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) {
1487 atomic_set(&req->ctx->cq_timeouts,
1488 atomic_read(&req->ctx->cq_timeouts) + 1);
1489 list_del_init(&req->timeout.list);
1490 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1491 io_put_req_deferred(req);
1495 static void io_queue_deferred(struct io_ring_ctx *ctx)
1497 while (!list_empty(&ctx->defer_list)) {
1498 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1499 struct io_defer_entry, list);
1501 if (req_need_defer(de->req, de->seq))
1503 list_del_init(&de->list);
1504 io_req_task_queue(de->req);
1509 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1510 __must_hold(&ctx->completion_lock)
1512 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1514 spin_lock_irq(&ctx->timeout_lock);
1515 while (!list_empty(&ctx->timeout_list)) {
1516 u32 events_needed, events_got;
1517 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1518 struct io_kiocb, timeout.list);
1520 if (io_is_timeout_noseq(req))
1524 * Since seq can easily wrap around over time, subtract
1525 * the last seq at which timeouts were flushed before comparing.
1526 * Assuming not more than 2^31-1 events have happened since,
1527 * these subtractions won't have wrapped, so we can check if
1528 * target is in [last_seq, current_seq] by comparing the two.
1530 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1531 events_got = seq - ctx->cq_last_tm_flush;
1532 if (events_got < events_needed)
1535 list_del_init(&req->timeout.list);
1536 io_kill_timeout(req, 0);
1538 ctx->cq_last_tm_flush = seq;
1539 spin_unlock_irq(&ctx->timeout_lock);
1542 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1544 if (ctx->off_timeout_used)
1545 io_flush_timeouts(ctx);
1546 if (ctx->drain_active)
1547 io_queue_deferred(ctx);
1550 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1552 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1553 __io_commit_cqring_flush(ctx);
1554 /* order cqe stores with ring update */
1555 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1558 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1560 struct io_rings *r = ctx->rings;
1562 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1565 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1567 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1570 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1572 struct io_rings *rings = ctx->rings;
1573 unsigned tail, mask = ctx->cq_entries - 1;
1576 * writes to the cq entry need to come after reading head; the
1577 * control dependency is enough as we're using WRITE_ONCE to
1580 if (__io_cqring_events(ctx) == ctx->cq_entries)
1583 tail = ctx->cached_cq_tail++;
1584 return &rings->cqes[tail & mask];
1587 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1589 if (likely(!ctx->cq_ev_fd))
1591 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1593 return !ctx->eventfd_async || io_wq_current_is_worker();
1597 * This should only get called when at least one event has been posted.
1598 * Some applications rely on the eventfd notification count only changing
1599 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1600 * 1:1 relationship between how many times this function is called (and
1601 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1603 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1606 * wake_up_all() may seem excessive, but io_wake_function() and
1607 * io_should_wake() handle the termination of the loop and only
1608 * wake as many waiters as we need to.
1610 if (wq_has_sleeper(&ctx->cq_wait))
1611 wake_up_all(&ctx->cq_wait);
1612 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1613 wake_up(&ctx->sq_data->wait);
1614 if (io_should_trigger_evfd(ctx))
1615 eventfd_signal(ctx->cq_ev_fd, 1);
1616 if (waitqueue_active(&ctx->poll_wait)) {
1617 wake_up_interruptible(&ctx->poll_wait);
1618 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1622 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1624 /* see waitqueue_active() comment */
1627 if (ctx->flags & IORING_SETUP_SQPOLL) {
1628 if (waitqueue_active(&ctx->cq_wait))
1629 wake_up_all(&ctx->cq_wait);
1631 if (io_should_trigger_evfd(ctx))
1632 eventfd_signal(ctx->cq_ev_fd, 1);
1633 if (waitqueue_active(&ctx->poll_wait)) {
1634 wake_up_interruptible(&ctx->poll_wait);
1635 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1639 /* Returns true if there are no backlogged entries after the flush */
1640 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1642 bool all_flushed, posted;
1644 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1648 spin_lock(&ctx->completion_lock);
1649 while (!list_empty(&ctx->cq_overflow_list)) {
1650 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1651 struct io_overflow_cqe *ocqe;
1655 ocqe = list_first_entry(&ctx->cq_overflow_list,
1656 struct io_overflow_cqe, list);
1658 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1660 io_account_cq_overflow(ctx);
1663 list_del(&ocqe->list);
1667 all_flushed = list_empty(&ctx->cq_overflow_list);
1669 clear_bit(0, &ctx->check_cq_overflow);
1670 WRITE_ONCE(ctx->rings->sq_flags,
1671 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1675 io_commit_cqring(ctx);
1676 spin_unlock(&ctx->completion_lock);
1678 io_cqring_ev_posted(ctx);
1682 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1686 if (test_bit(0, &ctx->check_cq_overflow)) {
1687 /* iopoll syncs against uring_lock, not completion_lock */
1688 if (ctx->flags & IORING_SETUP_IOPOLL)
1689 mutex_lock(&ctx->uring_lock);
1690 ret = __io_cqring_overflow_flush(ctx, false);
1691 if (ctx->flags & IORING_SETUP_IOPOLL)
1692 mutex_unlock(&ctx->uring_lock);
1698 /* must to be called somewhat shortly after putting a request */
1699 static inline void io_put_task(struct task_struct *task, int nr)
1701 struct io_uring_task *tctx = task->io_uring;
1703 if (likely(task == current)) {
1704 tctx->cached_refs += nr;
1706 percpu_counter_sub(&tctx->inflight, nr);
1707 if (unlikely(atomic_read(&tctx->in_idle)))
1708 wake_up(&tctx->wait);
1709 put_task_struct_many(task, nr);
1713 static void io_task_refs_refill(struct io_uring_task *tctx)
1715 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1717 percpu_counter_add(&tctx->inflight, refill);
1718 refcount_add(refill, ¤t->usage);
1719 tctx->cached_refs += refill;
1722 static inline void io_get_task_refs(int nr)
1724 struct io_uring_task *tctx = current->io_uring;
1726 tctx->cached_refs -= nr;
1727 if (unlikely(tctx->cached_refs < 0))
1728 io_task_refs_refill(tctx);
1731 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1732 long res, unsigned int cflags)
1734 struct io_overflow_cqe *ocqe;
1736 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1739 * If we're in ring overflow flush mode, or in task cancel mode,
1740 * or cannot allocate an overflow entry, then we need to drop it
1743 io_account_cq_overflow(ctx);
1746 if (list_empty(&ctx->cq_overflow_list)) {
1747 set_bit(0, &ctx->check_cq_overflow);
1748 WRITE_ONCE(ctx->rings->sq_flags,
1749 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1752 ocqe->cqe.user_data = user_data;
1753 ocqe->cqe.res = res;
1754 ocqe->cqe.flags = cflags;
1755 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1759 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1760 long res, unsigned int cflags)
1762 struct io_uring_cqe *cqe;
1764 trace_io_uring_complete(ctx, user_data, res, cflags);
1767 * If we can't get a cq entry, userspace overflowed the
1768 * submission (by quite a lot). Increment the overflow count in
1771 cqe = io_get_cqe(ctx);
1773 WRITE_ONCE(cqe->user_data, user_data);
1774 WRITE_ONCE(cqe->res, res);
1775 WRITE_ONCE(cqe->flags, cflags);
1778 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1781 /* not as hot to bloat with inlining */
1782 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1783 long res, unsigned int cflags)
1785 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1788 static void io_req_complete_post(struct io_kiocb *req, long res,
1789 unsigned int cflags)
1791 struct io_ring_ctx *ctx = req->ctx;
1793 spin_lock(&ctx->completion_lock);
1794 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1796 * If we're the last reference to this request, add to our locked
1799 if (req_ref_put_and_test(req)) {
1800 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1801 if (req->flags & IO_DISARM_MASK)
1802 io_disarm_next(req);
1804 io_req_task_queue(req->link);
1808 io_dismantle_req(req);
1809 io_put_task(req->task, 1);
1810 list_add(&req->inflight_entry, &ctx->locked_free_list);
1811 ctx->locked_free_nr++;
1813 if (!percpu_ref_tryget(&ctx->refs))
1816 io_commit_cqring(ctx);
1817 spin_unlock(&ctx->completion_lock);
1820 io_cqring_ev_posted(ctx);
1821 percpu_ref_put(&ctx->refs);
1825 static inline bool io_req_needs_clean(struct io_kiocb *req)
1827 return req->flags & IO_REQ_CLEAN_FLAGS;
1830 static void io_req_complete_state(struct io_kiocb *req, long res,
1831 unsigned int cflags)
1833 if (io_req_needs_clean(req))
1836 req->compl.cflags = cflags;
1837 req->flags |= REQ_F_COMPLETE_INLINE;
1840 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1841 long res, unsigned cflags)
1843 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1844 io_req_complete_state(req, res, cflags);
1846 io_req_complete_post(req, res, cflags);
1849 static inline void io_req_complete(struct io_kiocb *req, long res)
1851 __io_req_complete(req, 0, res, 0);
1854 static void io_req_complete_failed(struct io_kiocb *req, long res)
1857 io_req_complete_post(req, res, 0);
1860 static void io_req_complete_fail_submit(struct io_kiocb *req)
1863 * We don't submit, fail them all, for that replace hardlinks with
1864 * normal links. Extra REQ_F_LINK is tolerated.
1866 req->flags &= ~REQ_F_HARDLINK;
1867 req->flags |= REQ_F_LINK;
1868 io_req_complete_failed(req, req->result);
1872 * Don't initialise the fields below on every allocation, but do that in
1873 * advance and keep them valid across allocations.
1875 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1879 req->async_data = NULL;
1880 /* not necessary, but safer to zero */
1884 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1885 struct io_submit_state *state)
1887 spin_lock(&ctx->completion_lock);
1888 list_splice_init(&ctx->locked_free_list, &state->free_list);
1889 ctx->locked_free_nr = 0;
1890 spin_unlock(&ctx->completion_lock);
1893 /* Returns true IFF there are requests in the cache */
1894 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1896 struct io_submit_state *state = &ctx->submit_state;
1900 * If we have more than a batch's worth of requests in our IRQ side
1901 * locked cache, grab the lock and move them over to our submission
1904 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1905 io_flush_cached_locked_reqs(ctx, state);
1907 nr = state->free_reqs;
1908 while (!list_empty(&state->free_list)) {
1909 struct io_kiocb *req = list_first_entry(&state->free_list,
1910 struct io_kiocb, inflight_entry);
1912 list_del(&req->inflight_entry);
1913 state->reqs[nr++] = req;
1914 if (nr == ARRAY_SIZE(state->reqs))
1918 state->free_reqs = nr;
1923 * A request might get retired back into the request caches even before opcode
1924 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1925 * Because of that, io_alloc_req() should be called only under ->uring_lock
1926 * and with extra caution to not get a request that is still worked on.
1928 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1929 __must_hold(&ctx->uring_lock)
1931 struct io_submit_state *state = &ctx->submit_state;
1932 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1935 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1937 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1940 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1944 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1945 * retry single alloc to be on the safe side.
1947 if (unlikely(ret <= 0)) {
1948 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1949 if (!state->reqs[0])
1954 for (i = 0; i < ret; i++)
1955 io_preinit_req(state->reqs[i], ctx);
1956 state->free_reqs = ret;
1959 return state->reqs[state->free_reqs];
1962 static inline void io_put_file(struct file *file)
1968 static void io_dismantle_req(struct io_kiocb *req)
1970 unsigned int flags = req->flags;
1972 if (io_req_needs_clean(req))
1974 if (!(flags & REQ_F_FIXED_FILE))
1975 io_put_file(req->file);
1976 if (req->fixed_rsrc_refs)
1977 percpu_ref_put(req->fixed_rsrc_refs);
1978 if (req->async_data) {
1979 kfree(req->async_data);
1980 req->async_data = NULL;
1984 static void __io_free_req(struct io_kiocb *req)
1986 struct io_ring_ctx *ctx = req->ctx;
1988 io_dismantle_req(req);
1989 io_put_task(req->task, 1);
1991 spin_lock(&ctx->completion_lock);
1992 list_add(&req->inflight_entry, &ctx->locked_free_list);
1993 ctx->locked_free_nr++;
1994 spin_unlock(&ctx->completion_lock);
1996 percpu_ref_put(&ctx->refs);
1999 static inline void io_remove_next_linked(struct io_kiocb *req)
2001 struct io_kiocb *nxt = req->link;
2003 req->link = nxt->link;
2007 static bool io_kill_linked_timeout(struct io_kiocb *req)
2008 __must_hold(&req->ctx->completion_lock)
2009 __must_hold(&req->ctx->timeout_lock)
2011 struct io_kiocb *link = req->link;
2013 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2014 struct io_timeout_data *io = link->async_data;
2016 io_remove_next_linked(req);
2017 link->timeout.head = NULL;
2018 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2019 list_del(&link->timeout.list);
2020 io_cqring_fill_event(link->ctx, link->user_data,
2022 io_put_req_deferred(link);
2029 static void io_fail_links(struct io_kiocb *req)
2030 __must_hold(&req->ctx->completion_lock)
2032 struct io_kiocb *nxt, *link = req->link;
2036 long res = -ECANCELED;
2038 if (link->flags & REQ_F_FAIL)
2044 trace_io_uring_fail_link(req, link);
2045 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2046 io_put_req_deferred(link);
2051 static bool io_disarm_next(struct io_kiocb *req)
2052 __must_hold(&req->ctx->completion_lock)
2054 bool posted = false;
2056 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2057 struct io_kiocb *link = req->link;
2059 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2060 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2061 io_remove_next_linked(req);
2062 io_cqring_fill_event(link->ctx, link->user_data,
2064 io_put_req_deferred(link);
2067 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2068 struct io_ring_ctx *ctx = req->ctx;
2070 spin_lock_irq(&ctx->timeout_lock);
2071 posted = io_kill_linked_timeout(req);
2072 spin_unlock_irq(&ctx->timeout_lock);
2074 if (unlikely((req->flags & REQ_F_FAIL) &&
2075 !(req->flags & REQ_F_HARDLINK))) {
2076 posted |= (req->link != NULL);
2082 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2084 struct io_kiocb *nxt;
2087 * If LINK is set, we have dependent requests in this chain. If we
2088 * didn't fail this request, queue the first one up, moving any other
2089 * dependencies to the next request. In case of failure, fail the rest
2092 if (req->flags & IO_DISARM_MASK) {
2093 struct io_ring_ctx *ctx = req->ctx;
2096 spin_lock(&ctx->completion_lock);
2097 posted = io_disarm_next(req);
2099 io_commit_cqring(req->ctx);
2100 spin_unlock(&ctx->completion_lock);
2102 io_cqring_ev_posted(ctx);
2109 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2111 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2113 return __io_req_find_next(req);
2116 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2121 if (ctx->submit_state.compl_nr)
2122 io_submit_flush_completions(ctx);
2123 mutex_unlock(&ctx->uring_lock);
2126 percpu_ref_put(&ctx->refs);
2129 static void tctx_task_work(struct callback_head *cb)
2131 bool locked = false;
2132 struct io_ring_ctx *ctx = NULL;
2133 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2137 struct io_wq_work_node *node;
2139 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2140 io_submit_flush_completions(ctx);
2142 spin_lock_irq(&tctx->task_lock);
2143 node = tctx->task_list.first;
2144 INIT_WQ_LIST(&tctx->task_list);
2146 tctx->task_running = false;
2147 spin_unlock_irq(&tctx->task_lock);
2152 struct io_wq_work_node *next = node->next;
2153 struct io_kiocb *req = container_of(node, struct io_kiocb,
2156 if (req->ctx != ctx) {
2157 ctx_flush_and_put(ctx, &locked);
2159 /* if not contended, grab and improve batching */
2160 locked = mutex_trylock(&ctx->uring_lock);
2161 percpu_ref_get(&ctx->refs);
2163 req->io_task_work.func(req, &locked);
2170 ctx_flush_and_put(ctx, &locked);
2173 static void io_req_task_work_add(struct io_kiocb *req)
2175 struct task_struct *tsk = req->task;
2176 struct io_uring_task *tctx = tsk->io_uring;
2177 enum task_work_notify_mode notify;
2178 struct io_wq_work_node *node;
2179 unsigned long flags;
2182 WARN_ON_ONCE(!tctx);
2184 spin_lock_irqsave(&tctx->task_lock, flags);
2185 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2186 running = tctx->task_running;
2188 tctx->task_running = true;
2189 spin_unlock_irqrestore(&tctx->task_lock, flags);
2191 /* task_work already pending, we're done */
2196 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2197 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2198 * processing task_work. There's no reliable way to tell if TWA_RESUME
2201 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2202 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2203 wake_up_process(tsk);
2207 spin_lock_irqsave(&tctx->task_lock, flags);
2208 tctx->task_running = false;
2209 node = tctx->task_list.first;
2210 INIT_WQ_LIST(&tctx->task_list);
2211 spin_unlock_irqrestore(&tctx->task_lock, flags);
2214 req = container_of(node, struct io_kiocb, io_task_work.node);
2216 if (llist_add(&req->io_task_work.fallback_node,
2217 &req->ctx->fallback_llist))
2218 schedule_delayed_work(&req->ctx->fallback_work, 1);
2222 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2224 struct io_ring_ctx *ctx = req->ctx;
2226 /* not needed for normal modes, but SQPOLL depends on it */
2227 io_tw_lock(ctx, locked);
2228 io_req_complete_failed(req, req->result);
2231 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2233 struct io_ring_ctx *ctx = req->ctx;
2235 io_tw_lock(ctx, locked);
2236 /* req->task == current here, checking PF_EXITING is safe */
2237 if (likely(!(req->task->flags & PF_EXITING)))
2238 __io_queue_sqe(req);
2240 io_req_complete_failed(req, -EFAULT);
2243 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2246 req->io_task_work.func = io_req_task_cancel;
2247 io_req_task_work_add(req);
2250 static void io_req_task_queue(struct io_kiocb *req)
2252 req->io_task_work.func = io_req_task_submit;
2253 io_req_task_work_add(req);
2256 static void io_req_task_queue_reissue(struct io_kiocb *req)
2258 req->io_task_work.func = io_queue_async_work;
2259 io_req_task_work_add(req);
2262 static inline void io_queue_next(struct io_kiocb *req)
2264 struct io_kiocb *nxt = io_req_find_next(req);
2267 io_req_task_queue(nxt);
2270 static void io_free_req(struct io_kiocb *req)
2276 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2282 struct task_struct *task;
2287 static inline void io_init_req_batch(struct req_batch *rb)
2294 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2295 struct req_batch *rb)
2298 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2300 io_put_task(rb->task, rb->task_refs);
2303 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2304 struct io_submit_state *state)
2307 io_dismantle_req(req);
2309 if (req->task != rb->task) {
2311 io_put_task(rb->task, rb->task_refs);
2312 rb->task = req->task;
2318 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2319 state->reqs[state->free_reqs++] = req;
2321 list_add(&req->inflight_entry, &state->free_list);
2324 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2325 __must_hold(&ctx->uring_lock)
2327 struct io_submit_state *state = &ctx->submit_state;
2328 int i, nr = state->compl_nr;
2329 struct req_batch rb;
2331 spin_lock(&ctx->completion_lock);
2332 for (i = 0; i < nr; i++) {
2333 struct io_kiocb *req = state->compl_reqs[i];
2335 __io_cqring_fill_event(ctx, req->user_data, req->result,
2338 io_commit_cqring(ctx);
2339 spin_unlock(&ctx->completion_lock);
2340 io_cqring_ev_posted(ctx);
2342 io_init_req_batch(&rb);
2343 for (i = 0; i < nr; i++) {
2344 struct io_kiocb *req = state->compl_reqs[i];
2346 if (req_ref_put_and_test(req))
2347 io_req_free_batch(&rb, req, &ctx->submit_state);
2350 io_req_free_batch_finish(ctx, &rb);
2351 state->compl_nr = 0;
2355 * Drop reference to request, return next in chain (if there is one) if this
2356 * was the last reference to this request.
2358 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2360 struct io_kiocb *nxt = NULL;
2362 if (req_ref_put_and_test(req)) {
2363 nxt = io_req_find_next(req);
2369 static inline void io_put_req(struct io_kiocb *req)
2371 if (req_ref_put_and_test(req))
2375 static inline void io_put_req_deferred(struct io_kiocb *req)
2377 if (req_ref_put_and_test(req)) {
2378 req->io_task_work.func = io_free_req_work;
2379 io_req_task_work_add(req);
2383 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2385 /* See comment at the top of this file */
2387 return __io_cqring_events(ctx);
2390 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2392 struct io_rings *rings = ctx->rings;
2394 /* make sure SQ entry isn't read before tail */
2395 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2398 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2400 unsigned int cflags;
2402 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2403 cflags |= IORING_CQE_F_BUFFER;
2404 req->flags &= ~REQ_F_BUFFER_SELECTED;
2409 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2411 struct io_buffer *kbuf;
2413 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2415 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2416 return io_put_kbuf(req, kbuf);
2419 static inline bool io_run_task_work(void)
2421 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2422 __set_current_state(TASK_RUNNING);
2423 tracehook_notify_signal();
2431 * Find and free completed poll iocbs
2433 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2434 struct list_head *done)
2436 struct req_batch rb;
2437 struct io_kiocb *req;
2439 /* order with ->result store in io_complete_rw_iopoll() */
2442 io_init_req_batch(&rb);
2443 while (!list_empty(done)) {
2444 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2445 list_del(&req->inflight_entry);
2447 if (READ_ONCE(req->result) == -EAGAIN &&
2448 !(req->flags & REQ_F_DONT_REISSUE)) {
2449 req->iopoll_completed = 0;
2450 io_req_task_queue_reissue(req);
2454 __io_cqring_fill_event(ctx, req->user_data, req->result,
2455 io_put_rw_kbuf(req));
2458 if (req_ref_put_and_test(req))
2459 io_req_free_batch(&rb, req, &ctx->submit_state);
2462 io_commit_cqring(ctx);
2463 io_cqring_ev_posted_iopoll(ctx);
2464 io_req_free_batch_finish(ctx, &rb);
2467 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2470 struct io_kiocb *req, *tmp;
2475 * Only spin for completions if we don't have multiple devices hanging
2476 * off our complete list, and we're under the requested amount.
2478 spin = !ctx->poll_multi_queue && *nr_events < min;
2480 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2481 struct kiocb *kiocb = &req->rw.kiocb;
2485 * Move completed and retryable entries to our local lists.
2486 * If we find a request that requires polling, break out
2487 * and complete those lists first, if we have entries there.
2489 if (READ_ONCE(req->iopoll_completed)) {
2490 list_move_tail(&req->inflight_entry, &done);
2493 if (!list_empty(&done))
2496 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2497 if (unlikely(ret < 0))
2502 /* iopoll may have completed current req */
2503 if (READ_ONCE(req->iopoll_completed))
2504 list_move_tail(&req->inflight_entry, &done);
2507 if (!list_empty(&done))
2508 io_iopoll_complete(ctx, nr_events, &done);
2514 * We can't just wait for polled events to come to us, we have to actively
2515 * find and complete them.
2517 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2519 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2522 mutex_lock(&ctx->uring_lock);
2523 while (!list_empty(&ctx->iopoll_list)) {
2524 unsigned int nr_events = 0;
2526 io_do_iopoll(ctx, &nr_events, 0);
2528 /* let it sleep and repeat later if can't complete a request */
2532 * Ensure we allow local-to-the-cpu processing to take place,
2533 * in this case we need to ensure that we reap all events.
2534 * Also let task_work, etc. to progress by releasing the mutex
2536 if (need_resched()) {
2537 mutex_unlock(&ctx->uring_lock);
2539 mutex_lock(&ctx->uring_lock);
2542 mutex_unlock(&ctx->uring_lock);
2545 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2547 unsigned int nr_events = 0;
2551 * We disallow the app entering submit/complete with polling, but we
2552 * still need to lock the ring to prevent racing with polled issue
2553 * that got punted to a workqueue.
2555 mutex_lock(&ctx->uring_lock);
2557 * Don't enter poll loop if we already have events pending.
2558 * If we do, we can potentially be spinning for commands that
2559 * already triggered a CQE (eg in error).
2561 if (test_bit(0, &ctx->check_cq_overflow))
2562 __io_cqring_overflow_flush(ctx, false);
2563 if (io_cqring_events(ctx))
2567 * If a submit got punted to a workqueue, we can have the
2568 * application entering polling for a command before it gets
2569 * issued. That app will hold the uring_lock for the duration
2570 * of the poll right here, so we need to take a breather every
2571 * now and then to ensure that the issue has a chance to add
2572 * the poll to the issued list. Otherwise we can spin here
2573 * forever, while the workqueue is stuck trying to acquire the
2576 if (list_empty(&ctx->iopoll_list)) {
2577 u32 tail = ctx->cached_cq_tail;
2579 mutex_unlock(&ctx->uring_lock);
2581 mutex_lock(&ctx->uring_lock);
2583 /* some requests don't go through iopoll_list */
2584 if (tail != ctx->cached_cq_tail ||
2585 list_empty(&ctx->iopoll_list))
2588 ret = io_do_iopoll(ctx, &nr_events, min);
2589 } while (!ret && nr_events < min && !need_resched());
2591 mutex_unlock(&ctx->uring_lock);
2595 static void kiocb_end_write(struct io_kiocb *req)
2598 * Tell lockdep we inherited freeze protection from submission
2601 if (req->flags & REQ_F_ISREG) {
2602 struct super_block *sb = file_inode(req->file)->i_sb;
2604 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2610 static bool io_resubmit_prep(struct io_kiocb *req)
2612 struct io_async_rw *rw = req->async_data;
2615 return !io_req_prep_async(req);
2616 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2617 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2621 static bool io_rw_should_reissue(struct io_kiocb *req)
2623 umode_t mode = file_inode(req->file)->i_mode;
2624 struct io_ring_ctx *ctx = req->ctx;
2626 if (!S_ISBLK(mode) && !S_ISREG(mode))
2628 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2629 !(ctx->flags & IORING_SETUP_IOPOLL)))
2632 * If ref is dying, we might be running poll reap from the exit work.
2633 * Don't attempt to reissue from that path, just let it fail with
2636 if (percpu_ref_is_dying(&ctx->refs))
2639 * Play it safe and assume not safe to re-import and reissue if we're
2640 * not in the original thread group (or in task context).
2642 if (!same_thread_group(req->task, current) || !in_task())
2647 static bool io_resubmit_prep(struct io_kiocb *req)
2651 static bool io_rw_should_reissue(struct io_kiocb *req)
2657 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2659 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2660 kiocb_end_write(req);
2661 if (res != req->result) {
2662 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2663 io_rw_should_reissue(req)) {
2664 req->flags |= REQ_F_REISSUE;
2673 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2675 unsigned int cflags = io_put_rw_kbuf(req);
2676 long res = req->result;
2679 struct io_ring_ctx *ctx = req->ctx;
2680 struct io_submit_state *state = &ctx->submit_state;
2682 io_req_complete_state(req, res, cflags);
2683 state->compl_reqs[state->compl_nr++] = req;
2684 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2685 io_submit_flush_completions(ctx);
2687 io_req_complete_post(req, res, cflags);
2691 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2692 unsigned int issue_flags)
2694 if (__io_complete_rw_common(req, res))
2696 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2699 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2701 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2703 if (__io_complete_rw_common(req, res))
2706 req->io_task_work.func = io_req_task_complete;
2707 io_req_task_work_add(req);
2710 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2712 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2714 if (kiocb->ki_flags & IOCB_WRITE)
2715 kiocb_end_write(req);
2716 if (unlikely(res != req->result)) {
2717 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2718 io_resubmit_prep(req))) {
2720 req->flags |= REQ_F_DONT_REISSUE;
2724 WRITE_ONCE(req->result, res);
2725 /* order with io_iopoll_complete() checking ->result */
2727 WRITE_ONCE(req->iopoll_completed, 1);
2731 * After the iocb has been issued, it's safe to be found on the poll list.
2732 * Adding the kiocb to the list AFTER submission ensures that we don't
2733 * find it from a io_do_iopoll() thread before the issuer is done
2734 * accessing the kiocb cookie.
2736 static void io_iopoll_req_issued(struct io_kiocb *req)
2738 struct io_ring_ctx *ctx = req->ctx;
2739 const bool in_async = io_wq_current_is_worker();
2741 /* workqueue context doesn't hold uring_lock, grab it now */
2742 if (unlikely(in_async))
2743 mutex_lock(&ctx->uring_lock);
2746 * Track whether we have multiple files in our lists. This will impact
2747 * how we do polling eventually, not spinning if we're on potentially
2748 * different devices.
2750 if (list_empty(&ctx->iopoll_list)) {
2751 ctx->poll_multi_queue = false;
2752 } else if (!ctx->poll_multi_queue) {
2753 struct io_kiocb *list_req;
2754 unsigned int queue_num0, queue_num1;
2756 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2759 if (list_req->file != req->file) {
2760 ctx->poll_multi_queue = true;
2762 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2763 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2764 if (queue_num0 != queue_num1)
2765 ctx->poll_multi_queue = true;
2770 * For fast devices, IO may have already completed. If it has, add
2771 * it to the front so we find it first.
2773 if (READ_ONCE(req->iopoll_completed))
2774 list_add(&req->inflight_entry, &ctx->iopoll_list);
2776 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2778 if (unlikely(in_async)) {
2780 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2781 * in sq thread task context or in io worker task context. If
2782 * current task context is sq thread, we don't need to check
2783 * whether should wake up sq thread.
2785 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2786 wq_has_sleeper(&ctx->sq_data->wait))
2787 wake_up(&ctx->sq_data->wait);
2789 mutex_unlock(&ctx->uring_lock);
2793 static bool io_bdev_nowait(struct block_device *bdev)
2795 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2799 * If we tracked the file through the SCM inflight mechanism, we could support
2800 * any file. For now, just ensure that anything potentially problematic is done
2803 static bool __io_file_supports_nowait(struct file *file, int rw)
2805 umode_t mode = file_inode(file)->i_mode;
2807 if (S_ISBLK(mode)) {
2808 if (IS_ENABLED(CONFIG_BLOCK) &&
2809 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2815 if (S_ISREG(mode)) {
2816 if (IS_ENABLED(CONFIG_BLOCK) &&
2817 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2818 file->f_op != &io_uring_fops)
2823 /* any ->read/write should understand O_NONBLOCK */
2824 if (file->f_flags & O_NONBLOCK)
2827 if (!(file->f_mode & FMODE_NOWAIT))
2831 return file->f_op->read_iter != NULL;
2833 return file->f_op->write_iter != NULL;
2836 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2838 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2840 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2843 return __io_file_supports_nowait(req->file, rw);
2846 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2848 struct io_ring_ctx *ctx = req->ctx;
2849 struct kiocb *kiocb = &req->rw.kiocb;
2850 struct file *file = req->file;
2854 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2855 req->flags |= REQ_F_ISREG;
2857 kiocb->ki_pos = READ_ONCE(sqe->off);
2858 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2859 req->flags |= REQ_F_CUR_POS;
2860 kiocb->ki_pos = file->f_pos;
2862 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2863 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2864 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2868 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2869 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2870 req->flags |= REQ_F_NOWAIT;
2872 ioprio = READ_ONCE(sqe->ioprio);
2874 ret = ioprio_check_cap(ioprio);
2878 kiocb->ki_ioprio = ioprio;
2880 kiocb->ki_ioprio = get_current_ioprio();
2882 if (ctx->flags & IORING_SETUP_IOPOLL) {
2883 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2884 !kiocb->ki_filp->f_op->iopoll)
2887 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2888 kiocb->ki_complete = io_complete_rw_iopoll;
2889 req->iopoll_completed = 0;
2891 if (kiocb->ki_flags & IOCB_HIPRI)
2893 kiocb->ki_complete = io_complete_rw;
2896 if (req->opcode == IORING_OP_READ_FIXED ||
2897 req->opcode == IORING_OP_WRITE_FIXED) {
2899 io_req_set_rsrc_node(req);
2902 req->rw.addr = READ_ONCE(sqe->addr);
2903 req->rw.len = READ_ONCE(sqe->len);
2904 req->buf_index = READ_ONCE(sqe->buf_index);
2908 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2914 case -ERESTARTNOINTR:
2915 case -ERESTARTNOHAND:
2916 case -ERESTART_RESTARTBLOCK:
2918 * We can't just restart the syscall, since previously
2919 * submitted sqes may already be in progress. Just fail this
2925 kiocb->ki_complete(kiocb, ret, 0);
2929 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2930 unsigned int issue_flags)
2932 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2933 struct io_async_rw *io = req->async_data;
2934 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2936 /* add previously done IO, if any */
2937 if (io && io->bytes_done > 0) {
2939 ret = io->bytes_done;
2941 ret += io->bytes_done;
2944 if (req->flags & REQ_F_CUR_POS)
2945 req->file->f_pos = kiocb->ki_pos;
2946 if (ret >= 0 && check_reissue)
2947 __io_complete_rw(req, ret, 0, issue_flags);
2949 io_rw_done(kiocb, ret);
2951 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2952 req->flags &= ~REQ_F_REISSUE;
2953 if (io_resubmit_prep(req)) {
2954 io_req_task_queue_reissue(req);
2957 __io_req_complete(req, issue_flags, ret,
2958 io_put_rw_kbuf(req));
2963 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2964 struct io_mapped_ubuf *imu)
2966 size_t len = req->rw.len;
2967 u64 buf_end, buf_addr = req->rw.addr;
2970 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2972 /* not inside the mapped region */
2973 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2977 * May not be a start of buffer, set size appropriately
2978 * and advance us to the beginning.
2980 offset = buf_addr - imu->ubuf;
2981 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2985 * Don't use iov_iter_advance() here, as it's really slow for
2986 * using the latter parts of a big fixed buffer - it iterates
2987 * over each segment manually. We can cheat a bit here, because
2990 * 1) it's a BVEC iter, we set it up
2991 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2992 * first and last bvec
2994 * So just find our index, and adjust the iterator afterwards.
2995 * If the offset is within the first bvec (or the whole first
2996 * bvec, just use iov_iter_advance(). This makes it easier
2997 * since we can just skip the first segment, which may not
2998 * be PAGE_SIZE aligned.
3000 const struct bio_vec *bvec = imu->bvec;
3002 if (offset <= bvec->bv_len) {
3003 iov_iter_advance(iter, offset);
3005 unsigned long seg_skip;
3007 /* skip first vec */
3008 offset -= bvec->bv_len;
3009 seg_skip = 1 + (offset >> PAGE_SHIFT);
3011 iter->bvec = bvec + seg_skip;
3012 iter->nr_segs -= seg_skip;
3013 iter->count -= bvec->bv_len + offset;
3014 iter->iov_offset = offset & ~PAGE_MASK;
3021 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3023 struct io_ring_ctx *ctx = req->ctx;
3024 struct io_mapped_ubuf *imu = req->imu;
3025 u16 index, buf_index = req->buf_index;
3028 if (unlikely(buf_index >= ctx->nr_user_bufs))
3030 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3031 imu = READ_ONCE(ctx->user_bufs[index]);
3034 return __io_import_fixed(req, rw, iter, imu);
3037 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3040 mutex_unlock(&ctx->uring_lock);
3043 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3046 * "Normal" inline submissions always hold the uring_lock, since we
3047 * grab it from the system call. Same is true for the SQPOLL offload.
3048 * The only exception is when we've detached the request and issue it
3049 * from an async worker thread, grab the lock for that case.
3052 mutex_lock(&ctx->uring_lock);
3055 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3056 int bgid, struct io_buffer *kbuf,
3059 struct io_buffer *head;
3061 if (req->flags & REQ_F_BUFFER_SELECTED)
3064 io_ring_submit_lock(req->ctx, needs_lock);
3066 lockdep_assert_held(&req->ctx->uring_lock);
3068 head = xa_load(&req->ctx->io_buffers, bgid);
3070 if (!list_empty(&head->list)) {
3071 kbuf = list_last_entry(&head->list, struct io_buffer,
3073 list_del(&kbuf->list);
3076 xa_erase(&req->ctx->io_buffers, bgid);
3078 if (*len > kbuf->len)
3081 kbuf = ERR_PTR(-ENOBUFS);
3084 io_ring_submit_unlock(req->ctx, needs_lock);
3089 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3092 struct io_buffer *kbuf;
3095 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3096 bgid = req->buf_index;
3097 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3100 req->rw.addr = (u64) (unsigned long) kbuf;
3101 req->flags |= REQ_F_BUFFER_SELECTED;
3102 return u64_to_user_ptr(kbuf->addr);
3105 #ifdef CONFIG_COMPAT
3106 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3109 struct compat_iovec __user *uiov;
3110 compat_ssize_t clen;
3114 uiov = u64_to_user_ptr(req->rw.addr);
3115 if (!access_ok(uiov, sizeof(*uiov)))
3117 if (__get_user(clen, &uiov->iov_len))
3123 buf = io_rw_buffer_select(req, &len, needs_lock);
3125 return PTR_ERR(buf);
3126 iov[0].iov_base = buf;
3127 iov[0].iov_len = (compat_size_t) len;
3132 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3135 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3139 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3142 len = iov[0].iov_len;
3145 buf = io_rw_buffer_select(req, &len, needs_lock);
3147 return PTR_ERR(buf);
3148 iov[0].iov_base = buf;
3149 iov[0].iov_len = len;
3153 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3156 if (req->flags & REQ_F_BUFFER_SELECTED) {
3157 struct io_buffer *kbuf;
3159 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3160 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3161 iov[0].iov_len = kbuf->len;
3164 if (req->rw.len != 1)
3167 #ifdef CONFIG_COMPAT
3168 if (req->ctx->compat)
3169 return io_compat_import(req, iov, needs_lock);
3172 return __io_iov_buffer_select(req, iov, needs_lock);
3175 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3176 struct iov_iter *iter, bool needs_lock)
3178 void __user *buf = u64_to_user_ptr(req->rw.addr);
3179 size_t sqe_len = req->rw.len;
3180 u8 opcode = req->opcode;
3183 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3185 return io_import_fixed(req, rw, iter);
3188 /* buffer index only valid with fixed read/write, or buffer select */
3189 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3192 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3193 if (req->flags & REQ_F_BUFFER_SELECT) {
3194 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3196 return PTR_ERR(buf);
3197 req->rw.len = sqe_len;
3200 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3205 if (req->flags & REQ_F_BUFFER_SELECT) {
3206 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3208 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3213 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3217 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3219 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3223 * For files that don't have ->read_iter() and ->write_iter(), handle them
3224 * by looping over ->read() or ->write() manually.
3226 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3228 struct kiocb *kiocb = &req->rw.kiocb;
3229 struct file *file = req->file;
3233 * Don't support polled IO through this interface, and we can't
3234 * support non-blocking either. For the latter, this just causes
3235 * the kiocb to be handled from an async context.
3237 if (kiocb->ki_flags & IOCB_HIPRI)
3239 if (kiocb->ki_flags & IOCB_NOWAIT)
3242 while (iov_iter_count(iter)) {
3246 if (!iov_iter_is_bvec(iter)) {
3247 iovec = iov_iter_iovec(iter);
3249 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3250 iovec.iov_len = req->rw.len;
3254 nr = file->f_op->read(file, iovec.iov_base,
3255 iovec.iov_len, io_kiocb_ppos(kiocb));
3257 nr = file->f_op->write(file, iovec.iov_base,
3258 iovec.iov_len, io_kiocb_ppos(kiocb));
3267 if (nr != iovec.iov_len)
3271 iov_iter_advance(iter, nr);
3277 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3278 const struct iovec *fast_iov, struct iov_iter *iter)
3280 struct io_async_rw *rw = req->async_data;
3282 memcpy(&rw->iter, iter, sizeof(*iter));
3283 rw->free_iovec = iovec;
3285 /* can only be fixed buffers, no need to do anything */
3286 if (iov_iter_is_bvec(iter))
3289 unsigned iov_off = 0;
3291 rw->iter.iov = rw->fast_iov;
3292 if (iter->iov != fast_iov) {
3293 iov_off = iter->iov - fast_iov;
3294 rw->iter.iov += iov_off;
3296 if (rw->fast_iov != fast_iov)
3297 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3298 sizeof(struct iovec) * iter->nr_segs);
3300 req->flags |= REQ_F_NEED_CLEANUP;
3304 static inline int io_alloc_async_data(struct io_kiocb *req)
3306 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3307 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3308 return req->async_data == NULL;
3311 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3312 const struct iovec *fast_iov,
3313 struct iov_iter *iter, bool force)
3315 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3317 if (!req->async_data) {
3318 if (io_alloc_async_data(req)) {
3323 io_req_map_rw(req, iovec, fast_iov, iter);
3328 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3330 struct io_async_rw *iorw = req->async_data;
3331 struct iovec *iov = iorw->fast_iov;
3334 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3335 if (unlikely(ret < 0))
3338 iorw->bytes_done = 0;
3339 iorw->free_iovec = iov;
3341 req->flags |= REQ_F_NEED_CLEANUP;
3345 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3347 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3349 return io_prep_rw(req, sqe);
3353 * This is our waitqueue callback handler, registered through lock_page_async()
3354 * when we initially tried to do the IO with the iocb armed our waitqueue.
3355 * This gets called when the page is unlocked, and we generally expect that to
3356 * happen when the page IO is completed and the page is now uptodate. This will
3357 * queue a task_work based retry of the operation, attempting to copy the data
3358 * again. If the latter fails because the page was NOT uptodate, then we will
3359 * do a thread based blocking retry of the operation. That's the unexpected
3362 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3363 int sync, void *arg)
3365 struct wait_page_queue *wpq;
3366 struct io_kiocb *req = wait->private;
3367 struct wait_page_key *key = arg;
3369 wpq = container_of(wait, struct wait_page_queue, wait);
3371 if (!wake_page_match(wpq, key))
3374 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3375 list_del_init(&wait->entry);
3376 io_req_task_queue(req);
3381 * This controls whether a given IO request should be armed for async page
3382 * based retry. If we return false here, the request is handed to the async
3383 * worker threads for retry. If we're doing buffered reads on a regular file,
3384 * we prepare a private wait_page_queue entry and retry the operation. This
3385 * will either succeed because the page is now uptodate and unlocked, or it
3386 * will register a callback when the page is unlocked at IO completion. Through
3387 * that callback, io_uring uses task_work to setup a retry of the operation.
3388 * That retry will attempt the buffered read again. The retry will generally
3389 * succeed, or in rare cases where it fails, we then fall back to using the
3390 * async worker threads for a blocking retry.
3392 static bool io_rw_should_retry(struct io_kiocb *req)
3394 struct io_async_rw *rw = req->async_data;
3395 struct wait_page_queue *wait = &rw->wpq;
3396 struct kiocb *kiocb = &req->rw.kiocb;
3398 /* never retry for NOWAIT, we just complete with -EAGAIN */
3399 if (req->flags & REQ_F_NOWAIT)
3402 /* Only for buffered IO */
3403 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3407 * just use poll if we can, and don't attempt if the fs doesn't
3408 * support callback based unlocks
3410 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3413 wait->wait.func = io_async_buf_func;
3414 wait->wait.private = req;
3415 wait->wait.flags = 0;
3416 INIT_LIST_HEAD(&wait->wait.entry);
3417 kiocb->ki_flags |= IOCB_WAITQ;
3418 kiocb->ki_flags &= ~IOCB_NOWAIT;
3419 kiocb->ki_waitq = wait;
3423 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3425 if (req->file->f_op->read_iter)
3426 return call_read_iter(req->file, &req->rw.kiocb, iter);
3427 else if (req->file->f_op->read)
3428 return loop_rw_iter(READ, req, iter);
3433 static bool need_read_all(struct io_kiocb *req)
3435 return req->flags & REQ_F_ISREG ||
3436 S_ISBLK(file_inode(req->file)->i_mode);
3439 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3441 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3442 struct kiocb *kiocb = &req->rw.kiocb;
3443 struct iov_iter __iter, *iter = &__iter;
3444 struct io_async_rw *rw = req->async_data;
3445 ssize_t io_size, ret, ret2;
3446 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3452 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3456 io_size = iov_iter_count(iter);
3457 req->result = io_size;
3459 /* Ensure we clear previously set non-block flag */
3460 if (!force_nonblock)
3461 kiocb->ki_flags &= ~IOCB_NOWAIT;
3463 kiocb->ki_flags |= IOCB_NOWAIT;
3465 /* If the file doesn't support async, just async punt */
3466 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3467 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3468 return ret ?: -EAGAIN;
3471 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3472 if (unlikely(ret)) {
3477 ret = io_iter_do_read(req, iter);
3479 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3480 req->flags &= ~REQ_F_REISSUE;
3481 /* IOPOLL retry should happen for io-wq threads */
3482 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3484 /* no retry on NONBLOCK nor RWF_NOWAIT */
3485 if (req->flags & REQ_F_NOWAIT)
3487 /* some cases will consume bytes even on error returns */
3488 iov_iter_reexpand(iter, iter->count + iter->truncated);
3489 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3491 } else if (ret == -EIOCBQUEUED) {
3493 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3494 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3495 /* read all, failed, already did sync or don't want to retry */
3499 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3504 rw = req->async_data;
3505 /* now use our persistent iterator, if we aren't already */
3510 rw->bytes_done += ret;
3511 /* if we can retry, do so with the callbacks armed */
3512 if (!io_rw_should_retry(req)) {
3513 kiocb->ki_flags &= ~IOCB_WAITQ;
3518 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3519 * we get -EIOCBQUEUED, then we'll get a notification when the
3520 * desired page gets unlocked. We can also get a partial read
3521 * here, and if we do, then just retry at the new offset.
3523 ret = io_iter_do_read(req, iter);
3524 if (ret == -EIOCBQUEUED)
3526 /* we got some bytes, but not all. retry. */
3527 kiocb->ki_flags &= ~IOCB_WAITQ;
3528 } while (ret > 0 && ret < io_size);
3530 kiocb_done(kiocb, ret, issue_flags);
3532 /* it's faster to check here then delegate to kfree */
3538 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3540 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3542 return io_prep_rw(req, sqe);
3545 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3547 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3548 struct kiocb *kiocb = &req->rw.kiocb;
3549 struct iov_iter __iter, *iter = &__iter;
3550 struct io_async_rw *rw = req->async_data;
3551 ssize_t ret, ret2, io_size;
3552 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3558 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3562 io_size = iov_iter_count(iter);
3563 req->result = io_size;
3565 /* Ensure we clear previously set non-block flag */
3566 if (!force_nonblock)
3567 kiocb->ki_flags &= ~IOCB_NOWAIT;
3569 kiocb->ki_flags |= IOCB_NOWAIT;
3571 /* If the file doesn't support async, just async punt */
3572 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3575 /* file path doesn't support NOWAIT for non-direct_IO */
3576 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3577 (req->flags & REQ_F_ISREG))
3580 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3585 * Open-code file_start_write here to grab freeze protection,
3586 * which will be released by another thread in
3587 * io_complete_rw(). Fool lockdep by telling it the lock got
3588 * released so that it doesn't complain about the held lock when
3589 * we return to userspace.
3591 if (req->flags & REQ_F_ISREG) {
3592 sb_start_write(file_inode(req->file)->i_sb);
3593 __sb_writers_release(file_inode(req->file)->i_sb,
3596 kiocb->ki_flags |= IOCB_WRITE;
3598 if (req->file->f_op->write_iter)
3599 ret2 = call_write_iter(req->file, kiocb, iter);
3600 else if (req->file->f_op->write)
3601 ret2 = loop_rw_iter(WRITE, req, iter);
3605 if (req->flags & REQ_F_REISSUE) {
3606 req->flags &= ~REQ_F_REISSUE;
3611 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3612 * retry them without IOCB_NOWAIT.
3614 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3616 /* no retry on NONBLOCK nor RWF_NOWAIT */
3617 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3619 if (!force_nonblock || ret2 != -EAGAIN) {
3620 /* IOPOLL retry should happen for io-wq threads */
3621 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3624 kiocb_done(kiocb, ret2, issue_flags);
3627 /* some cases will consume bytes even on error returns */
3628 iov_iter_reexpand(iter, iter->count + iter->truncated);
3629 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3630 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3631 return ret ?: -EAGAIN;
3634 /* it's reportedly faster than delegating the null check to kfree() */
3640 static int io_renameat_prep(struct io_kiocb *req,
3641 const struct io_uring_sqe *sqe)
3643 struct io_rename *ren = &req->rename;
3644 const char __user *oldf, *newf;
3646 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3648 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3650 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3653 ren->old_dfd = READ_ONCE(sqe->fd);
3654 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3655 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3656 ren->new_dfd = READ_ONCE(sqe->len);
3657 ren->flags = READ_ONCE(sqe->rename_flags);
3659 ren->oldpath = getname(oldf);
3660 if (IS_ERR(ren->oldpath))
3661 return PTR_ERR(ren->oldpath);
3663 ren->newpath = getname(newf);
3664 if (IS_ERR(ren->newpath)) {
3665 putname(ren->oldpath);
3666 return PTR_ERR(ren->newpath);
3669 req->flags |= REQ_F_NEED_CLEANUP;
3673 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3675 struct io_rename *ren = &req->rename;
3678 if (issue_flags & IO_URING_F_NONBLOCK)
3681 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3682 ren->newpath, ren->flags);
3684 req->flags &= ~REQ_F_NEED_CLEANUP;
3687 io_req_complete(req, ret);
3691 static int io_unlinkat_prep(struct io_kiocb *req,
3692 const struct io_uring_sqe *sqe)
3694 struct io_unlink *un = &req->unlink;
3695 const char __user *fname;
3697 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3699 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3702 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3705 un->dfd = READ_ONCE(sqe->fd);
3707 un->flags = READ_ONCE(sqe->unlink_flags);
3708 if (un->flags & ~AT_REMOVEDIR)
3711 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3712 un->filename = getname(fname);
3713 if (IS_ERR(un->filename))
3714 return PTR_ERR(un->filename);
3716 req->flags |= REQ_F_NEED_CLEANUP;
3720 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3722 struct io_unlink *un = &req->unlink;
3725 if (issue_flags & IO_URING_F_NONBLOCK)
3728 if (un->flags & AT_REMOVEDIR)
3729 ret = do_rmdir(un->dfd, un->filename);
3731 ret = do_unlinkat(un->dfd, un->filename);
3733 req->flags &= ~REQ_F_NEED_CLEANUP;
3736 io_req_complete(req, ret);
3740 static int io_mkdirat_prep(struct io_kiocb *req,
3741 const struct io_uring_sqe *sqe)
3743 struct io_mkdir *mkd = &req->mkdir;
3744 const char __user *fname;
3746 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3748 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3751 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3754 mkd->dfd = READ_ONCE(sqe->fd);
3755 mkd->mode = READ_ONCE(sqe->len);
3757 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3758 mkd->filename = getname(fname);
3759 if (IS_ERR(mkd->filename))
3760 return PTR_ERR(mkd->filename);
3762 req->flags |= REQ_F_NEED_CLEANUP;
3766 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3768 struct io_mkdir *mkd = &req->mkdir;
3771 if (issue_flags & IO_URING_F_NONBLOCK)
3774 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3776 req->flags &= ~REQ_F_NEED_CLEANUP;
3779 io_req_complete(req, ret);
3783 static int io_symlinkat_prep(struct io_kiocb *req,
3784 const struct io_uring_sqe *sqe)
3786 struct io_symlink *sl = &req->symlink;
3787 const char __user *oldpath, *newpath;
3789 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3791 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3794 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3797 sl->new_dfd = READ_ONCE(sqe->fd);
3798 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3799 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3801 sl->oldpath = getname(oldpath);
3802 if (IS_ERR(sl->oldpath))
3803 return PTR_ERR(sl->oldpath);
3805 sl->newpath = getname(newpath);
3806 if (IS_ERR(sl->newpath)) {
3807 putname(sl->oldpath);
3808 return PTR_ERR(sl->newpath);
3811 req->flags |= REQ_F_NEED_CLEANUP;
3815 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3817 struct io_symlink *sl = &req->symlink;
3820 if (issue_flags & IO_URING_F_NONBLOCK)
3823 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3825 req->flags &= ~REQ_F_NEED_CLEANUP;
3828 io_req_complete(req, ret);
3832 static int io_linkat_prep(struct io_kiocb *req,
3833 const struct io_uring_sqe *sqe)
3835 struct io_hardlink *lnk = &req->hardlink;
3836 const char __user *oldf, *newf;
3838 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3840 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3842 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3845 lnk->old_dfd = READ_ONCE(sqe->fd);
3846 lnk->new_dfd = READ_ONCE(sqe->len);
3847 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3848 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3849 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3851 lnk->oldpath = getname(oldf);
3852 if (IS_ERR(lnk->oldpath))
3853 return PTR_ERR(lnk->oldpath);
3855 lnk->newpath = getname(newf);
3856 if (IS_ERR(lnk->newpath)) {
3857 putname(lnk->oldpath);
3858 return PTR_ERR(lnk->newpath);
3861 req->flags |= REQ_F_NEED_CLEANUP;
3865 static int io_linkat(struct io_kiocb *req, int issue_flags)
3867 struct io_hardlink *lnk = &req->hardlink;
3870 if (issue_flags & IO_URING_F_NONBLOCK)
3873 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3874 lnk->newpath, lnk->flags);
3876 req->flags &= ~REQ_F_NEED_CLEANUP;
3879 io_req_complete(req, ret);
3883 static int io_shutdown_prep(struct io_kiocb *req,
3884 const struct io_uring_sqe *sqe)
3886 #if defined(CONFIG_NET)
3887 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3889 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3890 sqe->buf_index || sqe->splice_fd_in))
3893 req->shutdown.how = READ_ONCE(sqe->len);
3900 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3902 #if defined(CONFIG_NET)
3903 struct socket *sock;
3906 if (issue_flags & IO_URING_F_NONBLOCK)
3909 sock = sock_from_file(req->file);
3910 if (unlikely(!sock))
3913 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3916 io_req_complete(req, ret);
3923 static int __io_splice_prep(struct io_kiocb *req,
3924 const struct io_uring_sqe *sqe)
3926 struct io_splice *sp = &req->splice;
3927 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3929 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3933 sp->len = READ_ONCE(sqe->len);
3934 sp->flags = READ_ONCE(sqe->splice_flags);
3936 if (unlikely(sp->flags & ~valid_flags))
3939 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3940 (sp->flags & SPLICE_F_FD_IN_FIXED));
3943 req->flags |= REQ_F_NEED_CLEANUP;
3947 static int io_tee_prep(struct io_kiocb *req,
3948 const struct io_uring_sqe *sqe)
3950 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3952 return __io_splice_prep(req, sqe);
3955 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3957 struct io_splice *sp = &req->splice;
3958 struct file *in = sp->file_in;
3959 struct file *out = sp->file_out;
3960 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3963 if (issue_flags & IO_URING_F_NONBLOCK)
3966 ret = do_tee(in, out, sp->len, flags);
3968 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3970 req->flags &= ~REQ_F_NEED_CLEANUP;
3974 io_req_complete(req, ret);
3978 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3980 struct io_splice *sp = &req->splice;
3982 sp->off_in = READ_ONCE(sqe->splice_off_in);
3983 sp->off_out = READ_ONCE(sqe->off);
3984 return __io_splice_prep(req, sqe);
3987 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3989 struct io_splice *sp = &req->splice;
3990 struct file *in = sp->file_in;
3991 struct file *out = sp->file_out;
3992 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3993 loff_t *poff_in, *poff_out;
3996 if (issue_flags & IO_URING_F_NONBLOCK)
3999 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4000 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4003 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4005 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4007 req->flags &= ~REQ_F_NEED_CLEANUP;
4011 io_req_complete(req, ret);
4016 * IORING_OP_NOP just posts a completion event, nothing else.
4018 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4020 struct io_ring_ctx *ctx = req->ctx;
4022 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4025 __io_req_complete(req, issue_flags, 0, 0);
4029 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4031 struct io_ring_ctx *ctx = req->ctx;
4036 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4038 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4042 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4043 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4046 req->sync.off = READ_ONCE(sqe->off);
4047 req->sync.len = READ_ONCE(sqe->len);
4051 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4053 loff_t end = req->sync.off + req->sync.len;
4056 /* fsync always requires a blocking context */
4057 if (issue_flags & IO_URING_F_NONBLOCK)
4060 ret = vfs_fsync_range(req->file, req->sync.off,
4061 end > 0 ? end : LLONG_MAX,
4062 req->sync.flags & IORING_FSYNC_DATASYNC);
4065 io_req_complete(req, ret);
4069 static int io_fallocate_prep(struct io_kiocb *req,
4070 const struct io_uring_sqe *sqe)
4072 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4075 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4078 req->sync.off = READ_ONCE(sqe->off);
4079 req->sync.len = READ_ONCE(sqe->addr);
4080 req->sync.mode = READ_ONCE(sqe->len);
4084 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4088 /* fallocate always requiring blocking context */
4089 if (issue_flags & IO_URING_F_NONBLOCK)
4091 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4095 io_req_complete(req, ret);
4099 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4101 const char __user *fname;
4104 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4106 if (unlikely(sqe->ioprio || sqe->buf_index))
4108 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4111 /* open.how should be already initialised */
4112 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4113 req->open.how.flags |= O_LARGEFILE;
4115 req->open.dfd = READ_ONCE(sqe->fd);
4116 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4117 req->open.filename = getname(fname);
4118 if (IS_ERR(req->open.filename)) {
4119 ret = PTR_ERR(req->open.filename);
4120 req->open.filename = NULL;
4124 req->open.file_slot = READ_ONCE(sqe->file_index);
4125 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4128 req->open.nofile = rlimit(RLIMIT_NOFILE);
4129 req->flags |= REQ_F_NEED_CLEANUP;
4133 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4135 u64 mode = READ_ONCE(sqe->len);
4136 u64 flags = READ_ONCE(sqe->open_flags);
4138 req->open.how = build_open_how(flags, mode);
4139 return __io_openat_prep(req, sqe);
4142 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4144 struct open_how __user *how;
4148 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4149 len = READ_ONCE(sqe->len);
4150 if (len < OPEN_HOW_SIZE_VER0)
4153 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4158 return __io_openat_prep(req, sqe);
4161 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4163 struct open_flags op;
4165 bool resolve_nonblock, nonblock_set;
4166 bool fixed = !!req->open.file_slot;
4169 ret = build_open_flags(&req->open.how, &op);
4172 nonblock_set = op.open_flag & O_NONBLOCK;
4173 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4174 if (issue_flags & IO_URING_F_NONBLOCK) {
4176 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4177 * it'll always -EAGAIN
4179 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4181 op.lookup_flags |= LOOKUP_CACHED;
4182 op.open_flag |= O_NONBLOCK;
4186 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4191 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4194 * We could hang on to this 'fd' on retrying, but seems like
4195 * marginal gain for something that is now known to be a slower
4196 * path. So just put it, and we'll get a new one when we retry.
4201 ret = PTR_ERR(file);
4202 /* only retry if RESOLVE_CACHED wasn't already set by application */
4203 if (ret == -EAGAIN &&
4204 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4209 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4210 file->f_flags &= ~O_NONBLOCK;
4211 fsnotify_open(file);
4214 fd_install(ret, file);
4216 ret = io_install_fixed_file(req, file, issue_flags,
4217 req->open.file_slot - 1);
4219 putname(req->open.filename);
4220 req->flags &= ~REQ_F_NEED_CLEANUP;
4223 __io_req_complete(req, issue_flags, ret, 0);
4227 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4229 return io_openat2(req, issue_flags);
4232 static int io_remove_buffers_prep(struct io_kiocb *req,
4233 const struct io_uring_sqe *sqe)
4235 struct io_provide_buf *p = &req->pbuf;
4238 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4242 tmp = READ_ONCE(sqe->fd);
4243 if (!tmp || tmp > USHRT_MAX)
4246 memset(p, 0, sizeof(*p));
4248 p->bgid = READ_ONCE(sqe->buf_group);
4252 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4253 int bgid, unsigned nbufs)
4257 /* shouldn't happen */
4261 /* the head kbuf is the list itself */
4262 while (!list_empty(&buf->list)) {
4263 struct io_buffer *nxt;
4265 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4266 list_del(&nxt->list);
4273 xa_erase(&ctx->io_buffers, bgid);
4278 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4280 struct io_provide_buf *p = &req->pbuf;
4281 struct io_ring_ctx *ctx = req->ctx;
4282 struct io_buffer *head;
4284 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4286 io_ring_submit_lock(ctx, !force_nonblock);
4288 lockdep_assert_held(&ctx->uring_lock);
4291 head = xa_load(&ctx->io_buffers, p->bgid);
4293 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4297 /* complete before unlock, IOPOLL may need the lock */
4298 __io_req_complete(req, issue_flags, ret, 0);
4299 io_ring_submit_unlock(ctx, !force_nonblock);
4303 static int io_provide_buffers_prep(struct io_kiocb *req,
4304 const struct io_uring_sqe *sqe)
4306 unsigned long size, tmp_check;
4307 struct io_provide_buf *p = &req->pbuf;
4310 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4313 tmp = READ_ONCE(sqe->fd);
4314 if (!tmp || tmp > USHRT_MAX)
4317 p->addr = READ_ONCE(sqe->addr);
4318 p->len = READ_ONCE(sqe->len);
4320 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4323 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4326 size = (unsigned long)p->len * p->nbufs;
4327 if (!access_ok(u64_to_user_ptr(p->addr), size))
4330 p->bgid = READ_ONCE(sqe->buf_group);
4331 tmp = READ_ONCE(sqe->off);
4332 if (tmp > USHRT_MAX)
4338 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4340 struct io_buffer *buf;
4341 u64 addr = pbuf->addr;
4342 int i, bid = pbuf->bid;
4344 for (i = 0; i < pbuf->nbufs; i++) {
4345 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4350 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4355 INIT_LIST_HEAD(&buf->list);
4358 list_add_tail(&buf->list, &(*head)->list);
4362 return i ? i : -ENOMEM;
4365 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4367 struct io_provide_buf *p = &req->pbuf;
4368 struct io_ring_ctx *ctx = req->ctx;
4369 struct io_buffer *head, *list;
4371 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4373 io_ring_submit_lock(ctx, !force_nonblock);
4375 lockdep_assert_held(&ctx->uring_lock);
4377 list = head = xa_load(&ctx->io_buffers, p->bgid);
4379 ret = io_add_buffers(p, &head);
4380 if (ret >= 0 && !list) {
4381 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4383 __io_remove_buffers(ctx, head, p->bgid, -1U);
4387 /* complete before unlock, IOPOLL may need the lock */
4388 __io_req_complete(req, issue_flags, ret, 0);
4389 io_ring_submit_unlock(ctx, !force_nonblock);
4393 static int io_epoll_ctl_prep(struct io_kiocb *req,
4394 const struct io_uring_sqe *sqe)
4396 #if defined(CONFIG_EPOLL)
4397 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4399 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4402 req->epoll.epfd = READ_ONCE(sqe->fd);
4403 req->epoll.op = READ_ONCE(sqe->len);
4404 req->epoll.fd = READ_ONCE(sqe->off);
4406 if (ep_op_has_event(req->epoll.op)) {
4407 struct epoll_event __user *ev;
4409 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4410 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4420 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4422 #if defined(CONFIG_EPOLL)
4423 struct io_epoll *ie = &req->epoll;
4425 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4427 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4428 if (force_nonblock && ret == -EAGAIN)
4433 __io_req_complete(req, issue_flags, ret, 0);
4440 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4442 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4443 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4445 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4448 req->madvise.addr = READ_ONCE(sqe->addr);
4449 req->madvise.len = READ_ONCE(sqe->len);
4450 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4457 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4459 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4460 struct io_madvise *ma = &req->madvise;
4463 if (issue_flags & IO_URING_F_NONBLOCK)
4466 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4469 io_req_complete(req, ret);
4476 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4478 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4480 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4483 req->fadvise.offset = READ_ONCE(sqe->off);
4484 req->fadvise.len = READ_ONCE(sqe->len);
4485 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4489 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4491 struct io_fadvise *fa = &req->fadvise;
4494 if (issue_flags & IO_URING_F_NONBLOCK) {
4495 switch (fa->advice) {
4496 case POSIX_FADV_NORMAL:
4497 case POSIX_FADV_RANDOM:
4498 case POSIX_FADV_SEQUENTIAL:
4505 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4508 __io_req_complete(req, issue_flags, ret, 0);
4512 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4514 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4516 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4518 if (req->flags & REQ_F_FIXED_FILE)
4521 req->statx.dfd = READ_ONCE(sqe->fd);
4522 req->statx.mask = READ_ONCE(sqe->len);
4523 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4524 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4525 req->statx.flags = READ_ONCE(sqe->statx_flags);
4530 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4532 struct io_statx *ctx = &req->statx;
4535 if (issue_flags & IO_URING_F_NONBLOCK)
4538 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4543 io_req_complete(req, ret);
4547 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4549 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4551 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4552 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4554 if (req->flags & REQ_F_FIXED_FILE)
4557 req->close.fd = READ_ONCE(sqe->fd);
4561 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4563 struct files_struct *files = current->files;
4564 struct io_close *close = &req->close;
4565 struct fdtable *fdt;
4566 struct file *file = NULL;
4569 spin_lock(&files->file_lock);
4570 fdt = files_fdtable(files);
4571 if (close->fd >= fdt->max_fds) {
4572 spin_unlock(&files->file_lock);
4575 file = fdt->fd[close->fd];
4576 if (!file || file->f_op == &io_uring_fops) {
4577 spin_unlock(&files->file_lock);
4582 /* if the file has a flush method, be safe and punt to async */
4583 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4584 spin_unlock(&files->file_lock);
4588 ret = __close_fd_get_file(close->fd, &file);
4589 spin_unlock(&files->file_lock);
4596 /* No ->flush() or already async, safely close from here */
4597 ret = filp_close(file, current->files);
4603 __io_req_complete(req, issue_flags, ret, 0);
4607 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4609 struct io_ring_ctx *ctx = req->ctx;
4611 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4613 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4617 req->sync.off = READ_ONCE(sqe->off);
4618 req->sync.len = READ_ONCE(sqe->len);
4619 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4623 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4627 /* sync_file_range always requires a blocking context */
4628 if (issue_flags & IO_URING_F_NONBLOCK)
4631 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4635 io_req_complete(req, ret);
4639 #if defined(CONFIG_NET)
4640 static int io_setup_async_msg(struct io_kiocb *req,
4641 struct io_async_msghdr *kmsg)
4643 struct io_async_msghdr *async_msg = req->async_data;
4647 if (io_alloc_async_data(req)) {
4648 kfree(kmsg->free_iov);
4651 async_msg = req->async_data;
4652 req->flags |= REQ_F_NEED_CLEANUP;
4653 memcpy(async_msg, kmsg, sizeof(*kmsg));
4654 async_msg->msg.msg_name = &async_msg->addr;
4655 /* if were using fast_iov, set it to the new one */
4656 if (!async_msg->free_iov)
4657 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4662 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4663 struct io_async_msghdr *iomsg)
4665 iomsg->msg.msg_name = &iomsg->addr;
4666 iomsg->free_iov = iomsg->fast_iov;
4667 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4668 req->sr_msg.msg_flags, &iomsg->free_iov);
4671 static int io_sendmsg_prep_async(struct io_kiocb *req)
4675 ret = io_sendmsg_copy_hdr(req, req->async_data);
4677 req->flags |= REQ_F_NEED_CLEANUP;
4681 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4683 struct io_sr_msg *sr = &req->sr_msg;
4685 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4688 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4689 sr->len = READ_ONCE(sqe->len);
4690 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4691 if (sr->msg_flags & MSG_DONTWAIT)
4692 req->flags |= REQ_F_NOWAIT;
4694 #ifdef CONFIG_COMPAT
4695 if (req->ctx->compat)
4696 sr->msg_flags |= MSG_CMSG_COMPAT;
4701 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4703 struct io_async_msghdr iomsg, *kmsg;
4704 struct socket *sock;
4709 sock = sock_from_file(req->file);
4710 if (unlikely(!sock))
4713 kmsg = req->async_data;
4715 ret = io_sendmsg_copy_hdr(req, &iomsg);
4721 flags = req->sr_msg.msg_flags;
4722 if (issue_flags & IO_URING_F_NONBLOCK)
4723 flags |= MSG_DONTWAIT;
4724 if (flags & MSG_WAITALL)
4725 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4727 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4728 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4729 return io_setup_async_msg(req, kmsg);
4730 if (ret == -ERESTARTSYS)
4733 /* fast path, check for non-NULL to avoid function call */
4735 kfree(kmsg->free_iov);
4736 req->flags &= ~REQ_F_NEED_CLEANUP;
4739 __io_req_complete(req, issue_flags, ret, 0);
4743 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4745 struct io_sr_msg *sr = &req->sr_msg;
4748 struct socket *sock;
4753 sock = sock_from_file(req->file);
4754 if (unlikely(!sock))
4757 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4761 msg.msg_name = NULL;
4762 msg.msg_control = NULL;
4763 msg.msg_controllen = 0;
4764 msg.msg_namelen = 0;
4766 flags = req->sr_msg.msg_flags;
4767 if (issue_flags & IO_URING_F_NONBLOCK)
4768 flags |= MSG_DONTWAIT;
4769 if (flags & MSG_WAITALL)
4770 min_ret = iov_iter_count(&msg.msg_iter);
4772 msg.msg_flags = flags;
4773 ret = sock_sendmsg(sock, &msg);
4774 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4776 if (ret == -ERESTARTSYS)
4781 __io_req_complete(req, issue_flags, ret, 0);
4785 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4786 struct io_async_msghdr *iomsg)
4788 struct io_sr_msg *sr = &req->sr_msg;
4789 struct iovec __user *uiov;
4793 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4794 &iomsg->uaddr, &uiov, &iov_len);
4798 if (req->flags & REQ_F_BUFFER_SELECT) {
4801 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4803 sr->len = iomsg->fast_iov[0].iov_len;
4804 iomsg->free_iov = NULL;
4806 iomsg->free_iov = iomsg->fast_iov;
4807 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4808 &iomsg->free_iov, &iomsg->msg.msg_iter,
4817 #ifdef CONFIG_COMPAT
4818 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4819 struct io_async_msghdr *iomsg)
4821 struct io_sr_msg *sr = &req->sr_msg;
4822 struct compat_iovec __user *uiov;
4827 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4832 uiov = compat_ptr(ptr);
4833 if (req->flags & REQ_F_BUFFER_SELECT) {
4834 compat_ssize_t clen;
4838 if (!access_ok(uiov, sizeof(*uiov)))
4840 if (__get_user(clen, &uiov->iov_len))
4845 iomsg->free_iov = NULL;
4847 iomsg->free_iov = iomsg->fast_iov;
4848 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4849 UIO_FASTIOV, &iomsg->free_iov,
4850 &iomsg->msg.msg_iter, true);
4859 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4860 struct io_async_msghdr *iomsg)
4862 iomsg->msg.msg_name = &iomsg->addr;
4864 #ifdef CONFIG_COMPAT
4865 if (req->ctx->compat)
4866 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4869 return __io_recvmsg_copy_hdr(req, iomsg);
4872 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4875 struct io_sr_msg *sr = &req->sr_msg;
4876 struct io_buffer *kbuf;
4878 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4883 req->flags |= REQ_F_BUFFER_SELECTED;
4887 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4889 return io_put_kbuf(req, req->sr_msg.kbuf);
4892 static int io_recvmsg_prep_async(struct io_kiocb *req)
4896 ret = io_recvmsg_copy_hdr(req, req->async_data);
4898 req->flags |= REQ_F_NEED_CLEANUP;
4902 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4904 struct io_sr_msg *sr = &req->sr_msg;
4906 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4909 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4910 sr->len = READ_ONCE(sqe->len);
4911 sr->bgid = READ_ONCE(sqe->buf_group);
4912 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4913 if (sr->msg_flags & MSG_DONTWAIT)
4914 req->flags |= REQ_F_NOWAIT;
4916 #ifdef CONFIG_COMPAT
4917 if (req->ctx->compat)
4918 sr->msg_flags |= MSG_CMSG_COMPAT;
4923 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4925 struct io_async_msghdr iomsg, *kmsg;
4926 struct socket *sock;
4927 struct io_buffer *kbuf;
4930 int ret, cflags = 0;
4931 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4933 sock = sock_from_file(req->file);
4934 if (unlikely(!sock))
4937 kmsg = req->async_data;
4939 ret = io_recvmsg_copy_hdr(req, &iomsg);
4945 if (req->flags & REQ_F_BUFFER_SELECT) {
4946 kbuf = io_recv_buffer_select(req, !force_nonblock);
4948 return PTR_ERR(kbuf);
4949 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4950 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4951 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4952 1, req->sr_msg.len);
4955 flags = req->sr_msg.msg_flags;
4957 flags |= MSG_DONTWAIT;
4958 if (flags & MSG_WAITALL)
4959 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4961 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4962 kmsg->uaddr, flags);
4963 if (force_nonblock && ret == -EAGAIN)
4964 return io_setup_async_msg(req, kmsg);
4965 if (ret == -ERESTARTSYS)
4968 if (req->flags & REQ_F_BUFFER_SELECTED)
4969 cflags = io_put_recv_kbuf(req);
4970 /* fast path, check for non-NULL to avoid function call */
4972 kfree(kmsg->free_iov);
4973 req->flags &= ~REQ_F_NEED_CLEANUP;
4974 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4976 __io_req_complete(req, issue_flags, ret, cflags);
4980 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4982 struct io_buffer *kbuf;
4983 struct io_sr_msg *sr = &req->sr_msg;
4985 void __user *buf = sr->buf;
4986 struct socket *sock;
4990 int ret, cflags = 0;
4991 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4993 sock = sock_from_file(req->file);
4994 if (unlikely(!sock))
4997 if (req->flags & REQ_F_BUFFER_SELECT) {
4998 kbuf = io_recv_buffer_select(req, !force_nonblock);
5000 return PTR_ERR(kbuf);
5001 buf = u64_to_user_ptr(kbuf->addr);
5004 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5008 msg.msg_name = NULL;
5009 msg.msg_control = NULL;
5010 msg.msg_controllen = 0;
5011 msg.msg_namelen = 0;
5012 msg.msg_iocb = NULL;
5015 flags = req->sr_msg.msg_flags;
5017 flags |= MSG_DONTWAIT;
5018 if (flags & MSG_WAITALL)
5019 min_ret = iov_iter_count(&msg.msg_iter);
5021 ret = sock_recvmsg(sock, &msg, flags);
5022 if (force_nonblock && ret == -EAGAIN)
5024 if (ret == -ERESTARTSYS)
5027 if (req->flags & REQ_F_BUFFER_SELECTED)
5028 cflags = io_put_recv_kbuf(req);
5029 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5031 __io_req_complete(req, issue_flags, ret, cflags);
5035 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5037 struct io_accept *accept = &req->accept;
5039 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5041 if (sqe->ioprio || sqe->len || sqe->buf_index)
5044 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5045 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5046 accept->flags = READ_ONCE(sqe->accept_flags);
5047 accept->nofile = rlimit(RLIMIT_NOFILE);
5049 accept->file_slot = READ_ONCE(sqe->file_index);
5050 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5051 (accept->flags & SOCK_CLOEXEC)))
5053 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5055 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5056 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5060 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5062 struct io_accept *accept = &req->accept;
5063 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5064 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5065 bool fixed = !!accept->file_slot;
5069 if (req->file->f_flags & O_NONBLOCK)
5070 req->flags |= REQ_F_NOWAIT;
5073 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5074 if (unlikely(fd < 0))
5077 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5082 ret = PTR_ERR(file);
5083 if (ret == -EAGAIN && force_nonblock)
5085 if (ret == -ERESTARTSYS)
5088 } else if (!fixed) {
5089 fd_install(fd, file);
5092 ret = io_install_fixed_file(req, file, issue_flags,
5093 accept->file_slot - 1);
5095 __io_req_complete(req, issue_flags, ret, 0);
5099 static int io_connect_prep_async(struct io_kiocb *req)
5101 struct io_async_connect *io = req->async_data;
5102 struct io_connect *conn = &req->connect;
5104 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5107 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5109 struct io_connect *conn = &req->connect;
5111 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5113 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5117 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5118 conn->addr_len = READ_ONCE(sqe->addr2);
5122 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5124 struct io_async_connect __io, *io;
5125 unsigned file_flags;
5127 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5129 if (req->async_data) {
5130 io = req->async_data;
5132 ret = move_addr_to_kernel(req->connect.addr,
5133 req->connect.addr_len,
5140 file_flags = force_nonblock ? O_NONBLOCK : 0;
5142 ret = __sys_connect_file(req->file, &io->address,
5143 req->connect.addr_len, file_flags);
5144 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5145 if (req->async_data)
5147 if (io_alloc_async_data(req)) {
5151 memcpy(req->async_data, &__io, sizeof(__io));
5154 if (ret == -ERESTARTSYS)
5159 __io_req_complete(req, issue_flags, ret, 0);
5162 #else /* !CONFIG_NET */
5163 #define IO_NETOP_FN(op) \
5164 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5166 return -EOPNOTSUPP; \
5169 #define IO_NETOP_PREP(op) \
5171 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5173 return -EOPNOTSUPP; \
5176 #define IO_NETOP_PREP_ASYNC(op) \
5178 static int io_##op##_prep_async(struct io_kiocb *req) \
5180 return -EOPNOTSUPP; \
5183 IO_NETOP_PREP_ASYNC(sendmsg);
5184 IO_NETOP_PREP_ASYNC(recvmsg);
5185 IO_NETOP_PREP_ASYNC(connect);
5186 IO_NETOP_PREP(accept);
5189 #endif /* CONFIG_NET */
5191 struct io_poll_table {
5192 struct poll_table_struct pt;
5193 struct io_kiocb *req;
5198 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5199 __poll_t mask, io_req_tw_func_t func)
5201 /* for instances that support it check for an event match first: */
5202 if (mask && !(mask & poll->events))
5205 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5207 list_del_init(&poll->wait.entry);
5210 req->io_task_work.func = func;
5213 * If this fails, then the task is exiting. When a task exits, the
5214 * work gets canceled, so just cancel this request as well instead
5215 * of executing it. We can't safely execute it anyway, as we may not
5216 * have the needed state needed for it anyway.
5218 io_req_task_work_add(req);
5222 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5223 __acquires(&req->ctx->completion_lock)
5225 struct io_ring_ctx *ctx = req->ctx;
5227 /* req->task == current here, checking PF_EXITING is safe */
5228 if (unlikely(req->task->flags & PF_EXITING))
5229 WRITE_ONCE(poll->canceled, true);
5231 if (!req->result && !READ_ONCE(poll->canceled)) {
5232 struct poll_table_struct pt = { ._key = poll->events };
5234 req->result = vfs_poll(req->file, &pt) & poll->events;
5237 spin_lock(&ctx->completion_lock);
5238 if (!req->result && !READ_ONCE(poll->canceled)) {
5239 add_wait_queue(poll->head, &poll->wait);
5246 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5248 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5249 if (req->opcode == IORING_OP_POLL_ADD)
5250 return req->async_data;
5251 return req->apoll->double_poll;
5254 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5256 if (req->opcode == IORING_OP_POLL_ADD)
5258 return &req->apoll->poll;
5261 static void io_poll_remove_double(struct io_kiocb *req)
5262 __must_hold(&req->ctx->completion_lock)
5264 struct io_poll_iocb *poll = io_poll_get_double(req);
5266 lockdep_assert_held(&req->ctx->completion_lock);
5268 if (poll && poll->head) {
5269 struct wait_queue_head *head = poll->head;
5271 spin_lock_irq(&head->lock);
5272 list_del_init(&poll->wait.entry);
5273 if (poll->wait.private)
5276 spin_unlock_irq(&head->lock);
5280 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5281 __must_hold(&req->ctx->completion_lock)
5283 struct io_ring_ctx *ctx = req->ctx;
5284 unsigned flags = IORING_CQE_F_MORE;
5287 if (READ_ONCE(req->poll.canceled)) {
5289 req->poll.events |= EPOLLONESHOT;
5291 error = mangle_poll(mask);
5293 if (req->poll.events & EPOLLONESHOT)
5295 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5296 req->poll.done = true;
5299 if (flags & IORING_CQE_F_MORE)
5302 return !(flags & IORING_CQE_F_MORE);
5305 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5306 __must_hold(&req->ctx->completion_lock)
5310 done = __io_poll_complete(req, mask);
5311 io_commit_cqring(req->ctx);
5315 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5317 struct io_ring_ctx *ctx = req->ctx;
5318 struct io_kiocb *nxt;
5320 if (io_poll_rewait(req, &req->poll)) {
5321 spin_unlock(&ctx->completion_lock);
5325 done = __io_poll_complete(req, req->result);
5327 io_poll_remove_double(req);
5328 hash_del(&req->hash_node);
5331 add_wait_queue(req->poll.head, &req->poll.wait);
5333 io_commit_cqring(ctx);
5334 spin_unlock(&ctx->completion_lock);
5335 io_cqring_ev_posted(ctx);
5338 nxt = io_put_req_find_next(req);
5340 io_req_task_submit(nxt, locked);
5345 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5346 int sync, void *key)
5348 struct io_kiocb *req = wait->private;
5349 struct io_poll_iocb *poll = io_poll_get_single(req);
5350 __poll_t mask = key_to_poll(key);
5351 unsigned long flags;
5353 /* for instances that support it check for an event match first: */
5354 if (mask && !(mask & poll->events))
5356 if (!(poll->events & EPOLLONESHOT))
5357 return poll->wait.func(&poll->wait, mode, sync, key);
5359 list_del_init(&wait->entry);
5364 spin_lock_irqsave(&poll->head->lock, flags);
5365 done = list_empty(&poll->wait.entry);
5367 list_del_init(&poll->wait.entry);
5368 /* make sure double remove sees this as being gone */
5369 wait->private = NULL;
5370 spin_unlock_irqrestore(&poll->head->lock, flags);
5372 /* use wait func handler, so it matches the rq type */
5373 poll->wait.func(&poll->wait, mode, sync, key);
5380 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5381 wait_queue_func_t wake_func)
5385 poll->canceled = false;
5386 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5387 /* mask in events that we always want/need */
5388 poll->events = events | IO_POLL_UNMASK;
5389 INIT_LIST_HEAD(&poll->wait.entry);
5390 init_waitqueue_func_entry(&poll->wait, wake_func);
5393 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5394 struct wait_queue_head *head,
5395 struct io_poll_iocb **poll_ptr)
5397 struct io_kiocb *req = pt->req;
5400 * The file being polled uses multiple waitqueues for poll handling
5401 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5404 if (unlikely(pt->nr_entries)) {
5405 struct io_poll_iocb *poll_one = poll;
5407 /* double add on the same waitqueue head, ignore */
5408 if (poll_one->head == head)
5410 /* already have a 2nd entry, fail a third attempt */
5412 if ((*poll_ptr)->head == head)
5414 pt->error = -EINVAL;
5418 * Can't handle multishot for double wait for now, turn it
5419 * into one-shot mode.
5421 if (!(poll_one->events & EPOLLONESHOT))
5422 poll_one->events |= EPOLLONESHOT;
5423 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5425 pt->error = -ENOMEM;
5428 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5430 poll->wait.private = req;
5437 if (poll->events & EPOLLEXCLUSIVE)
5438 add_wait_queue_exclusive(head, &poll->wait);
5440 add_wait_queue(head, &poll->wait);
5443 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5444 struct poll_table_struct *p)
5446 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5447 struct async_poll *apoll = pt->req->apoll;
5449 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5452 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5454 struct async_poll *apoll = req->apoll;
5455 struct io_ring_ctx *ctx = req->ctx;
5457 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5459 if (io_poll_rewait(req, &apoll->poll)) {
5460 spin_unlock(&ctx->completion_lock);
5464 hash_del(&req->hash_node);
5465 io_poll_remove_double(req);
5466 spin_unlock(&ctx->completion_lock);
5468 if (!READ_ONCE(apoll->poll.canceled))
5469 io_req_task_submit(req, locked);
5471 io_req_complete_failed(req, -ECANCELED);
5474 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5477 struct io_kiocb *req = wait->private;
5478 struct io_poll_iocb *poll = &req->apoll->poll;
5480 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5483 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5486 static void io_poll_req_insert(struct io_kiocb *req)
5488 struct io_ring_ctx *ctx = req->ctx;
5489 struct hlist_head *list;
5491 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5492 hlist_add_head(&req->hash_node, list);
5495 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5496 struct io_poll_iocb *poll,
5497 struct io_poll_table *ipt, __poll_t mask,
5498 wait_queue_func_t wake_func)
5499 __acquires(&ctx->completion_lock)
5501 struct io_ring_ctx *ctx = req->ctx;
5502 bool cancel = false;
5504 INIT_HLIST_NODE(&req->hash_node);
5505 io_init_poll_iocb(poll, mask, wake_func);
5506 poll->file = req->file;
5507 poll->wait.private = req;
5509 ipt->pt._key = mask;
5512 ipt->nr_entries = 0;
5514 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5515 if (unlikely(!ipt->nr_entries) && !ipt->error)
5516 ipt->error = -EINVAL;
5518 spin_lock(&ctx->completion_lock);
5519 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5520 io_poll_remove_double(req);
5521 if (likely(poll->head)) {
5522 spin_lock_irq(&poll->head->lock);
5523 if (unlikely(list_empty(&poll->wait.entry))) {
5529 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5530 list_del_init(&poll->wait.entry);
5532 WRITE_ONCE(poll->canceled, true);
5533 else if (!poll->done) /* actually waiting for an event */
5534 io_poll_req_insert(req);
5535 spin_unlock_irq(&poll->head->lock);
5547 static int io_arm_poll_handler(struct io_kiocb *req)
5549 const struct io_op_def *def = &io_op_defs[req->opcode];
5550 struct io_ring_ctx *ctx = req->ctx;
5551 struct async_poll *apoll;
5552 struct io_poll_table ipt;
5553 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5556 if (!req->file || !file_can_poll(req->file))
5557 return IO_APOLL_ABORTED;
5558 if (req->flags & REQ_F_POLLED)
5559 return IO_APOLL_ABORTED;
5560 if (!def->pollin && !def->pollout)
5561 return IO_APOLL_ABORTED;
5565 mask |= POLLIN | POLLRDNORM;
5567 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5568 if ((req->opcode == IORING_OP_RECVMSG) &&
5569 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5573 mask |= POLLOUT | POLLWRNORM;
5576 /* if we can't nonblock try, then no point in arming a poll handler */
5577 if (!io_file_supports_nowait(req, rw))
5578 return IO_APOLL_ABORTED;
5580 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5581 if (unlikely(!apoll))
5582 return IO_APOLL_ABORTED;
5583 apoll->double_poll = NULL;
5585 req->flags |= REQ_F_POLLED;
5586 ipt.pt._qproc = io_async_queue_proc;
5587 io_req_set_refcount(req);
5589 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5591 spin_unlock(&ctx->completion_lock);
5592 if (ret || ipt.error)
5593 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5595 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5596 mask, apoll->poll.events);
5600 static bool __io_poll_remove_one(struct io_kiocb *req,
5601 struct io_poll_iocb *poll, bool do_cancel)
5602 __must_hold(&req->ctx->completion_lock)
5604 bool do_complete = false;
5608 spin_lock_irq(&poll->head->lock);
5610 WRITE_ONCE(poll->canceled, true);
5611 if (!list_empty(&poll->wait.entry)) {
5612 list_del_init(&poll->wait.entry);
5615 spin_unlock_irq(&poll->head->lock);
5616 hash_del(&req->hash_node);
5620 static bool io_poll_remove_one(struct io_kiocb *req)
5621 __must_hold(&req->ctx->completion_lock)
5625 io_poll_remove_double(req);
5626 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5629 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5630 io_commit_cqring(req->ctx);
5632 io_put_req_deferred(req);
5638 * Returns true if we found and killed one or more poll requests
5640 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5643 struct hlist_node *tmp;
5644 struct io_kiocb *req;
5647 spin_lock(&ctx->completion_lock);
5648 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5649 struct hlist_head *list;
5651 list = &ctx->cancel_hash[i];
5652 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5653 if (io_match_task(req, tsk, cancel_all))
5654 posted += io_poll_remove_one(req);
5657 spin_unlock(&ctx->completion_lock);
5660 io_cqring_ev_posted(ctx);
5665 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5667 __must_hold(&ctx->completion_lock)
5669 struct hlist_head *list;
5670 struct io_kiocb *req;
5672 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5673 hlist_for_each_entry(req, list, hash_node) {
5674 if (sqe_addr != req->user_data)
5676 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5683 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5685 __must_hold(&ctx->completion_lock)
5687 struct io_kiocb *req;
5689 req = io_poll_find(ctx, sqe_addr, poll_only);
5692 if (io_poll_remove_one(req))
5698 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5703 events = READ_ONCE(sqe->poll32_events);
5705 events = swahw32(events);
5707 if (!(flags & IORING_POLL_ADD_MULTI))
5708 events |= EPOLLONESHOT;
5709 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5712 static int io_poll_update_prep(struct io_kiocb *req,
5713 const struct io_uring_sqe *sqe)
5715 struct io_poll_update *upd = &req->poll_update;
5718 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5720 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5722 flags = READ_ONCE(sqe->len);
5723 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5724 IORING_POLL_ADD_MULTI))
5726 /* meaningless without update */
5727 if (flags == IORING_POLL_ADD_MULTI)
5730 upd->old_user_data = READ_ONCE(sqe->addr);
5731 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5732 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5734 upd->new_user_data = READ_ONCE(sqe->off);
5735 if (!upd->update_user_data && upd->new_user_data)
5737 if (upd->update_events)
5738 upd->events = io_poll_parse_events(sqe, flags);
5739 else if (sqe->poll32_events)
5745 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5748 struct io_kiocb *req = wait->private;
5749 struct io_poll_iocb *poll = &req->poll;
5751 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5754 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5755 struct poll_table_struct *p)
5757 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5759 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5762 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5764 struct io_poll_iocb *poll = &req->poll;
5767 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5769 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5771 flags = READ_ONCE(sqe->len);
5772 if (flags & ~IORING_POLL_ADD_MULTI)
5775 io_req_set_refcount(req);
5776 poll->events = io_poll_parse_events(sqe, flags);
5780 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5782 struct io_poll_iocb *poll = &req->poll;
5783 struct io_ring_ctx *ctx = req->ctx;
5784 struct io_poll_table ipt;
5787 ipt.pt._qproc = io_poll_queue_proc;
5789 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5792 if (mask) { /* no async, we'd stolen it */
5794 io_poll_complete(req, mask);
5796 spin_unlock(&ctx->completion_lock);
5799 io_cqring_ev_posted(ctx);
5800 if (poll->events & EPOLLONESHOT)
5806 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5808 struct io_ring_ctx *ctx = req->ctx;
5809 struct io_kiocb *preq;
5813 spin_lock(&ctx->completion_lock);
5814 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5820 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5822 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5827 * Don't allow racy completion with singleshot, as we cannot safely
5828 * update those. For multishot, if we're racing with completion, just
5829 * let completion re-add it.
5831 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5832 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5836 /* we now have a detached poll request. reissue. */
5840 spin_unlock(&ctx->completion_lock);
5842 io_req_complete(req, ret);
5845 /* only mask one event flags, keep behavior flags */
5846 if (req->poll_update.update_events) {
5847 preq->poll.events &= ~0xffff;
5848 preq->poll.events |= req->poll_update.events & 0xffff;
5849 preq->poll.events |= IO_POLL_UNMASK;
5851 if (req->poll_update.update_user_data)
5852 preq->user_data = req->poll_update.new_user_data;
5853 spin_unlock(&ctx->completion_lock);
5855 /* complete update request, we're done with it */
5856 io_req_complete(req, ret);
5859 ret = io_poll_add(preq, issue_flags);
5862 io_req_complete(preq, ret);
5868 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5871 io_req_complete_post(req, -ETIME, 0);
5874 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5876 struct io_timeout_data *data = container_of(timer,
5877 struct io_timeout_data, timer);
5878 struct io_kiocb *req = data->req;
5879 struct io_ring_ctx *ctx = req->ctx;
5880 unsigned long flags;
5882 spin_lock_irqsave(&ctx->timeout_lock, flags);
5883 list_del_init(&req->timeout.list);
5884 atomic_set(&req->ctx->cq_timeouts,
5885 atomic_read(&req->ctx->cq_timeouts) + 1);
5886 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5888 req->io_task_work.func = io_req_task_timeout;
5889 io_req_task_work_add(req);
5890 return HRTIMER_NORESTART;
5893 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5895 __must_hold(&ctx->timeout_lock)
5897 struct io_timeout_data *io;
5898 struct io_kiocb *req;
5901 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5902 found = user_data == req->user_data;
5907 return ERR_PTR(-ENOENT);
5909 io = req->async_data;
5910 if (hrtimer_try_to_cancel(&io->timer) == -1)
5911 return ERR_PTR(-EALREADY);
5912 list_del_init(&req->timeout.list);
5916 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5917 __must_hold(&ctx->completion_lock)
5918 __must_hold(&ctx->timeout_lock)
5920 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5923 return PTR_ERR(req);
5926 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5927 io_put_req_deferred(req);
5931 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5933 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5934 case IORING_TIMEOUT_BOOTTIME:
5935 return CLOCK_BOOTTIME;
5936 case IORING_TIMEOUT_REALTIME:
5937 return CLOCK_REALTIME;
5939 /* can't happen, vetted at prep time */
5943 return CLOCK_MONOTONIC;
5947 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5948 struct timespec64 *ts, enum hrtimer_mode mode)
5949 __must_hold(&ctx->timeout_lock)
5951 struct io_timeout_data *io;
5952 struct io_kiocb *req;
5955 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
5956 found = user_data == req->user_data;
5963 io = req->async_data;
5964 if (hrtimer_try_to_cancel(&io->timer) == -1)
5966 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
5967 io->timer.function = io_link_timeout_fn;
5968 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
5972 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5973 struct timespec64 *ts, enum hrtimer_mode mode)
5974 __must_hold(&ctx->timeout_lock)
5976 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5977 struct io_timeout_data *data;
5980 return PTR_ERR(req);
5982 req->timeout.off = 0; /* noseq */
5983 data = req->async_data;
5984 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5985 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
5986 data->timer.function = io_timeout_fn;
5987 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5991 static int io_timeout_remove_prep(struct io_kiocb *req,
5992 const struct io_uring_sqe *sqe)
5994 struct io_timeout_rem *tr = &req->timeout_rem;
5996 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5998 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6000 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6003 tr->ltimeout = false;
6004 tr->addr = READ_ONCE(sqe->addr);
6005 tr->flags = READ_ONCE(sqe->timeout_flags);
6006 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6007 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6009 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6010 tr->ltimeout = true;
6011 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6013 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6015 } else if (tr->flags) {
6016 /* timeout removal doesn't support flags */
6023 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6025 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6030 * Remove or update an existing timeout command
6032 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6034 struct io_timeout_rem *tr = &req->timeout_rem;
6035 struct io_ring_ctx *ctx = req->ctx;
6038 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6039 spin_lock(&ctx->completion_lock);
6040 spin_lock_irq(&ctx->timeout_lock);
6041 ret = io_timeout_cancel(ctx, tr->addr);
6042 spin_unlock_irq(&ctx->timeout_lock);
6043 spin_unlock(&ctx->completion_lock);
6045 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6047 spin_lock_irq(&ctx->timeout_lock);
6049 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6051 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6052 spin_unlock_irq(&ctx->timeout_lock);
6057 io_req_complete_post(req, ret, 0);
6061 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6062 bool is_timeout_link)
6064 struct io_timeout_data *data;
6066 u32 off = READ_ONCE(sqe->off);
6068 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6070 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6073 if (off && is_timeout_link)
6075 flags = READ_ONCE(sqe->timeout_flags);
6076 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6078 /* more than one clock specified is invalid, obviously */
6079 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6082 INIT_LIST_HEAD(&req->timeout.list);
6083 req->timeout.off = off;
6084 if (unlikely(off && !req->ctx->off_timeout_used))
6085 req->ctx->off_timeout_used = true;
6087 if (!req->async_data && io_alloc_async_data(req))
6090 data = req->async_data;
6092 data->flags = flags;
6094 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6097 data->mode = io_translate_timeout_mode(flags);
6098 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6100 if (is_timeout_link) {
6101 struct io_submit_link *link = &req->ctx->submit_state.link;
6105 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6107 req->timeout.head = link->last;
6108 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6113 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6115 struct io_ring_ctx *ctx = req->ctx;
6116 struct io_timeout_data *data = req->async_data;
6117 struct list_head *entry;
6118 u32 tail, off = req->timeout.off;
6120 spin_lock_irq(&ctx->timeout_lock);
6123 * sqe->off holds how many events that need to occur for this
6124 * timeout event to be satisfied. If it isn't set, then this is
6125 * a pure timeout request, sequence isn't used.
6127 if (io_is_timeout_noseq(req)) {
6128 entry = ctx->timeout_list.prev;
6132 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6133 req->timeout.target_seq = tail + off;
6135 /* Update the last seq here in case io_flush_timeouts() hasn't.
6136 * This is safe because ->completion_lock is held, and submissions
6137 * and completions are never mixed in the same ->completion_lock section.
6139 ctx->cq_last_tm_flush = tail;
6142 * Insertion sort, ensuring the first entry in the list is always
6143 * the one we need first.
6145 list_for_each_prev(entry, &ctx->timeout_list) {
6146 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6149 if (io_is_timeout_noseq(nxt))
6151 /* nxt.seq is behind @tail, otherwise would've been completed */
6152 if (off >= nxt->timeout.target_seq - tail)
6156 list_add(&req->timeout.list, entry);
6157 data->timer.function = io_timeout_fn;
6158 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6159 spin_unlock_irq(&ctx->timeout_lock);
6163 struct io_cancel_data {
6164 struct io_ring_ctx *ctx;
6168 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6170 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6171 struct io_cancel_data *cd = data;
6173 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6176 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6177 struct io_ring_ctx *ctx)
6179 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6180 enum io_wq_cancel cancel_ret;
6183 if (!tctx || !tctx->io_wq)
6186 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6187 switch (cancel_ret) {
6188 case IO_WQ_CANCEL_OK:
6191 case IO_WQ_CANCEL_RUNNING:
6194 case IO_WQ_CANCEL_NOTFOUND:
6202 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6204 struct io_ring_ctx *ctx = req->ctx;
6207 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6209 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6213 spin_lock(&ctx->completion_lock);
6214 spin_lock_irq(&ctx->timeout_lock);
6215 ret = io_timeout_cancel(ctx, sqe_addr);
6216 spin_unlock_irq(&ctx->timeout_lock);
6219 ret = io_poll_cancel(ctx, sqe_addr, false);
6221 spin_unlock(&ctx->completion_lock);
6225 static int io_async_cancel_prep(struct io_kiocb *req,
6226 const struct io_uring_sqe *sqe)
6228 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6230 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6232 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6236 req->cancel.addr = READ_ONCE(sqe->addr);
6240 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6242 struct io_ring_ctx *ctx = req->ctx;
6243 u64 sqe_addr = req->cancel.addr;
6244 struct io_tctx_node *node;
6247 ret = io_try_cancel_userdata(req, sqe_addr);
6251 /* slow path, try all io-wq's */
6252 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6254 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6255 struct io_uring_task *tctx = node->task->io_uring;
6257 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6261 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6265 io_req_complete_post(req, ret, 0);
6269 static int io_rsrc_update_prep(struct io_kiocb *req,
6270 const struct io_uring_sqe *sqe)
6272 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6274 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6277 req->rsrc_update.offset = READ_ONCE(sqe->off);
6278 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6279 if (!req->rsrc_update.nr_args)
6281 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6285 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6287 struct io_ring_ctx *ctx = req->ctx;
6288 struct io_uring_rsrc_update2 up;
6291 if (issue_flags & IO_URING_F_NONBLOCK)
6294 up.offset = req->rsrc_update.offset;
6295 up.data = req->rsrc_update.arg;
6300 mutex_lock(&ctx->uring_lock);
6301 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6302 &up, req->rsrc_update.nr_args);
6303 mutex_unlock(&ctx->uring_lock);
6307 __io_req_complete(req, issue_flags, ret, 0);
6311 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6313 switch (req->opcode) {
6316 case IORING_OP_READV:
6317 case IORING_OP_READ_FIXED:
6318 case IORING_OP_READ:
6319 return io_read_prep(req, sqe);
6320 case IORING_OP_WRITEV:
6321 case IORING_OP_WRITE_FIXED:
6322 case IORING_OP_WRITE:
6323 return io_write_prep(req, sqe);
6324 case IORING_OP_POLL_ADD:
6325 return io_poll_add_prep(req, sqe);
6326 case IORING_OP_POLL_REMOVE:
6327 return io_poll_update_prep(req, sqe);
6328 case IORING_OP_FSYNC:
6329 return io_fsync_prep(req, sqe);
6330 case IORING_OP_SYNC_FILE_RANGE:
6331 return io_sfr_prep(req, sqe);
6332 case IORING_OP_SENDMSG:
6333 case IORING_OP_SEND:
6334 return io_sendmsg_prep(req, sqe);
6335 case IORING_OP_RECVMSG:
6336 case IORING_OP_RECV:
6337 return io_recvmsg_prep(req, sqe);
6338 case IORING_OP_CONNECT:
6339 return io_connect_prep(req, sqe);
6340 case IORING_OP_TIMEOUT:
6341 return io_timeout_prep(req, sqe, false);
6342 case IORING_OP_TIMEOUT_REMOVE:
6343 return io_timeout_remove_prep(req, sqe);
6344 case IORING_OP_ASYNC_CANCEL:
6345 return io_async_cancel_prep(req, sqe);
6346 case IORING_OP_LINK_TIMEOUT:
6347 return io_timeout_prep(req, sqe, true);
6348 case IORING_OP_ACCEPT:
6349 return io_accept_prep(req, sqe);
6350 case IORING_OP_FALLOCATE:
6351 return io_fallocate_prep(req, sqe);
6352 case IORING_OP_OPENAT:
6353 return io_openat_prep(req, sqe);
6354 case IORING_OP_CLOSE:
6355 return io_close_prep(req, sqe);
6356 case IORING_OP_FILES_UPDATE:
6357 return io_rsrc_update_prep(req, sqe);
6358 case IORING_OP_STATX:
6359 return io_statx_prep(req, sqe);
6360 case IORING_OP_FADVISE:
6361 return io_fadvise_prep(req, sqe);
6362 case IORING_OP_MADVISE:
6363 return io_madvise_prep(req, sqe);
6364 case IORING_OP_OPENAT2:
6365 return io_openat2_prep(req, sqe);
6366 case IORING_OP_EPOLL_CTL:
6367 return io_epoll_ctl_prep(req, sqe);
6368 case IORING_OP_SPLICE:
6369 return io_splice_prep(req, sqe);
6370 case IORING_OP_PROVIDE_BUFFERS:
6371 return io_provide_buffers_prep(req, sqe);
6372 case IORING_OP_REMOVE_BUFFERS:
6373 return io_remove_buffers_prep(req, sqe);
6375 return io_tee_prep(req, sqe);
6376 case IORING_OP_SHUTDOWN:
6377 return io_shutdown_prep(req, sqe);
6378 case IORING_OP_RENAMEAT:
6379 return io_renameat_prep(req, sqe);
6380 case IORING_OP_UNLINKAT:
6381 return io_unlinkat_prep(req, sqe);
6382 case IORING_OP_MKDIRAT:
6383 return io_mkdirat_prep(req, sqe);
6384 case IORING_OP_SYMLINKAT:
6385 return io_symlinkat_prep(req, sqe);
6386 case IORING_OP_LINKAT:
6387 return io_linkat_prep(req, sqe);
6390 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6395 static int io_req_prep_async(struct io_kiocb *req)
6397 if (!io_op_defs[req->opcode].needs_async_setup)
6399 if (WARN_ON_ONCE(req->async_data))
6401 if (io_alloc_async_data(req))
6404 switch (req->opcode) {
6405 case IORING_OP_READV:
6406 return io_rw_prep_async(req, READ);
6407 case IORING_OP_WRITEV:
6408 return io_rw_prep_async(req, WRITE);
6409 case IORING_OP_SENDMSG:
6410 return io_sendmsg_prep_async(req);
6411 case IORING_OP_RECVMSG:
6412 return io_recvmsg_prep_async(req);
6413 case IORING_OP_CONNECT:
6414 return io_connect_prep_async(req);
6416 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6421 static u32 io_get_sequence(struct io_kiocb *req)
6423 u32 seq = req->ctx->cached_sq_head;
6425 /* need original cached_sq_head, but it was increased for each req */
6426 io_for_each_link(req, req)
6431 static bool io_drain_req(struct io_kiocb *req)
6433 struct io_kiocb *pos;
6434 struct io_ring_ctx *ctx = req->ctx;
6435 struct io_defer_entry *de;
6439 if (req->flags & REQ_F_FAIL) {
6440 io_req_complete_fail_submit(req);
6445 * If we need to drain a request in the middle of a link, drain the
6446 * head request and the next request/link after the current link.
6447 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6448 * maintained for every request of our link.
6450 if (ctx->drain_next) {
6451 req->flags |= REQ_F_IO_DRAIN;
6452 ctx->drain_next = false;
6454 /* not interested in head, start from the first linked */
6455 io_for_each_link(pos, req->link) {
6456 if (pos->flags & REQ_F_IO_DRAIN) {
6457 ctx->drain_next = true;
6458 req->flags |= REQ_F_IO_DRAIN;
6463 /* Still need defer if there is pending req in defer list. */
6464 if (likely(list_empty_careful(&ctx->defer_list) &&
6465 !(req->flags & REQ_F_IO_DRAIN))) {
6466 ctx->drain_active = false;
6470 seq = io_get_sequence(req);
6471 /* Still a chance to pass the sequence check */
6472 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6475 ret = io_req_prep_async(req);
6478 io_prep_async_link(req);
6479 de = kmalloc(sizeof(*de), GFP_KERNEL);
6483 io_req_complete_failed(req, ret);
6487 spin_lock(&ctx->completion_lock);
6488 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6489 spin_unlock(&ctx->completion_lock);
6491 io_queue_async_work(req, NULL);
6495 trace_io_uring_defer(ctx, req, req->user_data);
6498 list_add_tail(&de->list, &ctx->defer_list);
6499 spin_unlock(&ctx->completion_lock);
6503 static void io_clean_op(struct io_kiocb *req)
6505 if (req->flags & REQ_F_BUFFER_SELECTED) {
6506 switch (req->opcode) {
6507 case IORING_OP_READV:
6508 case IORING_OP_READ_FIXED:
6509 case IORING_OP_READ:
6510 kfree((void *)(unsigned long)req->rw.addr);
6512 case IORING_OP_RECVMSG:
6513 case IORING_OP_RECV:
6514 kfree(req->sr_msg.kbuf);
6519 if (req->flags & REQ_F_NEED_CLEANUP) {
6520 switch (req->opcode) {
6521 case IORING_OP_READV:
6522 case IORING_OP_READ_FIXED:
6523 case IORING_OP_READ:
6524 case IORING_OP_WRITEV:
6525 case IORING_OP_WRITE_FIXED:
6526 case IORING_OP_WRITE: {
6527 struct io_async_rw *io = req->async_data;
6529 kfree(io->free_iovec);
6532 case IORING_OP_RECVMSG:
6533 case IORING_OP_SENDMSG: {
6534 struct io_async_msghdr *io = req->async_data;
6536 kfree(io->free_iov);
6539 case IORING_OP_SPLICE:
6541 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6542 io_put_file(req->splice.file_in);
6544 case IORING_OP_OPENAT:
6545 case IORING_OP_OPENAT2:
6546 if (req->open.filename)
6547 putname(req->open.filename);
6549 case IORING_OP_RENAMEAT:
6550 putname(req->rename.oldpath);
6551 putname(req->rename.newpath);
6553 case IORING_OP_UNLINKAT:
6554 putname(req->unlink.filename);
6556 case IORING_OP_MKDIRAT:
6557 putname(req->mkdir.filename);
6559 case IORING_OP_SYMLINKAT:
6560 putname(req->symlink.oldpath);
6561 putname(req->symlink.newpath);
6563 case IORING_OP_LINKAT:
6564 putname(req->hardlink.oldpath);
6565 putname(req->hardlink.newpath);
6569 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6570 kfree(req->apoll->double_poll);
6574 if (req->flags & REQ_F_INFLIGHT) {
6575 struct io_uring_task *tctx = req->task->io_uring;
6577 atomic_dec(&tctx->inflight_tracked);
6579 if (req->flags & REQ_F_CREDS)
6580 put_cred(req->creds);
6582 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6585 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6587 struct io_ring_ctx *ctx = req->ctx;
6588 const struct cred *creds = NULL;
6591 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6592 creds = override_creds(req->creds);
6594 switch (req->opcode) {
6596 ret = io_nop(req, issue_flags);
6598 case IORING_OP_READV:
6599 case IORING_OP_READ_FIXED:
6600 case IORING_OP_READ:
6601 ret = io_read(req, issue_flags);
6603 case IORING_OP_WRITEV:
6604 case IORING_OP_WRITE_FIXED:
6605 case IORING_OP_WRITE:
6606 ret = io_write(req, issue_flags);
6608 case IORING_OP_FSYNC:
6609 ret = io_fsync(req, issue_flags);
6611 case IORING_OP_POLL_ADD:
6612 ret = io_poll_add(req, issue_flags);
6614 case IORING_OP_POLL_REMOVE:
6615 ret = io_poll_update(req, issue_flags);
6617 case IORING_OP_SYNC_FILE_RANGE:
6618 ret = io_sync_file_range(req, issue_flags);
6620 case IORING_OP_SENDMSG:
6621 ret = io_sendmsg(req, issue_flags);
6623 case IORING_OP_SEND:
6624 ret = io_send(req, issue_flags);
6626 case IORING_OP_RECVMSG:
6627 ret = io_recvmsg(req, issue_flags);
6629 case IORING_OP_RECV:
6630 ret = io_recv(req, issue_flags);
6632 case IORING_OP_TIMEOUT:
6633 ret = io_timeout(req, issue_flags);
6635 case IORING_OP_TIMEOUT_REMOVE:
6636 ret = io_timeout_remove(req, issue_flags);
6638 case IORING_OP_ACCEPT:
6639 ret = io_accept(req, issue_flags);
6641 case IORING_OP_CONNECT:
6642 ret = io_connect(req, issue_flags);
6644 case IORING_OP_ASYNC_CANCEL:
6645 ret = io_async_cancel(req, issue_flags);
6647 case IORING_OP_FALLOCATE:
6648 ret = io_fallocate(req, issue_flags);
6650 case IORING_OP_OPENAT:
6651 ret = io_openat(req, issue_flags);
6653 case IORING_OP_CLOSE:
6654 ret = io_close(req, issue_flags);
6656 case IORING_OP_FILES_UPDATE:
6657 ret = io_files_update(req, issue_flags);
6659 case IORING_OP_STATX:
6660 ret = io_statx(req, issue_flags);
6662 case IORING_OP_FADVISE:
6663 ret = io_fadvise(req, issue_flags);
6665 case IORING_OP_MADVISE:
6666 ret = io_madvise(req, issue_flags);
6668 case IORING_OP_OPENAT2:
6669 ret = io_openat2(req, issue_flags);
6671 case IORING_OP_EPOLL_CTL:
6672 ret = io_epoll_ctl(req, issue_flags);
6674 case IORING_OP_SPLICE:
6675 ret = io_splice(req, issue_flags);
6677 case IORING_OP_PROVIDE_BUFFERS:
6678 ret = io_provide_buffers(req, issue_flags);
6680 case IORING_OP_REMOVE_BUFFERS:
6681 ret = io_remove_buffers(req, issue_flags);
6684 ret = io_tee(req, issue_flags);
6686 case IORING_OP_SHUTDOWN:
6687 ret = io_shutdown(req, issue_flags);
6689 case IORING_OP_RENAMEAT:
6690 ret = io_renameat(req, issue_flags);
6692 case IORING_OP_UNLINKAT:
6693 ret = io_unlinkat(req, issue_flags);
6695 case IORING_OP_MKDIRAT:
6696 ret = io_mkdirat(req, issue_flags);
6698 case IORING_OP_SYMLINKAT:
6699 ret = io_symlinkat(req, issue_flags);
6701 case IORING_OP_LINKAT:
6702 ret = io_linkat(req, issue_flags);
6710 revert_creds(creds);
6713 /* If the op doesn't have a file, we're not polling for it */
6714 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6715 io_iopoll_req_issued(req);
6720 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6722 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6724 req = io_put_req_find_next(req);
6725 return req ? &req->work : NULL;
6728 static void io_wq_submit_work(struct io_wq_work *work)
6730 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6731 struct io_kiocb *timeout;
6734 /* one will be dropped by ->io_free_work() after returning to io-wq */
6735 if (!(req->flags & REQ_F_REFCOUNT))
6736 __io_req_set_refcount(req, 2);
6740 timeout = io_prep_linked_timeout(req);
6742 io_queue_linked_timeout(timeout);
6744 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6745 if (work->flags & IO_WQ_WORK_CANCEL)
6750 ret = io_issue_sqe(req, 0);
6752 * We can get EAGAIN for polled IO even though we're
6753 * forcing a sync submission from here, since we can't
6754 * wait for request slots on the block side.
6762 /* avoid locking problems by failing it from a clean context */
6764 io_req_task_queue_fail(req, ret);
6767 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6770 return &table->files[i];
6773 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6776 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6778 return (struct file *) (slot->file_ptr & FFS_MASK);
6781 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6783 unsigned long file_ptr = (unsigned long) file;
6785 if (__io_file_supports_nowait(file, READ))
6786 file_ptr |= FFS_ASYNC_READ;
6787 if (__io_file_supports_nowait(file, WRITE))
6788 file_ptr |= FFS_ASYNC_WRITE;
6789 if (S_ISREG(file_inode(file)->i_mode))
6790 file_ptr |= FFS_ISREG;
6791 file_slot->file_ptr = file_ptr;
6794 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6795 struct io_kiocb *req, int fd)
6798 unsigned long file_ptr;
6800 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6802 fd = array_index_nospec(fd, ctx->nr_user_files);
6803 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6804 file = (struct file *) (file_ptr & FFS_MASK);
6805 file_ptr &= ~FFS_MASK;
6806 /* mask in overlapping REQ_F and FFS bits */
6807 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6808 io_req_set_rsrc_node(req);
6812 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6813 struct io_kiocb *req, int fd)
6815 struct file *file = fget(fd);
6817 trace_io_uring_file_get(ctx, fd);
6819 /* we don't allow fixed io_uring files */
6820 if (file && unlikely(file->f_op == &io_uring_fops))
6821 io_req_track_inflight(req);
6825 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6826 struct io_kiocb *req, int fd, bool fixed)
6829 return io_file_get_fixed(ctx, req, fd);
6831 return io_file_get_normal(ctx, req, fd);
6834 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6836 struct io_kiocb *prev = req->timeout.prev;
6840 ret = io_try_cancel_userdata(req, prev->user_data);
6841 io_req_complete_post(req, ret ?: -ETIME, 0);
6844 io_req_complete_post(req, -ETIME, 0);
6848 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6850 struct io_timeout_data *data = container_of(timer,
6851 struct io_timeout_data, timer);
6852 struct io_kiocb *prev, *req = data->req;
6853 struct io_ring_ctx *ctx = req->ctx;
6854 unsigned long flags;
6856 spin_lock_irqsave(&ctx->timeout_lock, flags);
6857 prev = req->timeout.head;
6858 req->timeout.head = NULL;
6861 * We don't expect the list to be empty, that will only happen if we
6862 * race with the completion of the linked work.
6865 io_remove_next_linked(prev);
6866 if (!req_ref_inc_not_zero(prev))
6869 list_del(&req->timeout.list);
6870 req->timeout.prev = prev;
6871 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6873 req->io_task_work.func = io_req_task_link_timeout;
6874 io_req_task_work_add(req);
6875 return HRTIMER_NORESTART;
6878 static void io_queue_linked_timeout(struct io_kiocb *req)
6880 struct io_ring_ctx *ctx = req->ctx;
6882 spin_lock_irq(&ctx->timeout_lock);
6884 * If the back reference is NULL, then our linked request finished
6885 * before we got a chance to setup the timer
6887 if (req->timeout.head) {
6888 struct io_timeout_data *data = req->async_data;
6890 data->timer.function = io_link_timeout_fn;
6891 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6893 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6895 spin_unlock_irq(&ctx->timeout_lock);
6896 /* drop submission reference */
6900 static void __io_queue_sqe(struct io_kiocb *req)
6901 __must_hold(&req->ctx->uring_lock)
6903 struct io_kiocb *linked_timeout;
6907 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6910 * We async punt it if the file wasn't marked NOWAIT, or if the file
6911 * doesn't support non-blocking read/write attempts
6914 if (req->flags & REQ_F_COMPLETE_INLINE) {
6915 struct io_ring_ctx *ctx = req->ctx;
6916 struct io_submit_state *state = &ctx->submit_state;
6918 state->compl_reqs[state->compl_nr++] = req;
6919 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6920 io_submit_flush_completions(ctx);
6924 linked_timeout = io_prep_linked_timeout(req);
6926 io_queue_linked_timeout(linked_timeout);
6927 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6928 linked_timeout = io_prep_linked_timeout(req);
6930 switch (io_arm_poll_handler(req)) {
6931 case IO_APOLL_READY:
6933 io_unprep_linked_timeout(req);
6935 case IO_APOLL_ABORTED:
6937 * Queued up for async execution, worker will release
6938 * submit reference when the iocb is actually submitted.
6940 io_queue_async_work(req, NULL);
6945 io_queue_linked_timeout(linked_timeout);
6947 io_req_complete_failed(req, ret);
6951 static inline void io_queue_sqe(struct io_kiocb *req)
6952 __must_hold(&req->ctx->uring_lock)
6954 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6957 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
6958 __io_queue_sqe(req);
6959 } else if (req->flags & REQ_F_FAIL) {
6960 io_req_complete_fail_submit(req);
6962 int ret = io_req_prep_async(req);
6965 io_req_complete_failed(req, ret);
6967 io_queue_async_work(req, NULL);
6972 * Check SQE restrictions (opcode and flags).
6974 * Returns 'true' if SQE is allowed, 'false' otherwise.
6976 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6977 struct io_kiocb *req,
6978 unsigned int sqe_flags)
6980 if (likely(!ctx->restricted))
6983 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6986 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6987 ctx->restrictions.sqe_flags_required)
6990 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6991 ctx->restrictions.sqe_flags_required))
6997 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6998 const struct io_uring_sqe *sqe)
6999 __must_hold(&ctx->uring_lock)
7001 struct io_submit_state *state;
7002 unsigned int sqe_flags;
7003 int personality, ret = 0;
7005 /* req is partially pre-initialised, see io_preinit_req() */
7006 req->opcode = READ_ONCE(sqe->opcode);
7007 /* same numerical values with corresponding REQ_F_*, safe to copy */
7008 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7009 req->user_data = READ_ONCE(sqe->user_data);
7011 req->fixed_rsrc_refs = NULL;
7012 req->task = current;
7014 /* enforce forwards compatibility on users */
7015 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7017 if (unlikely(req->opcode >= IORING_OP_LAST))
7019 if (!io_check_restriction(ctx, req, sqe_flags))
7022 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7023 !io_op_defs[req->opcode].buffer_select)
7025 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7026 ctx->drain_active = true;
7028 personality = READ_ONCE(sqe->personality);
7030 req->creds = xa_load(&ctx->personalities, personality);
7033 get_cred(req->creds);
7034 req->flags |= REQ_F_CREDS;
7036 state = &ctx->submit_state;
7039 * Plug now if we have more than 1 IO left after this, and the target
7040 * is potentially a read/write to block based storage.
7042 if (!state->plug_started && state->ios_left > 1 &&
7043 io_op_defs[req->opcode].plug) {
7044 blk_start_plug(&state->plug);
7045 state->plug_started = true;
7048 if (io_op_defs[req->opcode].needs_file) {
7049 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7050 (sqe_flags & IOSQE_FIXED_FILE));
7051 if (unlikely(!req->file))
7059 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7060 const struct io_uring_sqe *sqe)
7061 __must_hold(&ctx->uring_lock)
7063 struct io_submit_link *link = &ctx->submit_state.link;
7066 ret = io_init_req(ctx, req, sqe);
7067 if (unlikely(ret)) {
7069 /* fail even hard links since we don't submit */
7072 * we can judge a link req is failed or cancelled by if
7073 * REQ_F_FAIL is set, but the head is an exception since
7074 * it may be set REQ_F_FAIL because of other req's failure
7075 * so let's leverage req->result to distinguish if a head
7076 * is set REQ_F_FAIL because of its failure or other req's
7077 * failure so that we can set the correct ret code for it.
7078 * init result here to avoid affecting the normal path.
7080 if (!(link->head->flags & REQ_F_FAIL))
7081 req_fail_link_node(link->head, -ECANCELED);
7082 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7084 * the current req is a normal req, we should return
7085 * error and thus break the submittion loop.
7087 io_req_complete_failed(req, ret);
7090 req_fail_link_node(req, ret);
7092 ret = io_req_prep(req, sqe);
7097 /* don't need @sqe from now on */
7098 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7100 ctx->flags & IORING_SETUP_SQPOLL);
7103 * If we already have a head request, queue this one for async
7104 * submittal once the head completes. If we don't have a head but
7105 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7106 * submitted sync once the chain is complete. If none of those
7107 * conditions are true (normal request), then just queue it.
7110 struct io_kiocb *head = link->head;
7112 if (!(req->flags & REQ_F_FAIL)) {
7113 ret = io_req_prep_async(req);
7114 if (unlikely(ret)) {
7115 req_fail_link_node(req, ret);
7116 if (!(head->flags & REQ_F_FAIL))
7117 req_fail_link_node(head, -ECANCELED);
7120 trace_io_uring_link(ctx, req, head);
7121 link->last->link = req;
7124 /* last request of a link, enqueue the link */
7125 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7130 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7142 * Batched submission is done, ensure local IO is flushed out.
7144 static void io_submit_state_end(struct io_submit_state *state,
7145 struct io_ring_ctx *ctx)
7147 if (state->link.head)
7148 io_queue_sqe(state->link.head);
7149 if (state->compl_nr)
7150 io_submit_flush_completions(ctx);
7151 if (state->plug_started)
7152 blk_finish_plug(&state->plug);
7156 * Start submission side cache.
7158 static void io_submit_state_start(struct io_submit_state *state,
7159 unsigned int max_ios)
7161 state->plug_started = false;
7162 state->ios_left = max_ios;
7163 /* set only head, no need to init link_last in advance */
7164 state->link.head = NULL;
7167 static void io_commit_sqring(struct io_ring_ctx *ctx)
7169 struct io_rings *rings = ctx->rings;
7172 * Ensure any loads from the SQEs are done at this point,
7173 * since once we write the new head, the application could
7174 * write new data to them.
7176 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7180 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7181 * that is mapped by userspace. This means that care needs to be taken to
7182 * ensure that reads are stable, as we cannot rely on userspace always
7183 * being a good citizen. If members of the sqe are validated and then later
7184 * used, it's important that those reads are done through READ_ONCE() to
7185 * prevent a re-load down the line.
7187 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7189 unsigned head, mask = ctx->sq_entries - 1;
7190 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7193 * The cached sq head (or cq tail) serves two purposes:
7195 * 1) allows us to batch the cost of updating the user visible
7197 * 2) allows the kernel side to track the head on its own, even
7198 * though the application is the one updating it.
7200 head = READ_ONCE(ctx->sq_array[sq_idx]);
7201 if (likely(head < ctx->sq_entries))
7202 return &ctx->sq_sqes[head];
7204 /* drop invalid entries */
7206 WRITE_ONCE(ctx->rings->sq_dropped,
7207 READ_ONCE(ctx->rings->sq_dropped) + 1);
7211 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7212 __must_hold(&ctx->uring_lock)
7216 /* make sure SQ entry isn't read before tail */
7217 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7218 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7220 io_get_task_refs(nr);
7222 io_submit_state_start(&ctx->submit_state, nr);
7223 while (submitted < nr) {
7224 const struct io_uring_sqe *sqe;
7225 struct io_kiocb *req;
7227 req = io_alloc_req(ctx);
7228 if (unlikely(!req)) {
7230 submitted = -EAGAIN;
7233 sqe = io_get_sqe(ctx);
7234 if (unlikely(!sqe)) {
7235 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7238 /* will complete beyond this point, count as submitted */
7240 if (io_submit_sqe(ctx, req, sqe))
7244 if (unlikely(submitted != nr)) {
7245 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7246 int unused = nr - ref_used;
7248 current->io_uring->cached_refs += unused;
7249 percpu_ref_put_many(&ctx->refs, unused);
7252 io_submit_state_end(&ctx->submit_state, ctx);
7253 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7254 io_commit_sqring(ctx);
7259 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7261 return READ_ONCE(sqd->state);
7264 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7266 /* Tell userspace we may need a wakeup call */
7267 spin_lock(&ctx->completion_lock);
7268 WRITE_ONCE(ctx->rings->sq_flags,
7269 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7270 spin_unlock(&ctx->completion_lock);
7273 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7275 spin_lock(&ctx->completion_lock);
7276 WRITE_ONCE(ctx->rings->sq_flags,
7277 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7278 spin_unlock(&ctx->completion_lock);
7281 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7283 unsigned int to_submit;
7286 to_submit = io_sqring_entries(ctx);
7287 /* if we're handling multiple rings, cap submit size for fairness */
7288 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7289 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7291 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7292 unsigned nr_events = 0;
7293 const struct cred *creds = NULL;
7295 if (ctx->sq_creds != current_cred())
7296 creds = override_creds(ctx->sq_creds);
7298 mutex_lock(&ctx->uring_lock);
7299 if (!list_empty(&ctx->iopoll_list))
7300 io_do_iopoll(ctx, &nr_events, 0);
7303 * Don't submit if refs are dying, good for io_uring_register(),
7304 * but also it is relied upon by io_ring_exit_work()
7306 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7307 !(ctx->flags & IORING_SETUP_R_DISABLED))
7308 ret = io_submit_sqes(ctx, to_submit);
7309 mutex_unlock(&ctx->uring_lock);
7311 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7312 wake_up(&ctx->sqo_sq_wait);
7314 revert_creds(creds);
7320 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7322 struct io_ring_ctx *ctx;
7323 unsigned sq_thread_idle = 0;
7325 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7326 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7327 sqd->sq_thread_idle = sq_thread_idle;
7330 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7332 bool did_sig = false;
7333 struct ksignal ksig;
7335 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7336 signal_pending(current)) {
7337 mutex_unlock(&sqd->lock);
7338 if (signal_pending(current))
7339 did_sig = get_signal(&ksig);
7341 mutex_lock(&sqd->lock);
7343 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7346 static int io_sq_thread(void *data)
7348 struct io_sq_data *sqd = data;
7349 struct io_ring_ctx *ctx;
7350 unsigned long timeout = 0;
7351 char buf[TASK_COMM_LEN];
7354 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7355 set_task_comm(current, buf);
7357 if (sqd->sq_cpu != -1)
7358 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7360 set_cpus_allowed_ptr(current, cpu_online_mask);
7361 current->flags |= PF_NO_SETAFFINITY;
7363 mutex_lock(&sqd->lock);
7365 bool cap_entries, sqt_spin = false;
7367 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7368 if (io_sqd_handle_event(sqd))
7370 timeout = jiffies + sqd->sq_thread_idle;
7373 cap_entries = !list_is_singular(&sqd->ctx_list);
7374 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7375 int ret = __io_sq_thread(ctx, cap_entries);
7377 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7380 if (io_run_task_work())
7383 if (sqt_spin || !time_after(jiffies, timeout)) {
7386 timeout = jiffies + sqd->sq_thread_idle;
7390 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7391 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7392 bool needs_sched = true;
7394 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7395 io_ring_set_wakeup_flag(ctx);
7397 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7398 !list_empty_careful(&ctx->iopoll_list)) {
7399 needs_sched = false;
7402 if (io_sqring_entries(ctx)) {
7403 needs_sched = false;
7409 mutex_unlock(&sqd->lock);
7411 mutex_lock(&sqd->lock);
7413 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7414 io_ring_clear_wakeup_flag(ctx);
7417 finish_wait(&sqd->wait, &wait);
7418 timeout = jiffies + sqd->sq_thread_idle;
7421 io_uring_cancel_generic(true, sqd);
7423 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7424 io_ring_set_wakeup_flag(ctx);
7426 mutex_unlock(&sqd->lock);
7428 complete(&sqd->exited);
7432 struct io_wait_queue {
7433 struct wait_queue_entry wq;
7434 struct io_ring_ctx *ctx;
7436 unsigned nr_timeouts;
7439 static inline bool io_should_wake(struct io_wait_queue *iowq)
7441 struct io_ring_ctx *ctx = iowq->ctx;
7442 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7445 * Wake up if we have enough events, or if a timeout occurred since we
7446 * started waiting. For timeouts, we always want to return to userspace,
7447 * regardless of event count.
7449 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7452 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7453 int wake_flags, void *key)
7455 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7459 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7460 * the task, and the next invocation will do it.
7462 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7463 return autoremove_wake_function(curr, mode, wake_flags, key);
7467 static int io_run_task_work_sig(void)
7469 if (io_run_task_work())
7471 if (!signal_pending(current))
7473 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7474 return -ERESTARTSYS;
7478 /* when returns >0, the caller should retry */
7479 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7480 struct io_wait_queue *iowq,
7481 signed long *timeout)
7485 /* make sure we run task_work before checking for signals */
7486 ret = io_run_task_work_sig();
7487 if (ret || io_should_wake(iowq))
7489 /* let the caller flush overflows, retry */
7490 if (test_bit(0, &ctx->check_cq_overflow))
7493 *timeout = schedule_timeout(*timeout);
7494 return !*timeout ? -ETIME : 1;
7498 * Wait until events become available, if we don't already have some. The
7499 * application must reap them itself, as they reside on the shared cq ring.
7501 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7502 const sigset_t __user *sig, size_t sigsz,
7503 struct __kernel_timespec __user *uts)
7505 struct io_wait_queue iowq;
7506 struct io_rings *rings = ctx->rings;
7507 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7511 io_cqring_overflow_flush(ctx);
7512 if (io_cqring_events(ctx) >= min_events)
7514 if (!io_run_task_work())
7519 #ifdef CONFIG_COMPAT
7520 if (in_compat_syscall())
7521 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7525 ret = set_user_sigmask(sig, sigsz);
7532 struct timespec64 ts;
7534 if (get_timespec64(&ts, uts))
7536 timeout = timespec64_to_jiffies(&ts);
7539 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7540 iowq.wq.private = current;
7541 INIT_LIST_HEAD(&iowq.wq.entry);
7543 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7544 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7546 trace_io_uring_cqring_wait(ctx, min_events);
7548 /* if we can't even flush overflow, don't wait for more */
7549 if (!io_cqring_overflow_flush(ctx)) {
7553 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7554 TASK_INTERRUPTIBLE);
7555 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7556 finish_wait(&ctx->cq_wait, &iowq.wq);
7560 restore_saved_sigmask_unless(ret == -EINTR);
7562 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7565 static void io_free_page_table(void **table, size_t size)
7567 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7569 for (i = 0; i < nr_tables; i++)
7574 static void **io_alloc_page_table(size_t size)
7576 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7577 size_t init_size = size;
7580 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7584 for (i = 0; i < nr_tables; i++) {
7585 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7587 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7589 io_free_page_table(table, init_size);
7597 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7599 percpu_ref_exit(&ref_node->refs);
7603 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7605 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7606 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7607 unsigned long flags;
7608 bool first_add = false;
7610 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7613 while (!list_empty(&ctx->rsrc_ref_list)) {
7614 node = list_first_entry(&ctx->rsrc_ref_list,
7615 struct io_rsrc_node, node);
7616 /* recycle ref nodes in order */
7619 list_del(&node->node);
7620 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7622 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7625 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7628 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7630 struct io_rsrc_node *ref_node;
7632 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7636 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7641 INIT_LIST_HEAD(&ref_node->node);
7642 INIT_LIST_HEAD(&ref_node->rsrc_list);
7643 ref_node->done = false;
7647 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7648 struct io_rsrc_data *data_to_kill)
7650 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7651 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7654 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7656 rsrc_node->rsrc_data = data_to_kill;
7657 spin_lock_irq(&ctx->rsrc_ref_lock);
7658 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7659 spin_unlock_irq(&ctx->rsrc_ref_lock);
7661 atomic_inc(&data_to_kill->refs);
7662 percpu_ref_kill(&rsrc_node->refs);
7663 ctx->rsrc_node = NULL;
7666 if (!ctx->rsrc_node) {
7667 ctx->rsrc_node = ctx->rsrc_backup_node;
7668 ctx->rsrc_backup_node = NULL;
7672 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7674 if (ctx->rsrc_backup_node)
7676 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7677 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7680 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7684 /* As we may drop ->uring_lock, other task may have started quiesce */
7688 data->quiesce = true;
7690 ret = io_rsrc_node_switch_start(ctx);
7693 io_rsrc_node_switch(ctx, data);
7695 /* kill initial ref, already quiesced if zero */
7696 if (atomic_dec_and_test(&data->refs))
7698 mutex_unlock(&ctx->uring_lock);
7699 flush_delayed_work(&ctx->rsrc_put_work);
7700 ret = wait_for_completion_interruptible(&data->done);
7702 mutex_lock(&ctx->uring_lock);
7706 atomic_inc(&data->refs);
7707 /* wait for all works potentially completing data->done */
7708 flush_delayed_work(&ctx->rsrc_put_work);
7709 reinit_completion(&data->done);
7711 ret = io_run_task_work_sig();
7712 mutex_lock(&ctx->uring_lock);
7714 data->quiesce = false;
7719 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7721 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7722 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7724 return &data->tags[table_idx][off];
7727 static void io_rsrc_data_free(struct io_rsrc_data *data)
7729 size_t size = data->nr * sizeof(data->tags[0][0]);
7732 io_free_page_table((void **)data->tags, size);
7736 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7737 u64 __user *utags, unsigned nr,
7738 struct io_rsrc_data **pdata)
7740 struct io_rsrc_data *data;
7744 data = kzalloc(sizeof(*data), GFP_KERNEL);
7747 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7755 data->do_put = do_put;
7758 for (i = 0; i < nr; i++) {
7759 u64 *tag_slot = io_get_tag_slot(data, i);
7761 if (copy_from_user(tag_slot, &utags[i],
7767 atomic_set(&data->refs, 1);
7768 init_completion(&data->done);
7772 io_rsrc_data_free(data);
7776 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7778 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7779 GFP_KERNEL_ACCOUNT);
7780 return !!table->files;
7783 static void io_free_file_tables(struct io_file_table *table)
7785 kvfree(table->files);
7786 table->files = NULL;
7789 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7791 #if defined(CONFIG_UNIX)
7792 if (ctx->ring_sock) {
7793 struct sock *sock = ctx->ring_sock->sk;
7794 struct sk_buff *skb;
7796 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7802 for (i = 0; i < ctx->nr_user_files; i++) {
7805 file = io_file_from_index(ctx, i);
7810 io_free_file_tables(&ctx->file_table);
7811 io_rsrc_data_free(ctx->file_data);
7812 ctx->file_data = NULL;
7813 ctx->nr_user_files = 0;
7816 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7820 if (!ctx->file_data)
7822 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7824 __io_sqe_files_unregister(ctx);
7828 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7829 __releases(&sqd->lock)
7831 WARN_ON_ONCE(sqd->thread == current);
7834 * Do the dance but not conditional clear_bit() because it'd race with
7835 * other threads incrementing park_pending and setting the bit.
7837 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7838 if (atomic_dec_return(&sqd->park_pending))
7839 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7840 mutex_unlock(&sqd->lock);
7843 static void io_sq_thread_park(struct io_sq_data *sqd)
7844 __acquires(&sqd->lock)
7846 WARN_ON_ONCE(sqd->thread == current);
7848 atomic_inc(&sqd->park_pending);
7849 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7850 mutex_lock(&sqd->lock);
7852 wake_up_process(sqd->thread);
7855 static void io_sq_thread_stop(struct io_sq_data *sqd)
7857 WARN_ON_ONCE(sqd->thread == current);
7858 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7860 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7861 mutex_lock(&sqd->lock);
7863 wake_up_process(sqd->thread);
7864 mutex_unlock(&sqd->lock);
7865 wait_for_completion(&sqd->exited);
7868 static void io_put_sq_data(struct io_sq_data *sqd)
7870 if (refcount_dec_and_test(&sqd->refs)) {
7871 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7873 io_sq_thread_stop(sqd);
7878 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7880 struct io_sq_data *sqd = ctx->sq_data;
7883 io_sq_thread_park(sqd);
7884 list_del_init(&ctx->sqd_list);
7885 io_sqd_update_thread_idle(sqd);
7886 io_sq_thread_unpark(sqd);
7888 io_put_sq_data(sqd);
7889 ctx->sq_data = NULL;
7893 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7895 struct io_ring_ctx *ctx_attach;
7896 struct io_sq_data *sqd;
7899 f = fdget(p->wq_fd);
7901 return ERR_PTR(-ENXIO);
7902 if (f.file->f_op != &io_uring_fops) {
7904 return ERR_PTR(-EINVAL);
7907 ctx_attach = f.file->private_data;
7908 sqd = ctx_attach->sq_data;
7911 return ERR_PTR(-EINVAL);
7913 if (sqd->task_tgid != current->tgid) {
7915 return ERR_PTR(-EPERM);
7918 refcount_inc(&sqd->refs);
7923 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7926 struct io_sq_data *sqd;
7929 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7930 sqd = io_attach_sq_data(p);
7935 /* fall through for EPERM case, setup new sqd/task */
7936 if (PTR_ERR(sqd) != -EPERM)
7940 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7942 return ERR_PTR(-ENOMEM);
7944 atomic_set(&sqd->park_pending, 0);
7945 refcount_set(&sqd->refs, 1);
7946 INIT_LIST_HEAD(&sqd->ctx_list);
7947 mutex_init(&sqd->lock);
7948 init_waitqueue_head(&sqd->wait);
7949 init_completion(&sqd->exited);
7953 #if defined(CONFIG_UNIX)
7955 * Ensure the UNIX gc is aware of our file set, so we are certain that
7956 * the io_uring can be safely unregistered on process exit, even if we have
7957 * loops in the file referencing.
7959 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7961 struct sock *sk = ctx->ring_sock->sk;
7962 struct scm_fp_list *fpl;
7963 struct sk_buff *skb;
7966 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7970 skb = alloc_skb(0, GFP_KERNEL);
7979 fpl->user = get_uid(current_user());
7980 for (i = 0; i < nr; i++) {
7981 struct file *file = io_file_from_index(ctx, i + offset);
7985 fpl->fp[nr_files] = get_file(file);
7986 unix_inflight(fpl->user, fpl->fp[nr_files]);
7991 fpl->max = SCM_MAX_FD;
7992 fpl->count = nr_files;
7993 UNIXCB(skb).fp = fpl;
7994 skb->destructor = unix_destruct_scm;
7995 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7996 skb_queue_head(&sk->sk_receive_queue, skb);
7998 for (i = 0; i < nr_files; i++)
8009 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8010 * causes regular reference counting to break down. We rely on the UNIX
8011 * garbage collection to take care of this problem for us.
8013 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8015 unsigned left, total;
8019 left = ctx->nr_user_files;
8021 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8023 ret = __io_sqe_files_scm(ctx, this_files, total);
8027 total += this_files;
8033 while (total < ctx->nr_user_files) {
8034 struct file *file = io_file_from_index(ctx, total);
8044 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8050 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8052 struct file *file = prsrc->file;
8053 #if defined(CONFIG_UNIX)
8054 struct sock *sock = ctx->ring_sock->sk;
8055 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8056 struct sk_buff *skb;
8059 __skb_queue_head_init(&list);
8062 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8063 * remove this entry and rearrange the file array.
8065 skb = skb_dequeue(head);
8067 struct scm_fp_list *fp;
8069 fp = UNIXCB(skb).fp;
8070 for (i = 0; i < fp->count; i++) {
8073 if (fp->fp[i] != file)
8076 unix_notinflight(fp->user, fp->fp[i]);
8077 left = fp->count - 1 - i;
8079 memmove(&fp->fp[i], &fp->fp[i + 1],
8080 left * sizeof(struct file *));
8087 __skb_queue_tail(&list, skb);
8097 __skb_queue_tail(&list, skb);
8099 skb = skb_dequeue(head);
8102 if (skb_peek(&list)) {
8103 spin_lock_irq(&head->lock);
8104 while ((skb = __skb_dequeue(&list)) != NULL)
8105 __skb_queue_tail(head, skb);
8106 spin_unlock_irq(&head->lock);
8113 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8115 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8116 struct io_ring_ctx *ctx = rsrc_data->ctx;
8117 struct io_rsrc_put *prsrc, *tmp;
8119 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8120 list_del(&prsrc->list);
8123 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8125 io_ring_submit_lock(ctx, lock_ring);
8126 spin_lock(&ctx->completion_lock);
8127 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8129 io_commit_cqring(ctx);
8130 spin_unlock(&ctx->completion_lock);
8131 io_cqring_ev_posted(ctx);
8132 io_ring_submit_unlock(ctx, lock_ring);
8135 rsrc_data->do_put(ctx, prsrc);
8139 io_rsrc_node_destroy(ref_node);
8140 if (atomic_dec_and_test(&rsrc_data->refs))
8141 complete(&rsrc_data->done);
8144 static void io_rsrc_put_work(struct work_struct *work)
8146 struct io_ring_ctx *ctx;
8147 struct llist_node *node;
8149 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8150 node = llist_del_all(&ctx->rsrc_put_llist);
8153 struct io_rsrc_node *ref_node;
8154 struct llist_node *next = node->next;
8156 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8157 __io_rsrc_put_work(ref_node);
8162 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8163 unsigned nr_args, u64 __user *tags)
8165 __s32 __user *fds = (__s32 __user *) arg;
8174 if (nr_args > IORING_MAX_FIXED_FILES)
8176 if (nr_args > rlimit(RLIMIT_NOFILE))
8178 ret = io_rsrc_node_switch_start(ctx);
8181 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8187 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8190 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8191 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8195 /* allow sparse sets */
8198 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8205 if (unlikely(!file))
8209 * Don't allow io_uring instances to be registered. If UNIX
8210 * isn't enabled, then this causes a reference cycle and this
8211 * instance can never get freed. If UNIX is enabled we'll
8212 * handle it just fine, but there's still no point in allowing
8213 * a ring fd as it doesn't support regular read/write anyway.
8215 if (file->f_op == &io_uring_fops) {
8219 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8222 ret = io_sqe_files_scm(ctx);
8224 __io_sqe_files_unregister(ctx);
8228 io_rsrc_node_switch(ctx, NULL);
8231 for (i = 0; i < ctx->nr_user_files; i++) {
8232 file = io_file_from_index(ctx, i);
8236 io_free_file_tables(&ctx->file_table);
8237 ctx->nr_user_files = 0;
8239 io_rsrc_data_free(ctx->file_data);
8240 ctx->file_data = NULL;
8244 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8247 #if defined(CONFIG_UNIX)
8248 struct sock *sock = ctx->ring_sock->sk;
8249 struct sk_buff_head *head = &sock->sk_receive_queue;
8250 struct sk_buff *skb;
8253 * See if we can merge this file into an existing skb SCM_RIGHTS
8254 * file set. If there's no room, fall back to allocating a new skb
8255 * and filling it in.
8257 spin_lock_irq(&head->lock);
8258 skb = skb_peek(head);
8260 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8262 if (fpl->count < SCM_MAX_FD) {
8263 __skb_unlink(skb, head);
8264 spin_unlock_irq(&head->lock);
8265 fpl->fp[fpl->count] = get_file(file);
8266 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8268 spin_lock_irq(&head->lock);
8269 __skb_queue_head(head, skb);
8274 spin_unlock_irq(&head->lock);
8281 return __io_sqe_files_scm(ctx, 1, index);
8287 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8288 unsigned int issue_flags, u32 slot_index)
8290 struct io_ring_ctx *ctx = req->ctx;
8291 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8292 struct io_fixed_file *file_slot;
8295 io_ring_submit_lock(ctx, !force_nonblock);
8296 if (file->f_op == &io_uring_fops)
8299 if (!ctx->file_data)
8302 if (slot_index >= ctx->nr_user_files)
8305 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8306 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8308 if (file_slot->file_ptr)
8311 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8312 io_fixed_file_set(file_slot, file);
8313 ret = io_sqe_file_register(ctx, file, slot_index);
8315 file_slot->file_ptr = 0;
8321 io_ring_submit_unlock(ctx, !force_nonblock);
8327 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8328 struct io_rsrc_node *node, void *rsrc)
8330 struct io_rsrc_put *prsrc;
8332 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8336 prsrc->tag = *io_get_tag_slot(data, idx);
8338 list_add(&prsrc->list, &node->rsrc_list);
8342 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8343 struct io_uring_rsrc_update2 *up,
8346 u64 __user *tags = u64_to_user_ptr(up->tags);
8347 __s32 __user *fds = u64_to_user_ptr(up->data);
8348 struct io_rsrc_data *data = ctx->file_data;
8349 struct io_fixed_file *file_slot;
8353 bool needs_switch = false;
8355 if (!ctx->file_data)
8357 if (up->offset + nr_args > ctx->nr_user_files)
8360 for (done = 0; done < nr_args; done++) {
8363 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8364 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8368 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8372 if (fd == IORING_REGISTER_FILES_SKIP)
8375 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8376 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8378 if (file_slot->file_ptr) {
8379 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8380 err = io_queue_rsrc_removal(data, up->offset + done,
8381 ctx->rsrc_node, file);
8384 file_slot->file_ptr = 0;
8385 needs_switch = true;
8394 * Don't allow io_uring instances to be registered. If
8395 * UNIX isn't enabled, then this causes a reference
8396 * cycle and this instance can never get freed. If UNIX
8397 * is enabled we'll handle it just fine, but there's
8398 * still no point in allowing a ring fd as it doesn't
8399 * support regular read/write anyway.
8401 if (file->f_op == &io_uring_fops) {
8406 *io_get_tag_slot(data, up->offset + done) = tag;
8407 io_fixed_file_set(file_slot, file);
8408 err = io_sqe_file_register(ctx, file, i);
8410 file_slot->file_ptr = 0;
8418 io_rsrc_node_switch(ctx, data);
8419 return done ? done : err;
8422 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8423 struct task_struct *task)
8425 struct io_wq_hash *hash;
8426 struct io_wq_data data;
8427 unsigned int concurrency;
8429 mutex_lock(&ctx->uring_lock);
8430 hash = ctx->hash_map;
8432 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8434 mutex_unlock(&ctx->uring_lock);
8435 return ERR_PTR(-ENOMEM);
8437 refcount_set(&hash->refs, 1);
8438 init_waitqueue_head(&hash->wait);
8439 ctx->hash_map = hash;
8441 mutex_unlock(&ctx->uring_lock);
8445 data.free_work = io_wq_free_work;
8446 data.do_work = io_wq_submit_work;
8448 /* Do QD, or 4 * CPUS, whatever is smallest */
8449 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8451 return io_wq_create(concurrency, &data);
8454 static int io_uring_alloc_task_context(struct task_struct *task,
8455 struct io_ring_ctx *ctx)
8457 struct io_uring_task *tctx;
8460 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8461 if (unlikely(!tctx))
8464 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8465 if (unlikely(ret)) {
8470 tctx->io_wq = io_init_wq_offload(ctx, task);
8471 if (IS_ERR(tctx->io_wq)) {
8472 ret = PTR_ERR(tctx->io_wq);
8473 percpu_counter_destroy(&tctx->inflight);
8479 init_waitqueue_head(&tctx->wait);
8480 atomic_set(&tctx->in_idle, 0);
8481 atomic_set(&tctx->inflight_tracked, 0);
8482 task->io_uring = tctx;
8483 spin_lock_init(&tctx->task_lock);
8484 INIT_WQ_LIST(&tctx->task_list);
8485 init_task_work(&tctx->task_work, tctx_task_work);
8489 void __io_uring_free(struct task_struct *tsk)
8491 struct io_uring_task *tctx = tsk->io_uring;
8493 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8494 WARN_ON_ONCE(tctx->io_wq);
8495 WARN_ON_ONCE(tctx->cached_refs);
8497 percpu_counter_destroy(&tctx->inflight);
8499 tsk->io_uring = NULL;
8502 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8503 struct io_uring_params *p)
8507 /* Retain compatibility with failing for an invalid attach attempt */
8508 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8509 IORING_SETUP_ATTACH_WQ) {
8512 f = fdget(p->wq_fd);
8515 if (f.file->f_op != &io_uring_fops) {
8521 if (ctx->flags & IORING_SETUP_SQPOLL) {
8522 struct task_struct *tsk;
8523 struct io_sq_data *sqd;
8526 sqd = io_get_sq_data(p, &attached);
8532 ctx->sq_creds = get_current_cred();
8534 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8535 if (!ctx->sq_thread_idle)
8536 ctx->sq_thread_idle = HZ;
8538 io_sq_thread_park(sqd);
8539 list_add(&ctx->sqd_list, &sqd->ctx_list);
8540 io_sqd_update_thread_idle(sqd);
8541 /* don't attach to a dying SQPOLL thread, would be racy */
8542 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8543 io_sq_thread_unpark(sqd);
8550 if (p->flags & IORING_SETUP_SQ_AFF) {
8551 int cpu = p->sq_thread_cpu;
8554 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8561 sqd->task_pid = current->pid;
8562 sqd->task_tgid = current->tgid;
8563 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8570 ret = io_uring_alloc_task_context(tsk, ctx);
8571 wake_up_new_task(tsk);
8574 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8575 /* Can't have SQ_AFF without SQPOLL */
8582 complete(&ctx->sq_data->exited);
8584 io_sq_thread_finish(ctx);
8588 static inline void __io_unaccount_mem(struct user_struct *user,
8589 unsigned long nr_pages)
8591 atomic_long_sub(nr_pages, &user->locked_vm);
8594 static inline int __io_account_mem(struct user_struct *user,
8595 unsigned long nr_pages)
8597 unsigned long page_limit, cur_pages, new_pages;
8599 /* Don't allow more pages than we can safely lock */
8600 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8603 cur_pages = atomic_long_read(&user->locked_vm);
8604 new_pages = cur_pages + nr_pages;
8605 if (new_pages > page_limit)
8607 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8608 new_pages) != cur_pages);
8613 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8616 __io_unaccount_mem(ctx->user, nr_pages);
8618 if (ctx->mm_account)
8619 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8622 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8627 ret = __io_account_mem(ctx->user, nr_pages);
8632 if (ctx->mm_account)
8633 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8638 static void io_mem_free(void *ptr)
8645 page = virt_to_head_page(ptr);
8646 if (put_page_testzero(page))
8647 free_compound_page(page);
8650 static void *io_mem_alloc(size_t size)
8652 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8653 __GFP_NORETRY | __GFP_ACCOUNT;
8655 return (void *) __get_free_pages(gfp_flags, get_order(size));
8658 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8661 struct io_rings *rings;
8662 size_t off, sq_array_size;
8664 off = struct_size(rings, cqes, cq_entries);
8665 if (off == SIZE_MAX)
8669 off = ALIGN(off, SMP_CACHE_BYTES);
8677 sq_array_size = array_size(sizeof(u32), sq_entries);
8678 if (sq_array_size == SIZE_MAX)
8681 if (check_add_overflow(off, sq_array_size, &off))
8687 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8689 struct io_mapped_ubuf *imu = *slot;
8692 if (imu != ctx->dummy_ubuf) {
8693 for (i = 0; i < imu->nr_bvecs; i++)
8694 unpin_user_page(imu->bvec[i].bv_page);
8695 if (imu->acct_pages)
8696 io_unaccount_mem(ctx, imu->acct_pages);
8702 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8704 io_buffer_unmap(ctx, &prsrc->buf);
8708 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8712 for (i = 0; i < ctx->nr_user_bufs; i++)
8713 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8714 kfree(ctx->user_bufs);
8715 io_rsrc_data_free(ctx->buf_data);
8716 ctx->user_bufs = NULL;
8717 ctx->buf_data = NULL;
8718 ctx->nr_user_bufs = 0;
8721 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8728 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8730 __io_sqe_buffers_unregister(ctx);
8734 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8735 void __user *arg, unsigned index)
8737 struct iovec __user *src;
8739 #ifdef CONFIG_COMPAT
8741 struct compat_iovec __user *ciovs;
8742 struct compat_iovec ciov;
8744 ciovs = (struct compat_iovec __user *) arg;
8745 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8748 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8749 dst->iov_len = ciov.iov_len;
8753 src = (struct iovec __user *) arg;
8754 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8760 * Not super efficient, but this is just a registration time. And we do cache
8761 * the last compound head, so generally we'll only do a full search if we don't
8764 * We check if the given compound head page has already been accounted, to
8765 * avoid double accounting it. This allows us to account the full size of the
8766 * page, not just the constituent pages of a huge page.
8768 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8769 int nr_pages, struct page *hpage)
8773 /* check current page array */
8774 for (i = 0; i < nr_pages; i++) {
8775 if (!PageCompound(pages[i]))
8777 if (compound_head(pages[i]) == hpage)
8781 /* check previously registered pages */
8782 for (i = 0; i < ctx->nr_user_bufs; i++) {
8783 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8785 for (j = 0; j < imu->nr_bvecs; j++) {
8786 if (!PageCompound(imu->bvec[j].bv_page))
8788 if (compound_head(imu->bvec[j].bv_page) == hpage)
8796 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8797 int nr_pages, struct io_mapped_ubuf *imu,
8798 struct page **last_hpage)
8802 imu->acct_pages = 0;
8803 for (i = 0; i < nr_pages; i++) {
8804 if (!PageCompound(pages[i])) {
8809 hpage = compound_head(pages[i]);
8810 if (hpage == *last_hpage)
8812 *last_hpage = hpage;
8813 if (headpage_already_acct(ctx, pages, i, hpage))
8815 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8819 if (!imu->acct_pages)
8822 ret = io_account_mem(ctx, imu->acct_pages);
8824 imu->acct_pages = 0;
8828 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8829 struct io_mapped_ubuf **pimu,
8830 struct page **last_hpage)
8832 struct io_mapped_ubuf *imu = NULL;
8833 struct vm_area_struct **vmas = NULL;
8834 struct page **pages = NULL;
8835 unsigned long off, start, end, ubuf;
8837 int ret, pret, nr_pages, i;
8839 if (!iov->iov_base) {
8840 *pimu = ctx->dummy_ubuf;
8844 ubuf = (unsigned long) iov->iov_base;
8845 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8846 start = ubuf >> PAGE_SHIFT;
8847 nr_pages = end - start;
8852 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8856 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8861 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8866 mmap_read_lock(current->mm);
8867 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8869 if (pret == nr_pages) {
8870 /* don't support file backed memory */
8871 for (i = 0; i < nr_pages; i++) {
8872 struct vm_area_struct *vma = vmas[i];
8874 if (vma_is_shmem(vma))
8877 !is_file_hugepages(vma->vm_file)) {
8883 ret = pret < 0 ? pret : -EFAULT;
8885 mmap_read_unlock(current->mm);
8888 * if we did partial map, or found file backed vmas,
8889 * release any pages we did get
8892 unpin_user_pages(pages, pret);
8896 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8898 unpin_user_pages(pages, pret);
8902 off = ubuf & ~PAGE_MASK;
8903 size = iov->iov_len;
8904 for (i = 0; i < nr_pages; i++) {
8907 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8908 imu->bvec[i].bv_page = pages[i];
8909 imu->bvec[i].bv_len = vec_len;
8910 imu->bvec[i].bv_offset = off;
8914 /* store original address for later verification */
8916 imu->ubuf_end = ubuf + iov->iov_len;
8917 imu->nr_bvecs = nr_pages;
8928 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8930 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8931 return ctx->user_bufs ? 0 : -ENOMEM;
8934 static int io_buffer_validate(struct iovec *iov)
8936 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8939 * Don't impose further limits on the size and buffer
8940 * constraints here, we'll -EINVAL later when IO is
8941 * submitted if they are wrong.
8944 return iov->iov_len ? -EFAULT : 0;
8948 /* arbitrary limit, but we need something */
8949 if (iov->iov_len > SZ_1G)
8952 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8958 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8959 unsigned int nr_args, u64 __user *tags)
8961 struct page *last_hpage = NULL;
8962 struct io_rsrc_data *data;
8968 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8970 ret = io_rsrc_node_switch_start(ctx);
8973 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8976 ret = io_buffers_map_alloc(ctx, nr_args);
8978 io_rsrc_data_free(data);
8982 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8983 ret = io_copy_iov(ctx, &iov, arg, i);
8986 ret = io_buffer_validate(&iov);
8989 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8994 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9000 WARN_ON_ONCE(ctx->buf_data);
9002 ctx->buf_data = data;
9004 __io_sqe_buffers_unregister(ctx);
9006 io_rsrc_node_switch(ctx, NULL);
9010 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9011 struct io_uring_rsrc_update2 *up,
9012 unsigned int nr_args)
9014 u64 __user *tags = u64_to_user_ptr(up->tags);
9015 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9016 struct page *last_hpage = NULL;
9017 bool needs_switch = false;
9023 if (up->offset + nr_args > ctx->nr_user_bufs)
9026 for (done = 0; done < nr_args; done++) {
9027 struct io_mapped_ubuf *imu;
9028 int offset = up->offset + done;
9031 err = io_copy_iov(ctx, &iov, iovs, done);
9034 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9038 err = io_buffer_validate(&iov);
9041 if (!iov.iov_base && tag) {
9045 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9049 i = array_index_nospec(offset, ctx->nr_user_bufs);
9050 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9051 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9052 ctx->rsrc_node, ctx->user_bufs[i]);
9053 if (unlikely(err)) {
9054 io_buffer_unmap(ctx, &imu);
9057 ctx->user_bufs[i] = NULL;
9058 needs_switch = true;
9061 ctx->user_bufs[i] = imu;
9062 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9066 io_rsrc_node_switch(ctx, ctx->buf_data);
9067 return done ? done : err;
9070 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9072 __s32 __user *fds = arg;
9078 if (copy_from_user(&fd, fds, sizeof(*fds)))
9081 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9082 if (IS_ERR(ctx->cq_ev_fd)) {
9083 int ret = PTR_ERR(ctx->cq_ev_fd);
9085 ctx->cq_ev_fd = NULL;
9092 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9094 if (ctx->cq_ev_fd) {
9095 eventfd_ctx_put(ctx->cq_ev_fd);
9096 ctx->cq_ev_fd = NULL;
9103 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9105 struct io_buffer *buf;
9106 unsigned long index;
9108 xa_for_each(&ctx->io_buffers, index, buf)
9109 __io_remove_buffers(ctx, buf, index, -1U);
9112 static void io_req_cache_free(struct list_head *list)
9114 struct io_kiocb *req, *nxt;
9116 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9117 list_del(&req->inflight_entry);
9118 kmem_cache_free(req_cachep, req);
9122 static void io_req_caches_free(struct io_ring_ctx *ctx)
9124 struct io_submit_state *state = &ctx->submit_state;
9126 mutex_lock(&ctx->uring_lock);
9128 if (state->free_reqs) {
9129 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9130 state->free_reqs = 0;
9133 io_flush_cached_locked_reqs(ctx, state);
9134 io_req_cache_free(&state->free_list);
9135 mutex_unlock(&ctx->uring_lock);
9138 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9140 if (data && !atomic_dec_and_test(&data->refs))
9141 wait_for_completion(&data->done);
9144 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9146 io_sq_thread_finish(ctx);
9148 if (ctx->mm_account) {
9149 mmdrop(ctx->mm_account);
9150 ctx->mm_account = NULL;
9153 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9154 io_wait_rsrc_data(ctx->buf_data);
9155 io_wait_rsrc_data(ctx->file_data);
9157 mutex_lock(&ctx->uring_lock);
9159 __io_sqe_buffers_unregister(ctx);
9161 __io_sqe_files_unregister(ctx);
9163 __io_cqring_overflow_flush(ctx, true);
9164 mutex_unlock(&ctx->uring_lock);
9165 io_eventfd_unregister(ctx);
9166 io_destroy_buffers(ctx);
9168 put_cred(ctx->sq_creds);
9170 /* there are no registered resources left, nobody uses it */
9172 io_rsrc_node_destroy(ctx->rsrc_node);
9173 if (ctx->rsrc_backup_node)
9174 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9175 flush_delayed_work(&ctx->rsrc_put_work);
9177 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9178 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9180 #if defined(CONFIG_UNIX)
9181 if (ctx->ring_sock) {
9182 ctx->ring_sock->file = NULL; /* so that iput() is called */
9183 sock_release(ctx->ring_sock);
9186 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9188 io_mem_free(ctx->rings);
9189 io_mem_free(ctx->sq_sqes);
9191 percpu_ref_exit(&ctx->refs);
9192 free_uid(ctx->user);
9193 io_req_caches_free(ctx);
9195 io_wq_put_hash(ctx->hash_map);
9196 kfree(ctx->cancel_hash);
9197 kfree(ctx->dummy_ubuf);
9201 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9203 struct io_ring_ctx *ctx = file->private_data;
9206 poll_wait(file, &ctx->poll_wait, wait);
9208 * synchronizes with barrier from wq_has_sleeper call in
9212 if (!io_sqring_full(ctx))
9213 mask |= EPOLLOUT | EPOLLWRNORM;
9216 * Don't flush cqring overflow list here, just do a simple check.
9217 * Otherwise there could possible be ABBA deadlock:
9220 * lock(&ctx->uring_lock);
9222 * lock(&ctx->uring_lock);
9225 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9226 * pushs them to do the flush.
9228 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9229 mask |= EPOLLIN | EPOLLRDNORM;
9234 static int io_uring_fasync(int fd, struct file *file, int on)
9236 struct io_ring_ctx *ctx = file->private_data;
9238 return fasync_helper(fd, file, on, &ctx->cq_fasync);
9241 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9243 const struct cred *creds;
9245 creds = xa_erase(&ctx->personalities, id);
9254 struct io_tctx_exit {
9255 struct callback_head task_work;
9256 struct completion completion;
9257 struct io_ring_ctx *ctx;
9260 static void io_tctx_exit_cb(struct callback_head *cb)
9262 struct io_uring_task *tctx = current->io_uring;
9263 struct io_tctx_exit *work;
9265 work = container_of(cb, struct io_tctx_exit, task_work);
9267 * When @in_idle, we're in cancellation and it's racy to remove the
9268 * node. It'll be removed by the end of cancellation, just ignore it.
9270 if (!atomic_read(&tctx->in_idle))
9271 io_uring_del_tctx_node((unsigned long)work->ctx);
9272 complete(&work->completion);
9275 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9277 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9279 return req->ctx == data;
9282 static void io_ring_exit_work(struct work_struct *work)
9284 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9285 unsigned long timeout = jiffies + HZ * 60 * 5;
9286 unsigned long interval = HZ / 20;
9287 struct io_tctx_exit exit;
9288 struct io_tctx_node *node;
9292 * If we're doing polled IO and end up having requests being
9293 * submitted async (out-of-line), then completions can come in while
9294 * we're waiting for refs to drop. We need to reap these manually,
9295 * as nobody else will be looking for them.
9298 io_uring_try_cancel_requests(ctx, NULL, true);
9300 struct io_sq_data *sqd = ctx->sq_data;
9301 struct task_struct *tsk;
9303 io_sq_thread_park(sqd);
9305 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9306 io_wq_cancel_cb(tsk->io_uring->io_wq,
9307 io_cancel_ctx_cb, ctx, true);
9308 io_sq_thread_unpark(sqd);
9311 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9312 /* there is little hope left, don't run it too often */
9315 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9317 init_completion(&exit.completion);
9318 init_task_work(&exit.task_work, io_tctx_exit_cb);
9321 * Some may use context even when all refs and requests have been put,
9322 * and they are free to do so while still holding uring_lock or
9323 * completion_lock, see io_req_task_submit(). Apart from other work,
9324 * this lock/unlock section also waits them to finish.
9326 mutex_lock(&ctx->uring_lock);
9327 while (!list_empty(&ctx->tctx_list)) {
9328 WARN_ON_ONCE(time_after(jiffies, timeout));
9330 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9332 /* don't spin on a single task if cancellation failed */
9333 list_rotate_left(&ctx->tctx_list);
9334 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9335 if (WARN_ON_ONCE(ret))
9337 wake_up_process(node->task);
9339 mutex_unlock(&ctx->uring_lock);
9340 wait_for_completion(&exit.completion);
9341 mutex_lock(&ctx->uring_lock);
9343 mutex_unlock(&ctx->uring_lock);
9344 spin_lock(&ctx->completion_lock);
9345 spin_unlock(&ctx->completion_lock);
9347 io_ring_ctx_free(ctx);
9350 /* Returns true if we found and killed one or more timeouts */
9351 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9354 struct io_kiocb *req, *tmp;
9357 spin_lock(&ctx->completion_lock);
9358 spin_lock_irq(&ctx->timeout_lock);
9359 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9360 if (io_match_task(req, tsk, cancel_all)) {
9361 io_kill_timeout(req, -ECANCELED);
9365 spin_unlock_irq(&ctx->timeout_lock);
9367 io_commit_cqring(ctx);
9368 spin_unlock(&ctx->completion_lock);
9370 io_cqring_ev_posted(ctx);
9371 return canceled != 0;
9374 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9376 unsigned long index;
9377 struct creds *creds;
9379 mutex_lock(&ctx->uring_lock);
9380 percpu_ref_kill(&ctx->refs);
9382 __io_cqring_overflow_flush(ctx, true);
9383 xa_for_each(&ctx->personalities, index, creds)
9384 io_unregister_personality(ctx, index);
9385 mutex_unlock(&ctx->uring_lock);
9387 io_kill_timeouts(ctx, NULL, true);
9388 io_poll_remove_all(ctx, NULL, true);
9390 /* if we failed setting up the ctx, we might not have any rings */
9391 io_iopoll_try_reap_events(ctx);
9393 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9395 * Use system_unbound_wq to avoid spawning tons of event kworkers
9396 * if we're exiting a ton of rings at the same time. It just adds
9397 * noise and overhead, there's no discernable change in runtime
9398 * over using system_wq.
9400 queue_work(system_unbound_wq, &ctx->exit_work);
9403 static int io_uring_release(struct inode *inode, struct file *file)
9405 struct io_ring_ctx *ctx = file->private_data;
9407 file->private_data = NULL;
9408 io_ring_ctx_wait_and_kill(ctx);
9412 struct io_task_cancel {
9413 struct task_struct *task;
9417 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9419 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9420 struct io_task_cancel *cancel = data;
9423 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9424 struct io_ring_ctx *ctx = req->ctx;
9426 /* protect against races with linked timeouts */
9427 spin_lock(&ctx->completion_lock);
9428 ret = io_match_task(req, cancel->task, cancel->all);
9429 spin_unlock(&ctx->completion_lock);
9431 ret = io_match_task(req, cancel->task, cancel->all);
9436 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9437 struct task_struct *task, bool cancel_all)
9439 struct io_defer_entry *de;
9442 spin_lock(&ctx->completion_lock);
9443 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9444 if (io_match_task(de->req, task, cancel_all)) {
9445 list_cut_position(&list, &ctx->defer_list, &de->list);
9449 spin_unlock(&ctx->completion_lock);
9450 if (list_empty(&list))
9453 while (!list_empty(&list)) {
9454 de = list_first_entry(&list, struct io_defer_entry, list);
9455 list_del_init(&de->list);
9456 io_req_complete_failed(de->req, -ECANCELED);
9462 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9464 struct io_tctx_node *node;
9465 enum io_wq_cancel cret;
9468 mutex_lock(&ctx->uring_lock);
9469 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9470 struct io_uring_task *tctx = node->task->io_uring;
9473 * io_wq will stay alive while we hold uring_lock, because it's
9474 * killed after ctx nodes, which requires to take the lock.
9476 if (!tctx || !tctx->io_wq)
9478 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9479 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9481 mutex_unlock(&ctx->uring_lock);
9486 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9487 struct task_struct *task,
9490 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9491 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9494 enum io_wq_cancel cret;
9498 ret |= io_uring_try_cancel_iowq(ctx);
9499 } else if (tctx && tctx->io_wq) {
9501 * Cancels requests of all rings, not only @ctx, but
9502 * it's fine as the task is in exit/exec.
9504 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9506 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9509 /* SQPOLL thread does its own polling */
9510 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9511 (ctx->sq_data && ctx->sq_data->thread == current)) {
9512 while (!list_empty_careful(&ctx->iopoll_list)) {
9513 io_iopoll_try_reap_events(ctx);
9518 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9519 ret |= io_poll_remove_all(ctx, task, cancel_all);
9520 ret |= io_kill_timeouts(ctx, task, cancel_all);
9522 ret |= io_run_task_work();
9529 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9531 struct io_uring_task *tctx = current->io_uring;
9532 struct io_tctx_node *node;
9535 if (unlikely(!tctx)) {
9536 ret = io_uring_alloc_task_context(current, ctx);
9539 tctx = current->io_uring;
9541 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9542 node = kmalloc(sizeof(*node), GFP_KERNEL);
9546 node->task = current;
9548 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9555 mutex_lock(&ctx->uring_lock);
9556 list_add(&node->ctx_node, &ctx->tctx_list);
9557 mutex_unlock(&ctx->uring_lock);
9564 * Note that this task has used io_uring. We use it for cancelation purposes.
9566 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9568 struct io_uring_task *tctx = current->io_uring;
9570 if (likely(tctx && tctx->last == ctx))
9572 return __io_uring_add_tctx_node(ctx);
9576 * Remove this io_uring_file -> task mapping.
9578 static void io_uring_del_tctx_node(unsigned long index)
9580 struct io_uring_task *tctx = current->io_uring;
9581 struct io_tctx_node *node;
9585 node = xa_erase(&tctx->xa, index);
9589 WARN_ON_ONCE(current != node->task);
9590 WARN_ON_ONCE(list_empty(&node->ctx_node));
9592 mutex_lock(&node->ctx->uring_lock);
9593 list_del(&node->ctx_node);
9594 mutex_unlock(&node->ctx->uring_lock);
9596 if (tctx->last == node->ctx)
9601 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9603 struct io_wq *wq = tctx->io_wq;
9604 struct io_tctx_node *node;
9605 unsigned long index;
9607 xa_for_each(&tctx->xa, index, node)
9608 io_uring_del_tctx_node(index);
9611 * Must be after io_uring_del_task_file() (removes nodes under
9612 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9614 io_wq_put_and_exit(wq);
9619 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9622 return atomic_read(&tctx->inflight_tracked);
9623 return percpu_counter_sum(&tctx->inflight);
9626 static void io_uring_drop_tctx_refs(struct task_struct *task)
9628 struct io_uring_task *tctx = task->io_uring;
9629 unsigned int refs = tctx->cached_refs;
9632 tctx->cached_refs = 0;
9633 percpu_counter_sub(&tctx->inflight, refs);
9634 put_task_struct_many(task, refs);
9639 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9640 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9642 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9644 struct io_uring_task *tctx = current->io_uring;
9645 struct io_ring_ctx *ctx;
9649 WARN_ON_ONCE(sqd && sqd->thread != current);
9651 if (!current->io_uring)
9654 io_wq_exit_start(tctx->io_wq);
9656 atomic_inc(&tctx->in_idle);
9658 io_uring_drop_tctx_refs(current);
9659 /* read completions before cancelations */
9660 inflight = tctx_inflight(tctx, !cancel_all);
9665 struct io_tctx_node *node;
9666 unsigned long index;
9668 xa_for_each(&tctx->xa, index, node) {
9669 /* sqpoll task will cancel all its requests */
9670 if (node->ctx->sq_data)
9672 io_uring_try_cancel_requests(node->ctx, current,
9676 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9677 io_uring_try_cancel_requests(ctx, current,
9681 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9682 io_uring_drop_tctx_refs(current);
9684 * If we've seen completions, retry without waiting. This
9685 * avoids a race where a completion comes in before we did
9686 * prepare_to_wait().
9688 if (inflight == tctx_inflight(tctx, !cancel_all))
9690 finish_wait(&tctx->wait, &wait);
9692 atomic_dec(&tctx->in_idle);
9694 io_uring_clean_tctx(tctx);
9696 /* for exec all current's requests should be gone, kill tctx */
9697 __io_uring_free(current);
9701 void __io_uring_cancel(bool cancel_all)
9703 io_uring_cancel_generic(cancel_all, NULL);
9706 static void *io_uring_validate_mmap_request(struct file *file,
9707 loff_t pgoff, size_t sz)
9709 struct io_ring_ctx *ctx = file->private_data;
9710 loff_t offset = pgoff << PAGE_SHIFT;
9715 case IORING_OFF_SQ_RING:
9716 case IORING_OFF_CQ_RING:
9719 case IORING_OFF_SQES:
9723 return ERR_PTR(-EINVAL);
9726 page = virt_to_head_page(ptr);
9727 if (sz > page_size(page))
9728 return ERR_PTR(-EINVAL);
9735 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9737 size_t sz = vma->vm_end - vma->vm_start;
9741 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9743 return PTR_ERR(ptr);
9745 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9746 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9749 #else /* !CONFIG_MMU */
9751 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9753 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9756 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9758 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9761 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9762 unsigned long addr, unsigned long len,
9763 unsigned long pgoff, unsigned long flags)
9767 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9769 return PTR_ERR(ptr);
9771 return (unsigned long) ptr;
9774 #endif /* !CONFIG_MMU */
9776 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9781 if (!io_sqring_full(ctx))
9783 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9785 if (!io_sqring_full(ctx))
9788 } while (!signal_pending(current));
9790 finish_wait(&ctx->sqo_sq_wait, &wait);
9794 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9795 struct __kernel_timespec __user **ts,
9796 const sigset_t __user **sig)
9798 struct io_uring_getevents_arg arg;
9801 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9802 * is just a pointer to the sigset_t.
9804 if (!(flags & IORING_ENTER_EXT_ARG)) {
9805 *sig = (const sigset_t __user *) argp;
9811 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9812 * timespec and sigset_t pointers if good.
9814 if (*argsz != sizeof(arg))
9816 if (copy_from_user(&arg, argp, sizeof(arg)))
9818 *sig = u64_to_user_ptr(arg.sigmask);
9819 *argsz = arg.sigmask_sz;
9820 *ts = u64_to_user_ptr(arg.ts);
9824 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9825 u32, min_complete, u32, flags, const void __user *, argp,
9828 struct io_ring_ctx *ctx;
9835 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9836 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9840 if (unlikely(!f.file))
9844 if (unlikely(f.file->f_op != &io_uring_fops))
9848 ctx = f.file->private_data;
9849 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9853 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9857 * For SQ polling, the thread will do all submissions and completions.
9858 * Just return the requested submit count, and wake the thread if
9862 if (ctx->flags & IORING_SETUP_SQPOLL) {
9863 io_cqring_overflow_flush(ctx);
9865 if (unlikely(ctx->sq_data->thread == NULL)) {
9869 if (flags & IORING_ENTER_SQ_WAKEUP)
9870 wake_up(&ctx->sq_data->wait);
9871 if (flags & IORING_ENTER_SQ_WAIT) {
9872 ret = io_sqpoll_wait_sq(ctx);
9876 submitted = to_submit;
9877 } else if (to_submit) {
9878 ret = io_uring_add_tctx_node(ctx);
9881 mutex_lock(&ctx->uring_lock);
9882 submitted = io_submit_sqes(ctx, to_submit);
9883 mutex_unlock(&ctx->uring_lock);
9885 if (submitted != to_submit)
9888 if (flags & IORING_ENTER_GETEVENTS) {
9889 const sigset_t __user *sig;
9890 struct __kernel_timespec __user *ts;
9892 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9896 min_complete = min(min_complete, ctx->cq_entries);
9899 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9900 * space applications don't need to do io completion events
9901 * polling again, they can rely on io_sq_thread to do polling
9902 * work, which can reduce cpu usage and uring_lock contention.
9904 if (ctx->flags & IORING_SETUP_IOPOLL &&
9905 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9906 ret = io_iopoll_check(ctx, min_complete);
9908 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9913 percpu_ref_put(&ctx->refs);
9916 return submitted ? submitted : ret;
9919 #ifdef CONFIG_PROC_FS
9920 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9921 const struct cred *cred)
9923 struct user_namespace *uns = seq_user_ns(m);
9924 struct group_info *gi;
9929 seq_printf(m, "%5d\n", id);
9930 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9931 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9932 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9933 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9934 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9935 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9936 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9937 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9938 seq_puts(m, "\n\tGroups:\t");
9939 gi = cred->group_info;
9940 for (g = 0; g < gi->ngroups; g++) {
9941 seq_put_decimal_ull(m, g ? " " : "",
9942 from_kgid_munged(uns, gi->gid[g]));
9944 seq_puts(m, "\n\tCapEff:\t");
9945 cap = cred->cap_effective;
9946 CAP_FOR_EACH_U32(__capi)
9947 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9952 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9954 struct io_sq_data *sq = NULL;
9959 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9960 * since fdinfo case grabs it in the opposite direction of normal use
9961 * cases. If we fail to get the lock, we just don't iterate any
9962 * structures that could be going away outside the io_uring mutex.
9964 has_lock = mutex_trylock(&ctx->uring_lock);
9966 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9972 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9973 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9974 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9975 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9976 struct file *f = io_file_from_index(ctx, i);
9979 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9981 seq_printf(m, "%5u: <none>\n", i);
9983 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9984 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9985 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9986 unsigned int len = buf->ubuf_end - buf->ubuf;
9988 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9990 if (has_lock && !xa_empty(&ctx->personalities)) {
9991 unsigned long index;
9992 const struct cred *cred;
9994 seq_printf(m, "Personalities:\n");
9995 xa_for_each(&ctx->personalities, index, cred)
9996 io_uring_show_cred(m, index, cred);
9998 seq_printf(m, "PollList:\n");
9999 spin_lock(&ctx->completion_lock);
10000 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10001 struct hlist_head *list = &ctx->cancel_hash[i];
10002 struct io_kiocb *req;
10004 hlist_for_each_entry(req, list, hash_node)
10005 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10006 req->task->task_works != NULL);
10008 spin_unlock(&ctx->completion_lock);
10010 mutex_unlock(&ctx->uring_lock);
10013 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10015 struct io_ring_ctx *ctx = f->private_data;
10017 if (percpu_ref_tryget(&ctx->refs)) {
10018 __io_uring_show_fdinfo(ctx, m);
10019 percpu_ref_put(&ctx->refs);
10024 static const struct file_operations io_uring_fops = {
10025 .release = io_uring_release,
10026 .mmap = io_uring_mmap,
10028 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10029 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10031 .poll = io_uring_poll,
10032 .fasync = io_uring_fasync,
10033 #ifdef CONFIG_PROC_FS
10034 .show_fdinfo = io_uring_show_fdinfo,
10038 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10039 struct io_uring_params *p)
10041 struct io_rings *rings;
10042 size_t size, sq_array_offset;
10044 /* make sure these are sane, as we already accounted them */
10045 ctx->sq_entries = p->sq_entries;
10046 ctx->cq_entries = p->cq_entries;
10048 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10049 if (size == SIZE_MAX)
10052 rings = io_mem_alloc(size);
10056 ctx->rings = rings;
10057 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10058 rings->sq_ring_mask = p->sq_entries - 1;
10059 rings->cq_ring_mask = p->cq_entries - 1;
10060 rings->sq_ring_entries = p->sq_entries;
10061 rings->cq_ring_entries = p->cq_entries;
10063 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10064 if (size == SIZE_MAX) {
10065 io_mem_free(ctx->rings);
10070 ctx->sq_sqes = io_mem_alloc(size);
10071 if (!ctx->sq_sqes) {
10072 io_mem_free(ctx->rings);
10080 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10084 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10088 ret = io_uring_add_tctx_node(ctx);
10093 fd_install(fd, file);
10098 * Allocate an anonymous fd, this is what constitutes the application
10099 * visible backing of an io_uring instance. The application mmaps this
10100 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10101 * we have to tie this fd to a socket for file garbage collection purposes.
10103 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10106 #if defined(CONFIG_UNIX)
10109 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10112 return ERR_PTR(ret);
10115 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10116 O_RDWR | O_CLOEXEC);
10117 #if defined(CONFIG_UNIX)
10118 if (IS_ERR(file)) {
10119 sock_release(ctx->ring_sock);
10120 ctx->ring_sock = NULL;
10122 ctx->ring_sock->file = file;
10128 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10129 struct io_uring_params __user *params)
10131 struct io_ring_ctx *ctx;
10137 if (entries > IORING_MAX_ENTRIES) {
10138 if (!(p->flags & IORING_SETUP_CLAMP))
10140 entries = IORING_MAX_ENTRIES;
10144 * Use twice as many entries for the CQ ring. It's possible for the
10145 * application to drive a higher depth than the size of the SQ ring,
10146 * since the sqes are only used at submission time. This allows for
10147 * some flexibility in overcommitting a bit. If the application has
10148 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10149 * of CQ ring entries manually.
10151 p->sq_entries = roundup_pow_of_two(entries);
10152 if (p->flags & IORING_SETUP_CQSIZE) {
10154 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10155 * to a power-of-two, if it isn't already. We do NOT impose
10156 * any cq vs sq ring sizing.
10158 if (!p->cq_entries)
10160 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10161 if (!(p->flags & IORING_SETUP_CLAMP))
10163 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10165 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10166 if (p->cq_entries < p->sq_entries)
10169 p->cq_entries = 2 * p->sq_entries;
10172 ctx = io_ring_ctx_alloc(p);
10175 ctx->compat = in_compat_syscall();
10176 if (!capable(CAP_IPC_LOCK))
10177 ctx->user = get_uid(current_user());
10180 * This is just grabbed for accounting purposes. When a process exits,
10181 * the mm is exited and dropped before the files, hence we need to hang
10182 * on to this mm purely for the purposes of being able to unaccount
10183 * memory (locked/pinned vm). It's not used for anything else.
10185 mmgrab(current->mm);
10186 ctx->mm_account = current->mm;
10188 ret = io_allocate_scq_urings(ctx, p);
10192 ret = io_sq_offload_create(ctx, p);
10195 /* always set a rsrc node */
10196 ret = io_rsrc_node_switch_start(ctx);
10199 io_rsrc_node_switch(ctx, NULL);
10201 memset(&p->sq_off, 0, sizeof(p->sq_off));
10202 p->sq_off.head = offsetof(struct io_rings, sq.head);
10203 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10204 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10205 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10206 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10207 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10208 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10210 memset(&p->cq_off, 0, sizeof(p->cq_off));
10211 p->cq_off.head = offsetof(struct io_rings, cq.head);
10212 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10213 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10214 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10215 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10216 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10217 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10219 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10220 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10221 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10222 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10223 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10224 IORING_FEAT_RSRC_TAGS;
10226 if (copy_to_user(params, p, sizeof(*p))) {
10231 file = io_uring_get_file(ctx);
10232 if (IS_ERR(file)) {
10233 ret = PTR_ERR(file);
10238 * Install ring fd as the very last thing, so we don't risk someone
10239 * having closed it before we finish setup
10241 ret = io_uring_install_fd(ctx, file);
10243 /* fput will clean it up */
10248 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10251 io_ring_ctx_wait_and_kill(ctx);
10256 * Sets up an aio uring context, and returns the fd. Applications asks for a
10257 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10258 * params structure passed in.
10260 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10262 struct io_uring_params p;
10265 if (copy_from_user(&p, params, sizeof(p)))
10267 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10272 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10273 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10274 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10275 IORING_SETUP_R_DISABLED))
10278 return io_uring_create(entries, &p, params);
10281 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10282 struct io_uring_params __user *, params)
10284 return io_uring_setup(entries, params);
10287 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10289 struct io_uring_probe *p;
10293 size = struct_size(p, ops, nr_args);
10294 if (size == SIZE_MAX)
10296 p = kzalloc(size, GFP_KERNEL);
10301 if (copy_from_user(p, arg, size))
10304 if (memchr_inv(p, 0, size))
10307 p->last_op = IORING_OP_LAST - 1;
10308 if (nr_args > IORING_OP_LAST)
10309 nr_args = IORING_OP_LAST;
10311 for (i = 0; i < nr_args; i++) {
10313 if (!io_op_defs[i].not_supported)
10314 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10319 if (copy_to_user(arg, p, size))
10326 static int io_register_personality(struct io_ring_ctx *ctx)
10328 const struct cred *creds;
10332 creds = get_current_cred();
10334 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10335 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10343 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10344 unsigned int nr_args)
10346 struct io_uring_restriction *res;
10350 /* Restrictions allowed only if rings started disabled */
10351 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10354 /* We allow only a single restrictions registration */
10355 if (ctx->restrictions.registered)
10358 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10361 size = array_size(nr_args, sizeof(*res));
10362 if (size == SIZE_MAX)
10365 res = memdup_user(arg, size);
10367 return PTR_ERR(res);
10371 for (i = 0; i < nr_args; i++) {
10372 switch (res[i].opcode) {
10373 case IORING_RESTRICTION_REGISTER_OP:
10374 if (res[i].register_op >= IORING_REGISTER_LAST) {
10379 __set_bit(res[i].register_op,
10380 ctx->restrictions.register_op);
10382 case IORING_RESTRICTION_SQE_OP:
10383 if (res[i].sqe_op >= IORING_OP_LAST) {
10388 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10390 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10391 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10393 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10394 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10403 /* Reset all restrictions if an error happened */
10405 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10407 ctx->restrictions.registered = true;
10413 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10415 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10418 if (ctx->restrictions.registered)
10419 ctx->restricted = 1;
10421 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10422 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10423 wake_up(&ctx->sq_data->wait);
10427 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10428 struct io_uring_rsrc_update2 *up,
10436 if (check_add_overflow(up->offset, nr_args, &tmp))
10438 err = io_rsrc_node_switch_start(ctx);
10443 case IORING_RSRC_FILE:
10444 return __io_sqe_files_update(ctx, up, nr_args);
10445 case IORING_RSRC_BUFFER:
10446 return __io_sqe_buffers_update(ctx, up, nr_args);
10451 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10454 struct io_uring_rsrc_update2 up;
10458 memset(&up, 0, sizeof(up));
10459 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10461 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10464 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10465 unsigned size, unsigned type)
10467 struct io_uring_rsrc_update2 up;
10469 if (size != sizeof(up))
10471 if (copy_from_user(&up, arg, sizeof(up)))
10473 if (!up.nr || up.resv)
10475 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10478 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10479 unsigned int size, unsigned int type)
10481 struct io_uring_rsrc_register rr;
10483 /* keep it extendible */
10484 if (size != sizeof(rr))
10487 memset(&rr, 0, sizeof(rr));
10488 if (copy_from_user(&rr, arg, size))
10490 if (!rr.nr || rr.resv || rr.resv2)
10494 case IORING_RSRC_FILE:
10495 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10496 rr.nr, u64_to_user_ptr(rr.tags));
10497 case IORING_RSRC_BUFFER:
10498 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10499 rr.nr, u64_to_user_ptr(rr.tags));
10504 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10507 struct io_uring_task *tctx = current->io_uring;
10508 cpumask_var_t new_mask;
10511 if (!tctx || !tctx->io_wq)
10514 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10517 cpumask_clear(new_mask);
10518 if (len > cpumask_size())
10519 len = cpumask_size();
10521 if (copy_from_user(new_mask, arg, len)) {
10522 free_cpumask_var(new_mask);
10526 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10527 free_cpumask_var(new_mask);
10531 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10533 struct io_uring_task *tctx = current->io_uring;
10535 if (!tctx || !tctx->io_wq)
10538 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10541 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10544 struct io_uring_task *tctx = NULL;
10545 struct io_sq_data *sqd = NULL;
10546 __u32 new_count[2];
10549 if (copy_from_user(new_count, arg, sizeof(new_count)))
10551 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10552 if (new_count[i] > INT_MAX)
10555 if (ctx->flags & IORING_SETUP_SQPOLL) {
10556 sqd = ctx->sq_data;
10559 * Observe the correct sqd->lock -> ctx->uring_lock
10560 * ordering. Fine to drop uring_lock here, we hold
10561 * a ref to the ctx.
10563 mutex_unlock(&ctx->uring_lock);
10564 mutex_lock(&sqd->lock);
10565 mutex_lock(&ctx->uring_lock);
10566 tctx = sqd->thread->io_uring;
10569 tctx = current->io_uring;
10573 if (!tctx || !tctx->io_wq)
10576 ret = io_wq_max_workers(tctx->io_wq, new_count);
10581 mutex_unlock(&sqd->lock);
10583 if (copy_to_user(arg, new_count, sizeof(new_count)))
10589 mutex_unlock(&sqd->lock);
10593 static bool io_register_op_must_quiesce(int op)
10596 case IORING_REGISTER_BUFFERS:
10597 case IORING_UNREGISTER_BUFFERS:
10598 case IORING_REGISTER_FILES:
10599 case IORING_UNREGISTER_FILES:
10600 case IORING_REGISTER_FILES_UPDATE:
10601 case IORING_REGISTER_PROBE:
10602 case IORING_REGISTER_PERSONALITY:
10603 case IORING_UNREGISTER_PERSONALITY:
10604 case IORING_REGISTER_FILES2:
10605 case IORING_REGISTER_FILES_UPDATE2:
10606 case IORING_REGISTER_BUFFERS2:
10607 case IORING_REGISTER_BUFFERS_UPDATE:
10608 case IORING_REGISTER_IOWQ_AFF:
10609 case IORING_UNREGISTER_IOWQ_AFF:
10610 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10617 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10621 percpu_ref_kill(&ctx->refs);
10624 * Drop uring mutex before waiting for references to exit. If another
10625 * thread is currently inside io_uring_enter() it might need to grab the
10626 * uring_lock to make progress. If we hold it here across the drain
10627 * wait, then we can deadlock. It's safe to drop the mutex here, since
10628 * no new references will come in after we've killed the percpu ref.
10630 mutex_unlock(&ctx->uring_lock);
10632 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10635 ret = io_run_task_work_sig();
10636 } while (ret >= 0);
10637 mutex_lock(&ctx->uring_lock);
10640 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10644 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10645 void __user *arg, unsigned nr_args)
10646 __releases(ctx->uring_lock)
10647 __acquires(ctx->uring_lock)
10652 * We're inside the ring mutex, if the ref is already dying, then
10653 * someone else killed the ctx or is already going through
10654 * io_uring_register().
10656 if (percpu_ref_is_dying(&ctx->refs))
10659 if (ctx->restricted) {
10660 if (opcode >= IORING_REGISTER_LAST)
10662 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10663 if (!test_bit(opcode, ctx->restrictions.register_op))
10667 if (io_register_op_must_quiesce(opcode)) {
10668 ret = io_ctx_quiesce(ctx);
10674 case IORING_REGISTER_BUFFERS:
10675 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10677 case IORING_UNREGISTER_BUFFERS:
10679 if (arg || nr_args)
10681 ret = io_sqe_buffers_unregister(ctx);
10683 case IORING_REGISTER_FILES:
10684 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10686 case IORING_UNREGISTER_FILES:
10688 if (arg || nr_args)
10690 ret = io_sqe_files_unregister(ctx);
10692 case IORING_REGISTER_FILES_UPDATE:
10693 ret = io_register_files_update(ctx, arg, nr_args);
10695 case IORING_REGISTER_EVENTFD:
10696 case IORING_REGISTER_EVENTFD_ASYNC:
10700 ret = io_eventfd_register(ctx, arg);
10703 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10704 ctx->eventfd_async = 1;
10706 ctx->eventfd_async = 0;
10708 case IORING_UNREGISTER_EVENTFD:
10710 if (arg || nr_args)
10712 ret = io_eventfd_unregister(ctx);
10714 case IORING_REGISTER_PROBE:
10716 if (!arg || nr_args > 256)
10718 ret = io_probe(ctx, arg, nr_args);
10720 case IORING_REGISTER_PERSONALITY:
10722 if (arg || nr_args)
10724 ret = io_register_personality(ctx);
10726 case IORING_UNREGISTER_PERSONALITY:
10730 ret = io_unregister_personality(ctx, nr_args);
10732 case IORING_REGISTER_ENABLE_RINGS:
10734 if (arg || nr_args)
10736 ret = io_register_enable_rings(ctx);
10738 case IORING_REGISTER_RESTRICTIONS:
10739 ret = io_register_restrictions(ctx, arg, nr_args);
10741 case IORING_REGISTER_FILES2:
10742 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10744 case IORING_REGISTER_FILES_UPDATE2:
10745 ret = io_register_rsrc_update(ctx, arg, nr_args,
10748 case IORING_REGISTER_BUFFERS2:
10749 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10751 case IORING_REGISTER_BUFFERS_UPDATE:
10752 ret = io_register_rsrc_update(ctx, arg, nr_args,
10753 IORING_RSRC_BUFFER);
10755 case IORING_REGISTER_IOWQ_AFF:
10757 if (!arg || !nr_args)
10759 ret = io_register_iowq_aff(ctx, arg, nr_args);
10761 case IORING_UNREGISTER_IOWQ_AFF:
10763 if (arg || nr_args)
10765 ret = io_unregister_iowq_aff(ctx);
10767 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10769 if (!arg || nr_args != 2)
10771 ret = io_register_iowq_max_workers(ctx, arg);
10778 if (io_register_op_must_quiesce(opcode)) {
10779 /* bring the ctx back to life */
10780 percpu_ref_reinit(&ctx->refs);
10781 reinit_completion(&ctx->ref_comp);
10786 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10787 void __user *, arg, unsigned int, nr_args)
10789 struct io_ring_ctx *ctx;
10798 if (f.file->f_op != &io_uring_fops)
10801 ctx = f.file->private_data;
10803 io_run_task_work();
10805 mutex_lock(&ctx->uring_lock);
10806 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10807 mutex_unlock(&ctx->uring_lock);
10808 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10809 ctx->cq_ev_fd != NULL, ret);
10815 static int __init io_uring_init(void)
10817 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10818 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10819 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10822 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10823 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10824 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10825 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10826 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10827 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10828 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10829 BUILD_BUG_SQE_ELEM(8, __u64, off);
10830 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10831 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10832 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10833 BUILD_BUG_SQE_ELEM(24, __u32, len);
10834 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10835 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10836 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10837 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10838 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10839 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10840 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10841 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10842 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10843 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10844 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10845 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10846 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10847 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10848 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10849 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10850 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10851 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10852 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10853 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10854 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10856 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10857 sizeof(struct io_uring_rsrc_update));
10858 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10859 sizeof(struct io_uring_rsrc_update2));
10861 /* ->buf_index is u16 */
10862 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10864 /* should fit into one byte */
10865 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10867 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10868 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
10870 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10874 __initcall(io_uring_init);