *imu)
2964 size_t len = req->rw.len;
2965 u64 buf_end, buf_addr = req->rw.addr;
2968 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2970 /* not inside the mapped region */
2971 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2975 * May not be a start of buffer, set size appropriately
2976 * and advance us to the beginning.
2978 offset = buf_addr - imu->ubuf;
2979 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2983 * Don't use iov_iter_advance() here, as it's really slow for
2984 * using the latter parts of a big fixed buffer - it iterates
2985 * over each segment manually. We can cheat a bit here, because
2988 * 1) it's a BVEC iter, we set it up
2989 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2990 * first and last bvec
2992 * So just find our index, and adjust the iterator afterwards.
2993 * If the offset is within the first bvec (or the whole first
2994 * bvec, just use iov_iter_advance(). This makes it easier
2995 * since we can just skip the first segment, which may not
2996 * be PAGE_SIZE aligned.
2998 const struct bio_vec *bvec = imu->bvec;
3000 if (offset <= bvec->bv_len) {
3001 iov_iter_advance(iter, offset);
3003 unsigned long seg_skip;
3005 /* skip first vec */
3006 offset -= bvec->bv_len;
3007 seg_skip = 1 + (offset >> PAGE_SHIFT);
3009 iter->bvec = bvec + seg_skip;
3010 iter->nr_segs -= seg_skip;
3011 iter->count -= bvec->bv_len + offset;
3012 iter->iov_offset = offset & ~PAGE_MASK;
3019 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3021 struct io_ring_ctx *ctx = req->ctx;
3022 struct io_mapped_ubuf *imu = req->imu;
3023 u16 index, buf_index = req->buf_index;
3026 if (unlikely(buf_index >= ctx->nr_user_bufs))
3028 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3029 imu = READ_ONCE(ctx->user_bufs[index]);
3032 return __io_import_fixed(req, rw, iter, imu);
3035 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3038 mutex_unlock(&ctx->uring_lock);
3041 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3044 * "Normal" inline submissions always hold the uring_lock, since we
3045 * grab it from the system call. Same is true for the SQPOLL offload.
3046 * The only exception is when we've detached the request and issue it
3047 * from an async worker thread, grab the lock for that case.
3050 mutex_lock(&ctx->uring_lock);
3053 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3054 int bgid, struct io_buffer *kbuf,
3057 struct io_buffer *head;
3059 if (req->flags & REQ_F_BUFFER_SELECTED)
3062 io_ring_submit_lock(req->ctx, needs_lock);
3064 lockdep_assert_held(&req->ctx->uring_lock);
3066 head = xa_load(&req->ctx->io_buffers, bgid);
3068 if (!list_empty(&head->list)) {
3069 kbuf = list_last_entry(&head->list, struct io_buffer,
3071 list_del(&kbuf->list);
3074 xa_erase(&req->ctx->io_buffers, bgid);
3076 if (*len > kbuf->len)
3079 kbuf = ERR_PTR(-ENOBUFS);
3082 io_ring_submit_unlock(req->ctx, needs_lock);
3087 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3090 struct io_buffer *kbuf;
3093 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3094 bgid = req->buf_index;
3095 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3098 req->rw.addr = (u64) (unsigned long) kbuf;
3099 req->flags |= REQ_F_BUFFER_SELECTED;
3100 return u64_to_user_ptr(kbuf->addr);
3103 #ifdef CONFIG_COMPAT
3104 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3107 struct compat_iovec __user *uiov;
3108 compat_ssize_t clen;
3112 uiov = u64_to_user_ptr(req->rw.addr);
3113 if (!access_ok(uiov, sizeof(*uiov)))
3115 if (__get_user(clen, &uiov->iov_len))
3121 buf = io_rw_buffer_select(req, &len, needs_lock);
3123 return PTR_ERR(buf);
3124 iov[0].iov_base = buf;
3125 iov[0].iov_len = (compat_size_t) len;
3130 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3133 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3137 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3140 len = iov[0].iov_len;
3143 buf = io_rw_buffer_select(req, &len, needs_lock);
3145 return PTR_ERR(buf);
3146 iov[0].iov_base = buf;
3147 iov[0].iov_len = len;
3151 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3154 if (req->flags & REQ_F_BUFFER_SELECTED) {
3155 struct io_buffer *kbuf;
3157 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3158 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3159 iov[0].iov_len = kbuf->len;
3162 if (req->rw.len != 1)
3165 #ifdef CONFIG_COMPAT
3166 if (req->ctx->compat)
3167 return io_compat_import(req, iov, needs_lock);
3170 return __io_iov_buffer_select(req, iov, needs_lock);
3173 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3174 struct iov_iter *iter, bool needs_lock)
3176 void __user *buf = u64_to_user_ptr(req->rw.addr);
3177 size_t sqe_len = req->rw.len;
3178 u8 opcode = req->opcode;
3181 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3183 return io_import_fixed(req, rw, iter);
3186 /* buffer index only valid with fixed read/write, or buffer select */
3187 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3190 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3191 if (req->flags & REQ_F_BUFFER_SELECT) {
3192 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3194 return PTR_ERR(buf);
3195 req->rw.len = sqe_len;
3198 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3203 if (req->flags & REQ_F_BUFFER_SELECT) {
3204 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3206 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3211 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3215 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3217 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3221 * For files that don't have ->read_iter() and ->write_iter(), handle them
3222 * by looping over ->read() or ->write() manually.
3224 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3226 struct kiocb *kiocb = &req->rw.kiocb;
3227 struct file *file = req->file;
3231 * Don't support polled IO through this interface, and we can't
3232 * support non-blocking either. For the latter, this just causes
3233 * the kiocb to be handled from an async context.
3235 if (kiocb->ki_flags & IOCB_HIPRI)
3237 if (kiocb->ki_flags & IOCB_NOWAIT)
3240 while (iov_iter_count(iter)) {
3244 if (!iov_iter_is_bvec(iter)) {
3245 iovec = iov_iter_iovec(iter);
3247 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3248 iovec.iov_len = req->rw.len;
3252 nr = file->f_op->read(file, iovec.iov_base,
3253 iovec.iov_len, io_kiocb_ppos(kiocb));
3255 nr = file->f_op->write(file, iovec.iov_base,
3256 iovec.iov_len, io_kiocb_ppos(kiocb));
3265 if (nr != iovec.iov_len)
3269 iov_iter_advance(iter, nr);
3275 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3276 const struct iovec *fast_iov, struct iov_iter *iter)
3278 struct io_async_rw *rw = req->async_data;
3280 memcpy(&rw->iter, iter, sizeof(*iter));
3281 rw->free_iovec = iovec;
3283 /* can only be fixed buffers, no need to do anything */
3284 if (iov_iter_is_bvec(iter))
3287 unsigned iov_off = 0;
3289 rw->iter.iov = rw->fast_iov;
3290 if (iter->iov != fast_iov) {
3291 iov_off = iter->iov - fast_iov;
3292 rw->iter.iov += iov_off;
3294 if (rw->fast_iov != fast_iov)
3295 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3296 sizeof(struct iovec) * iter->nr_segs);
3298 req->flags |= REQ_F_NEED_CLEANUP;
3302 static inline int io_alloc_async_data(struct io_kiocb *req)
3304 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3305 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3306 return req->async_data == NULL;
3309 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3310 const struct iovec *fast_iov,
3311 struct iov_iter *iter, bool force)
3313 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3315 if (!req->async_data) {
3316 if (io_alloc_async_data(req)) {
3321 io_req_map_rw(req, iovec, fast_iov, iter);
3326 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3328 struct io_async_rw *iorw = req->async_data;
3329 struct iovec *iov = iorw->fast_iov;
3332 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3333 if (unlikely(ret < 0))
3336 iorw->bytes_done = 0;
3337 iorw->free_iovec = iov;
3339 req->flags |= REQ_F_NEED_CLEANUP;
3343 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3345 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3347 return io_prep_rw(req, sqe);
3351 * This is our waitqueue callback handler, registered through lock_page_async()
3352 * when we initially tried to do the IO with the iocb armed our waitqueue.
3353 * This gets called when the page is unlocked, and we generally expect that to
3354 * happen when the page IO is completed and the page is now uptodate. This will
3355 * queue a task_work based retry of the operation, attempting to copy the data
3356 * again. If the latter fails because the page was NOT uptodate, then we will
3357 * do a thread based blocking retry of the operation. That's the unexpected
3360 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3361 int sync, void *arg)
3363 struct wait_page_queue *wpq;
3364 struct io_kiocb *req = wait->private;
3365 struct wait_page_key *key = arg;
3367 wpq = container_of(wait, struct wait_page_queue, wait);
3369 if (!wake_page_match(wpq, key))
3372 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3373 list_del_init(&wait->entry);
3374 io_req_task_queue(req);
3379 * This controls whether a given IO request should be armed for async page
3380 * based retry. If we return false here, the request is handed to the async
3381 * worker threads for retry. If we're doing buffered reads on a regular file,
3382 * we prepare a private wait_page_queue entry and retry the operation. This
3383 * will either succeed because the page is now uptodate and unlocked, or it
3384 * will register a callback when the page is unlocked at IO completion. Through
3385 * that callback, io_uring uses task_work to setup a retry of the operation.
3386 * That retry will attempt the buffered read again. The retry will generally
3387 * succeed, or in rare cases where it fails, we then fall back to using the
3388 * async worker threads for a blocking retry.
3390 static bool io_rw_should_retry(struct io_kiocb *req)
3392 struct io_async_rw *rw = req->async_data;
3393 struct wait_page_queue *wait = &rw->wpq;
3394 struct kiocb *kiocb = &req->rw.kiocb;
3396 /* never retry for NOWAIT, we just complete with -EAGAIN */
3397 if (req->flags & REQ_F_NOWAIT)
3400 /* Only for buffered IO */
3401 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3405 * just use poll if we can, and don't attempt if the fs doesn't
3406 * support callback based unlocks
3408 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3411 wait->wait.func = io_async_buf_func;
3412 wait->wait.private = req;
3413 wait->wait.flags = 0;
3414 INIT_LIST_HEAD(&wait->wait.entry);
3415 kiocb->ki_flags |= IOCB_WAITQ;
3416 kiocb->ki_flags &= ~IOCB_NOWAIT;
3417 kiocb->ki_waitq = wait;
3421 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3423 if (req->file->f_op->read_iter)
3424 return call_read_iter(req->file, &req->rw.kiocb, iter);
3425 else if (req->file->f_op->read)
3426 return loop_rw_iter(READ, req, iter);
3431 static bool need_read_all(struct io_kiocb *req)
3433 return req->flags & REQ_F_ISREG ||
3434 S_ISBLK(file_inode(req->file)->i_mode);
3437 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3439 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3440 struct kiocb *kiocb = &req->rw.kiocb;
3441 struct iov_iter __iter, *iter = &__iter;
3442 struct io_async_rw *rw = req->async_data;
3443 ssize_t io_size, ret, ret2;
3444 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3450 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3454 io_size = iov_iter_count(iter);
3455 req->result = io_size;
3457 /* Ensure we clear previously set non-block flag */
3458 if (!force_nonblock)
3459 kiocb->ki_flags &= ~IOCB_NOWAIT;
3461 kiocb->ki_flags |= IOCB_NOWAIT;
3463 /* If the file doesn't support async, just async punt */
3464 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3465 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3466 return ret ?: -EAGAIN;
3469 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3470 if (unlikely(ret)) {
3475 ret = io_iter_do_read(req, iter);
3477 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3478 req->flags &= ~REQ_F_REISSUE;
3479 /* IOPOLL retry should happen for io-wq threads */
3480 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3482 /* no retry on NONBLOCK nor RWF_NOWAIT */
3483 if (req->flags & REQ_F_NOWAIT)
3485 /* some cases will consume bytes even on error returns */
3486 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3488 } else if (ret == -EIOCBQUEUED) {
3490 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3491 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3492 /* read all, failed, already did sync or don't want to retry */
3496 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3501 rw = req->async_data;
3502 /* now use our persistent iterator, if we aren't already */
3507 rw->bytes_done += ret;
3508 /* if we can retry, do so with the callbacks armed */
3509 if (!io_rw_should_retry(req)) {
3510 kiocb->ki_flags &= ~IOCB_WAITQ;
3515 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3516 * we get -EIOCBQUEUED, then we'll get a notification when the
3517 * desired page gets unlocked. We can also get a partial read
3518 * here, and if we do, then just retry at the new offset.
3520 ret = io_iter_do_read(req, iter);
3521 if (ret == -EIOCBQUEUED)
3523 /* we got some bytes, but not all. retry. */
3524 kiocb->ki_flags &= ~IOCB_WAITQ;
3525 } while (ret > 0 && ret < io_size);
3527 kiocb_done(kiocb, ret, issue_flags);
3529 /* it's faster to check here then delegate to kfree */
3535 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3537 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3539 return io_prep_rw(req, sqe);
3542 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3544 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3545 struct kiocb *kiocb = &req->rw.kiocb;
3546 struct iov_iter __iter, *iter = &__iter;
3547 struct io_async_rw *rw = req->async_data;
3548 ssize_t ret, ret2, io_size;
3549 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3555 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3559 io_size = iov_iter_count(iter);
3560 req->result = io_size;
3562 /* Ensure we clear previously set non-block flag */
3563 if (!force_nonblock)
3564 kiocb->ki_flags &= ~IOCB_NOWAIT;
3566 kiocb->ki_flags |= IOCB_NOWAIT;
3568 /* If the file doesn't support async, just async punt */
3569 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3572 /* file path doesn't support NOWAIT for non-direct_IO */
3573 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3574 (req->flags & REQ_F_ISREG))
3577 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3582 * Open-code file_start_write here to grab freeze protection,
3583 * which will be released by another thread in
3584 * io_complete_rw(). Fool lockdep by telling it the lock got
3585 * released so that it doesn't complain about the held lock when
3586 * we return to userspace.
3588 if (req->flags & REQ_F_ISREG) {
3589 sb_start_write(file_inode(req->file)->i_sb);
3590 __sb_writers_release(file_inode(req->file)->i_sb,
3593 kiocb->ki_flags |= IOCB_WRITE;
3595 if (req->file->f_op->write_iter)
3596 ret2 = call_write_iter(req->file, kiocb, iter);
3597 else if (req->file->f_op->write)
3598 ret2 = loop_rw_iter(WRITE, req, iter);
3602 if (req->flags & REQ_F_REISSUE) {
3603 req->flags &= ~REQ_F_REISSUE;
3608 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3609 * retry them without IOCB_NOWAIT.
3611 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3613 /* no retry on NONBLOCK nor RWF_NOWAIT */
3614 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3616 if (!force_nonblock || ret2 != -EAGAIN) {
3617 /* IOPOLL retry should happen for io-wq threads */
3618 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3621 kiocb_done(kiocb, ret2, issue_flags);
3624 /* some cases will consume bytes even on error returns */
3625 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3626 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3627 return ret ?: -EAGAIN;
3630 /* it's reportedly faster than delegating the null check to kfree() */
3636 static int io_renameat_prep(struct io_kiocb *req,
3637 const struct io_uring_sqe *sqe)
3639 struct io_rename *ren = &req->rename;
3640 const char __user *oldf, *newf;
3642 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3644 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3646 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3649 ren->old_dfd = READ_ONCE(sqe->fd);
3650 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3651 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3652 ren->new_dfd = READ_ONCE(sqe->len);
3653 ren->flags = READ_ONCE(sqe->rename_flags);
3655 ren->oldpath = getname(oldf);
3656 if (IS_ERR(ren->oldpath))
3657 return PTR_ERR(ren->oldpath);
3659 ren->newpath = getname(newf);
3660 if (IS_ERR(ren->newpath)) {
3661 putname(ren->oldpath);
3662 return PTR_ERR(ren->newpath);
3665 req->flags |= REQ_F_NEED_CLEANUP;
3669 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3671 struct io_rename *ren = &req->rename;
3674 if (issue_flags & IO_URING_F_NONBLOCK)
3677 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3678 ren->newpath, ren->flags);
3680 req->flags &= ~REQ_F_NEED_CLEANUP;
3683 io_req_complete(req, ret);
3687 static int io_unlinkat_prep(struct io_kiocb *req,
3688 const struct io_uring_sqe *sqe)
3690 struct io_unlink *un = &req->unlink;
3691 const char __user *fname;
3693 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3695 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3698 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3701 un->dfd = READ_ONCE(sqe->fd);
3703 un->flags = READ_ONCE(sqe->unlink_flags);
3704 if (un->flags & ~AT_REMOVEDIR)
3707 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3708 un->filename = getname(fname);
3709 if (IS_ERR(un->filename))
3710 return PTR_ERR(un->filename);
3712 req->flags |= REQ_F_NEED_CLEANUP;
3716 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3718 struct io_unlink *un = &req->unlink;
3721 if (issue_flags & IO_URING_F_NONBLOCK)
3724 if (un->flags & AT_REMOVEDIR)
3725 ret = do_rmdir(un->dfd, un->filename);
3727 ret = do_unlinkat(un->dfd, un->filename);
3729 req->flags &= ~REQ_F_NEED_CLEANUP;
3732 io_req_complete(req, ret);
3736 static int io_mkdirat_prep(struct io_kiocb *req,
3737 const struct io_uring_sqe *sqe)
3739 struct io_mkdir *mkd = &req->mkdir;
3740 const char __user *fname;
3742 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3744 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3747 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3750 mkd->dfd = READ_ONCE(sqe->fd);
3751 mkd->mode = READ_ONCE(sqe->len);
3753 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3754 mkd->filename = getname(fname);
3755 if (IS_ERR(mkd->filename))
3756 return PTR_ERR(mkd->filename);
3758 req->flags |= REQ_F_NEED_CLEANUP;
3762 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3764 struct io_mkdir *mkd = &req->mkdir;
3767 if (issue_flags & IO_URING_F_NONBLOCK)
3770 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3772 req->flags &= ~REQ_F_NEED_CLEANUP;
3775 io_req_complete(req, ret);
3779 static int io_symlinkat_prep(struct io_kiocb *req,
3780 const struct io_uring_sqe *sqe)
3782 struct io_symlink *sl = &req->symlink;
3783 const char __user *oldpath, *newpath;
3785 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3787 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3790 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3793 sl->new_dfd = READ_ONCE(sqe->fd);
3794 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3795 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3797 sl->oldpath = getname(oldpath);
3798 if (IS_ERR(sl->oldpath))
3799 return PTR_ERR(sl->oldpath);
3801 sl->newpath = getname(newpath);
3802 if (IS_ERR(sl->newpath)) {
3803 putname(sl->oldpath);
3804 return PTR_ERR(sl->newpath);
3807 req->flags |= REQ_F_NEED_CLEANUP;
3811 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3813 struct io_symlink *sl = &req->symlink;
3816 if (issue_flags & IO_URING_F_NONBLOCK)
3819 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3821 req->flags &= ~REQ_F_NEED_CLEANUP;
3824 io_req_complete(req, ret);
3828 static int io_linkat_prep(struct io_kiocb *req,
3829 const struct io_uring_sqe *sqe)
3831 struct io_hardlink *lnk = &req->hardlink;
3832 const char __user *oldf, *newf;
3834 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3836 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3838 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3841 lnk->old_dfd = READ_ONCE(sqe->fd);
3842 lnk->new_dfd = READ_ONCE(sqe->len);
3843 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3844 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3845 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3847 lnk->oldpath = getname(oldf);
3848 if (IS_ERR(lnk->oldpath))
3849 return PTR_ERR(lnk->oldpath);
3851 lnk->newpath = getname(newf);
3852 if (IS_ERR(lnk->newpath)) {
3853 putname(lnk->oldpath);
3854 return PTR_ERR(lnk->newpath);
3857 req->flags |= REQ_F_NEED_CLEANUP;
3861 static int io_linkat(struct io_kiocb *req, int issue_flags)
3863 struct io_hardlink *lnk = &req->hardlink;
3866 if (issue_flags & IO_URING_F_NONBLOCK)
3869 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3870 lnk->newpath, lnk->flags);
3872 req->flags &= ~REQ_F_NEED_CLEANUP;
3875 io_req_complete(req, ret);
3879 static int io_shutdown_prep(struct io_kiocb *req,
3880 const struct io_uring_sqe *sqe)
3882 #if defined(CONFIG_NET)
3883 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3885 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3886 sqe->buf_index || sqe->splice_fd_in))
3889 req->shutdown.how = READ_ONCE(sqe->len);
3896 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3898 #if defined(CONFIG_NET)
3899 struct socket *sock;
3902 if (issue_flags & IO_URING_F_NONBLOCK)
3905 sock = sock_from_file(req->file);
3906 if (unlikely(!sock))
3909 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3912 io_req_complete(req, ret);
3919 static int __io_splice_prep(struct io_kiocb *req,
3920 const struct io_uring_sqe *sqe)
3922 struct io_splice *sp = &req->splice;
3923 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3925 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3929 sp->len = READ_ONCE(sqe->len);
3930 sp->flags = READ_ONCE(sqe->splice_flags);
3932 if (unlikely(sp->flags & ~valid_flags))
3935 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3936 (sp->flags & SPLICE_F_FD_IN_FIXED));
3939 req->flags |= REQ_F_NEED_CLEANUP;
3943 static int io_tee_prep(struct io_kiocb *req,
3944 const struct io_uring_sqe *sqe)
3946 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3948 return __io_splice_prep(req, sqe);
3951 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3953 struct io_splice *sp = &req->splice;
3954 struct file *in = sp->file_in;
3955 struct file *out = sp->file_out;
3956 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3959 if (issue_flags & IO_URING_F_NONBLOCK)
3962 ret = do_tee(in, out, sp->len, flags);
3964 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3966 req->flags &= ~REQ_F_NEED_CLEANUP;
3970 io_req_complete(req, ret);
3974 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3976 struct io_splice *sp = &req->splice;
3978 sp->off_in = READ_ONCE(sqe->splice_off_in);
3979 sp->off_out = READ_ONCE(sqe->off);
3980 return __io_splice_prep(req, sqe);
3983 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3985 struct io_splice *sp = &req->splice;
3986 struct file *in = sp->file_in;
3987 struct file *out = sp->file_out;
3988 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3989 loff_t *poff_in, *poff_out;
3992 if (issue_flags & IO_URING_F_NONBLOCK)
3995 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3996 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3999 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4001 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4003 req->flags &= ~REQ_F_NEED_CLEANUP;
4007 io_req_complete(req, ret);
4012 * IORING_OP_NOP just posts a completion event, nothing else.
4014 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4016 struct io_ring_ctx *ctx = req->ctx;
4018 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4021 __io_req_complete(req, issue_flags, 0, 0);
4025 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4027 struct io_ring_ctx *ctx = req->ctx;
4032 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4034 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4038 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4039 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4042 req->sync.off = READ_ONCE(sqe->off);
4043 req->sync.len = READ_ONCE(sqe->len);
4047 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4049 loff_t end = req->sync.off + req->sync.len;
4052 /* fsync always requires a blocking context */
4053 if (issue_flags & IO_URING_F_NONBLOCK)
4056 ret = vfs_fsync_range(req->file, req->sync.off,
4057 end > 0 ? end : LLONG_MAX,
4058 req->sync.flags & IORING_FSYNC_DATASYNC);
4061 io_req_complete(req, ret);
4065 static int io_fallocate_prep(struct io_kiocb *req,
4066 const struct io_uring_sqe *sqe)
4068 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4071 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4074 req->sync.off = READ_ONCE(sqe->off);
4075 req->sync.len = READ_ONCE(sqe->addr);
4076 req->sync.mode = READ_ONCE(sqe->len);
4080 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4084 /* fallocate always requiring blocking context */
4085 if (issue_flags & IO_URING_F_NONBLOCK)
4087 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4091 io_req_complete(req, ret);
4095 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4097 const char __user *fname;
4100 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4102 if (unlikely(sqe->ioprio || sqe->buf_index))
4104 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4107 /* open.how should be already initialised */
4108 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4109 req->open.how.flags |= O_LARGEFILE;
4111 req->open.dfd = READ_ONCE(sqe->fd);
4112 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4113 req->open.filename = getname(fname);
4114 if (IS_ERR(req->open.filename)) {
4115 ret = PTR_ERR(req->open.filename);
4116 req->open.filename = NULL;
4120 req->open.file_slot = READ_ONCE(sqe->file_index);
4121 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4124 req->open.nofile = rlimit(RLIMIT_NOFILE);
4125 req->flags |= REQ_F_NEED_CLEANUP;
4129 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4131 u64 mode = READ_ONCE(sqe->len);
4132 u64 flags = READ_ONCE(sqe->open_flags);
4134 req->open.how = build_open_how(flags, mode);
4135 return __io_openat_prep(req, sqe);
4138 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4140 struct open_how __user *how;
4144 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4145 len = READ_ONCE(sqe->len);
4146 if (len < OPEN_HOW_SIZE_VER0)
4149 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4154 return __io_openat_prep(req, sqe);
4157 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4159 struct open_flags op;
4161 bool resolve_nonblock, nonblock_set;
4162 bool fixed = !!req->open.file_slot;
4165 ret = build_open_flags(&req->open.how, &op);
4168 nonblock_set = op.open_flag & O_NONBLOCK;
4169 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4170 if (issue_flags & IO_URING_F_NONBLOCK) {
4172 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4173 * it'll always -EAGAIN
4175 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4177 op.lookup_flags |= LOOKUP_CACHED;
4178 op.open_flag |= O_NONBLOCK;
4182 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4187 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4190 * We could hang on to this 'fd' on retrying, but seems like
4191 * marginal gain for something that is now known to be a slower
4192 * path. So just put it, and we'll get a new one when we retry.
4197 ret = PTR_ERR(file);
4198 /* only retry if RESOLVE_CACHED wasn't already set by application */
4199 if (ret == -EAGAIN &&
4200 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4205 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4206 file->f_flags &= ~O_NONBLOCK;
4207 fsnotify_open(file);
4210 fd_install(ret, file);
4212 ret = io_install_fixed_file(req, file, issue_flags,
4213 req->open.file_slot - 1);
4215 putname(req->open.filename);
4216 req->flags &= ~REQ_F_NEED_CLEANUP;
4219 __io_req_complete(req, issue_flags, ret, 0);
4223 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4225 return io_openat2(req, issue_flags);
4228 static int io_remove_buffers_prep(struct io_kiocb *req,
4229 const struct io_uring_sqe *sqe)
4231 struct io_provide_buf *p = &req->pbuf;
4234 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4238 tmp = READ_ONCE(sqe->fd);
4239 if (!tmp || tmp > USHRT_MAX)
4242 memset(p, 0, sizeof(*p));
4244 p->bgid = READ_ONCE(sqe->buf_group);
4248 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4249 int bgid, unsigned nbufs)
4253 /* shouldn't happen */
4257 /* the head kbuf is the list itself */
4258 while (!list_empty(&buf->list)) {
4259 struct io_buffer *nxt;
4261 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4262 list_del(&nxt->list);
4269 xa_erase(&ctx->io_buffers, bgid);
4274 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4276 struct io_provide_buf *p = &req->pbuf;
4277 struct io_ring_ctx *ctx = req->ctx;
4278 struct io_buffer *head;
4280 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4282 io_ring_submit_lock(ctx, !force_nonblock);
4284 lockdep_assert_held(&ctx->uring_lock);
4287 head = xa_load(&ctx->io_buffers, p->bgid);
4289 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4293 /* complete before unlock, IOPOLL may need the lock */
4294 __io_req_complete(req, issue_flags, ret, 0);
4295 io_ring_submit_unlock(ctx, !force_nonblock);
4299 static int io_provide_buffers_prep(struct io_kiocb *req,
4300 const struct io_uring_sqe *sqe)
4302 unsigned long size, tmp_check;
4303 struct io_provide_buf *p = &req->pbuf;
4306 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4309 tmp = READ_ONCE(sqe->fd);
4310 if (!tmp || tmp > USHRT_MAX)
4313 p->addr = READ_ONCE(sqe->addr);
4314 p->len = READ_ONCE(sqe->len);
4316 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4319 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4322 size = (unsigned long)p->len * p->nbufs;
4323 if (!access_ok(u64_to_user_ptr(p->addr), size))
4326 p->bgid = READ_ONCE(sqe->buf_group);
4327 tmp = READ_ONCE(sqe->off);
4328 if (tmp > USHRT_MAX)
4334 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4336 struct io_buffer *buf;
4337 u64 addr = pbuf->addr;
4338 int i, bid = pbuf->bid;
4340 for (i = 0; i < pbuf->nbufs; i++) {
4341 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4346 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4351 INIT_LIST_HEAD(&buf->list);
4354 list_add_tail(&buf->list, &(*head)->list);
4358 return i ? i : -ENOMEM;
4361 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4363 struct io_provide_buf *p = &req->pbuf;
4364 struct io_ring_ctx *ctx = req->ctx;
4365 struct io_buffer *head, *list;
4367 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4369 io_ring_submit_lock(ctx, !force_nonblock);
4371 lockdep_assert_held(&ctx->uring_lock);
4373 list = head = xa_load(&ctx->io_buffers, p->bgid);
4375 ret = io_add_buffers(p, &head);
4376 if (ret >= 0 && !list) {
4377 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4379 __io_remove_buffers(ctx, head, p->bgid, -1U);
4383 /* complete before unlock, IOPOLL may need the lock */
4384 __io_req_complete(req, issue_flags, ret, 0);
4385 io_ring_submit_unlock(ctx, !force_nonblock);
4389 static int io_epoll_ctl_prep(struct io_kiocb *req,
4390 const struct io_uring_sqe *sqe)
4392 #if defined(CONFIG_EPOLL)
4393 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4395 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4398 req->epoll.epfd = READ_ONCE(sqe->fd);
4399 req->epoll.op = READ_ONCE(sqe->len);
4400 req->epoll.fd = READ_ONCE(sqe->off);
4402 if (ep_op_has_event(req->epoll.op)) {
4403 struct epoll_event __user *ev;
4405 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4406 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4416 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4418 #if defined(CONFIG_EPOLL)
4419 struct io_epoll *ie = &req->epoll;
4421 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4423 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4424 if (force_nonblock && ret == -EAGAIN)
4429 __io_req_complete(req, issue_flags, ret, 0);
4436 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4438 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4439 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4441 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4444 req->madvise.addr = READ_ONCE(sqe->addr);
4445 req->madvise.len = READ_ONCE(sqe->len);
4446 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4453 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4455 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4456 struct io_madvise *ma = &req->madvise;
4459 if (issue_flags & IO_URING_F_NONBLOCK)
4462 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4465 io_req_complete(req, ret);
4472 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4474 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4476 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4479 req->fadvise.offset = READ_ONCE(sqe->off);
4480 req->fadvise.len = READ_ONCE(sqe->len);
4481 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4485 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4487 struct io_fadvise *fa = &req->fadvise;
4490 if (issue_flags & IO_URING_F_NONBLOCK) {
4491 switch (fa->advice) {
4492 case POSIX_FADV_NORMAL:
4493 case POSIX_FADV_RANDOM:
4494 case POSIX_FADV_SEQUENTIAL:
4501 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4504 __io_req_complete(req, issue_flags, ret, 0);
4508 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4510 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4512 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4514 if (req->flags & REQ_F_FIXED_FILE)
4517 req->statx.dfd = READ_ONCE(sqe->fd);
4518 req->statx.mask = READ_ONCE(sqe->len);
4519 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4520 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4521 req->statx.flags = READ_ONCE(sqe->statx_flags);
4526 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4528 struct io_statx *ctx = &req->statx;
4531 if (issue_flags & IO_URING_F_NONBLOCK)
4534 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4539 io_req_complete(req, ret);
4543 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4547 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4548 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4550 if (req->flags & REQ_F_FIXED_FILE)
4553 req->close.fd = READ_ONCE(sqe->fd);
4557 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4559 struct files_struct *files = current->files;
4560 struct io_close *close = &req->close;
4561 struct fdtable *fdt;
4562 struct file *file = NULL;
4565 spin_lock(&files->file_lock);
4566 fdt = files_fdtable(files);
4567 if (close->fd >= fdt->max_fds) {
4568 spin_unlock(&files->file_lock);
4571 file = fdt->fd[close->fd];
4572 if (!file || file->f_op == &io_uring_fops) {
4573 spin_unlock(&files->file_lock);
4578 /* if the file has a flush method, be safe and punt to async */
4579 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4580 spin_unlock(&files->file_lock);
4584 ret = __close_fd_get_file(close->fd, &file);
4585 spin_unlock(&files->file_lock);
4592 /* No ->flush() or already async, safely close from here */
4593 ret = filp_close(file, current->files);
4599 __io_req_complete(req, issue_flags, ret, 0);
4603 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4605 struct io_ring_ctx *ctx = req->ctx;
4607 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4609 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4613 req->sync.off = READ_ONCE(sqe->off);
4614 req->sync.len = READ_ONCE(sqe->len);
4615 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4619 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4623 /* sync_file_range always requires a blocking context */
4624 if (issue_flags & IO_URING_F_NONBLOCK)
4627 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4631 io_req_complete(req, ret);
4635 #if defined(CONFIG_NET)
4636 static int io_setup_async_msg(struct io_kiocb *req,
4637 struct io_async_msghdr *kmsg)
4639 struct io_async_msghdr *async_msg = req->async_data;
4643 if (io_alloc_async_data(req)) {
4644 kfree(kmsg->free_iov);
4647 async_msg = req->async_data;
4648 req->flags |= REQ_F_NEED_CLEANUP;
4649 memcpy(async_msg, kmsg, sizeof(*kmsg));
4650 async_msg->msg.msg_name = &async_msg->addr;
4651 /* if were using fast_iov, set it to the new one */
4652 if (!async_msg->free_iov)
4653 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4658 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4659 struct io_async_msghdr *iomsg)
4661 iomsg->msg.msg_name = &iomsg->addr;
4662 iomsg->free_iov = iomsg->fast_iov;
4663 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4664 req->sr_msg.msg_flags, &iomsg->free_iov);
4667 static int io_sendmsg_prep_async(struct io_kiocb *req)
4671 ret = io_sendmsg_copy_hdr(req, req->async_data);
4673 req->flags |= REQ_F_NEED_CLEANUP;
4677 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4679 struct io_sr_msg *sr = &req->sr_msg;
4681 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4684 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4685 sr->len = READ_ONCE(sqe->len);
4686 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4687 if (sr->msg_flags & MSG_DONTWAIT)
4688 req->flags |= REQ_F_NOWAIT;
4690 #ifdef CONFIG_COMPAT
4691 if (req->ctx->compat)
4692 sr->msg_flags |= MSG_CMSG_COMPAT;
4697 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4699 struct io_async_msghdr iomsg, *kmsg;
4700 struct socket *sock;
4705 sock = sock_from_file(req->file);
4706 if (unlikely(!sock))
4709 kmsg = req->async_data;
4711 ret = io_sendmsg_copy_hdr(req, &iomsg);
4717 flags = req->sr_msg.msg_flags;
4718 if (issue_flags & IO_URING_F_NONBLOCK)
4719 flags |= MSG_DONTWAIT;
4720 if (flags & MSG_WAITALL)
4721 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4723 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4724 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4725 return io_setup_async_msg(req, kmsg);
4726 if (ret == -ERESTARTSYS)
4729 /* fast path, check for non-NULL to avoid function call */
4731 kfree(kmsg->free_iov);
4732 req->flags &= ~REQ_F_NEED_CLEANUP;
4735 __io_req_complete(req, issue_flags, ret, 0);
4739 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4741 struct io_sr_msg *sr = &req->sr_msg;
4744 struct socket *sock;
4749 sock = sock_from_file(req->file);
4750 if (unlikely(!sock))
4753 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4757 msg.msg_name = NULL;
4758 msg.msg_control = NULL;
4759 msg.msg_controllen = 0;
4760 msg.msg_namelen = 0;
4762 flags = req->sr_msg.msg_flags;
4763 if (issue_flags & IO_URING_F_NONBLOCK)
4764 flags |= MSG_DONTWAIT;
4765 if (flags & MSG_WAITALL)
4766 min_ret = iov_iter_count(&msg.msg_iter);
4768 msg.msg_flags = flags;
4769 ret = sock_sendmsg(sock, &msg);
4770 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4772 if (ret == -ERESTARTSYS)
4777 __io_req_complete(req, issue_flags, ret, 0);
4781 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4782 struct io_async_msghdr *iomsg)
4784 struct io_sr_msg *sr = &req->sr_msg;
4785 struct iovec __user *uiov;
4789 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4790 &iomsg->uaddr, &uiov, &iov_len);
4794 if (req->flags & REQ_F_BUFFER_SELECT) {
4797 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4799 sr->len = iomsg->fast_iov[0].iov_len;
4800 iomsg->free_iov = NULL;
4802 iomsg->free_iov = iomsg->fast_iov;
4803 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4804 &iomsg->free_iov, &iomsg->msg.msg_iter,
4813 #ifdef CONFIG_COMPAT
4814 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4815 struct io_async_msghdr *iomsg)
4817 struct io_sr_msg *sr = &req->sr_msg;
4818 struct compat_iovec __user *uiov;
4823 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4828 uiov = compat_ptr(ptr);
4829 if (req->flags & REQ_F_BUFFER_SELECT) {
4830 compat_ssize_t clen;
4834 if (!access_ok(uiov, sizeof(*uiov)))
4836 if (__get_user(clen, &uiov->iov_len))
4841 iomsg->free_iov = NULL;
4843 iomsg->free_iov = iomsg->fast_iov;
4844 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4845 UIO_FASTIOV, &iomsg->free_iov,
4846 &iomsg->msg.msg_iter, true);
4855 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4856 struct io_async_msghdr *iomsg)
4858 iomsg->msg.msg_name = &iomsg->addr;
4860 #ifdef CONFIG_COMPAT
4861 if (req->ctx->compat)
4862 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4865 return __io_recvmsg_copy_hdr(req, iomsg);
4868 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4871 struct io_sr_msg *sr = &req->sr_msg;
4872 struct io_buffer *kbuf;
4874 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4879 req->flags |= REQ_F_BUFFER_SELECTED;
4883 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4885 return io_put_kbuf(req, req->sr_msg.kbuf);
4888 static int io_recvmsg_prep_async(struct io_kiocb *req)
4892 ret = io_recvmsg_copy_hdr(req, req->async_data);
4894 req->flags |= REQ_F_NEED_CLEANUP;
4898 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4900 struct io_sr_msg *sr = &req->sr_msg;
4902 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4905 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4906 sr->len = READ_ONCE(sqe->len);
4907 sr->bgid = READ_ONCE(sqe->buf_group);
4908 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4909 if (sr->msg_flags & MSG_DONTWAIT)
4910 req->flags |= REQ_F_NOWAIT;
4912 #ifdef CONFIG_COMPAT
4913 if (req->ctx->compat)
4914 sr->msg_flags |= MSG_CMSG_COMPAT;
4919 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4921 struct io_async_msghdr iomsg, *kmsg;
4922 struct socket *sock;
4923 struct io_buffer *kbuf;
4926 int ret, cflags = 0;
4927 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4929 sock = sock_from_file(req->file);
4930 if (unlikely(!sock))
4933 kmsg = req->async_data;
4935 ret = io_recvmsg_copy_hdr(req, &iomsg);
4941 if (req->flags & REQ_F_BUFFER_SELECT) {
4942 kbuf = io_recv_buffer_select(req, !force_nonblock);
4944 return PTR_ERR(kbuf);
4945 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4946 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4947 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4948 1, req->sr_msg.len);
4951 flags = req->sr_msg.msg_flags;
4953 flags |= MSG_DONTWAIT;
4954 if (flags & MSG_WAITALL)
4955 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4957 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4958 kmsg->uaddr, flags);
4959 if (force_nonblock && ret == -EAGAIN)
4960 return io_setup_async_msg(req, kmsg);
4961 if (ret == -ERESTARTSYS)
4964 if (req->flags & REQ_F_BUFFER_SELECTED)
4965 cflags = io_put_recv_kbuf(req);
4966 /* fast path, check for non-NULL to avoid function call */
4968 kfree(kmsg->free_iov);
4969 req->flags &= ~REQ_F_NEED_CLEANUP;
4970 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4972 __io_req_complete(req, issue_flags, ret, cflags);
4976 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4978 struct io_buffer *kbuf;
4979 struct io_sr_msg *sr = &req->sr_msg;
4981 void __user *buf = sr->buf;
4982 struct socket *sock;
4986 int ret, cflags = 0;
4987 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4989 sock = sock_from_file(req->file);
4990 if (unlikely(!sock))
4993 if (req->flags & REQ_F_BUFFER_SELECT) {
4994 kbuf = io_recv_buffer_select(req, !force_nonblock);
4996 return PTR_ERR(kbuf);
4997 buf = u64_to_user_ptr(kbuf->addr);
5000 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5004 msg.msg_name = NULL;
5005 msg.msg_control = NULL;
5006 msg.msg_controllen = 0;
5007 msg.msg_namelen = 0;
5008 msg.msg_iocb = NULL;
5011 flags = req->sr_msg.msg_flags;
5013 flags |= MSG_DONTWAIT;
5014 if (flags & MSG_WAITALL)
5015 min_ret = iov_iter_count(&msg.msg_iter);
5017 ret = sock_recvmsg(sock, &msg, flags);
5018 if (force_nonblock && ret == -EAGAIN)
5020 if (ret == -ERESTARTSYS)
5023 if (req->flags & REQ_F_BUFFER_SELECTED)
5024 cflags = io_put_recv_kbuf(req);
5025 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5027 __io_req_complete(req, issue_flags, ret, cflags);
5031 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5033 struct io_accept *accept = &req->accept;
5035 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5037 if (sqe->ioprio || sqe->len || sqe->buf_index)
5040 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5041 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5042 accept->flags = READ_ONCE(sqe->accept_flags);
5043 accept->nofile = rlimit(RLIMIT_NOFILE);
5045 accept->file_slot = READ_ONCE(sqe->file_index);
5046 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5047 (accept->flags & SOCK_CLOEXEC)))
5049 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5051 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5052 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5056 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5058 struct io_accept *accept = &req->accept;
5059 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5060 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5061 bool fixed = !!accept->file_slot;
5065 if (req->file->f_flags & O_NONBLOCK)
5066 req->flags |= REQ_F_NOWAIT;
5069 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5070 if (unlikely(fd < 0))
5073 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5078 ret = PTR_ERR(file);
5079 if (ret == -EAGAIN && force_nonblock)
5081 if (ret == -ERESTARTSYS)
5084 } else if (!fixed) {
5085 fd_install(fd, file);
5088 ret = io_install_fixed_file(req, file, issue_flags,
5089 accept->file_slot - 1);
5091 __io_req_complete(req, issue_flags, ret, 0);
5095 static int io_connect_prep_async(struct io_kiocb *req)
5097 struct io_async_connect *io = req->async_data;
5098 struct io_connect *conn = &req->connect;
5100 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5103 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5105 struct io_connect *conn = &req->connect;
5107 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5109 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5113 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5114 conn->addr_len = READ_ONCE(sqe->addr2);
5118 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5120 struct io_async_connect __io, *io;
5121 unsigned file_flags;
5123 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5125 if (req->async_data) {
5126 io = req->async_data;
5128 ret = move_addr_to_kernel(req->connect.addr,
5129 req->connect.addr_len,
5136 file_flags = force_nonblock ? O_NONBLOCK : 0;
5138 ret = __sys_connect_file(req->file, &io->address,
5139 req->connect.addr_len, file_flags);
5140 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5141 if (req->async_data)
5143 if (io_alloc_async_data(req)) {
5147 memcpy(req->async_data, &__io, sizeof(__io));
5150 if (ret == -ERESTARTSYS)
5155 __io_req_complete(req, issue_flags, ret, 0);
5158 #else /* !CONFIG_NET */
5159 #define IO_NETOP_FN(op) \
5160 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5162 return -EOPNOTSUPP; \
5165 #define IO_NETOP_PREP(op) \
5167 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5169 return -EOPNOTSUPP; \
5172 #define IO_NETOP_PREP_ASYNC(op) \
5174 static int io_##op##_prep_async(struct io_kiocb *req) \
5176 return -EOPNOTSUPP; \
5179 IO_NETOP_PREP_ASYNC(sendmsg);
5180 IO_NETOP_PREP_ASYNC(recvmsg);
5181 IO_NETOP_PREP_ASYNC(connect);
5182 IO_NETOP_PREP(accept);
5185 #endif /* CONFIG_NET */
5187 struct io_poll_table {
5188 struct poll_table_struct pt;
5189 struct io_kiocb *req;
5194 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5195 __poll_t mask, io_req_tw_func_t func)
5197 /* for instances that support it check for an event match first: */
5198 if (mask && !(mask & poll->events))
5201 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5203 list_del_init(&poll->wait.entry);
5206 req->io_task_work.func = func;
5209 * If this fails, then the task is exiting. When a task exits, the
5210 * work gets canceled, so just cancel this request as well instead
5211 * of executing it. We can't safely execute it anyway, as we may not
5212 * have the needed state needed for it anyway.
5214 io_req_task_work_add(req);
5218 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5219 __acquires(&req->ctx->completion_lock)
5221 struct io_ring_ctx *ctx = req->ctx;
5223 /* req->task == current here, checking PF_EXITING is safe */
5224 if (unlikely(req->task->flags & PF_EXITING))
5225 WRITE_ONCE(poll->canceled, true);
5227 if (!req->result && !READ_ONCE(poll->canceled)) {
5228 struct poll_table_struct pt = { ._key = poll->events };
5230 req->result = vfs_poll(req->file, &pt) & poll->events;
5233 spin_lock(&ctx->completion_lock);
5234 if (!req->result && !READ_ONCE(poll->canceled)) {
5235 add_wait_queue(poll->head, &poll->wait);
5242 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5244 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5245 if (req->opcode == IORING_OP_POLL_ADD)
5246 return req->async_data;
5247 return req->apoll->double_poll;
5250 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5252 if (req->opcode == IORING_OP_POLL_ADD)
5254 return &req->apoll->poll;
5257 static void io_poll_remove_double(struct io_kiocb *req)
5258 __must_hold(&req->ctx->completion_lock)
5260 struct io_poll_iocb *poll = io_poll_get_double(req);
5262 lockdep_assert_held(&req->ctx->completion_lock);
5264 if (poll && poll->head) {
5265 struct wait_queue_head *head = poll->head;
5267 spin_lock_irq(&head->lock);
5268 list_del_init(&poll->wait.entry);
5269 if (poll->wait.private)
5272 spin_unlock_irq(&head->lock);
5276 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5277 __must_hold(&req->ctx->completion_lock)
5279 struct io_ring_ctx *ctx = req->ctx;
5280 unsigned flags = IORING_CQE_F_MORE;
5283 if (READ_ONCE(req->poll.canceled)) {
5285 req->poll.events |= EPOLLONESHOT;
5287 error = mangle_poll(mask);
5289 if (req->poll.events & EPOLLONESHOT)
5291 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5292 req->poll.done = true;
5295 if (flags & IORING_CQE_F_MORE)
5298 return !(flags & IORING_CQE_F_MORE);
5301 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5302 __must_hold(&req->ctx->completion_lock)
5306 done = __io_poll_complete(req, mask);
5307 io_commit_cqring(req->ctx);
5311 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5313 struct io_ring_ctx *ctx = req->ctx;
5314 struct io_kiocb *nxt;
5316 if (io_poll_rewait(req, &req->poll)) {
5317 spin_unlock(&ctx->completion_lock);
5321 done = __io_poll_complete(req, req->result);
5323 io_poll_remove_double(req);
5324 hash_del(&req->hash_node);
5327 add_wait_queue(req->poll.head, &req->poll.wait);
5329 io_commit_cqring(ctx);
5330 spin_unlock(&ctx->completion_lock);
5331 io_cqring_ev_posted(ctx);
5334 nxt = io_put_req_find_next(req);
5336 io_req_task_submit(nxt, locked);
5341 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5342 int sync, void *key)
5344 struct io_kiocb *req = wait->private;
5345 struct io_poll_iocb *poll = io_poll_get_single(req);
5346 __poll_t mask = key_to_poll(key);
5347 unsigned long flags;
5349 /* for instances that support it check for an event match first: */
5350 if (mask && !(mask & poll->events))
5352 if (!(poll->events & EPOLLONESHOT))
5353 return poll->wait.func(&poll->wait, mode, sync, key);
5355 list_del_init(&wait->entry);
5360 spin_lock_irqsave(&poll->head->lock, flags);
5361 done = list_empty(&poll->wait.entry);
5363 list_del_init(&poll->wait.entry);
5364 /* make sure double remove sees this as being gone */
5365 wait->private = NULL;
5366 spin_unlock_irqrestore(&poll->head->lock, flags);
5368 /* use wait func handler, so it matches the rq type */
5369 poll->wait.func(&poll->wait, mode, sync, key);
5376 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5377 wait_queue_func_t wake_func)
5381 poll->canceled = false;
5382 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5383 /* mask in events that we always want/need */
5384 poll->events = events | IO_POLL_UNMASK;
5385 INIT_LIST_HEAD(&poll->wait.entry);
5386 init_waitqueue_func_entry(&poll->wait, wake_func);
5389 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5390 struct wait_queue_head *head,
5391 struct io_poll_iocb **poll_ptr)
5393 struct io_kiocb *req = pt->req;
5396 * The file being polled uses multiple waitqueues for poll handling
5397 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5400 if (unlikely(pt->nr_entries)) {
5401 struct io_poll_iocb *poll_one = poll;
5403 /* double add on the same waitqueue head, ignore */
5404 if (poll_one->head == head)
5406 /* already have a 2nd entry, fail a third attempt */
5408 if ((*poll_ptr)->head == head)
5410 pt->error = -EINVAL;
5414 * Can't handle multishot for double wait for now, turn it
5415 * into one-shot mode.
5417 if (!(poll_one->events & EPOLLONESHOT))
5418 poll_one->events |= EPOLLONESHOT;
5419 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5421 pt->error = -ENOMEM;
5424 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5426 poll->wait.private = req;
5433 if (poll->events & EPOLLEXCLUSIVE)
5434 add_wait_queue_exclusive(head, &poll->wait);
5436 add_wait_queue(head, &poll->wait);
5439 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5440 struct poll_table_struct *p)
5442 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5443 struct async_poll *apoll = pt->req->apoll;
5445 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5448 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5450 struct async_poll *apoll = req->apoll;
5451 struct io_ring_ctx *ctx = req->ctx;
5453 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5455 if (io_poll_rewait(req, &apoll->poll)) {
5456 spin_unlock(&ctx->completion_lock);
5460 hash_del(&req->hash_node);
5461 io_poll_remove_double(req);
5462 spin_unlock(&ctx->completion_lock);
5464 if (!READ_ONCE(apoll->poll.canceled))
5465 io_req_task_submit(req, locked);
5467 io_req_complete_failed(req, -ECANCELED);
5470 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5473 struct io_kiocb *req = wait->private;
5474 struct io_poll_iocb *poll = &req->apoll->poll;
5476 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5479 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5482 static void io_poll_req_insert(struct io_kiocb *req)
5484 struct io_ring_ctx *ctx = req->ctx;
5485 struct hlist_head *list;
5487 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5488 hlist_add_head(&req->hash_node, list);
5491 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5492 struct io_poll_iocb *poll,
5493 struct io_poll_table *ipt, __poll_t mask,
5494 wait_queue_func_t wake_func)
5495 __acquires(&ctx->completion_lock)
5497 struct io_ring_ctx *ctx = req->ctx;
5498 bool cancel = false;
5500 INIT_HLIST_NODE(&req->hash_node);
5501 io_init_poll_iocb(poll, mask, wake_func);
5502 poll->file = req->file;
5503 poll->wait.private = req;
5505 ipt->pt._key = mask;
5508 ipt->nr_entries = 0;
5510 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5511 if (unlikely(!ipt->nr_entries) && !ipt->error)
5512 ipt->error = -EINVAL;
5514 spin_lock(&ctx->completion_lock);
5515 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5516 io_poll_remove_double(req);
5517 if (likely(poll->head)) {
5518 spin_lock_irq(&poll->head->lock);
5519 if (unlikely(list_empty(&poll->wait.entry))) {
5525 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5526 list_del_init(&poll->wait.entry);
5528 WRITE_ONCE(poll->canceled, true);
5529 else if (!poll->done) /* actually waiting for an event */
5530 io_poll_req_insert(req);
5531 spin_unlock_irq(&poll->head->lock);
5543 static int io_arm_poll_handler(struct io_kiocb *req)
5545 const struct io_op_def *def = &io_op_defs[req->opcode];
5546 struct io_ring_ctx *ctx = req->ctx;
5547 struct async_poll *apoll;
5548 struct io_poll_table ipt;
5549 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5552 if (!req->file || !file_can_poll(req->file))
5553 return IO_APOLL_ABORTED;
5554 if (req->flags & REQ_F_POLLED)
5555 return IO_APOLL_ABORTED;
5556 if (!def->pollin && !def->pollout)
5557 return IO_APOLL_ABORTED;
5561 mask |= POLLIN | POLLRDNORM;
5563 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5564 if ((req->opcode == IORING_OP_RECVMSG) &&
5565 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5569 mask |= POLLOUT | POLLWRNORM;
5572 /* if we can't nonblock try, then no point in arming a poll handler */
5573 if (!io_file_supports_nowait(req, rw))
5574 return IO_APOLL_ABORTED;
5576 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5577 if (unlikely(!apoll))
5578 return IO_APOLL_ABORTED;
5579 apoll->double_poll = NULL;
5581 req->flags |= REQ_F_POLLED;
5582 ipt.pt._qproc = io_async_queue_proc;
5583 io_req_set_refcount(req);
5585 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5587 spin_unlock(&ctx->completion_lock);
5588 if (ret || ipt.error)
5589 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5591 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5592 mask, apoll->poll.events);
5596 static bool __io_poll_remove_one(struct io_kiocb *req,
5597 struct io_poll_iocb *poll, bool do_cancel)
5598 __must_hold(&req->ctx->completion_lock)
5600 bool do_complete = false;
5604 spin_lock_irq(&poll->head->lock);
5606 WRITE_ONCE(poll->canceled, true);
5607 if (!list_empty(&poll->wait.entry)) {
5608 list_del_init(&poll->wait.entry);
5611 spin_unlock_irq(&poll->head->lock);
5612 hash_del(&req->hash_node);
5616 static bool io_poll_remove_one(struct io_kiocb *req)
5617 __must_hold(&req->ctx->completion_lock)
5621 io_poll_remove_double(req);
5622 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5625 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5626 io_commit_cqring(req->ctx);
5628 io_put_req_deferred(req);
5634 * Returns true if we found and killed one or more poll requests
5636 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5639 struct hlist_node *tmp;
5640 struct io_kiocb *req;
5643 spin_lock(&ctx->completion_lock);
5644 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5645 struct hlist_head *list;
5647 list = &ctx->cancel_hash[i];
5648 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5649 if (io_match_task(req, tsk, cancel_all))
5650 posted += io_poll_remove_one(req);
5653 spin_unlock(&ctx->completion_lock);
5656 io_cqring_ev_posted(ctx);
5661 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5663 __must_hold(&ctx->completion_lock)
5665 struct hlist_head *list;
5666 struct io_kiocb *req;
5668 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5669 hlist_for_each_entry(req, list, hash_node) {
5670 if (sqe_addr != req->user_data)
5672 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5679 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5681 __must_hold(&ctx->completion_lock)
5683 struct io_kiocb *req;
5685 req = io_poll_find(ctx, sqe_addr, poll_only);
5688 if (io_poll_remove_one(req))
5694 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5699 events = READ_ONCE(sqe->poll32_events);
5701 events = swahw32(events);
5703 if (!(flags & IORING_POLL_ADD_MULTI))
5704 events |= EPOLLONESHOT;
5705 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5708 static int io_poll_update_prep(struct io_kiocb *req,
5709 const struct io_uring_sqe *sqe)
5711 struct io_poll_update *upd = &req->poll_update;
5714 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5716 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5718 flags = READ_ONCE(sqe->len);
5719 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5720 IORING_POLL_ADD_MULTI))
5722 /* meaningless without update */
5723 if (flags == IORING_POLL_ADD_MULTI)
5726 upd->old_user_data = READ_ONCE(sqe->addr);
5727 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5728 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5730 upd->new_user_data = READ_ONCE(sqe->off);
5731 if (!upd->update_user_data && upd->new_user_data)
5733 if (upd->update_events)
5734 upd->events = io_poll_parse_events(sqe, flags);
5735 else if (sqe->poll32_events)
5741 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5744 struct io_kiocb *req = wait->private;
5745 struct io_poll_iocb *poll = &req->poll;
5747 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5750 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5751 struct poll_table_struct *p)
5753 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5755 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5758 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5760 struct io_poll_iocb *poll = &req->poll;
5763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5765 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5767 flags = READ_ONCE(sqe->len);
5768 if (flags & ~IORING_POLL_ADD_MULTI)
5771 io_req_set_refcount(req);
5772 poll->events = io_poll_parse_events(sqe, flags);
5776 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5778 struct io_poll_iocb *poll = &req->poll;
5779 struct io_ring_ctx *ctx = req->ctx;
5780 struct io_poll_table ipt;
5783 ipt.pt._qproc = io_poll_queue_proc;
5785 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5788 if (mask) { /* no async, we'd stolen it */
5790 io_poll_complete(req, mask);
5792 spin_unlock(&ctx->completion_lock);
5795 io_cqring_ev_posted(ctx);
5796 if (poll->events & EPOLLONESHOT)
5802 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5804 struct io_ring_ctx *ctx = req->ctx;
5805 struct io_kiocb *preq;
5809 spin_lock(&ctx->completion_lock);
5810 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5816 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5818 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5823 * Don't allow racy completion with singleshot, as we cannot safely
5824 * update those. For multishot, if we're racing with completion, just
5825 * let completion re-add it.
5827 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5828 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5832 /* we now have a detached poll request. reissue. */
5836 spin_unlock(&ctx->completion_lock);
5838 io_req_complete(req, ret);
5841 /* only mask one event flags, keep behavior flags */
5842 if (req->poll_update.update_events) {
5843 preq->poll.events &= ~0xffff;
5844 preq->poll.events |= req->poll_update.events & 0xffff;
5845 preq->poll.events |= IO_POLL_UNMASK;
5847 if (req->poll_update.update_user_data)
5848 preq->user_data = req->poll_update.new_user_data;
5849 spin_unlock(&ctx->completion_lock);
5851 /* complete update request, we're done with it */
5852 io_req_complete(req, ret);
5855 ret = io_poll_add(preq, issue_flags);
5858 io_req_complete(preq, ret);
5864 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5867 io_req_complete_post(req, -ETIME, 0);
5870 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5872 struct io_timeout_data *data = container_of(timer,
5873 struct io_timeout_data, timer);
5874 struct io_kiocb *req = data->req;
5875 struct io_ring_ctx *ctx = req->ctx;
5876 unsigned long flags;
5878 spin_lock_irqsave(&ctx->timeout_lock, flags);
5879 list_del_init(&req->timeout.list);
5880 atomic_set(&req->ctx->cq_timeouts,
5881 atomic_read(&req->ctx->cq_timeouts) + 1);
5882 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5884 req->io_task_work.func = io_req_task_timeout;
5885 io_req_task_work_add(req);
5886 return HRTIMER_NORESTART;
5889 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5891 __must_hold(&ctx->timeout_lock)
5893 struct io_timeout_data *io;
5894 struct io_kiocb *req;
5897 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5898 found = user_data == req->user_data;
5903 return ERR_PTR(-ENOENT);
5905 io = req->async_data;
5906 if (hrtimer_try_to_cancel(&io->timer) == -1)
5907 return ERR_PTR(-EALREADY);
5908 list_del_init(&req->timeout.list);
5912 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5913 __must_hold(&ctx->completion_lock)
5914 __must_hold(&ctx->timeout_lock)
5916 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5919 return PTR_ERR(req);
5922 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5923 io_put_req_deferred(req);
5927 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5929 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5930 case IORING_TIMEOUT_BOOTTIME:
5931 return CLOCK_BOOTTIME;
5932 case IORING_TIMEOUT_REALTIME:
5933 return CLOCK_REALTIME;
5935 /* can't happen, vetted at prep time */
5939 return CLOCK_MONOTONIC;
5943 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5944 struct timespec64 *ts, enum hrtimer_mode mode)
5945 __must_hold(&ctx->timeout_lock)
5947 struct io_timeout_data *io;
5948 struct io_kiocb *req;
5951 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
5952 found = user_data == req->user_data;
5959 io = req->async_data;
5960 if (hrtimer_try_to_cancel(&io->timer) == -1)
5962 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
5963 io->timer.function = io_link_timeout_fn;
5964 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
5968 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5969 struct timespec64 *ts, enum hrtimer_mode mode)
5970 __must_hold(&ctx->timeout_lock)
5972 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5973 struct io_timeout_data *data;
5976 return PTR_ERR(req);
5978 req->timeout.off = 0; /* noseq */
5979 data = req->async_data;
5980 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5981 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
5982 data->timer.function = io_timeout_fn;
5983 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5987 static int io_timeout_remove_prep(struct io_kiocb *req,
5988 const struct io_uring_sqe *sqe)
5990 struct io_timeout_rem *tr = &req->timeout_rem;
5992 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5994 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5996 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
5999 tr->ltimeout = false;
6000 tr->addr = READ_ONCE(sqe->addr);
6001 tr->flags = READ_ONCE(sqe->timeout_flags);
6002 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6003 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6005 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6006 tr->ltimeout = true;
6007 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6009 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6011 } else if (tr->flags) {
6012 /* timeout removal doesn't support flags */
6019 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6021 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6026 * Remove or update an existing timeout command
6028 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6030 struct io_timeout_rem *tr = &req->timeout_rem;
6031 struct io_ring_ctx *ctx = req->ctx;
6034 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6035 spin_lock(&ctx->completion_lock);
6036 spin_lock_irq(&ctx->timeout_lock);
6037 ret = io_timeout_cancel(ctx, tr->addr);
6038 spin_unlock_irq(&ctx->timeout_lock);
6039 spin_unlock(&ctx->completion_lock);
6041 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6043 spin_lock_irq(&ctx->timeout_lock);
6045 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6047 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6048 spin_unlock_irq(&ctx->timeout_lock);
6053 io_req_complete_post(req, ret, 0);
6057 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6058 bool is_timeout_link)
6060 struct io_timeout_data *data;
6062 u32 off = READ_ONCE(sqe->off);
6064 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6066 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6069 if (off && is_timeout_link)
6071 flags = READ_ONCE(sqe->timeout_flags);
6072 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6074 /* more than one clock specified is invalid, obviously */
6075 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6078 INIT_LIST_HEAD(&req->timeout.list);
6079 req->timeout.off = off;
6080 if (unlikely(off && !req->ctx->off_timeout_used))
6081 req->ctx->off_timeout_used = true;
6083 if (!req->async_data && io_alloc_async_data(req))
6086 data = req->async_data;
6088 data->flags = flags;
6090 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6093 data->mode = io_translate_timeout_mode(flags);
6094 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6096 if (is_timeout_link) {
6097 struct io_submit_link *link = &req->ctx->submit_state.link;
6101 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6103 req->timeout.head = link->last;
6104 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6109 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6111 struct io_ring_ctx *ctx = req->ctx;
6112 struct io_timeout_data *data = req->async_data;
6113 struct list_head *entry;
6114 u32 tail, off = req->timeout.off;
6116 spin_lock_irq(&ctx->timeout_lock);
6119 * sqe->off holds how many events that need to occur for this
6120 * timeout event to be satisfied. If it isn't set, then this is
6121 * a pure timeout request, sequence isn't used.
6123 if (io_is_timeout_noseq(req)) {
6124 entry = ctx->timeout_list.prev;
6128 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6129 req->timeout.target_seq = tail + off;
6131 /* Update the last seq here in case io_flush_timeouts() hasn't.
6132 * This is safe because ->completion_lock is held, and submissions
6133 * and completions are never mixed in the same ->completion_lock section.
6135 ctx->cq_last_tm_flush = tail;
6138 * Insertion sort, ensuring the first entry in the list is always
6139 * the one we need first.
6141 list_for_each_prev(entry, &ctx->timeout_list) {
6142 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6145 if (io_is_timeout_noseq(nxt))
6147 /* nxt.seq is behind @tail, otherwise would've been completed */
6148 if (off >= nxt->timeout.target_seq - tail)
6152 list_add(&req->timeout.list, entry);
6153 data->timer.function = io_timeout_fn;
6154 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6155 spin_unlock_irq(&ctx->timeout_lock);
6159 struct io_cancel_data {
6160 struct io_ring_ctx *ctx;
6164 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6166 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6167 struct io_cancel_data *cd = data;
6169 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6172 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6173 struct io_ring_ctx *ctx)
6175 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6176 enum io_wq_cancel cancel_ret;
6179 if (!tctx || !tctx->io_wq)
6182 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6183 switch (cancel_ret) {
6184 case IO_WQ_CANCEL_OK:
6187 case IO_WQ_CANCEL_RUNNING:
6190 case IO_WQ_CANCEL_NOTFOUND:
6198 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6200 struct io_ring_ctx *ctx = req->ctx;
6203 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6205 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6209 spin_lock(&ctx->completion_lock);
6210 spin_lock_irq(&ctx->timeout_lock);
6211 ret = io_timeout_cancel(ctx, sqe_addr);
6212 spin_unlock_irq(&ctx->timeout_lock);
6215 ret = io_poll_cancel(ctx, sqe_addr, false);
6217 spin_unlock(&ctx->completion_lock);
6221 static int io_async_cancel_prep(struct io_kiocb *req,
6222 const struct io_uring_sqe *sqe)
6224 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6226 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6228 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6232 req->cancel.addr = READ_ONCE(sqe->addr);
6236 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6238 struct io_ring_ctx *ctx = req->ctx;
6239 u64 sqe_addr = req->cancel.addr;
6240 struct io_tctx_node *node;
6243 ret = io_try_cancel_userdata(req, sqe_addr);
6247 /* slow path, try all io-wq's */
6248 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6250 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6251 struct io_uring_task *tctx = node->task->io_uring;
6253 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6257 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6261 io_req_complete_post(req, ret, 0);
6265 static int io_rsrc_update_prep(struct io_kiocb *req,
6266 const struct io_uring_sqe *sqe)
6268 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6270 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6273 req->rsrc_update.offset = READ_ONCE(sqe->off);
6274 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6275 if (!req->rsrc_update.nr_args)
6277 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6281 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6283 struct io_ring_ctx *ctx = req->ctx;
6284 struct io_uring_rsrc_update2 up;
6287 if (issue_flags & IO_URING_F_NONBLOCK)
6290 up.offset = req->rsrc_update.offset;
6291 up.data = req->rsrc_update.arg;
6296 mutex_lock(&ctx->uring_lock);
6297 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6298 &up, req->rsrc_update.nr_args);
6299 mutex_unlock(&ctx->uring_lock);
6303 __io_req_complete(req, issue_flags, ret, 0);
6307 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6309 switch (req->opcode) {
6312 case IORING_OP_READV:
6313 case IORING_OP_READ_FIXED:
6314 case IORING_OP_READ:
6315 return io_read_prep(req, sqe);
6316 case IORING_OP_WRITEV:
6317 case IORING_OP_WRITE_FIXED:
6318 case IORING_OP_WRITE:
6319 return io_write_prep(req, sqe);
6320 case IORING_OP_POLL_ADD:
6321 return io_poll_add_prep(req, sqe);
6322 case IORING_OP_POLL_REMOVE:
6323 return io_poll_update_prep(req, sqe);
6324 case IORING_OP_FSYNC:
6325 return io_fsync_prep(req, sqe);
6326 case IORING_OP_SYNC_FILE_RANGE:
6327 return io_sfr_prep(req, sqe);
6328 case IORING_OP_SENDMSG:
6329 case IORING_OP_SEND:
6330 return io_sendmsg_prep(req, sqe);
6331 case IORING_OP_RECVMSG:
6332 case IORING_OP_RECV:
6333 return io_recvmsg_prep(req, sqe);
6334 case IORING_OP_CONNECT:
6335 return io_connect_prep(req, sqe);
6336 case IORING_OP_TIMEOUT:
6337 return io_timeout_prep(req, sqe, false);
6338 case IORING_OP_TIMEOUT_REMOVE:
6339 return io_timeout_remove_prep(req, sqe);
6340 case IORING_OP_ASYNC_CANCEL:
6341 return io_async_cancel_prep(req, sqe);
6342 case IORING_OP_LINK_TIMEOUT:
6343 return io_timeout_prep(req, sqe, true);
6344 case IORING_OP_ACCEPT:
6345 return io_accept_prep(req, sqe);
6346 case IORING_OP_FALLOCATE:
6347 return io_fallocate_prep(req, sqe);
6348 case IORING_OP_OPENAT:
6349 return io_openat_prep(req, sqe);
6350 case IORING_OP_CLOSE:
6351 return io_close_prep(req, sqe);
6352 case IORING_OP_FILES_UPDATE:
6353 return io_rsrc_update_prep(req, sqe);
6354 case IORING_OP_STATX:
6355 return io_statx_prep(req, sqe);
6356 case IORING_OP_FADVISE:
6357 return io_fadvise_prep(req, sqe);
6358 case IORING_OP_MADVISE:
6359 return io_madvise_prep(req, sqe);
6360 case IORING_OP_OPENAT2:
6361 return io_openat2_prep(req, sqe);
6362 case IORING_OP_EPOLL_CTL:
6363 return io_epoll_ctl_prep(req, sqe);
6364 case IORING_OP_SPLICE:
6365 return io_splice_prep(req, sqe);
6366 case IORING_OP_PROVIDE_BUFFERS:
6367 return io_provide_buffers_prep(req, sqe);
6368 case IORING_OP_REMOVE_BUFFERS:
6369 return io_remove_buffers_prep(req, sqe);
6371 return io_tee_prep(req, sqe);
6372 case IORING_OP_SHUTDOWN:
6373 return io_shutdown_prep(req, sqe);
6374 case IORING_OP_RENAMEAT:
6375 return io_renameat_prep(req, sqe);
6376 case IORING_OP_UNLINKAT:
6377 return io_unlinkat_prep(req, sqe);
6378 case IORING_OP_MKDIRAT:
6379 return io_mkdirat_prep(req, sqe);
6380 case IORING_OP_SYMLINKAT:
6381 return io_symlinkat_prep(req, sqe);
6382 case IORING_OP_LINKAT:
6383 return io_linkat_prep(req, sqe);
6386 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6391 static int io_req_prep_async(struct io_kiocb *req)
6393 if (!io_op_defs[req->opcode].needs_async_setup)
6395 if (WARN_ON_ONCE(req->async_data))
6397 if (io_alloc_async_data(req))
6400 switch (req->opcode) {
6401 case IORING_OP_READV:
6402 return io_rw_prep_async(req, READ);
6403 case IORING_OP_WRITEV:
6404 return io_rw_prep_async(req, WRITE);
6405 case IORING_OP_SENDMSG:
6406 return io_sendmsg_prep_async(req);
6407 case IORING_OP_RECVMSG:
6408 return io_recvmsg_prep_async(req);
6409 case IORING_OP_CONNECT:
6410 return io_connect_prep_async(req);
6412 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6417 static u32 io_get_sequence(struct io_kiocb *req)
6419 u32 seq = req->ctx->cached_sq_head;
6421 /* need original cached_sq_head, but it was increased for each req */
6422 io_for_each_link(req, req)
6427 static bool io_drain_req(struct io_kiocb *req)
6429 struct io_kiocb *pos;
6430 struct io_ring_ctx *ctx = req->ctx;
6431 struct io_defer_entry *de;
6435 if (req->flags & REQ_F_FAIL) {
6436 io_req_complete_fail_submit(req);
6441 * If we need to drain a request in the middle of a link, drain the
6442 * head request and the next request/link after the current link.
6443 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6444 * maintained for every request of our link.
6446 if (ctx->drain_next) {
6447 req->flags |= REQ_F_IO_DRAIN;
6448 ctx->drain_next = false;
6450 /* not interested in head, start from the first linked */
6451 io_for_each_link(pos, req->link) {
6452 if (pos->flags & REQ_F_IO_DRAIN) {
6453 ctx->drain_next = true;
6454 req->flags |= REQ_F_IO_DRAIN;
6459 /* Still need defer if there is pending req in defer list. */
6460 if (likely(list_empty_careful(&ctx->defer_list) &&
6461 !(req->flags & REQ_F_IO_DRAIN))) {
6462 ctx->drain_active = false;
6466 seq = io_get_sequence(req);
6467 /* Still a chance to pass the sequence check */
6468 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6471 ret = io_req_prep_async(req);
6474 io_prep_async_link(req);
6475 de = kmalloc(sizeof(*de), GFP_KERNEL);
6479 io_req_complete_failed(req, ret);
6483 spin_lock(&ctx->completion_lock);
6484 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6485 spin_unlock(&ctx->completion_lock);
6487 io_queue_async_work(req, NULL);
6491 trace_io_uring_defer(ctx, req, req->user_data);
6494 list_add_tail(&de->list, &ctx->defer_list);
6495 spin_unlock(&ctx->completion_lock);
6499 static void io_clean_op(struct io_kiocb *req)
6501 if (req->flags & REQ_F_BUFFER_SELECTED) {
6502 switch (req->opcode) {
6503 case IORING_OP_READV:
6504 case IORING_OP_READ_FIXED:
6505 case IORING_OP_READ:
6506 kfree((void *)(unsigned long)req->rw.addr);
6508 case IORING_OP_RECVMSG:
6509 case IORING_OP_RECV:
6510 kfree(req->sr_msg.kbuf);
6515 if (req->flags & REQ_F_NEED_CLEANUP) {
6516 switch (req->opcode) {
6517 case IORING_OP_READV:
6518 case IORING_OP_READ_FIXED:
6519 case IORING_OP_READ:
6520 case IORING_OP_WRITEV:
6521 case IORING_OP_WRITE_FIXED:
6522 case IORING_OP_WRITE: {
6523 struct io_async_rw *io = req->async_data;
6525 kfree(io->free_iovec);
6528 case IORING_OP_RECVMSG:
6529 case IORING_OP_SENDMSG: {
6530 struct io_async_msghdr *io = req->async_data;
6532 kfree(io->free_iov);
6535 case IORING_OP_SPLICE:
6537 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6538 io_put_file(req->splice.file_in);
6540 case IORING_OP_OPENAT:
6541 case IORING_OP_OPENAT2:
6542 if (req->open.filename)
6543 putname(req->open.filename);
6545 case IORING_OP_RENAMEAT:
6546 putname(req->rename.oldpath);
6547 putname(req->rename.newpath);
6549 case IORING_OP_UNLINKAT:
6550 putname(req->unlink.filename);
6552 case IORING_OP_MKDIRAT:
6553 putname(req->mkdir.filename);
6555 case IORING_OP_SYMLINKAT:
6556 putname(req->symlink.oldpath);
6557 putname(req->symlink.newpath);
6559 case IORING_OP_LINKAT:
6560 putname(req->hardlink.oldpath);
6561 putname(req->hardlink.newpath);
6565 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6566 kfree(req->apoll->double_poll);
6570 if (req->flags & REQ_F_INFLIGHT) {
6571 struct io_uring_task *tctx = req->task->io_uring;
6573 atomic_dec(&tctx->inflight_tracked);
6575 if (req->flags & REQ_F_CREDS)
6576 put_cred(req->creds);
6578 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6581 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6583 struct io_ring_ctx *ctx = req->ctx;
6584 const struct cred *creds = NULL;
6587 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6588 creds = override_creds(req->creds);
6590 switch (req->opcode) {
6592 ret = io_nop(req, issue_flags);
6594 case IORING_OP_READV:
6595 case IORING_OP_READ_FIXED:
6596 case IORING_OP_READ:
6597 ret = io_read(req, issue_flags);
6599 case IORING_OP_WRITEV:
6600 case IORING_OP_WRITE_FIXED:
6601 case IORING_OP_WRITE:
6602 ret = io_write(req, issue_flags);
6604 case IORING_OP_FSYNC:
6605 ret = io_fsync(req, issue_flags);
6607 case IORING_OP_POLL_ADD:
6608 ret = io_poll_add(req, issue_flags);
6610 case IORING_OP_POLL_REMOVE:
6611 ret = io_poll_update(req, issue_flags);
6613 case IORING_OP_SYNC_FILE_RANGE:
6614 ret = io_sync_file_range(req, issue_flags);
6616 case IORING_OP_SENDMSG:
6617 ret = io_sendmsg(req, issue_flags);
6619 case IORING_OP_SEND:
6620 ret = io_send(req, issue_flags);
6622 case IORING_OP_RECVMSG:
6623 ret = io_recvmsg(req, issue_flags);
6625 case IORING_OP_RECV:
6626 ret = io_recv(req, issue_flags);
6628 case IORING_OP_TIMEOUT:
6629 ret = io_timeout(req, issue_flags);
6631 case IORING_OP_TIMEOUT_REMOVE:
6632 ret = io_timeout_remove(req, issue_flags);
6634 case IORING_OP_ACCEPT:
6635 ret = io_accept(req, issue_flags);
6637 case IORING_OP_CONNECT:
6638 ret = io_connect(req, issue_flags);
6640 case IORING_OP_ASYNC_CANCEL:
6641 ret = io_async_cancel(req, issue_flags);
6643 case IORING_OP_FALLOCATE:
6644 ret = io_fallocate(req, issue_flags);
6646 case IORING_OP_OPENAT:
6647 ret = io_openat(req, issue_flags);
6649 case IORING_OP_CLOSE:
6650 ret = io_close(req, issue_flags);
6652 case IORING_OP_FILES_UPDATE:
6653 ret = io_files_update(req, issue_flags);
6655 case IORING_OP_STATX:
6656 ret = io_statx(req, issue_flags);
6658 case IORING_OP_FADVISE:
6659 ret = io_fadvise(req, issue_flags);
6661 case IORING_OP_MADVISE:
6662 ret = io_madvise(req, issue_flags);
6664 case IORING_OP_OPENAT2:
6665 ret = io_openat2(req, issue_flags);
6667 case IORING_OP_EPOLL_CTL:
6668 ret = io_epoll_ctl(req, issue_flags);
6670 case IORING_OP_SPLICE:
6671 ret = io_splice(req, issue_flags);
6673 case IORING_OP_PROVIDE_BUFFERS:
6674 ret = io_provide_buffers(req, issue_flags);
6676 case IORING_OP_REMOVE_BUFFERS:
6677 ret = io_remove_buffers(req, issue_flags);
6680 ret = io_tee(req, issue_flags);
6682 case IORING_OP_SHUTDOWN:
6683 ret = io_shutdown(req, issue_flags);
6685 case IORING_OP_RENAMEAT:
6686 ret = io_renameat(req, issue_flags);
6688 case IORING_OP_UNLINKAT:
6689 ret = io_unlinkat(req, issue_flags);
6691 case IORING_OP_MKDIRAT:
6692 ret = io_mkdirat(req, issue_flags);
6694 case IORING_OP_SYMLINKAT:
6695 ret = io_symlinkat(req, issue_flags);
6697 case IORING_OP_LINKAT:
6698 ret = io_linkat(req, issue_flags);
6706 revert_creds(creds);
6709 /* If the op doesn't have a file, we're not polling for it */
6710 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6711 io_iopoll_req_issued(req);
6716 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6718 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6720 req = io_put_req_find_next(req);
6721 return req ? &req->work : NULL;
6724 static void io_wq_submit_work(struct io_wq_work *work)
6726 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6727 struct io_kiocb *timeout;
6730 /* one will be dropped by ->io_free_work() after returning to io-wq */
6731 if (!(req->flags & REQ_F_REFCOUNT))
6732 __io_req_set_refcount(req, 2);
6736 timeout = io_prep_linked_timeout(req);
6738 io_queue_linked_timeout(timeout);
6740 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6741 if (work->flags & IO_WQ_WORK_CANCEL)
6746 ret = io_issue_sqe(req, 0);
6748 * We can get EAGAIN for polled IO even though we're
6749 * forcing a sync submission from here, since we can't
6750 * wait for request slots on the block side.
6758 /* avoid locking problems by failing it from a clean context */
6760 io_req_task_queue_fail(req, ret);
6763 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6766 return &table->files[i];
6769 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6772 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6774 return (struct file *) (slot->file_ptr & FFS_MASK);
6777 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6779 unsigned long file_ptr = (unsigned long) file;
6781 if (__io_file_supports_nowait(file, READ))
6782 file_ptr |= FFS_ASYNC_READ;
6783 if (__io_file_supports_nowait(file, WRITE))
6784 file_ptr |= FFS_ASYNC_WRITE;
6785 if (S_ISREG(file_inode(file)->i_mode))
6786 file_ptr |= FFS_ISREG;
6787 file_slot->file_ptr = file_ptr;
6790 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6791 struct io_kiocb *req, int fd)
6794 unsigned long file_ptr;
6796 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6798 fd = array_index_nospec(fd, ctx->nr_user_files);
6799 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6800 file = (struct file *) (file_ptr & FFS_MASK);
6801 file_ptr &= ~FFS_MASK;
6802 /* mask in overlapping REQ_F and FFS bits */
6803 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6804 io_req_set_rsrc_node(req);
6808 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6809 struct io_kiocb *req, int fd)
6811 struct file *file = fget(fd);
6813 trace_io_uring_file_get(ctx, fd);
6815 /* we don't allow fixed io_uring files */
6816 if (file && unlikely(file->f_op == &io_uring_fops))
6817 io_req_track_inflight(req);
6821 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6822 struct io_kiocb *req, int fd, bool fixed)
6825 return io_file_get_fixed(ctx, req, fd);
6827 return io_file_get_normal(ctx, req, fd);
6830 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6832 struct io_kiocb *prev = req->timeout.prev;
6836 ret = io_try_cancel_userdata(req, prev->user_data);
6837 io_req_complete_post(req, ret ?: -ETIME, 0);
6840 io_req_complete_post(req, -ETIME, 0);
6844 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6846 struct io_timeout_data *data = container_of(timer,
6847 struct io_timeout_data, timer);
6848 struct io_kiocb *prev, *req = data->req;
6849 struct io_ring_ctx *ctx = req->ctx;
6850 unsigned long flags;
6852 spin_lock_irqsave(&ctx->timeout_lock, flags);
6853 prev = req->timeout.head;
6854 req->timeout.head = NULL;
6857 * We don't expect the list to be empty, that will only happen if we
6858 * race with the completion of the linked work.
6861 io_remove_next_linked(prev);
6862 if (!req_ref_inc_not_zero(prev))
6865 list_del(&req->timeout.list);
6866 req->timeout.prev = prev;
6867 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6869 req->io_task_work.func = io_req_task_link_timeout;
6870 io_req_task_work_add(req);
6871 return HRTIMER_NORESTART;
6874 static void io_queue_linked_timeout(struct io_kiocb *req)
6876 struct io_ring_ctx *ctx = req->ctx;
6878 spin_lock_irq(&ctx->timeout_lock);
6880 * If the back reference is NULL, then our linked request finished
6881 * before we got a chance to setup the timer
6883 if (req->timeout.head) {
6884 struct io_timeout_data *data = req->async_data;
6886 data->timer.function = io_link_timeout_fn;
6887 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6889 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6891 spin_unlock_irq(&ctx->timeout_lock);
6892 /* drop submission reference */
6896 static void __io_queue_sqe(struct io_kiocb *req)
6897 __must_hold(&req->ctx->uring_lock)
6899 struct io_kiocb *linked_timeout;
6903 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6906 * We async punt it if the file wasn't marked NOWAIT, or if the file
6907 * doesn't support non-blocking read/write attempts
6910 if (req->flags & REQ_F_COMPLETE_INLINE) {
6911 struct io_ring_ctx *ctx = req->ctx;
6912 struct io_submit_state *state = &ctx->submit_state;
6914 state->compl_reqs[state->compl_nr++] = req;
6915 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6916 io_submit_flush_completions(ctx);
6920 linked_timeout = io_prep_linked_timeout(req);
6922 io_queue_linked_timeout(linked_timeout);
6923 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6924 linked_timeout = io_prep_linked_timeout(req);
6926 switch (io_arm_poll_handler(req)) {
6927 case IO_APOLL_READY:
6929 io_unprep_linked_timeout(req);
6931 case IO_APOLL_ABORTED:
6933 * Queued up for async execution, worker will release
6934 * submit reference when the iocb is actually submitted.
6936 io_queue_async_work(req, NULL);
6941 io_queue_linked_timeout(linked_timeout);
6943 io_req_complete_failed(req, ret);
6947 static inline void io_queue_sqe(struct io_kiocb *req)
6948 __must_hold(&req->ctx->uring_lock)
6950 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6953 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
6954 __io_queue_sqe(req);
6955 } else if (req->flags & REQ_F_FAIL) {
6956 io_req_complete_fail_submit(req);
6958 int ret = io_req_prep_async(req);
6961 io_req_complete_failed(req, ret);
6963 io_queue_async_work(req, NULL);
6968 * Check SQE restrictions (opcode and flags).
6970 * Returns 'true' if SQE is allowed, 'false' otherwise.
6972 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6973 struct io_kiocb *req,
6974 unsigned int sqe_flags)
6976 if (likely(!ctx->restricted))
6979 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6982 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6983 ctx->restrictions.sqe_flags_required)
6986 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6987 ctx->restrictions.sqe_flags_required))
6993 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6994 const struct io_uring_sqe *sqe)
6995 __must_hold(&ctx->uring_lock)
6997 struct io_submit_state *state;
6998 unsigned int sqe_flags;
6999 int personality, ret = 0;
7001 /* req is partially pre-initialised, see io_preinit_req() */
7002 req->opcode = READ_ONCE(sqe->opcode);
7003 /* same numerical values with corresponding REQ_F_*, safe to copy */
7004 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7005 req->user_data = READ_ONCE(sqe->user_data);
7007 req->fixed_rsrc_refs = NULL;
7008 req->task = current;
7010 /* enforce forwards compatibility on users */
7011 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7013 if (unlikely(req->opcode >= IORING_OP_LAST))
7015 if (!io_check_restriction(ctx, req, sqe_flags))
7018 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7019 !io_op_defs[req->opcode].buffer_select)
7021 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7022 ctx->drain_active = true;
7024 personality = READ_ONCE(sqe->personality);
7026 req->creds = xa_load(&ctx->personalities, personality);
7029 get_cred(req->creds);
7030 req->flags |= REQ_F_CREDS;
7032 state = &ctx->submit_state;
7035 * Plug now if we have more than 1 IO left after this, and the target
7036 * is potentially a read/write to block based storage.
7038 if (!state->plug_started && state->ios_left > 1 &&
7039 io_op_defs[req->opcode].plug) {
7040 blk_start_plug(&state->plug);
7041 state->plug_started = true;
7044 if (io_op_defs[req->opcode].needs_file) {
7045 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7046 (sqe_flags & IOSQE_FIXED_FILE));
7047 if (unlikely(!req->file))
7055 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7056 const struct io_uring_sqe *sqe)
7057 __must_hold(&ctx->uring_lock)
7059 struct io_submit_link *link = &ctx->submit_state.link;
7062 ret = io_init_req(ctx, req, sqe);
7063 if (unlikely(ret)) {
7065 /* fail even hard links since we don't submit */
7068 * we can judge a link req is failed or cancelled by if
7069 * REQ_F_FAIL is set, but the head is an exception since
7070 * it may be set REQ_F_FAIL because of other req's failure
7071 * so let's leverage req->result to distinguish if a head
7072 * is set REQ_F_FAIL because of its failure or other req's
7073 * failure so that we can set the correct ret code for it.
7074 * init result here to avoid affecting the normal path.
7076 if (!(link->head->flags & REQ_F_FAIL))
7077 req_fail_link_node(link->head, -ECANCELED);
7078 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7080 * the current req is a normal req, we should return
7081 * error and thus break the submittion loop.
7083 io_req_complete_failed(req, ret);
7086 req_fail_link_node(req, ret);
7088 ret = io_req_prep(req, sqe);
7093 /* don't need @sqe from now on */
7094 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7096 ctx->flags & IORING_SETUP_SQPOLL);
7099 * If we already have a head request, queue this one for async
7100 * submittal once the head completes. If we don't have a head but
7101 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7102 * submitted sync once the chain is complete. If none of those
7103 * conditions are true (normal request), then just queue it.
7106 struct io_kiocb *head = link->head;
7108 if (!(req->flags & REQ_F_FAIL)) {
7109 ret = io_req_prep_async(req);
7110 if (unlikely(ret)) {
7111 req_fail_link_node(req, ret);
7112 if (!(head->flags & REQ_F_FAIL))
7113 req_fail_link_node(head, -ECANCELED);
7116 trace_io_uring_link(ctx, req, head);
7117 link->last->link = req;
7120 /* last request of a link, enqueue the link */
7121 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7126 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7138 * Batched submission is done, ensure local IO is flushed out.
7140 static void io_submit_state_end(struct io_submit_state *state,
7141 struct io_ring_ctx *ctx)
7143 if (state->link.head)
7144 io_queue_sqe(state->link.head);
7145 if (state->compl_nr)
7146 io_submit_flush_completions(ctx);
7147 if (state->plug_started)
7148 blk_finish_plug(&state->plug);
7152 * Start submission side cache.
7154 static void io_submit_state_start(struct io_submit_state *state,
7155 unsigned int max_ios)
7157 state->plug_started = false;
7158 state->ios_left = max_ios;
7159 /* set only head, no need to init link_last in advance */
7160 state->link.head = NULL;
7163 static void io_commit_sqring(struct io_ring_ctx *ctx)
7165 struct io_rings *rings = ctx->rings;
7168 * Ensure any loads from the SQEs are done at this point,
7169 * since once we write the new head, the application could
7170 * write new data to them.
7172 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7176 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7177 * that is mapped by userspace. This means that care needs to be taken to
7178 * ensure that reads are stable, as we cannot rely on userspace always
7179 * being a good citizen. If members of the sqe are validated and then later
7180 * used, it's important that those reads are done through READ_ONCE() to
7181 * prevent a re-load down the line.
7183 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7185 unsigned head, mask = ctx->sq_entries - 1;
7186 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7189 * The cached sq head (or cq tail) serves two purposes:
7191 * 1) allows us to batch the cost of updating the user visible
7193 * 2) allows the kernel side to track the head on its own, even
7194 * though the application is the one updating it.
7196 head = READ_ONCE(ctx->sq_array[sq_idx]);
7197 if (likely(head < ctx->sq_entries))
7198 return &ctx->sq_sqes[head];
7200 /* drop invalid entries */
7202 WRITE_ONCE(ctx->rings->sq_dropped,
7203 READ_ONCE(ctx->rings->sq_dropped) + 1);
7207 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7208 __must_hold(&ctx->uring_lock)
7212 /* make sure SQ entry isn't read before tail */
7213 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7214 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7216 io_get_task_refs(nr);
7218 io_submit_state_start(&ctx->submit_state, nr);
7219 while (submitted < nr) {
7220 const struct io_uring_sqe *sqe;
7221 struct io_kiocb *req;
7223 req = io_alloc_req(ctx);
7224 if (unlikely(!req)) {
7226 submitted = -EAGAIN;
7229 sqe = io_get_sqe(ctx);
7230 if (unlikely(!sqe)) {
7231 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7234 /* will complete beyond this point, count as submitted */
7236 if (io_submit_sqe(ctx, req, sqe))
7240 if (unlikely(submitted != nr)) {
7241 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7242 int unused = nr - ref_used;
7244 current->io_uring->cached_refs += unused;
7245 percpu_ref_put_many(&ctx->refs, unused);
7248 io_submit_state_end(&ctx->submit_state, ctx);
7249 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7250 io_commit_sqring(ctx);
7255 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7257 return READ_ONCE(sqd->state);
7260 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7262 /* Tell userspace we may need a wakeup call */
7263 spin_lock(&ctx->completion_lock);
7264 WRITE_ONCE(ctx->rings->sq_flags,
7265 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7266 spin_unlock(&ctx->completion_lock);
7269 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7271 spin_lock(&ctx->completion_lock);
7272 WRITE_ONCE(ctx->rings->sq_flags,
7273 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7274 spin_unlock(&ctx->completion_lock);
7277 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7279 unsigned int to_submit;
7282 to_submit = io_sqring_entries(ctx);
7283 /* if we're handling multiple rings, cap submit size for fairness */
7284 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7285 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7287 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7288 unsigned nr_events = 0;
7289 const struct cred *creds = NULL;
7291 if (ctx->sq_creds != current_cred())
7292 creds = override_creds(ctx->sq_creds);
7294 mutex_lock(&ctx->uring_lock);
7295 if (!list_empty(&ctx->iopoll_list))
7296 io_do_iopoll(ctx, &nr_events, 0);
7299 * Don't submit if refs are dying, good for io_uring_register(),
7300 * but also it is relied upon by io_ring_exit_work()
7302 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7303 !(ctx->flags & IORING_SETUP_R_DISABLED))
7304 ret = io_submit_sqes(ctx, to_submit);
7305 mutex_unlock(&ctx->uring_lock);
7307 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7308 wake_up(&ctx->sqo_sq_wait);
7310 revert_creds(creds);
7316 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7318 struct io_ring_ctx *ctx;
7319 unsigned sq_thread_idle = 0;
7321 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7322 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7323 sqd->sq_thread_idle = sq_thread_idle;
7326 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7328 bool did_sig = false;
7329 struct ksignal ksig;
7331 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7332 signal_pending(current)) {
7333 mutex_unlock(&sqd->lock);
7334 if (signal_pending(current))
7335 did_sig = get_signal(&ksig);
7337 mutex_lock(&sqd->lock);
7339 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7342 static int io_sq_thread(void *data)
7344 struct io_sq_data *sqd = data;
7345 struct io_ring_ctx *ctx;
7346 unsigned long timeout = 0;
7347 char buf[TASK_COMM_LEN];
7350 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7351 set_task_comm(current, buf);
7353 if (sqd->sq_cpu != -1)
7354 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7356 set_cpus_allowed_ptr(current, cpu_online_mask);
7357 current->flags |= PF_NO_SETAFFINITY;
7359 mutex_lock(&sqd->lock);
7361 bool cap_entries, sqt_spin = false;
7363 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7364 if (io_sqd_handle_event(sqd))
7366 timeout = jiffies + sqd->sq_thread_idle;
7369 cap_entries = !list_is_singular(&sqd->ctx_list);
7370 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7371 int ret = __io_sq_thread(ctx, cap_entries);
7373 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7376 if (io_run_task_work())
7379 if (sqt_spin || !time_after(jiffies, timeout)) {
7382 timeout = jiffies + sqd->sq_thread_idle;
7386 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7387 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7388 bool needs_sched = true;
7390 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7391 io_ring_set_wakeup_flag(ctx);
7393 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7394 !list_empty_careful(&ctx->iopoll_list)) {
7395 needs_sched = false;
7398 if (io_sqring_entries(ctx)) {
7399 needs_sched = false;
7405 mutex_unlock(&sqd->lock);
7407 mutex_lock(&sqd->lock);
7409 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7410 io_ring_clear_wakeup_flag(ctx);
7413 finish_wait(&sqd->wait, &wait);
7414 timeout = jiffies + sqd->sq_thread_idle;
7417 io_uring_cancel_generic(true, sqd);
7419 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7420 io_ring_set_wakeup_flag(ctx);
7422 mutex_unlock(&sqd->lock);
7424 complete(&sqd->exited);
7428 struct io_wait_queue {
7429 struct wait_queue_entry wq;
7430 struct io_ring_ctx *ctx;
7432 unsigned nr_timeouts;
7435 static inline bool io_should_wake(struct io_wait_queue *iowq)
7437 struct io_ring_ctx *ctx = iowq->ctx;
7438 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7441 * Wake up if we have enough events, or if a timeout occurred since we
7442 * started waiting. For timeouts, we always want to return to userspace,
7443 * regardless of event count.
7445 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7448 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7449 int wake_flags, void *key)
7451 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7455 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7456 * the task, and the next invocation will do it.
7458 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7459 return autoremove_wake_function(curr, mode, wake_flags, key);
7463 static int io_run_task_work_sig(void)
7465 if (io_run_task_work())
7467 if (!signal_pending(current))
7469 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7470 return -ERESTARTSYS;
7474 /* when returns >0, the caller should retry */
7475 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7476 struct io_wait_queue *iowq,
7477 signed long *timeout)
7481 /* make sure we run task_work before checking for signals */
7482 ret = io_run_task_work_sig();
7483 if (ret || io_should_wake(iowq))
7485 /* let the caller flush overflows, retry */
7486 if (test_bit(0, &ctx->check_cq_overflow))
7489 *timeout = schedule_timeout(*timeout);
7490 return !*timeout ? -ETIME : 1;
7494 * Wait until events become available, if we don't already have some. The
7495 * application must reap them itself, as they reside on the shared cq ring.
7497 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7498 const sigset_t __user *sig, size_t sigsz,
7499 struct __kernel_timespec __user *uts)
7501 struct io_wait_queue iowq;
7502 struct io_rings *rings = ctx->rings;
7503 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7507 io_cqring_overflow_flush(ctx);
7508 if (io_cqring_events(ctx) >= min_events)
7510 if (!io_run_task_work())
7515 #ifdef CONFIG_COMPAT
7516 if (in_compat_syscall())
7517 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7521 ret = set_user_sigmask(sig, sigsz);
7528 struct timespec64 ts;
7530 if (get_timespec64(&ts, uts))
7532 timeout = timespec64_to_jiffies(&ts);
7535 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7536 iowq.wq.private = current;
7537 INIT_LIST_HEAD(&iowq.wq.entry);
7539 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7540 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7542 trace_io_uring_cqring_wait(ctx, min_events);
7544 /* if we can't even flush overflow, don't wait for more */
7545 if (!io_cqring_overflow_flush(ctx)) {
7549 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7550 TASK_INTERRUPTIBLE);
7551 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7552 finish_wait(&ctx->cq_wait, &iowq.wq);
7556 restore_saved_sigmask_unless(ret == -EINTR);
7558 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7561 static void io_free_page_table(void **table, size_t size)
7563 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7565 for (i = 0; i < nr_tables; i++)
7570 static void **io_alloc_page_table(size_t size)
7572 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7573 size_t init_size = size;
7576 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7580 for (i = 0; i < nr_tables; i++) {
7581 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7583 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7585 io_free_page_table(table, init_size);
7593 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7595 percpu_ref_exit(&ref_node->refs);
7599 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7601 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7602 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7603 unsigned long flags;
7604 bool first_add = false;
7606 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7609 while (!list_empty(&ctx->rsrc_ref_list)) {
7610 node = list_first_entry(&ctx->rsrc_ref_list,
7611 struct io_rsrc_node, node);
7612 /* recycle ref nodes in order */
7615 list_del(&node->node);
7616 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7618 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7621 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7624 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7626 struct io_rsrc_node *ref_node;
7628 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7632 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7637 INIT_LIST_HEAD(&ref_node->node);
7638 INIT_LIST_HEAD(&ref_node->rsrc_list);
7639 ref_node->done = false;
7643 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7644 struct io_rsrc_data *data_to_kill)
7646 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7647 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7650 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7652 rsrc_node->rsrc_data = data_to_kill;
7653 spin_lock_irq(&ctx->rsrc_ref_lock);
7654 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7655 spin_unlock_irq(&ctx->rsrc_ref_lock);
7657 atomic_inc(&data_to_kill->refs);
7658 percpu_ref_kill(&rsrc_node->refs);
7659 ctx->rsrc_node = NULL;
7662 if (!ctx->rsrc_node) {
7663 ctx->rsrc_node = ctx->rsrc_backup_node;
7664 ctx->rsrc_backup_node = NULL;
7668 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7670 if (ctx->rsrc_backup_node)
7672 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7673 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7676 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7680 /* As we may drop ->uring_lock, other task may have started quiesce */
7684 data->quiesce = true;
7686 ret = io_rsrc_node_switch_start(ctx);
7689 io_rsrc_node_switch(ctx, data);
7691 /* kill initial ref, already quiesced if zero */
7692 if (atomic_dec_and_test(&data->refs))
7694 mutex_unlock(&ctx->uring_lock);
7695 flush_delayed_work(&ctx->rsrc_put_work);
7696 ret = wait_for_completion_interruptible(&data->done);
7698 mutex_lock(&ctx->uring_lock);
7702 atomic_inc(&data->refs);
7703 /* wait for all works potentially completing data->done */
7704 flush_delayed_work(&ctx->rsrc_put_work);
7705 reinit_completion(&data->done);
7707 ret = io_run_task_work_sig();
7708 mutex_lock(&ctx->uring_lock);
7710 data->quiesce = false;
7715 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7717 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7718 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7720 return &data->tags[table_idx][off];
7723 static void io_rsrc_data_free(struct io_rsrc_data *data)
7725 size_t size = data->nr * sizeof(data->tags[0][0]);
7728 io_free_page_table((void **)data->tags, size);
7732 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7733 u64 __user *utags, unsigned nr,
7734 struct io_rsrc_data **pdata)
7736 struct io_rsrc_data *data;
7740 data = kzalloc(sizeof(*data), GFP_KERNEL);
7743 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7751 data->do_put = do_put;
7754 for (i = 0; i < nr; i++) {
7755 u64 *tag_slot = io_get_tag_slot(data, i);
7757 if (copy_from_user(tag_slot, &utags[i],
7763 atomic_set(&data->refs, 1);
7764 init_completion(&data->done);
7768 io_rsrc_data_free(data);
7772 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7774 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7775 GFP_KERNEL_ACCOUNT);
7776 return !!table->files;
7779 static void io_free_file_tables(struct io_file_table *table)
7781 kvfree(table->files);
7782 table->files = NULL;
7785 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7787 #if defined(CONFIG_UNIX)
7788 if (ctx->ring_sock) {
7789 struct sock *sock = ctx->ring_sock->sk;
7790 struct sk_buff *skb;
7792 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7798 for (i = 0; i < ctx->nr_user_files; i++) {
7801 file = io_file_from_index(ctx, i);
7806 io_free_file_tables(&ctx->file_table);
7807 io_rsrc_data_free(ctx->file_data);
7808 ctx->file_data = NULL;
7809 ctx->nr_user_files = 0;
7812 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7816 if (!ctx->file_data)
7818 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7820 __io_sqe_files_unregister(ctx);
7824 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7825 __releases(&sqd->lock)
7827 WARN_ON_ONCE(sqd->thread == current);
7830 * Do the dance but not conditional clear_bit() because it'd race with
7831 * other threads incrementing park_pending and setting the bit.
7833 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7834 if (atomic_dec_return(&sqd->park_pending))
7835 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7836 mutex_unlock(&sqd->lock);
7839 static void io_sq_thread_park(struct io_sq_data *sqd)
7840 __acquires(&sqd->lock)
7842 WARN_ON_ONCE(sqd->thread == current);
7844 atomic_inc(&sqd->park_pending);
7845 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7846 mutex_lock(&sqd->lock);
7848 wake_up_process(sqd->thread);
7851 static void io_sq_thread_stop(struct io_sq_data *sqd)
7853 WARN_ON_ONCE(sqd->thread == current);
7854 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7856 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7857 mutex_lock(&sqd->lock);
7859 wake_up_process(sqd->thread);
7860 mutex_unlock(&sqd->lock);
7861 wait_for_completion(&sqd->exited);
7864 static void io_put_sq_data(struct io_sq_data *sqd)
7866 if (refcount_dec_and_test(&sqd->refs)) {
7867 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7869 io_sq_thread_stop(sqd);
7874 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7876 struct io_sq_data *sqd = ctx->sq_data;
7879 io_sq_thread_park(sqd);
7880 list_del_init(&ctx->sqd_list);
7881 io_sqd_update_thread_idle(sqd);
7882 io_sq_thread_unpark(sqd);
7884 io_put_sq_data(sqd);
7885 ctx->sq_data = NULL;
7889 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7891 struct io_ring_ctx *ctx_attach;
7892 struct io_sq_data *sqd;
7895 f = fdget(p->wq_fd);
7897 return ERR_PTR(-ENXIO);
7898 if (f.file->f_op != &io_uring_fops) {
7900 return ERR_PTR(-EINVAL);
7903 ctx_attach = f.file->private_data;
7904 sqd = ctx_attach->sq_data;
7907 return ERR_PTR(-EINVAL);
7909 if (sqd->task_tgid != current->tgid) {
7911 return ERR_PTR(-EPERM);
7914 refcount_inc(&sqd->refs);
7919 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7922 struct io_sq_data *sqd;
7925 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7926 sqd = io_attach_sq_data(p);
7931 /* fall through for EPERM case, setup new sqd/task */
7932 if (PTR_ERR(sqd) != -EPERM)
7936 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7938 return ERR_PTR(-ENOMEM);
7940 atomic_set(&sqd->park_pending, 0);
7941 refcount_set(&sqd->refs, 1);
7942 INIT_LIST_HEAD(&sqd->ctx_list);
7943 mutex_init(&sqd->lock);
7944 init_waitqueue_head(&sqd->wait);
7945 init_completion(&sqd->exited);
7949 #if defined(CONFIG_UNIX)
7951 * Ensure the UNIX gc is aware of our file set, so we are certain that
7952 * the io_uring can be safely unregistered on process exit, even if we have
7953 * loops in the file referencing.
7955 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7957 struct sock *sk = ctx->ring_sock->sk;
7958 struct scm_fp_list *fpl;
7959 struct sk_buff *skb;
7962 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7966 skb = alloc_skb(0, GFP_KERNEL);
7975 fpl->user = get_uid(current_user());
7976 for (i = 0; i < nr; i++) {
7977 struct file *file = io_file_from_index(ctx, i + offset);
7981 fpl->fp[nr_files] = get_file(file);
7982 unix_inflight(fpl->user, fpl->fp[nr_files]);
7987 fpl->max = SCM_MAX_FD;
7988 fpl->count = nr_files;
7989 UNIXCB(skb).fp = fpl;
7990 skb->destructor = unix_destruct_scm;
7991 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7992 skb_queue_head(&sk->sk_receive_queue, skb);
7994 for (i = 0; i < nr_files; i++)
8005 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8006 * causes regular reference counting to break down. We rely on the UNIX
8007 * garbage collection to take care of this problem for us.
8009 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8011 unsigned left, total;
8015 left = ctx->nr_user_files;
8017 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8019 ret = __io_sqe_files_scm(ctx, this_files, total);
8023 total += this_files;
8029 while (total < ctx->nr_user_files) {
8030 struct file *file = io_file_from_index(ctx, total);
8040 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8046 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8048 struct file *file = prsrc->file;
8049 #if defined(CONFIG_UNIX)
8050 struct sock *sock = ctx->ring_sock->sk;
8051 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8052 struct sk_buff *skb;
8055 __skb_queue_head_init(&list);
8058 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8059 * remove this entry and rearrange the file array.
8061 skb = skb_dequeue(head);
8063 struct scm_fp_list *fp;
8065 fp = UNIXCB(skb).fp;
8066 for (i = 0; i < fp->count; i++) {
8069 if (fp->fp[i] != file)
8072 unix_notinflight(fp->user, fp->fp[i]);
8073 left = fp->count - 1 - i;
8075 memmove(&fp->fp[i], &fp->fp[i + 1],
8076 left * sizeof(struct file *));
8083 __skb_queue_tail(&list, skb);
8093 __skb_queue_tail(&list, skb);
8095 skb = skb_dequeue(head);
8098 if (skb_peek(&list)) {
8099 spin_lock_irq(&head->lock);
8100 while ((skb = __skb_dequeue(&list)) != NULL)
8101 __skb_queue_tail(head, skb);
8102 spin_unlock_irq(&head->lock);
8109 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8111 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8112 struct io_ring_ctx *ctx = rsrc_data->ctx;
8113 struct io_rsrc_put *prsrc, *tmp;
8115 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8116 list_del(&prsrc->list);
8119 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8121 io_ring_submit_lock(ctx, lock_ring);
8122 spin_lock(&ctx->completion_lock);
8123 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8125 io_commit_cqring(ctx);
8126 spin_unlock(&ctx->completion_lock);
8127 io_cqring_ev_posted(ctx);
8128 io_ring_submit_unlock(ctx, lock_ring);
8131 rsrc_data->do_put(ctx, prsrc);
8135 io_rsrc_node_destroy(ref_node);
8136 if (atomic_dec_and_test(&rsrc_data->refs))
8137 complete(&rsrc_data->done);
8140 static void io_rsrc_put_work(struct work_struct *work)
8142 struct io_ring_ctx *ctx;
8143 struct llist_node *node;
8145 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8146 node = llist_del_all(&ctx->rsrc_put_llist);
8149 struct io_rsrc_node *ref_node;
8150 struct llist_node *next = node->next;
8152 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8153 __io_rsrc_put_work(ref_node);
8158 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8159 unsigned nr_args, u64 __user *tags)
8161 __s32 __user *fds = (__s32 __user *) arg;
8170 if (nr_args > IORING_MAX_FIXED_FILES)
8172 if (nr_args > rlimit(RLIMIT_NOFILE))
8174 ret = io_rsrc_node_switch_start(ctx);
8177 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8183 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8186 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8187 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8191 /* allow sparse sets */
8194 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8201 if (unlikely(!file))
8205 * Don't allow io_uring instances to be registered. If UNIX
8206 * isn't enabled, then this causes a reference cycle and this
8207 * instance can never get freed. If UNIX is enabled we'll
8208 * handle it just fine, but there's still no point in allowing
8209 * a ring fd as it doesn't support regular read/write anyway.
8211 if (file->f_op == &io_uring_fops) {
8215 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8218 ret = io_sqe_files_scm(ctx);
8220 __io_sqe_files_unregister(ctx);
8224 io_rsrc_node_switch(ctx, NULL);
8227 for (i = 0; i < ctx->nr_user_files; i++) {
8228 file = io_file_from_index(ctx, i);
8232 io_free_file_tables(&ctx->file_table);
8233 ctx->nr_user_files = 0;
8235 io_rsrc_data_free(ctx->file_data);
8236 ctx->file_data = NULL;
8240 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8243 #if defined(CONFIG_UNIX)
8244 struct sock *sock = ctx->ring_sock->sk;
8245 struct sk_buff_head *head = &sock->sk_receive_queue;
8246 struct sk_buff *skb;
8249 * See if we can merge this file into an existing skb SCM_RIGHTS
8250 * file set. If there's no room, fall back to allocating a new skb
8251 * and filling it in.
8253 spin_lock_irq(&head->lock);
8254 skb = skb_peek(head);
8256 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8258 if (fpl->count < SCM_MAX_FD) {
8259 __skb_unlink(skb, head);
8260 spin_unlock_irq(&head->lock);
8261 fpl->fp[fpl->count] = get_file(file);
8262 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8264 spin_lock_irq(&head->lock);
8265 __skb_queue_head(head, skb);
8270 spin_unlock_irq(&head->lock);
8277 return __io_sqe_files_scm(ctx, 1, index);
8283 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8284 unsigned int issue_flags, u32 slot_index)
8286 struct io_ring_ctx *ctx = req->ctx;
8287 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8288 struct io_fixed_file *file_slot;
8291 io_ring_submit_lock(ctx, !force_nonblock);
8292 if (file->f_op == &io_uring_fops)
8295 if (!ctx->file_data)
8298 if (slot_index >= ctx->nr_user_files)
8301 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8302 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8304 if (file_slot->file_ptr)
8307 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8308 io_fixed_file_set(file_slot, file);
8309 ret = io_sqe_file_register(ctx, file, slot_index);
8311 file_slot->file_ptr = 0;
8317 io_ring_submit_unlock(ctx, !force_nonblock);
8323 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8324 struct io_rsrc_node *node, void *rsrc)
8326 struct io_rsrc_put *prsrc;
8328 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8332 prsrc->tag = *io_get_tag_slot(data, idx);
8334 list_add(&prsrc->list, &node->rsrc_list);
8338 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8339 struct io_uring_rsrc_update2 *up,
8342 u64 __user *tags = u64_to_user_ptr(up->tags);
8343 __s32 __user *fds = u64_to_user_ptr(up->data);
8344 struct io_rsrc_data *data = ctx->file_data;
8345 struct io_fixed_file *file_slot;
8349 bool needs_switch = false;
8351 if (!ctx->file_data)
8353 if (up->offset + nr_args > ctx->nr_user_files)
8356 for (done = 0; done < nr_args; done++) {
8359 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8360 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8364 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8368 if (fd == IORING_REGISTER_FILES_SKIP)
8371 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8372 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8374 if (file_slot->file_ptr) {
8375 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8376 err = io_queue_rsrc_removal(data, up->offset + done,
8377 ctx->rsrc_node, file);
8380 file_slot->file_ptr = 0;
8381 needs_switch = true;
8390 * Don't allow io_uring instances to be registered. If
8391 * UNIX isn't enabled, then this causes a reference
8392 * cycle and this instance can never get freed. If UNIX
8393 * is enabled we'll handle it just fine, but there's
8394 * still no point in allowing a ring fd as it doesn't
8395 * support regular read/write anyway.
8397 if (file->f_op == &io_uring_fops) {
8402 *io_get_tag_slot(data, up->offset + done) = tag;
8403 io_fixed_file_set(file_slot, file);
8404 err = io_sqe_file_register(ctx, file, i);
8406 file_slot->file_ptr = 0;
8414 io_rsrc_node_switch(ctx, data);
8415 return done ? done : err;
8418 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8419 struct task_struct *task)
8421 struct io_wq_hash *hash;
8422 struct io_wq_data data;
8423 unsigned int concurrency;
8425 mutex_lock(&ctx->uring_lock);
8426 hash = ctx->hash_map;
8428 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8430 mutex_unlock(&ctx->uring_lock);
8431 return ERR_PTR(-ENOMEM);
8433 refcount_set(&hash->refs, 1);
8434 init_waitqueue_head(&hash->wait);
8435 ctx->hash_map = hash;
8437 mutex_unlock(&ctx->uring_lock);
8441 data.free_work = io_wq_free_work;
8442 data.do_work = io_wq_submit_work;
8444 /* Do QD, or 4 * CPUS, whatever is smallest */
8445 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8447 return io_wq_create(concurrency, &data);
8450 static int io_uring_alloc_task_context(struct task_struct *task,
8451 struct io_ring_ctx *ctx)
8453 struct io_uring_task *tctx;
8456 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8457 if (unlikely(!tctx))
8460 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8461 if (unlikely(ret)) {
8466 tctx->io_wq = io_init_wq_offload(ctx, task);
8467 if (IS_ERR(tctx->io_wq)) {
8468 ret = PTR_ERR(tctx->io_wq);
8469 percpu_counter_destroy(&tctx->inflight);
8475 init_waitqueue_head(&tctx->wait);
8476 atomic_set(&tctx->in_idle, 0);
8477 atomic_set(&tctx->inflight_tracked, 0);
8478 task->io_uring = tctx;
8479 spin_lock_init(&tctx->task_lock);
8480 INIT_WQ_LIST(&tctx->task_list);
8481 init_task_work(&tctx->task_work, tctx_task_work);
8485 void __io_uring_free(struct task_struct *tsk)
8487 struct io_uring_task *tctx = tsk->io_uring;
8489 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8490 WARN_ON_ONCE(tctx->io_wq);
8491 WARN_ON_ONCE(tctx->cached_refs);
8493 percpu_counter_destroy(&tctx->inflight);
8495 tsk->io_uring = NULL;
8498 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8499 struct io_uring_params *p)
8503 /* Retain compatibility with failing for an invalid attach attempt */
8504 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8505 IORING_SETUP_ATTACH_WQ) {
8508 f = fdget(p->wq_fd);
8511 if (f.file->f_op != &io_uring_fops) {
8517 if (ctx->flags & IORING_SETUP_SQPOLL) {
8518 struct task_struct *tsk;
8519 struct io_sq_data *sqd;
8522 sqd = io_get_sq_data(p, &attached);
8528 ctx->sq_creds = get_current_cred();
8530 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8531 if (!ctx->sq_thread_idle)
8532 ctx->sq_thread_idle = HZ;
8534 io_sq_thread_park(sqd);
8535 list_add(&ctx->sqd_list, &sqd->ctx_list);
8536 io_sqd_update_thread_idle(sqd);
8537 /* don't attach to a dying SQPOLL thread, would be racy */
8538 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8539 io_sq_thread_unpark(sqd);
8546 if (p->flags & IORING_SETUP_SQ_AFF) {
8547 int cpu = p->sq_thread_cpu;
8550 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8557 sqd->task_pid = current->pid;
8558 sqd->task_tgid = current->tgid;
8559 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8566 ret = io_uring_alloc_task_context(tsk, ctx);
8567 wake_up_new_task(tsk);
8570 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8571 /* Can't have SQ_AFF without SQPOLL */
8578 complete(&ctx->sq_data->exited);
8580 io_sq_thread_finish(ctx);
8584 static inline void __io_unaccount_mem(struct user_struct *user,
8585 unsigned long nr_pages)
8587 atomic_long_sub(nr_pages, &user->locked_vm);
8590 static inline int __io_account_mem(struct user_struct *user,
8591 unsigned long nr_pages)
8593 unsigned long page_limit, cur_pages, new_pages;
8595 /* Don't allow more pages than we can safely lock */
8596 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8599 cur_pages = atomic_long_read(&user->locked_vm);
8600 new_pages = cur_pages + nr_pages;
8601 if (new_pages > page_limit)
8603 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8604 new_pages) != cur_pages);
8609 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8612 __io_unaccount_mem(ctx->user, nr_pages);
8614 if (ctx->mm_account)
8615 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8618 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8623 ret = __io_account_mem(ctx->user, nr_pages);
8628 if (ctx->mm_account)
8629 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8634 static void io_mem_free(void *ptr)
8641 page = virt_to_head_page(ptr);
8642 if (put_page_testzero(page))
8643 free_compound_page(page);
8646 static void *io_mem_alloc(size_t size)
8648 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8649 __GFP_NORETRY | __GFP_ACCOUNT;
8651 return (void *) __get_free_pages(gfp_flags, get_order(size));
8654 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8657 struct io_rings *rings;
8658 size_t off, sq_array_size;
8660 off = struct_size(rings, cqes, cq_entries);
8661 if (off == SIZE_MAX)
8665 off = ALIGN(off, SMP_CACHE_BYTES);
8673 sq_array_size = array_size(sizeof(u32), sq_entries);
8674 if (sq_array_size == SIZE_MAX)
8677 if (check_add_overflow(off, sq_array_size, &off))
8683 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8685 struct io_mapped_ubuf *imu = *slot;
8688 if (imu != ctx->dummy_ubuf) {
8689 for (i = 0; i < imu->nr_bvecs; i++)
8690 unpin_user_page(imu->bvec[i].bv_page);
8691 if (imu->acct_pages)
8692 io_unaccount_mem(ctx, imu->acct_pages);
8698 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8700 io_buffer_unmap(ctx, &prsrc->buf);
8704 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8708 for (i = 0; i < ctx->nr_user_bufs; i++)
8709 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8710 kfree(ctx->user_bufs);
8711 io_rsrc_data_free(ctx->buf_data);
8712 ctx->user_bufs = NULL;
8713 ctx->buf_data = NULL;
8714 ctx->nr_user_bufs = 0;
8717 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8724 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8726 __io_sqe_buffers_unregister(ctx);
8730 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8731 void __user *arg, unsigned index)
8733 struct iovec __user *src;
8735 #ifdef CONFIG_COMPAT
8737 struct compat_iovec __user *ciovs;
8738 struct compat_iovec ciov;
8740 ciovs = (struct compat_iovec __user *) arg;
8741 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8744 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8745 dst->iov_len = ciov.iov_len;
8749 src = (struct iovec __user *) arg;
8750 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8756 * Not super efficient, but this is just a registration time. And we do cache
8757 * the last compound head, so generally we'll only do a full search if we don't
8760 * We check if the given compound head page has already been accounted, to
8761 * avoid double accounting it. This allows us to account the full size of the
8762 * page, not just the constituent pages of a huge page.
8764 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8765 int nr_pages, struct page *hpage)
8769 /* check current page array */
8770 for (i = 0; i < nr_pages; i++) {
8771 if (!PageCompound(pages[i]))
8773 if (compound_head(pages[i]) == hpage)
8777 /* check previously registered pages */
8778 for (i = 0; i < ctx->nr_user_bufs; i++) {
8779 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8781 for (j = 0; j < imu->nr_bvecs; j++) {
8782 if (!PageCompound(imu->bvec[j].bv_page))
8784 if (compound_head(imu->bvec[j].bv_page) == hpage)
8792 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8793 int nr_pages, struct io_mapped_ubuf *imu,
8794 struct page **last_hpage)
8798 imu->acct_pages = 0;
8799 for (i = 0; i < nr_pages; i++) {
8800 if (!PageCompound(pages[i])) {
8805 hpage = compound_head(pages[i]);
8806 if (hpage == *last_hpage)
8808 *last_hpage = hpage;
8809 if (headpage_already_acct(ctx, pages, i, hpage))
8811 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8815 if (!imu->acct_pages)
8818 ret = io_account_mem(ctx, imu->acct_pages);
8820 imu->acct_pages = 0;
8824 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8825 struct io_mapped_ubuf **pimu,
8826 struct page **last_hpage)
8828 struct io_mapped_ubuf *imu = NULL;
8829 struct vm_area_struct **vmas = NULL;
8830 struct page **pages = NULL;
8831 unsigned long off, start, end, ubuf;
8833 int ret, pret, nr_pages, i;
8835 if (!iov->iov_base) {
8836 *pimu = ctx->dummy_ubuf;
8840 ubuf = (unsigned long) iov->iov_base;
8841 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8842 start = ubuf >> PAGE_SHIFT;
8843 nr_pages = end - start;
8848 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8852 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8857 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8862 mmap_read_lock(current->mm);
8863 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8865 if (pret == nr_pages) {
8866 /* don't support file backed memory */
8867 for (i = 0; i < nr_pages; i++) {
8868 struct vm_area_struct *vma = vmas[i];
8870 if (vma_is_shmem(vma))
8873 !is_file_hugepages(vma->vm_file)) {
8879 ret = pret < 0 ? pret : -EFAULT;
8881 mmap_read_unlock(current->mm);
8884 * if we did partial map, or found file backed vmas,
8885 * release any pages we did get
8888 unpin_user_pages(pages, pret);
8892 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8894 unpin_user_pages(pages, pret);
8898 off = ubuf & ~PAGE_MASK;
8899 size = iov->iov_len;
8900 for (i = 0; i < nr_pages; i++) {
8903 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8904 imu->bvec[i].bv_page = pages[i];
8905 imu->bvec[i].bv_len = vec_len;
8906 imu->bvec[i].bv_offset = off;
8910 /* store original address for later verification */
8912 imu->ubuf_end = ubuf + iov->iov_len;
8913 imu->nr_bvecs = nr_pages;
8924 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8926 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8927 return ctx->user_bufs ? 0 : -ENOMEM;
8930 static int io_buffer_validate(struct iovec *iov)
8932 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8935 * Don't impose further limits on the size and buffer
8936 * constraints here, we'll -EINVAL later when IO is
8937 * submitted if they are wrong.
8940 return iov->iov_len ? -EFAULT : 0;
8944 /* arbitrary limit, but we need something */
8945 if (iov->iov_len > SZ_1G)
8948 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8954 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8955 unsigned int nr_args, u64 __user *tags)
8957 struct page *last_hpage = NULL;
8958 struct io_rsrc_data *data;
8964 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8966 ret = io_rsrc_node_switch_start(ctx);
8969 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8972 ret = io_buffers_map_alloc(ctx, nr_args);
8974 io_rsrc_data_free(data);
8978 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8979 ret = io_copy_iov(ctx, &iov, arg, i);
8982 ret = io_buffer_validate(&iov);
8985 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8990 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8996 WARN_ON_ONCE(ctx->buf_data);
8998 ctx->buf_data = data;
9000 __io_sqe_buffers_unregister(ctx);
9002 io_rsrc_node_switch(ctx, NULL);
9006 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9007 struct io_uring_rsrc_update2 *up,
9008 unsigned int nr_args)
9010 u64 __user *tags = u64_to_user_ptr(up->tags);
9011 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9012 struct page *last_hpage = NULL;
9013 bool needs_switch = false;
9019 if (up->offset + nr_args > ctx->nr_user_bufs)
9022 for (done = 0; done < nr_args; done++) {
9023 struct io_mapped_ubuf *imu;
9024 int offset = up->offset + done;
9027 err = io_copy_iov(ctx, &iov, iovs, done);
9030 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9034 err = io_buffer_validate(&iov);
9037 if (!iov.iov_base && tag) {
9041 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9045 i = array_index_nospec(offset, ctx->nr_user_bufs);
9046 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9047 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9048 ctx->rsrc_node, ctx->user_bufs[i]);
9049 if (unlikely(err)) {
9050 io_buffer_unmap(ctx, &imu);
9053 ctx->user_bufs[i] = NULL;
9054 needs_switch = true;
9057 ctx->user_bufs[i] = imu;
9058 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9062 io_rsrc_node_switch(ctx, ctx->buf_data);
9063 return done ? done : err;
9066 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9068 __s32 __user *fds = arg;
9074 if (copy_from_user(&fd, fds, sizeof(*fds)))
9077 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9078 if (IS_ERR(ctx->cq_ev_fd)) {
9079 int ret = PTR_ERR(ctx->cq_ev_fd);
9081 ctx->cq_ev_fd = NULL;
9088 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9090 if (ctx->cq_ev_fd) {
9091 eventfd_ctx_put(ctx->cq_ev_fd);
9092 ctx->cq_ev_fd = NULL;
9099 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9101 struct io_buffer *buf;
9102 unsigned long index;
9104 xa_for_each(&ctx->io_buffers, index, buf)
9105 __io_remove_buffers(ctx, buf, index, -1U);
9108 static void io_req_cache_free(struct list_head *list)
9110 struct io_kiocb *req, *nxt;
9112 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9113 list_del(&req->inflight_entry);
9114 kmem_cache_free(req_cachep, req);
9118 static void io_req_caches_free(struct io_ring_ctx *ctx)
9120 struct io_submit_state *state = &ctx->submit_state;
9122 mutex_lock(&ctx->uring_lock);
9124 if (state->free_reqs) {
9125 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9126 state->free_reqs = 0;
9129 io_flush_cached_locked_reqs(ctx, state);
9130 io_req_cache_free(&state->free_list);
9131 mutex_unlock(&ctx->uring_lock);
9134 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9136 if (data && !atomic_dec_and_test(&data->refs))
9137 wait_for_completion(&data->done);
9140 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9142 io_sq_thread_finish(ctx);
9144 if (ctx->mm_account) {
9145 mmdrop(ctx->mm_account);
9146 ctx->mm_account = NULL;
9149 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9150 io_wait_rsrc_data(ctx->buf_data);
9151 io_wait_rsrc_data(ctx->file_data);
9153 mutex_lock(&ctx->uring_lock);
9155 __io_sqe_buffers_unregister(ctx);
9157 __io_sqe_files_unregister(ctx);
9159 __io_cqring_overflow_flush(ctx, true);
9160 mutex_unlock(&ctx->uring_lock);
9161 io_eventfd_unregister(ctx);
9162 io_destroy_buffers(ctx);
9164 put_cred(ctx->sq_creds);
9166 /* there are no registered resources left, nobody uses it */
9168 io_rsrc_node_destroy(ctx->rsrc_node);
9169 if (ctx->rsrc_backup_node)
9170 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9171 flush_delayed_work(&ctx->rsrc_put_work);
9173 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9174 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9176 #if defined(CONFIG_UNIX)
9177 if (ctx->ring_sock) {
9178 ctx->ring_sock->file = NULL; /* so that iput() is called */
9179 sock_release(ctx->ring_sock);
9182 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9184 io_mem_free(ctx->rings);
9185 io_mem_free(ctx->sq_sqes);
9187 percpu_ref_exit(&ctx->refs);
9188 free_uid(ctx->user);
9189 io_req_caches_free(ctx);
9191 io_wq_put_hash(ctx->hash_map);
9192 kfree(ctx->cancel_hash);
9193 kfree(ctx->dummy_ubuf);
9197 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9199 struct io_ring_ctx *ctx = file->private_data;
9202 poll_wait(file, &ctx->poll_wait, wait);
9204 * synchronizes with barrier from wq_has_sleeper call in
9208 if (!io_sqring_full(ctx))
9209 mask |= EPOLLOUT | EPOLLWRNORM;
9212 * Don't flush cqring overflow list here, just do a simple check.
9213 * Otherwise there could possible be ABBA deadlock:
9216 * lock(&ctx->uring_lock);
9218 * lock(&ctx->uring_lock);
9221 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9222 * pushs them to do the flush.
9224 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9225 mask |= EPOLLIN | EPOLLRDNORM;
9230 static int io_uring_fasync(int fd, struct file *file, int on)
9232 struct io_ring_ctx *ctx = file->private_data;
9234 return fasync_helper(fd, file, on, &ctx->cq_fasync);
9237 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9239 const struct cred *creds;
9241 creds = xa_erase(&ctx->personalities, id);
9250 struct io_tctx_exit {
9251 struct callback_head task_work;
9252 struct completion completion;
9253 struct io_ring_ctx *ctx;
9256 static void io_tctx_exit_cb(struct callback_head *cb)
9258 struct io_uring_task *tctx = current->io_uring;
9259 struct io_tctx_exit *work;
9261 work = container_of(cb, struct io_tctx_exit, task_work);
9263 * When @in_idle, we're in cancellation and it's racy to remove the
9264 * node. It'll be removed by the end of cancellation, just ignore it.
9266 if (!atomic_read(&tctx->in_idle))
9267 io_uring_del_tctx_node((unsigned long)work->ctx);
9268 complete(&work->completion);
9271 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9273 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9275 return req->ctx == data;
9278 static void io_ring_exit_work(struct work_struct *work)
9280 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9281 unsigned long timeout = jiffies + HZ * 60 * 5;
9282 unsigned long interval = HZ / 20;
9283 struct io_tctx_exit exit;
9284 struct io_tctx_node *node;
9288 * If we're doing polled IO and end up having requests being
9289 * submitted async (out-of-line), then completions can come in while
9290 * we're waiting for refs to drop. We need to reap these manually,
9291 * as nobody else will be looking for them.
9294 io_uring_try_cancel_requests(ctx, NULL, true);
9296 struct io_sq_data *sqd = ctx->sq_data;
9297 struct task_struct *tsk;
9299 io_sq_thread_park(sqd);
9301 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9302 io_wq_cancel_cb(tsk->io_uring->io_wq,
9303 io_cancel_ctx_cb, ctx, true);
9304 io_sq_thread_unpark(sqd);
9307 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9308 /* there is little hope left, don't run it too often */
9311 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9313 init_completion(&exit.completion);
9314 init_task_work(&exit.task_work, io_tctx_exit_cb);
9317 * Some may use context even when all refs and requests have been put,
9318 * and they are free to do so while still holding uring_lock or
9319 * completion_lock, see io_req_task_submit(). Apart from other work,
9320 * this lock/unlock section also waits them to finish.
9322 mutex_lock(&ctx->uring_lock);
9323 while (!list_empty(&ctx->tctx_list)) {
9324 WARN_ON_ONCE(time_after(jiffies, timeout));
9326 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9328 /* don't spin on a single task if cancellation failed */
9329 list_rotate_left(&ctx->tctx_list);
9330 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9331 if (WARN_ON_ONCE(ret))
9333 wake_up_process(node->task);
9335 mutex_unlock(&ctx->uring_lock);
9336 wait_for_completion(&exit.completion);
9337 mutex_lock(&ctx->uring_lock);
9339 mutex_unlock(&ctx->uring_lock);
9340 spin_lock(&ctx->completion_lock);
9341 spin_unlock(&ctx->completion_lock);
9343 io_ring_ctx_free(ctx);
9346 /* Returns true if we found and killed one or more timeouts */
9347 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9350 struct io_kiocb *req, *tmp;
9353 spin_lock(&ctx->completion_lock);
9354 spin_lock_irq(&ctx->timeout_lock);
9355 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9356 if (io_match_task(req, tsk, cancel_all)) {
9357 io_kill_timeout(req, -ECANCELED);
9361 spin_unlock_irq(&ctx->timeout_lock);
9363 io_commit_cqring(ctx);
9364 spin_unlock(&ctx->completion_lock);
9366 io_cqring_ev_posted(ctx);
9367 return canceled != 0;
9370 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9372 unsigned long index;
9373 struct creds *creds;
9375 mutex_lock(&ctx->uring_lock);
9376 percpu_ref_kill(&ctx->refs);
9378 __io_cqring_overflow_flush(ctx, true);
9379 xa_for_each(&ctx->personalities, index, creds)
9380 io_unregister_personality(ctx, index);
9381 mutex_unlock(&ctx->uring_lock);
9383 io_kill_timeouts(ctx, NULL, true);
9384 io_poll_remove_all(ctx, NULL, true);
9386 /* if we failed setting up the ctx, we might not have any rings */
9387 io_iopoll_try_reap_events(ctx);
9389 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9391 * Use system_unbound_wq to avoid spawning tons of event kworkers
9392 * if we're exiting a ton of rings at the same time. It just adds
9393 * noise and overhead, there's no discernable change in runtime
9394 * over using system_wq.
9396 queue_work(system_unbound_wq, &ctx->exit_work);
9399 static int io_uring_release(struct inode *inode, struct file *file)
9401 struct io_ring_ctx *ctx = file->private_data;
9403 file->private_data = NULL;
9404 io_ring_ctx_wait_and_kill(ctx);
9408 struct io_task_cancel {
9409 struct task_struct *task;
9413 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9415 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9416 struct io_task_cancel *cancel = data;
9419 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9420 struct io_ring_ctx *ctx = req->ctx;
9422 /* protect against races with linked timeouts */
9423 spin_lock(&ctx->completion_lock);
9424 ret = io_match_task(req, cancel->task, cancel->all);
9425 spin_unlock(&ctx->completion_lock);
9427 ret = io_match_task(req, cancel->task, cancel->all);
9432 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9433 struct task_struct *task, bool cancel_all)
9435 struct io_defer_entry *de;
9438 spin_lock(&ctx->completion_lock);
9439 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9440 if (io_match_task(de->req, task, cancel_all)) {
9441 list_cut_position(&list, &ctx->defer_list, &de->list);
9445 spin_unlock(&ctx->completion_lock);
9446 if (list_empty(&list))
9449 while (!list_empty(&list)) {
9450 de = list_first_entry(&list, struct io_defer_entry, list);
9451 list_del_init(&de->list);
9452 io_req_complete_failed(de->req, -ECANCELED);
9458 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9460 struct io_tctx_node *node;
9461 enum io_wq_cancel cret;
9464 mutex_lock(&ctx->uring_lock);
9465 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9466 struct io_uring_task *tctx = node->task->io_uring;
9469 * io_wq will stay alive while we hold uring_lock, because it's
9470 * killed after ctx nodes, which requires to take the lock.
9472 if (!tctx || !tctx->io_wq)
9474 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9475 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9477 mutex_unlock(&ctx->uring_lock);
9482 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9483 struct task_struct *task,
9486 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9487 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9490 enum io_wq_cancel cret;
9494 ret |= io_uring_try_cancel_iowq(ctx);
9495 } else if (tctx && tctx->io_wq) {
9497 * Cancels requests of all rings, not only @ctx, but
9498 * it's fine as the task is in exit/exec.
9500 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9502 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9505 /* SQPOLL thread does its own polling */
9506 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9507 (ctx->sq_data && ctx->sq_data->thread == current)) {
9508 while (!list_empty_careful(&ctx->iopoll_list)) {
9509 io_iopoll_try_reap_events(ctx);
9514 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9515 ret |= io_poll_remove_all(ctx, task, cancel_all);
9516 ret |= io_kill_timeouts(ctx, task, cancel_all);
9518 ret |= io_run_task_work();
9525 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9527 struct io_uring_task *tctx = current->io_uring;
9528 struct io_tctx_node *node;
9531 if (unlikely(!tctx)) {
9532 ret = io_uring_alloc_task_context(current, ctx);
9535 tctx = current->io_uring;
9537 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9538 node = kmalloc(sizeof(*node), GFP_KERNEL);
9542 node->task = current;
9544 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9551 mutex_lock(&ctx->uring_lock);
9552 list_add(&node->ctx_node, &ctx->tctx_list);
9553 mutex_unlock(&ctx->uring_lock);
9560 * Note that this task has used io_uring. We use it for cancelation purposes.
9562 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9564 struct io_uring_task *tctx = current->io_uring;
9566 if (likely(tctx && tctx->last == ctx))
9568 return __io_uring_add_tctx_node(ctx);
9572 * Remove this io_uring_file -> task mapping.
9574 static void io_uring_del_tctx_node(unsigned long index)
9576 struct io_uring_task *tctx = current->io_uring;
9577 struct io_tctx_node *node;
9581 node = xa_erase(&tctx->xa, index);
9585 WARN_ON_ONCE(current != node->task);
9586 WARN_ON_ONCE(list_empty(&node->ctx_node));
9588 mutex_lock(&node->ctx->uring_lock);
9589 list_del(&node->ctx_node);
9590 mutex_unlock(&node->ctx->uring_lock);
9592 if (tctx->last == node->ctx)
9597 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9599 struct io_wq *wq = tctx->io_wq;
9600 struct io_tctx_node *node;
9601 unsigned long index;
9603 xa_for_each(&tctx->xa, index, node)
9604 io_uring_del_tctx_node(index);
9607 * Must be after io_uring_del_task_file() (removes nodes under
9608 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9610 io_wq_put_and_exit(wq);
9615 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9618 return atomic_read(&tctx->inflight_tracked);
9619 return percpu_counter_sum(&tctx->inflight);
9622 static void io_uring_drop_tctx_refs(struct task_struct *task)
9624 struct io_uring_task *tctx = task->io_uring;
9625 unsigned int refs = tctx->cached_refs;
9628 tctx->cached_refs = 0;
9629 percpu_counter_sub(&tctx->inflight, refs);
9630 put_task_struct_many(task, refs);
9635 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9636 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9638 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9640 struct io_uring_task *tctx = current->io_uring;
9641 struct io_ring_ctx *ctx;
9645 WARN_ON_ONCE(sqd && sqd->thread != current);
9647 if (!current->io_uring)
9650 io_wq_exit_start(tctx->io_wq);
9652 atomic_inc(&tctx->in_idle);
9654 io_uring_drop_tctx_refs(current);
9655 /* read completions before cancelations */
9656 inflight = tctx_inflight(tctx, !cancel_all);
9661 struct io_tctx_node *node;
9662 unsigned long index;
9664 xa_for_each(&tctx->xa, index, node) {
9665 /* sqpoll task will cancel all its requests */
9666 if (node->ctx->sq_data)
9668 io_uring_try_cancel_requests(node->ctx, current,
9672 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9673 io_uring_try_cancel_requests(ctx, current,
9677 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9678 io_uring_drop_tctx_refs(current);
9680 * If we've seen completions, retry without waiting. This
9681 * avoids a race where a completion comes in before we did
9682 * prepare_to_wait().
9684 if (inflight == tctx_inflight(tctx, !cancel_all))
9686 finish_wait(&tctx->wait, &wait);
9688 atomic_dec(&tctx->in_idle);
9690 io_uring_clean_tctx(tctx);
9692 /* for exec all current's requests should be gone, kill tctx */
9693 __io_uring_free(current);
9697 void __io_uring_cancel(bool cancel_all)
9699 io_uring_cancel_generic(cancel_all, NULL);
9702 static void *io_uring_validate_mmap_request(struct file *file,
9703 loff_t pgoff, size_t sz)
9705 struct io_ring_ctx *ctx = file->private_data;
9706 loff_t offset = pgoff << PAGE_SHIFT;
9711 case IORING_OFF_SQ_RING:
9712 case IORING_OFF_CQ_RING:
9715 case IORING_OFF_SQES:
9719 return ERR_PTR(-EINVAL);
9722 page = virt_to_head_page(ptr);
9723 if (sz > page_size(page))
9724 return ERR_PTR(-EINVAL);
9731 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9733 size_t sz = vma->vm_end - vma->vm_start;
9737 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9739 return PTR_ERR(ptr);
9741 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9742 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9745 #else /* !CONFIG_MMU */
9747 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9749 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9752 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9754 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9757 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9758 unsigned long addr, unsigned long len,
9759 unsigned long pgoff, unsigned long flags)
9763 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9765 return PTR_ERR(ptr);
9767 return (unsigned long) ptr;
9770 #endif /* !CONFIG_MMU */
9772 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9777 if (!io_sqring_full(ctx))
9779 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9781 if (!io_sqring_full(ctx))
9784 } while (!signal_pending(current));
9786 finish_wait(&ctx->sqo_sq_wait, &wait);
9790 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9791 struct __kernel_timespec __user **ts,
9792 const sigset_t __user **sig)
9794 struct io_uring_getevents_arg arg;
9797 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9798 * is just a pointer to the sigset_t.
9800 if (!(flags & IORING_ENTER_EXT_ARG)) {
9801 *sig = (const sigset_t __user *) argp;
9807 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9808 * timespec and sigset_t pointers if good.
9810 if (*argsz != sizeof(arg))
9812 if (copy_from_user(&arg, argp, sizeof(arg)))
9814 *sig = u64_to_user_ptr(arg.sigmask);
9815 *argsz = arg.sigmask_sz;
9816 *ts = u64_to_user_ptr(arg.ts);
9820 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9821 u32, min_complete, u32, flags, const void __user *, argp,
9824 struct io_ring_ctx *ctx;
9831 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9832 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9836 if (unlikely(!f.file))
9840 if (unlikely(f.file->f_op != &io_uring_fops))
9844 ctx = f.file->private_data;
9845 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9849 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9853 * For SQ polling, the thread will do all submissions and completions.
9854 * Just return the requested submit count, and wake the thread if
9858 if (ctx->flags & IORING_SETUP_SQPOLL) {
9859 io_cqring_overflow_flush(ctx);
9861 if (unlikely(ctx->sq_data->thread == NULL)) {
9865 if (flags & IORING_ENTER_SQ_WAKEUP)
9866 wake_up(&ctx->sq_data->wait);
9867 if (flags & IORING_ENTER_SQ_WAIT) {
9868 ret = io_sqpoll_wait_sq(ctx);
9872 submitted = to_submit;
9873 } else if (to_submit) {
9874 ret = io_uring_add_tctx_node(ctx);
9877 mutex_lock(&ctx->uring_lock);
9878 submitted = io_submit_sqes(ctx, to_submit);
9879 mutex_unlock(&ctx->uring_lock);
9881 if (submitted != to_submit)
9884 if (flags & IORING_ENTER_GETEVENTS) {
9885 const sigset_t __user *sig;
9886 struct __kernel_timespec __user *ts;
9888 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9892 min_complete = min(min_complete, ctx->cq_entries);
9895 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9896 * space applications don't need to do io completion events
9897 * polling again, they can rely on io_sq_thread to do polling
9898 * work, which can reduce cpu usage and uring_lock contention.
9900 if (ctx->flags & IORING_SETUP_IOPOLL &&
9901 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9902 ret = io_iopoll_check(ctx, min_complete);
9904 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9909 percpu_ref_put(&ctx->refs);
9912 return submitted ? submitted : ret;
9915 #ifdef CONFIG_PROC_FS
9916 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9917 const struct cred *cred)
9919 struct user_namespace *uns = seq_user_ns(m);
9920 struct group_info *gi;
9925 seq_printf(m, "%5d\n", id);
9926 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9927 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9928 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9929 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9930 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9931 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9932 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9933 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9934 seq_puts(m, "\n\tGroups:\t");
9935 gi = cred->group_info;
9936 for (g = 0; g < gi->ngroups; g++) {
9937 seq_put_decimal_ull(m, g ? " " : "",
9938 from_kgid_munged(uns, gi->gid[g]));
9940 seq_puts(m, "\n\tCapEff:\t");
9941 cap = cred->cap_effective;
9942 CAP_FOR_EACH_U32(__capi)
9943 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9948 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9950 struct io_sq_data *sq = NULL;
9955 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9956 * since fdinfo case grabs it in the opposite direction of normal use
9957 * cases. If we fail to get the lock, we just don't iterate any
9958 * structures that could be going away outside the io_uring mutex.
9960 has_lock = mutex_trylock(&ctx->uring_lock);
9962 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9968 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9969 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9970 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9971 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9972 struct file *f = io_file_from_index(ctx, i);
9975 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9977 seq_printf(m, "%5u: <none>\n", i);
9979 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9980 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9981 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9982 unsigned int len = buf->ubuf_end - buf->ubuf;
9984 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9986 if (has_lock && !xa_empty(&ctx->personalities)) {
9987 unsigned long index;
9988 const struct cred *cred;
9990 seq_printf(m, "Personalities:\n");
9991 xa_for_each(&ctx->personalities, index, cred)
9992 io_uring_show_cred(m, index, cred);
9994 seq_printf(m, "PollList:\n");
9995 spin_lock(&ctx->completion_lock);
9996 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9997 struct hlist_head *list = &ctx->cancel_hash[i];
9998 struct io_kiocb *req;
10000 hlist_for_each_entry(req, list, hash_node)
10001 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10002 req->task->task_works != NULL);
10004 spin_unlock(&ctx->completion_lock);
10006 mutex_unlock(&ctx->uring_lock);
10009 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10011 struct io_ring_ctx *ctx = f->private_data;
10013 if (percpu_ref_tryget(&ctx->refs)) {
10014 __io_uring_show_fdinfo(ctx, m);
10015 percpu_ref_put(&ctx->refs);
10020 static const struct file_operations io_uring_fops = {
10021 .release = io_uring_release,
10022 .mmap = io_uring_mmap,
10024 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10025 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10027 .poll = io_uring_poll,
10028 .fasync = io_uring_fasync,
10029 #ifdef CONFIG_PROC_FS
10030 .show_fdinfo = io_uring_show_fdinfo,
10034 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10035 struct io_uring_params *p)
10037 struct io_rings *rings;
10038 size_t size, sq_array_offset;
10040 /* make sure these are sane, as we already accounted them */
10041 ctx->sq_entries = p->sq_entries;
10042 ctx->cq_entries = p->cq_entries;
10044 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10045 if (size == SIZE_MAX)
10048 rings = io_mem_alloc(size);
10052 ctx->rings = rings;
10053 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10054 rings->sq_ring_mask = p->sq_entries - 1;
10055 rings->cq_ring_mask = p->cq_entries - 1;
10056 rings->sq_ring_entries = p->sq_entries;
10057 rings->cq_ring_entries = p->cq_entries;
10059 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10060 if (size == SIZE_MAX) {
10061 io_mem_free(ctx->rings);
10066 ctx->sq_sqes = io_mem_alloc(size);
10067 if (!ctx->sq_sqes) {
10068 io_mem_free(ctx->rings);
10076 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10080 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10084 ret = io_uring_add_tctx_node(ctx);
10089 fd_install(fd, file);
10094 * Allocate an anonymous fd, this is what constitutes the application
10095 * visible backing of an io_uring instance. The application mmaps this
10096 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10097 * we have to tie this fd to a socket for file garbage collection purposes.
10099 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10102 #if defined(CONFIG_UNIX)
10105 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10108 return ERR_PTR(ret);
10111 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10112 O_RDWR | O_CLOEXEC);
10113 #if defined(CONFIG_UNIX)
10114 if (IS_ERR(file)) {
10115 sock_release(ctx->ring_sock);
10116 ctx->ring_sock = NULL;
10118 ctx->ring_sock->file = file;
10124 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10125 struct io_uring_params __user *params)
10127 struct io_ring_ctx *ctx;
10133 if (entries > IORING_MAX_ENTRIES) {
10134 if (!(p->flags & IORING_SETUP_CLAMP))
10136 entries = IORING_MAX_ENTRIES;
10140 * Use twice as many entries for the CQ ring. It's possible for the
10141 * application to drive a higher depth than the size of the SQ ring,
10142 * since the sqes are only used at submission time. This allows for
10143 * some flexibility in overcommitting a bit. If the application has
10144 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10145 * of CQ ring entries manually.
10147 p->sq_entries = roundup_pow_of_two(entries);
10148 if (p->flags & IORING_SETUP_CQSIZE) {
10150 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10151 * to a power-of-two, if it isn't already. We do NOT impose
10152 * any cq vs sq ring sizing.
10154 if (!p->cq_entries)
10156 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10157 if (!(p->flags & IORING_SETUP_CLAMP))
10159 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10161 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10162 if (p->cq_entries < p->sq_entries)
10165 p->cq_entries = 2 * p->sq_entries;
10168 ctx = io_ring_ctx_alloc(p);
10171 ctx->compat = in_compat_syscall();
10172 if (!capable(CAP_IPC_LOCK))
10173 ctx->user = get_uid(current_user());
10176 * This is just grabbed for accounting purposes. When a process exits,
10177 * the mm is exited and dropped before the files, hence we need to hang
10178 * on to this mm purely for the purposes of being able to unaccount
10179 * memory (locked/pinned vm). It's not used for anything else.
10181 mmgrab(current->mm);
10182 ctx->mm_account = current->mm;
10184 ret = io_allocate_scq_urings(ctx, p);
10188 ret = io_sq_offload_create(ctx, p);
10191 /* always set a rsrc node */
10192 ret = io_rsrc_node_switch_start(ctx);
10195 io_rsrc_node_switch(ctx, NULL);
10197 memset(&p->sq_off, 0, sizeof(p->sq_off));
10198 p->sq_off.head = offsetof(struct io_rings, sq.head);
10199 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10200 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10201 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10202 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10203 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10204 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10206 memset(&p->cq_off, 0, sizeof(p->cq_off));
10207 p->cq_off.head = offsetof(struct io_rings, cq.head);
10208 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10209 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10210 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10211 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10212 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10213 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10215 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10216 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10217 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10218 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10219 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10220 IORING_FEAT_RSRC_TAGS;
10222 if (copy_to_user(params, p, sizeof(*p))) {
10227 file = io_uring_get_file(ctx);
10228 if (IS_ERR(file)) {
10229 ret = PTR_ERR(file);
10234 * Install ring fd as the very last thing, so we don't risk someone
10235 * having closed it before we finish setup
10237 ret = io_uring_install_fd(ctx, file);
10239 /* fput will clean it up */
10244 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10247 io_ring_ctx_wait_and_kill(ctx);
10252 * Sets up an aio uring context, and returns the fd. Applications asks for a
10253 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10254 * params structure passed in.
10256 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10258 struct io_uring_params p;
10261 if (copy_from_user(&p, params, sizeof(p)))
10263 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10268 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10269 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10270 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10271 IORING_SETUP_R_DISABLED))
10274 return io_uring_create(entries, &p, params);
10277 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10278 struct io_uring_params __user *, params)
10280 return io_uring_setup(entries, params);
10283 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10285 struct io_uring_probe *p;
10289 size = struct_size(p, ops, nr_args);
10290 if (size == SIZE_MAX)
10292 p = kzalloc(size, GFP_KERNEL);
10297 if (copy_from_user(p, arg, size))
10300 if (memchr_inv(p, 0, size))
10303 p->last_op = IORING_OP_LAST - 1;
10304 if (nr_args > IORING_OP_LAST)
10305 nr_args = IORING_OP_LAST;
10307 for (i = 0; i < nr_args; i++) {
10309 if (!io_op_defs[i].not_supported)
10310 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10315 if (copy_to_user(arg, p, size))
10322 static int io_register_personality(struct io_ring_ctx *ctx)
10324 const struct cred *creds;
10328 creds = get_current_cred();
10330 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10331 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10339 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10340 unsigned int nr_args)
10342 struct io_uring_restriction *res;
10346 /* Restrictions allowed only if rings started disabled */
10347 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10350 /* We allow only a single restrictions registration */
10351 if (ctx->restrictions.registered)
10354 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10357 size = array_size(nr_args, sizeof(*res));
10358 if (size == SIZE_MAX)
10361 res = memdup_user(arg, size);
10363 return PTR_ERR(res);
10367 for (i = 0; i < nr_args; i++) {
10368 switch (res[i].opcode) {
10369 case IORING_RESTRICTION_REGISTER_OP:
10370 if (res[i].register_op >= IORING_REGISTER_LAST) {
10375 __set_bit(res[i].register_op,
10376 ctx->restrictions.register_op);
10378 case IORING_RESTRICTION_SQE_OP:
10379 if (res[i].sqe_op >= IORING_OP_LAST) {
10384 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10386 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10387 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10389 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10390 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10399 /* Reset all restrictions if an error happened */
10401 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10403 ctx->restrictions.registered = true;
10409 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10411 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10414 if (ctx->restrictions.registered)
10415 ctx->restricted = 1;
10417 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10418 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10419 wake_up(&ctx->sq_data->wait);
10423 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10424 struct io_uring_rsrc_update2 *up,
10432 if (check_add_overflow(up->offset, nr_args, &tmp))
10434 err = io_rsrc_node_switch_start(ctx);
10439 case IORING_RSRC_FILE:
10440 return __io_sqe_files_update(ctx, up, nr_args);
10441 case IORING_RSRC_BUFFER:
10442 return __io_sqe_buffers_update(ctx, up, nr_args);
10447 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10450 struct io_uring_rsrc_update2 up;
10454 memset(&up, 0, sizeof(up));
10455 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10457 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10460 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10461 unsigned size, unsigned type)
10463 struct io_uring_rsrc_update2 up;
10465 if (size != sizeof(up))
10467 if (copy_from_user(&up, arg, sizeof(up)))
10469 if (!up.nr || up.resv)
10471 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10474 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10475 unsigned int size, unsigned int type)
10477 struct io_uring_rsrc_register rr;
10479 /* keep it extendible */
10480 if (size != sizeof(rr))
10483 memset(&rr, 0, sizeof(rr));
10484 if (copy_from_user(&rr, arg, size))
10486 if (!rr.nr || rr.resv || rr.resv2)
10490 case IORING_RSRC_FILE:
10491 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10492 rr.nr, u64_to_user_ptr(rr.tags));
10493 case IORING_RSRC_BUFFER:
10494 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10495 rr.nr, u64_to_user_ptr(rr.tags));
10500 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10503 struct io_uring_task *tctx = current->io_uring;
10504 cpumask_var_t new_mask;
10507 if (!tctx || !tctx->io_wq)
10510 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10513 cpumask_clear(new_mask);
10514 if (len > cpumask_size())
10515 len = cpumask_size();
10517 if (copy_from_user(new_mask, arg, len)) {
10518 free_cpumask_var(new_mask);
10522 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10523 free_cpumask_var(new_mask);
10527 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10529 struct io_uring_task *tctx = current->io_uring;
10531 if (!tctx || !tctx->io_wq)
10534 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10537 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10540 struct io_uring_task *tctx = NULL;
10541 struct io_sq_data *sqd = NULL;
10542 __u32 new_count[2];
10545 if (copy_from_user(new_count, arg, sizeof(new_count)))
10547 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10548 if (new_count[i] > INT_MAX)
10551 if (ctx->flags & IORING_SETUP_SQPOLL) {
10552 sqd = ctx->sq_data;
10555 * Observe the correct sqd->lock -> ctx->uring_lock
10556 * ordering. Fine to drop uring_lock here, we hold
10557 * a ref to the ctx.
10559 mutex_unlock(&ctx->uring_lock);
10560 mutex_lock(&sqd->lock);
10561 mutex_lock(&ctx->uring_lock);
10562 tctx = sqd->thread->io_uring;
10565 tctx = current->io_uring;
10569 if (!tctx || !tctx->io_wq)
10572 ret = io_wq_max_workers(tctx->io_wq, new_count);
10577 mutex_unlock(&sqd->lock);
10579 if (copy_to_user(arg, new_count, sizeof(new_count)))
10585 mutex_unlock(&sqd->lock);
10589 static bool io_register_op_must_quiesce(int op)
10592 case IORING_REGISTER_BUFFERS:
10593 case IORING_UNREGISTER_BUFFERS:
10594 case IORING_REGISTER_FILES:
10595 case IORING_UNREGISTER_FILES:
10596 case IORING_REGISTER_FILES_UPDATE:
10597 case IORING_REGISTER_PROBE:
10598 case IORING_REGISTER_PERSONALITY:
10599 case IORING_UNREGISTER_PERSONALITY:
10600 case IORING_REGISTER_FILES2:
10601 case IORING_REGISTER_FILES_UPDATE2:
10602 case IORING_REGISTER_BUFFERS2:
10603 case IORING_REGISTER_BUFFERS_UPDATE:
10604 case IORING_REGISTER_IOWQ_AFF:
10605 case IORING_UNREGISTER_IOWQ_AFF:
10606 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10613 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10617 percpu_ref_kill(&ctx->refs);
10620 * Drop uring mutex before waiting for references to exit. If another
10621 * thread is currently inside io_uring_enter() it might need to grab the
10622 * uring_lock to make progress. If we hold it here across the drain
10623 * wait, then we can deadlock. It's safe to drop the mutex here, since
10624 * no new references will come in after we've killed the percpu ref.
10626 mutex_unlock(&ctx->uring_lock);
10628 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10631 ret = io_run_task_work_sig();
10632 } while (ret >= 0);
10633 mutex_lock(&ctx->uring_lock);
10636 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10640 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10641 void __user *arg, unsigned nr_args)
10642 __releases(ctx->uring_lock)
10643 __acquires(ctx->uring_lock)
10648 * We're inside the ring mutex, if the ref is already dying, then
10649 * someone else killed the ctx or is already going through
10650 * io_uring_register().
10652 if (percpu_ref_is_dying(&ctx->refs))
10655 if (ctx->restricted) {
10656 if (opcode >= IORING_REGISTER_LAST)
10658 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10659 if (!test_bit(opcode, ctx->restrictions.register_op))
10663 if (io_register_op_must_quiesce(opcode)) {
10664 ret = io_ctx_quiesce(ctx);
10670 case IORING_REGISTER_BUFFERS:
10671 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10673 case IORING_UNREGISTER_BUFFERS:
10675 if (arg || nr_args)
10677 ret = io_sqe_buffers_unregister(ctx);
10679 case IORING_REGISTER_FILES:
10680 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10682 case IORING_UNREGISTER_FILES:
10684 if (arg || nr_args)
10686 ret = io_sqe_files_unregister(ctx);
10688 case IORING_REGISTER_FILES_UPDATE:
10689 ret = io_register_files_update(ctx, arg, nr_args);
10691 case IORING_REGISTER_EVENTFD:
10692 case IORING_REGISTER_EVENTFD_ASYNC:
10696 ret = io_eventfd_register(ctx, arg);
10699 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10700 ctx->eventfd_async = 1;
10702 ctx->eventfd_async = 0;
10704 case IORING_UNREGISTER_EVENTFD:
10706 if (arg || nr_args)
10708 ret = io_eventfd_unregister(ctx);
10710 case IORING_REGISTER_PROBE:
10712 if (!arg || nr_args > 256)
10714 ret = io_probe(ctx, arg, nr_args);
10716 case IORING_REGISTER_PERSONALITY:
10718 if (arg || nr_args)
10720 ret = io_register_personality(ctx);
10722 case IORING_UNREGISTER_PERSONALITY:
10726 ret = io_unregister_personality(ctx, nr_args);
10728 case IORING_REGISTER_ENABLE_RINGS:
10730 if (arg || nr_args)
10732 ret = io_register_enable_rings(ctx);
10734 case IORING_REGISTER_RESTRICTIONS:
10735 ret = io_register_restrictions(ctx, arg, nr_args);
10737 case IORING_REGISTER_FILES2:
10738 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10740 case IORING_REGISTER_FILES_UPDATE2:
10741 ret = io_register_rsrc_update(ctx, arg, nr_args,
10744 case IORING_REGISTER_BUFFERS2:
10745 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10747 case IORING_REGISTER_BUFFERS_UPDATE:
10748 ret = io_register_rsrc_update(ctx, arg, nr_args,
10749 IORING_RSRC_BUFFER);
10751 case IORING_REGISTER_IOWQ_AFF:
10753 if (!arg || !nr_args)
10755 ret = io_register_iowq_aff(ctx, arg, nr_args);
10757 case IORING_UNREGISTER_IOWQ_AFF:
10759 if (arg || nr_args)
10761 ret = io_unregister_iowq_aff(ctx);
10763 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10765 if (!arg || nr_args != 2)
10767 ret = io_register_iowq_max_workers(ctx, arg);
10774 if (io_register_op_must_quiesce(opcode)) {
10775 /* bring the ctx back to life */
10776 percpu_ref_reinit(&ctx->refs);
10777 reinit_completion(&ctx->ref_comp);
10782 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10783 void __user *, arg, unsigned int, nr_args)
10785 struct io_ring_ctx *ctx;
10794 if (f.file->f_op != &io_uring_fops)
10797 ctx = f.file->private_data;
10799 io_run_task_work();
10801 mutex_lock(&ctx->uring_lock);
10802 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10803 mutex_unlock(&ctx->uring_lock);
10804 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10805 ctx->cq_ev_fd != NULL, ret);
10811 static int __init io_uring_init(void)
10813 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10814 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10815 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10818 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10819 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10820 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10821 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10822 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10823 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10824 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10825 BUILD_BUG_SQE_ELEM(8, __u64, off);
10826 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10827 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10828 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10829 BUILD_BUG_SQE_ELEM(24, __u32, len);
10830 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10831 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10832 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10833 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10834 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10835 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10836 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10837 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10838 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10839 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10840 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10841 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10842 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10843 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10844 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10845 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10846 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10847 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10848 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10849 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10850 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10852 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10853 sizeof(struct io_uring_rsrc_update));
10854 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10855 sizeof(struct io_uring_rsrc_update2));
10857 /* ->buf_index is u16 */
10858 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10860 /* should fit into one byte */
10861 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10863 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10864 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10866 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10870 __initcall(io_uring_init);